2015-04-15 15:51:25 -07:00
|
|
|
// Copyright 2015 The Go Authors. All rights reserved.
|
|
|
|
// Use of this source code is governed by a BSD-style
|
|
|
|
// license that can be found in the LICENSE file.
|
|
|
|
|
|
|
|
package gc
|
|
|
|
|
|
|
|
import (
|
2015-08-10 12:15:52 -07:00
|
|
|
"bytes"
|
2015-06-12 11:01:13 -07:00
|
|
|
"fmt"
|
2015-08-10 12:15:52 -07:00
|
|
|
"html"
|
2015-07-22 13:13:53 -07:00
|
|
|
"os"
|
2015-07-10 10:47:28 -06:00
|
|
|
"strings"
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-05-28 13:49:20 -07:00
|
|
|
"cmd/compile/internal/ssa"
|
2015-05-12 11:06:44 -07:00
|
|
|
"cmd/internal/obj"
|
2016-04-06 12:01:40 -07:00
|
|
|
"cmd/internal/sys"
|
2015-04-15 15:51:25 -07:00
|
|
|
)
|
|
|
|
|
2016-03-03 22:06:57 -08:00
|
|
|
var ssaEnabled = true
|
|
|
|
|
2016-01-28 13:46:30 -08:00
|
|
|
var ssaConfig *ssa.Config
|
|
|
|
var ssaExp ssaExport
|
|
|
|
|
[dev.ssa] cmd/compile: enhance command line option processing for SSA
The -d compiler flag can also specify ssa phase and flag,
for example -d=ssa/generic_cse/time,ssa/generic_cse/stats
Spaces in the phase names can be specified with an
underscore. Flags currently parsed (not necessarily
recognized by the phases yet) are:
on, off, mem, time, debug, stats, and test
On, off and time are handled in the harness,
debug, stats, and test are interpreted by the phase itself.
The pass is now attached to the Func being compiled, and a
new method logStats(key, ...value) on *Func to encourage a
semi-standardized format for that output. Output fields
are separated by tabs to ease digestion by awk and
spreadsheets. For example,
if f.pass.stats > 0 {
f.logStat("CSE REWRITES", rewrites)
}
Change-Id: I16db2b5af64c50ca9a47efeb51d961147a903abc
Reviewed-on: https://go-review.googlesource.com/19885
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
2016-02-25 13:10:51 -05:00
|
|
|
func initssa() *ssa.Config {
|
|
|
|
ssaExp.unimplemented = false
|
|
|
|
ssaExp.mustImplement = true
|
|
|
|
if ssaConfig == nil {
|
2016-04-06 12:01:40 -07:00
|
|
|
ssaConfig = ssa.NewConfig(Thearch.LinkArch.Name, &ssaExp, Ctxt, Debug['N'] == 0)
|
2016-07-26 11:51:33 -07:00
|
|
|
if Thearch.LinkArch.Name == "386" {
|
|
|
|
ssaConfig.Set387(Thearch.Use387)
|
|
|
|
}
|
[dev.ssa] cmd/compile: enhance command line option processing for SSA
The -d compiler flag can also specify ssa phase and flag,
for example -d=ssa/generic_cse/time,ssa/generic_cse/stats
Spaces in the phase names can be specified with an
underscore. Flags currently parsed (not necessarily
recognized by the phases yet) are:
on, off, mem, time, debug, stats, and test
On, off and time are handled in the harness,
debug, stats, and test are interpreted by the phase itself.
The pass is now attached to the Func being compiled, and a
new method logStats(key, ...value) on *Func to encourage a
semi-standardized format for that output. Output fields
are separated by tabs to ease digestion by awk and
spreadsheets. For example,
if f.pass.stats > 0 {
f.logStat("CSE REWRITES", rewrites)
}
Change-Id: I16db2b5af64c50ca9a47efeb51d961147a903abc
Reviewed-on: https://go-review.googlesource.com/19885
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
2016-02-25 13:10:51 -05:00
|
|
|
}
|
|
|
|
return ssaConfig
|
|
|
|
}
|
|
|
|
|
2015-12-11 20:41:52 -08:00
|
|
|
func shouldssa(fn *Node) bool {
|
2016-04-06 12:01:40 -07:00
|
|
|
switch Thearch.LinkArch.Name {
|
2016-03-21 22:57:26 -07:00
|
|
|
default:
|
|
|
|
// Only available for testing.
|
|
|
|
if os.Getenv("SSATEST") == "" {
|
|
|
|
return false
|
|
|
|
}
|
2016-08-25 15:52:45 -04:00
|
|
|
case "amd64", "amd64p32", "arm", "386", "arm64", "ppc64le", "mips64", "mips64le":
|
2016-06-07 14:18:04 -04:00
|
|
|
// Generally available.
|
2015-12-11 20:41:52 -08:00
|
|
|
}
|
2016-03-03 22:06:57 -08:00
|
|
|
if !ssaEnabled {
|
|
|
|
return false
|
|
|
|
}
|
2015-06-12 11:01:13 -07:00
|
|
|
|
2015-10-19 11:36:07 -04:00
|
|
|
// Environment variable control of SSA CG
|
|
|
|
// 1. IF GOSSAFUNC == current function name THEN
|
|
|
|
// compile this function with SSA and log output to ssa.html
|
|
|
|
|
2015-10-26 17:34:06 -04:00
|
|
|
// 2. IF GOSSAHASH == "" THEN
|
2015-10-19 11:36:07 -04:00
|
|
|
// compile this function (and everything else) with SSA
|
|
|
|
|
2015-10-26 17:34:06 -04:00
|
|
|
// 3. IF GOSSAHASH == "n" or "N"
|
2015-10-19 11:36:07 -04:00
|
|
|
// IF GOSSAPKG == current package name THEN
|
|
|
|
// compile this function (and everything in this package) with SSA
|
|
|
|
// ELSE
|
|
|
|
// use the old back end for this function.
|
|
|
|
// This is for compatibility with existing test harness and should go away.
|
|
|
|
|
|
|
|
// 4. IF GOSSAHASH is a suffix of the binary-rendered SHA1 hash of the function name THEN
|
|
|
|
// compile this function with SSA
|
|
|
|
// ELSE
|
|
|
|
// compile this function with the old back end.
|
|
|
|
|
2015-10-26 17:34:06 -04:00
|
|
|
// Plan is for 3 to be removed when the tests are revised.
|
|
|
|
// SSA is now default, and is disabled by setting
|
|
|
|
// GOSSAHASH to n or N, or selectively with strings of
|
|
|
|
// 0 and 1.
|
2015-10-19 11:36:07 -04:00
|
|
|
|
2015-12-11 20:41:52 -08:00
|
|
|
name := fn.Func.Nname.Sym.Name
|
|
|
|
|
|
|
|
funcname := os.Getenv("GOSSAFUNC")
|
|
|
|
if funcname != "" {
|
|
|
|
// If GOSSAFUNC is set, compile only that function.
|
|
|
|
return name == funcname
|
|
|
|
}
|
|
|
|
|
|
|
|
pkg := os.Getenv("GOSSAPKG")
|
|
|
|
if pkg != "" {
|
|
|
|
// If GOSSAPKG is set, compile only that package.
|
|
|
|
return localpkg.Name == pkg
|
|
|
|
}
|
|
|
|
|
[dev.ssa] cmd/compile: enhance command line option processing for SSA
The -d compiler flag can also specify ssa phase and flag,
for example -d=ssa/generic_cse/time,ssa/generic_cse/stats
Spaces in the phase names can be specified with an
underscore. Flags currently parsed (not necessarily
recognized by the phases yet) are:
on, off, mem, time, debug, stats, and test
On, off and time are handled in the harness,
debug, stats, and test are interpreted by the phase itself.
The pass is now attached to the Func being compiled, and a
new method logStats(key, ...value) on *Func to encourage a
semi-standardized format for that output. Output fields
are separated by tabs to ease digestion by awk and
spreadsheets. For example,
if f.pass.stats > 0 {
f.logStat("CSE REWRITES", rewrites)
}
Change-Id: I16db2b5af64c50ca9a47efeb51d961147a903abc
Reviewed-on: https://go-review.googlesource.com/19885
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
2016-02-25 13:10:51 -05:00
|
|
|
return initssa().DebugHashMatch("GOSSAHASH", name)
|
2015-12-11 20:41:52 -08:00
|
|
|
}
|
|
|
|
|
|
|
|
// buildssa builds an SSA function.
|
|
|
|
func buildssa(fn *Node) *ssa.Func {
|
|
|
|
name := fn.Func.Nname.Sym.Name
|
2016-03-01 13:47:48 -08:00
|
|
|
printssa := name == os.Getenv("GOSSAFUNC")
|
2015-12-11 20:41:52 -08:00
|
|
|
if printssa {
|
2015-07-20 18:42:45 -07:00
|
|
|
fmt.Println("generating SSA for", name)
|
2016-03-04 13:16:48 -08:00
|
|
|
dumplist("buildssa-enter", fn.Func.Enter)
|
|
|
|
dumplist("buildssa-body", fn.Nbody)
|
|
|
|
dumplist("buildssa-exit", fn.Func.Exit)
|
2015-06-12 11:01:13 -07:00
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-05-18 16:44:20 -07:00
|
|
|
var s state
|
2015-05-30 01:03:06 -04:00
|
|
|
s.pushLine(fn.Lineno)
|
|
|
|
defer s.popLine()
|
|
|
|
|
2016-02-27 17:49:31 -08:00
|
|
|
if fn.Func.Pragma&CgoUnsafeArgs != 0 {
|
|
|
|
s.cgoUnsafeArgs = true
|
|
|
|
}
|
2016-03-16 21:51:17 -07:00
|
|
|
if fn.Func.Pragma&Nowritebarrier != 0 {
|
|
|
|
s.noWB = true
|
|
|
|
}
|
|
|
|
defer func() {
|
|
|
|
if s.WBLineno != 0 {
|
|
|
|
fn.Func.WBLineno = s.WBLineno
|
|
|
|
}
|
|
|
|
}()
|
2015-04-15 15:51:25 -07:00
|
|
|
// TODO(khr): build config just once at the start of the compiler binary
|
2015-06-12 11:01:13 -07:00
|
|
|
|
2016-01-28 13:46:30 -08:00
|
|
|
ssaExp.log = printssa
|
[dev.ssa] cmd/compile: enhance command line option processing for SSA
The -d compiler flag can also specify ssa phase and flag,
for example -d=ssa/generic_cse/time,ssa/generic_cse/stats
Spaces in the phase names can be specified with an
underscore. Flags currently parsed (not necessarily
recognized by the phases yet) are:
on, off, mem, time, debug, stats, and test
On, off and time are handled in the harness,
debug, stats, and test are interpreted by the phase itself.
The pass is now attached to the Func being compiled, and a
new method logStats(key, ...value) on *Func to encourage a
semi-standardized format for that output. Output fields
are separated by tabs to ease digestion by awk and
spreadsheets. For example,
if f.pass.stats > 0 {
f.logStat("CSE REWRITES", rewrites)
}
Change-Id: I16db2b5af64c50ca9a47efeb51d961147a903abc
Reviewed-on: https://go-review.googlesource.com/19885
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
2016-02-25 13:10:51 -05:00
|
|
|
|
|
|
|
s.config = initssa()
|
2015-04-15 15:51:25 -07:00
|
|
|
s.f = s.config.NewFunc()
|
2015-06-12 11:01:13 -07:00
|
|
|
s.f.Name = name
|
2015-10-08 12:39:56 -04:00
|
|
|
s.exitCode = fn.Func.Exit
|
2015-11-09 21:35:40 -08:00
|
|
|
s.panics = map[funcLine]*ssa.Block{}
|
2015-06-12 11:01:13 -07:00
|
|
|
|
2015-08-10 12:15:52 -07:00
|
|
|
if name == os.Getenv("GOSSAFUNC") {
|
|
|
|
// TODO: tempfile? it is handy to have the location
|
|
|
|
// of this file be stable, so you can just reload in the browser.
|
2016-01-13 11:14:57 -08:00
|
|
|
s.config.HTML = ssa.NewHTMLWriter("ssa.html", s.config, name)
|
2015-08-10 12:15:52 -07:00
|
|
|
// TODO: generate and print a mapping from nodes to values and blocks
|
|
|
|
}
|
|
|
|
defer func() {
|
2015-12-11 20:41:52 -08:00
|
|
|
if !printssa {
|
2015-08-10 12:15:52 -07:00
|
|
|
s.config.HTML.Close()
|
|
|
|
}
|
|
|
|
}()
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
// Allocate starting block
|
|
|
|
s.f.Entry = s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
|
2015-05-18 16:44:20 -07:00
|
|
|
// Allocate starting values
|
2015-07-20 15:39:14 -07:00
|
|
|
s.labels = map[string]*ssaLabel{}
|
|
|
|
s.labeledNodes = map[*Node]*ssaLabel{}
|
2015-11-02 08:10:26 -08:00
|
|
|
s.startmem = s.entryNewValue0(ssa.OpInitMem, ssa.TypeMem)
|
2015-07-30 11:03:05 -07:00
|
|
|
s.sp = s.entryNewValue0(ssa.OpSP, Types[TUINTPTR]) // TODO: use generic pointer type (unsafe.Pointer?) instead
|
|
|
|
s.sb = s.entryNewValue0(ssa.OpSB, Types[TUINTPTR])
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-09-11 16:40:05 -04:00
|
|
|
s.startBlock(s.f.Entry)
|
|
|
|
s.vars[&memVar] = s.startmem
|
|
|
|
|
2015-10-15 20:25:32 -05:00
|
|
|
s.varsyms = map[*Node]interface{}{}
|
|
|
|
|
2015-06-19 21:02:28 -07:00
|
|
|
// Generate addresses of local declarations
|
|
|
|
s.decladdrs = map[*Node]*ssa.Value{}
|
2016-02-29 13:31:48 -08:00
|
|
|
for _, n := range fn.Func.Dcl {
|
2015-06-19 21:02:28 -07:00
|
|
|
switch n.Class {
|
2016-02-27 17:49:31 -08:00
|
|
|
case PPARAM, PPARAMOUT:
|
2015-10-15 20:25:32 -05:00
|
|
|
aux := s.lookupSymbol(n, &ssa.ArgSymbol{Typ: n.Type, Node: n})
|
2015-06-19 21:02:28 -07:00
|
|
|
s.decladdrs[n] = s.entryNewValue1A(ssa.OpAddr, Ptrto(n.Type), aux, s.sp)
|
2016-02-27 17:49:31 -08:00
|
|
|
if n.Class == PPARAMOUT && s.canSSA(n) {
|
|
|
|
// Save ssa-able PPARAMOUT variables so we can
|
|
|
|
// store them back to the stack at the end of
|
|
|
|
// the function.
|
|
|
|
s.returns = append(s.returns, n)
|
|
|
|
}
|
2016-04-21 19:28:28 -07:00
|
|
|
if n.Class == PPARAM && s.canSSA(n) && n.Type.IsPtrShaped() {
|
|
|
|
s.ptrargs = append(s.ptrargs, n)
|
|
|
|
n.SetNotLiveAtEnd(true) // SSA takes care of this explicitly
|
|
|
|
}
|
2016-02-27 17:49:31 -08:00
|
|
|
case PAUTO:
|
2015-08-24 02:16:19 -07:00
|
|
|
// processed at each use, to prevent Addr coming
|
|
|
|
// before the decl.
|
cmd/compile: fix liveness computation for heap-escaped parameters
The liveness computation of parameters generally was never
correct, but forcing all parameters to be live throughout the
function covered up that problem. The new SSA back end is
too clever: even though it currently keeps the parameter values live
throughout the function, it may find optimizations that mean
the current values are not written back to the original parameter
stack slots immediately or ever (for example if a parameter is set
to nil, SSA constant propagation may replace all later uses of the
parameter with a constant nil, eliminating the need to write the nil
value back to the stack slot), so the liveness code must now
track the actual operations on the stack slots, exposing these
problems.
One small problem in the handling of arguments is that nodarg
can return ONAME PPARAM nodes with adjusted offsets, so that
there are actually multiple *Node pointers for the same parameter
in the instruction stream. This might be possible to correct, but
not in this CL. For now, we fix this by using n.Orig instead of n
when considering PPARAM and PPARAMOUT nodes.
The major problem in the handling of arguments is general
confusion in the liveness code about the meaning of PPARAM|PHEAP
and PPARAMOUT|PHEAP nodes, especially as contrasted with PAUTO|PHEAP.
The difference between these two is that when a local variable "moves"
to the heap, it's really just allocated there to start with; in contrast,
when an argument moves to the heap, the actual data has to be copied
there from the stack at the beginning of the function, and when a
result "moves" to the heap the value in the heap has to be copied
back to the stack when the function returns
This general confusion is also present in the SSA back end.
The PHEAP bit worked decently when I first introduced it 7 years ago (!)
in 391425ae. The back end did nothing sophisticated, and in particular
there was no analysis at all: no escape analysis, no liveness analysis,
and certainly no SSA back end. But the complications caused in the
various downstream consumers suggest that this should be a detail
kept mainly in the front end.
This CL therefore eliminates both the PHEAP bit and even the idea of
"heap variables" from the back ends.
First, it replaces the PPARAM|PHEAP, PPARAMOUT|PHEAP, and PAUTO|PHEAP
variable classes with the single PAUTOHEAP, a pseudo-class indicating
a variable maintained on the heap and available by indirecting a
local variable kept on the stack (a plain PAUTO).
Second, walkexpr replaces all references to PAUTOHEAP variables
with indirections of the corresponding PAUTO variable.
The back ends and the liveness code now just see plain indirected
variables. This may actually produce better code, but the real goal
here is to eliminate these little-used and somewhat suspect code
paths in the back end analyses.
The OPARAM node type goes away too.
A followup CL will do the same to PPARAMREF. I'm not sure that
the back ends (SSA in particular) are handling those right either,
and with the framework established in this CL that change is trivial
and the result clearly more correct.
Fixes #15747.
Change-Id: I2770b1ce3cbc93981bfc7166be66a9da12013d74
Reviewed-on: https://go-review.googlesource.com/23393
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-05-25 01:33:24 -04:00
|
|
|
case PAUTOHEAP:
|
|
|
|
// moved to heap - already handled by frontend
|
2015-09-06 13:42:26 -07:00
|
|
|
case PFUNC:
|
|
|
|
// local function - already handled by frontend
|
2015-07-01 20:37:25 +01:00
|
|
|
default:
|
cmd/compile: fix liveness computation for heap-escaped parameters
The liveness computation of parameters generally was never
correct, but forcing all parameters to be live throughout the
function covered up that problem. The new SSA back end is
too clever: even though it currently keeps the parameter values live
throughout the function, it may find optimizations that mean
the current values are not written back to the original parameter
stack slots immediately or ever (for example if a parameter is set
to nil, SSA constant propagation may replace all later uses of the
parameter with a constant nil, eliminating the need to write the nil
value back to the stack slot), so the liveness code must now
track the actual operations on the stack slots, exposing these
problems.
One small problem in the handling of arguments is that nodarg
can return ONAME PPARAM nodes with adjusted offsets, so that
there are actually multiple *Node pointers for the same parameter
in the instruction stream. This might be possible to correct, but
not in this CL. For now, we fix this by using n.Orig instead of n
when considering PPARAM and PPARAMOUT nodes.
The major problem in the handling of arguments is general
confusion in the liveness code about the meaning of PPARAM|PHEAP
and PPARAMOUT|PHEAP nodes, especially as contrasted with PAUTO|PHEAP.
The difference between these two is that when a local variable "moves"
to the heap, it's really just allocated there to start with; in contrast,
when an argument moves to the heap, the actual data has to be copied
there from the stack at the beginning of the function, and when a
result "moves" to the heap the value in the heap has to be copied
back to the stack when the function returns
This general confusion is also present in the SSA back end.
The PHEAP bit worked decently when I first introduced it 7 years ago (!)
in 391425ae. The back end did nothing sophisticated, and in particular
there was no analysis at all: no escape analysis, no liveness analysis,
and certainly no SSA back end. But the complications caused in the
various downstream consumers suggest that this should be a detail
kept mainly in the front end.
This CL therefore eliminates both the PHEAP bit and even the idea of
"heap variables" from the back ends.
First, it replaces the PPARAM|PHEAP, PPARAMOUT|PHEAP, and PAUTO|PHEAP
variable classes with the single PAUTOHEAP, a pseudo-class indicating
a variable maintained on the heap and available by indirecting a
local variable kept on the stack (a plain PAUTO).
Second, walkexpr replaces all references to PAUTOHEAP variables
with indirections of the corresponding PAUTO variable.
The back ends and the liveness code now just see plain indirected
variables. This may actually produce better code, but the real goal
here is to eliminate these little-used and somewhat suspect code
paths in the back end analyses.
The OPARAM node type goes away too.
A followup CL will do the same to PPARAMREF. I'm not sure that
the back ends (SSA in particular) are handling those right either,
and with the framework established in this CL that change is trivial
and the result clearly more correct.
Fixes #15747.
Change-Id: I2770b1ce3cbc93981bfc7166be66a9da12013d74
Reviewed-on: https://go-review.googlesource.com/23393
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-05-25 01:33:24 -04:00
|
|
|
s.Unimplementedf("local variable with class %s unimplemented", classnames[n.Class])
|
2015-06-19 21:02:28 -07:00
|
|
|
}
|
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
|
|
|
|
// Convert the AST-based IR to the SSA-based IR
|
2016-02-29 13:31:48 -08:00
|
|
|
s.stmts(fn.Func.Enter)
|
2016-03-01 12:50:17 -08:00
|
|
|
s.stmts(fn.Nbody)
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-05-18 16:44:20 -07:00
|
|
|
// fallthrough to exit
|
2015-09-08 16:04:37 -07:00
|
|
|
if s.curBlock != nil {
|
2016-03-09 19:27:57 -08:00
|
|
|
s.pushLine(fn.Func.Endlineno)
|
|
|
|
s.exit()
|
|
|
|
s.popLine()
|
2015-05-18 16:44:20 -07:00
|
|
|
}
|
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
// Check that we used all labels
|
|
|
|
for name, lab := range s.labels {
|
2016-05-27 15:33:11 -07:00
|
|
|
if !lab.used() && !lab.reported && !lab.defNode.Used {
|
2016-03-02 11:01:25 -08:00
|
|
|
yyerrorl(lab.defNode.Lineno, "label %v defined and not used", name)
|
2015-07-20 15:39:14 -07:00
|
|
|
lab.reported = true
|
|
|
|
}
|
|
|
|
if lab.used() && !lab.defined() && !lab.reported {
|
2016-03-02 11:01:25 -08:00
|
|
|
yyerrorl(lab.useNode.Lineno, "label %v not defined", name)
|
2015-07-20 15:39:14 -07:00
|
|
|
lab.reported = true
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check any forward gotos. Non-forward gotos have already been checked.
|
|
|
|
for _, n := range s.fwdGotos {
|
|
|
|
lab := s.labels[n.Left.Sym.Name]
|
|
|
|
// If the label is undefined, we have already have printed an error.
|
|
|
|
if lab.defined() {
|
|
|
|
s.checkgoto(n, lab.defNode)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if nerrors > 0 {
|
2016-01-28 16:11:56 -08:00
|
|
|
s.f.Free()
|
2015-12-11 20:41:52 -08:00
|
|
|
return nil
|
2015-07-20 15:39:14 -07:00
|
|
|
}
|
|
|
|
|
cmd/compile: use sparse algorithm for phis in large program
This adds a sparse method for locating nearest ancestors
in a dominator tree, and checks blocks with more than one
predecessor for differences and inserts phi functions where
there are.
Uses reversed post order to cut number of passes, running
it from first def to last use ("last use" for paramout and
mem is end-of-program; last use for a phi input from a
backedge is the source of the back edge)
Includes a cutover from old algorithm to new to avoid paying
large constant factor for small programs. This keeps normal
builds running at about the same time, while not running
over-long on large machine-generated inputs.
Add "phase" flags for ssa/build -- ssa/build/stats prints
number of blocks, values (before and after linking references
and inserting phis, so expansion can be measured), and their
product; the product governs the cutover, where a good value
seems to be somewhere between 1 and 5 million.
Among the files compiled by make.bash, this is the shape of
the tail of the distribution for #blocks, #vars, and their
product:
#blocks #vars product
max 6171 28180 173,898,780
99.9% 1641 6548 10,401,878
99% 463 1909 873,721
95% 152 639 95,235
90% 84 359 30,021
The old algorithm is indeed usually fastest, for 99%ile
values of usually.
The fix to LookupVarOutgoing
( https://go-review.googlesource.com/#/c/22790/ )
deals with some of the same problems addressed by this CL,
but on at least one bug ( #15537 ) this change is still
a significant help.
With this CL:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 4m35.200s
user 13m16.644s
sys 0m36.712s
and pprof reports 3.4GB allocated in one of the larger profiles
With tip:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 10m36.569s
user 25m52.286s
sys 4m3.696s
and pprof reports 8.3GB allocated in the same larger profile
With this CL, most of the compilation time on the benchmarked
input is spent in register/stack allocation (cumulative 53%)
and in the sparse lookup algorithm itself (cumulative 20%).
Fixes #15537.
Change-Id: Ia0299dda6a291534d8b08e5f9883216ded677a00
Reviewed-on: https://go-review.googlesource.com/22342
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-04-21 13:24:58 -04:00
|
|
|
prelinkNumvars := s.f.NumValues()
|
|
|
|
sparseDefState := s.locatePotentialPhiFunctions(fn)
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
// Link up variable uses to variable definitions
|
cmd/compile: use sparse algorithm for phis in large program
This adds a sparse method for locating nearest ancestors
in a dominator tree, and checks blocks with more than one
predecessor for differences and inserts phi functions where
there are.
Uses reversed post order to cut number of passes, running
it from first def to last use ("last use" for paramout and
mem is end-of-program; last use for a phi input from a
backedge is the source of the back edge)
Includes a cutover from old algorithm to new to avoid paying
large constant factor for small programs. This keeps normal
builds running at about the same time, while not running
over-long on large machine-generated inputs.
Add "phase" flags for ssa/build -- ssa/build/stats prints
number of blocks, values (before and after linking references
and inserting phis, so expansion can be measured), and their
product; the product governs the cutover, where a good value
seems to be somewhere between 1 and 5 million.
Among the files compiled by make.bash, this is the shape of
the tail of the distribution for #blocks, #vars, and their
product:
#blocks #vars product
max 6171 28180 173,898,780
99.9% 1641 6548 10,401,878
99% 463 1909 873,721
95% 152 639 95,235
90% 84 359 30,021
The old algorithm is indeed usually fastest, for 99%ile
values of usually.
The fix to LookupVarOutgoing
( https://go-review.googlesource.com/#/c/22790/ )
deals with some of the same problems addressed by this CL,
but on at least one bug ( #15537 ) this change is still
a significant help.
With this CL:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 4m35.200s
user 13m16.644s
sys 0m36.712s
and pprof reports 3.4GB allocated in one of the larger profiles
With tip:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 10m36.569s
user 25m52.286s
sys 4m3.696s
and pprof reports 8.3GB allocated in the same larger profile
With this CL, most of the compilation time on the benchmarked
input is spent in register/stack allocation (cumulative 53%)
and in the sparse lookup algorithm itself (cumulative 20%).
Fixes #15537.
Change-Id: Ia0299dda6a291534d8b08e5f9883216ded677a00
Reviewed-on: https://go-review.googlesource.com/22342
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-04-21 13:24:58 -04:00
|
|
|
s.linkForwardReferences(sparseDefState)
|
|
|
|
|
|
|
|
if ssa.BuildStats > 0 {
|
|
|
|
s.f.LogStat("build", s.f.NumBlocks(), "blocks", prelinkNumvars, "vars_before",
|
|
|
|
s.f.NumValues(), "vars_after", prelinkNumvars*s.f.NumBlocks(), "ssa_phi_loc_cutoff_score")
|
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-10-08 12:39:56 -04:00
|
|
|
// Don't carry reference this around longer than necessary
|
2016-02-29 13:31:48 -08:00
|
|
|
s.exitCode = Nodes{}
|
2015-10-08 12:39:56 -04:00
|
|
|
|
2015-07-17 16:47:43 +00:00
|
|
|
// Main call to ssa package to compile function
|
|
|
|
ssa.Compile(s.f)
|
|
|
|
|
2015-12-11 20:41:52 -08:00
|
|
|
return s.f
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
|
2015-05-18 16:44:20 -07:00
|
|
|
type state struct {
|
2015-04-15 15:51:25 -07:00
|
|
|
// configuration (arch) information
|
|
|
|
config *ssa.Config
|
|
|
|
|
|
|
|
// function we're building
|
|
|
|
f *ssa.Func
|
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
// labels and labeled control flow nodes (OFOR, OSWITCH, OSELECT) in f
|
|
|
|
labels map[string]*ssaLabel
|
|
|
|
labeledNodes map[*Node]*ssaLabel
|
|
|
|
|
|
|
|
// gotos that jump forward; required for deferred checkgoto calls
|
|
|
|
fwdGotos []*Node
|
2015-10-08 12:39:56 -04:00
|
|
|
// Code that must precede any return
|
|
|
|
// (e.g., copying value of heap-escaped paramout back to true paramout)
|
2016-02-29 13:31:48 -08:00
|
|
|
exitCode Nodes
|
2015-07-20 15:39:14 -07:00
|
|
|
|
|
|
|
// unlabeled break and continue statement tracking
|
|
|
|
breakTo *ssa.Block // current target for plain break statement
|
|
|
|
continueTo *ssa.Block // current target for plain continue statement
|
2015-04-15 15:51:25 -07:00
|
|
|
|
|
|
|
// current location where we're interpreting the AST
|
|
|
|
curBlock *ssa.Block
|
|
|
|
|
2015-06-19 21:02:28 -07:00
|
|
|
// variable assignments in the current block (map from variable symbol to ssa value)
|
|
|
|
// *Node is the unique identifier (an ONAME Node) for the variable.
|
|
|
|
vars map[*Node]*ssa.Value
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// all defined variables at the end of each block. Indexed by block ID.
|
2015-06-19 21:02:28 -07:00
|
|
|
defvars []map[*Node]*ssa.Value
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-08-24 02:16:19 -07:00
|
|
|
// addresses of PPARAM and PPARAMOUT variables.
|
2015-06-19 21:02:28 -07:00
|
|
|
decladdrs map[*Node]*ssa.Value
|
2015-05-18 16:44:20 -07:00
|
|
|
|
2015-10-15 20:25:32 -05:00
|
|
|
// symbols for PEXTERN, PAUTO and PPARAMOUT variables so they can be reused.
|
|
|
|
varsyms map[*Node]interface{}
|
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// starting values. Memory, stack pointer, and globals pointer
|
2015-05-18 16:44:20 -07:00
|
|
|
startmem *ssa.Value
|
|
|
|
sp *ssa.Value
|
2015-06-19 21:02:28 -07:00
|
|
|
sb *ssa.Value
|
2015-05-30 01:03:06 -04:00
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// line number stack. The current line number is top of stack
|
2015-05-30 01:03:06 -04:00
|
|
|
line []int32
|
2015-11-09 21:35:40 -08:00
|
|
|
|
|
|
|
// list of panic calls by function name and line number.
|
|
|
|
// Used to deduplicate panic calls.
|
|
|
|
panics map[funcLine]*ssa.Block
|
2016-01-14 16:02:23 -08:00
|
|
|
|
|
|
|
// list of FwdRef values.
|
|
|
|
fwdRefs []*ssa.Value
|
2016-02-27 17:49:31 -08:00
|
|
|
|
cmd/compile: fix liveness computation for heap-escaped parameters
The liveness computation of parameters generally was never
correct, but forcing all parameters to be live throughout the
function covered up that problem. The new SSA back end is
too clever: even though it currently keeps the parameter values live
throughout the function, it may find optimizations that mean
the current values are not written back to the original parameter
stack slots immediately or ever (for example if a parameter is set
to nil, SSA constant propagation may replace all later uses of the
parameter with a constant nil, eliminating the need to write the nil
value back to the stack slot), so the liveness code must now
track the actual operations on the stack slots, exposing these
problems.
One small problem in the handling of arguments is that nodarg
can return ONAME PPARAM nodes with adjusted offsets, so that
there are actually multiple *Node pointers for the same parameter
in the instruction stream. This might be possible to correct, but
not in this CL. For now, we fix this by using n.Orig instead of n
when considering PPARAM and PPARAMOUT nodes.
The major problem in the handling of arguments is general
confusion in the liveness code about the meaning of PPARAM|PHEAP
and PPARAMOUT|PHEAP nodes, especially as contrasted with PAUTO|PHEAP.
The difference between these two is that when a local variable "moves"
to the heap, it's really just allocated there to start with; in contrast,
when an argument moves to the heap, the actual data has to be copied
there from the stack at the beginning of the function, and when a
result "moves" to the heap the value in the heap has to be copied
back to the stack when the function returns
This general confusion is also present in the SSA back end.
The PHEAP bit worked decently when I first introduced it 7 years ago (!)
in 391425ae. The back end did nothing sophisticated, and in particular
there was no analysis at all: no escape analysis, no liveness analysis,
and certainly no SSA back end. But the complications caused in the
various downstream consumers suggest that this should be a detail
kept mainly in the front end.
This CL therefore eliminates both the PHEAP bit and even the idea of
"heap variables" from the back ends.
First, it replaces the PPARAM|PHEAP, PPARAMOUT|PHEAP, and PAUTO|PHEAP
variable classes with the single PAUTOHEAP, a pseudo-class indicating
a variable maintained on the heap and available by indirecting a
local variable kept on the stack (a plain PAUTO).
Second, walkexpr replaces all references to PAUTOHEAP variables
with indirections of the corresponding PAUTO variable.
The back ends and the liveness code now just see plain indirected
variables. This may actually produce better code, but the real goal
here is to eliminate these little-used and somewhat suspect code
paths in the back end analyses.
The OPARAM node type goes away too.
A followup CL will do the same to PPARAMREF. I'm not sure that
the back ends (SSA in particular) are handling those right either,
and with the framework established in this CL that change is trivial
and the result clearly more correct.
Fixes #15747.
Change-Id: I2770b1ce3cbc93981bfc7166be66a9da12013d74
Reviewed-on: https://go-review.googlesource.com/23393
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-05-25 01:33:24 -04:00
|
|
|
// list of PPARAMOUT (return) variables.
|
2016-02-27 17:49:31 -08:00
|
|
|
returns []*Node
|
|
|
|
|
2016-04-21 19:28:28 -07:00
|
|
|
// list of PPARAM SSA-able pointer-shaped args. We ensure these are live
|
|
|
|
// throughout the function to help users avoid premature finalizers.
|
|
|
|
ptrargs []*Node
|
|
|
|
|
2016-02-27 17:49:31 -08:00
|
|
|
cgoUnsafeArgs bool
|
2016-03-16 21:51:17 -07:00
|
|
|
noWB bool
|
|
|
|
WBLineno int32 // line number of first write barrier. 0=no write barriers
|
2015-11-09 21:35:40 -08:00
|
|
|
}
|
|
|
|
|
|
|
|
type funcLine struct {
|
|
|
|
f *Node
|
|
|
|
line int32
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
type ssaLabel struct {
|
|
|
|
target *ssa.Block // block identified by this label
|
|
|
|
breakTarget *ssa.Block // block to break to in control flow node identified by this label
|
|
|
|
continueTarget *ssa.Block // block to continue to in control flow node identified by this label
|
|
|
|
defNode *Node // label definition Node (OLABEL)
|
|
|
|
// Label use Node (OGOTO, OBREAK, OCONTINUE).
|
|
|
|
// Used only for error detection and reporting.
|
|
|
|
// There might be multiple uses, but we only need to track one.
|
|
|
|
useNode *Node
|
|
|
|
reported bool // reported indicates whether an error has already been reported for this label
|
|
|
|
}
|
|
|
|
|
|
|
|
// defined reports whether the label has a definition (OLABEL node).
|
|
|
|
func (l *ssaLabel) defined() bool { return l.defNode != nil }
|
|
|
|
|
|
|
|
// used reports whether the label has a use (OGOTO, OBREAK, or OCONTINUE node).
|
|
|
|
func (l *ssaLabel) used() bool { return l.useNode != nil }
|
|
|
|
|
|
|
|
// label returns the label associated with sym, creating it if necessary.
|
|
|
|
func (s *state) label(sym *Sym) *ssaLabel {
|
|
|
|
lab := s.labels[sym.Name]
|
|
|
|
if lab == nil {
|
|
|
|
lab = new(ssaLabel)
|
|
|
|
s.labels[sym.Name] = lab
|
|
|
|
}
|
|
|
|
return lab
|
|
|
|
}
|
|
|
|
|
2016-01-13 11:14:57 -08:00
|
|
|
func (s *state) Logf(msg string, args ...interface{}) { s.config.Logf(msg, args...) }
|
2016-01-29 14:44:15 -05:00
|
|
|
func (s *state) Log() bool { return s.config.Log() }
|
2016-01-13 11:14:57 -08:00
|
|
|
func (s *state) Fatalf(msg string, args ...interface{}) { s.config.Fatalf(s.peekLine(), msg, args...) }
|
|
|
|
func (s *state) Unimplementedf(msg string, args ...interface{}) {
|
|
|
|
s.config.Unimplementedf(s.peekLine(), msg, args...)
|
|
|
|
}
|
2016-03-13 23:04:31 -05:00
|
|
|
func (s *state) Warnl(line int32, msg string, args ...interface{}) { s.config.Warnl(line, msg, args...) }
|
|
|
|
func (s *state) Debug_checknil() bool { return s.config.Debug_checknil() }
|
2015-06-12 11:01:13 -07:00
|
|
|
|
2015-09-17 10:31:16 -07:00
|
|
|
var (
|
|
|
|
// dummy node for the memory variable
|
2015-10-22 14:22:38 -07:00
|
|
|
memVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "mem"}}
|
2015-09-17 10:31:16 -07:00
|
|
|
|
|
|
|
// dummy nodes for temporary variables
|
cmd/compile: avoid a spill in append fast path
Instead of spilling newlen, recalculate it.
This removes a spill from the fast path,
at the cost of a cheap recalculation
on the (rare) growth path.
This uses 8 bytes less of stack space.
It generates two more bytes of code,
but that is due to suboptimal register allocation;
see far below.
Runtime append microbenchmarks are all over the map,
presumably due to incidental code movement.
Sample code:
func s(b []byte) []byte {
b = append(b, 1, 2, 3)
return b
}
Before:
"".s t=1 size=160 args=0x30 locals=0x48
0x0000 00000 (append.go:8) TEXT "".s(SB), $72-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 149
0x0013 00019 (append.go:8) SUBQ $72, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+88(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ DX, "".autotmp_0+64(SP)
0x0025 00037 (append.go:9) MOVQ "".b+96(FP), BX
0x002a 00042 (append.go:9) CMPQ DX, BX
0x002d 00045 (append.go:9) JGT $0, 86
0x002f 00047 (append.go:8) MOVQ "".b+80(FP), AX
0x0034 00052 (append.go:9) MOVB $1, (AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x003d 00061 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x0042 00066 (append.go:10) MOVQ AX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ DX, "".~r1+112(FP)
0x004c 00076 (append.go:10) MOVQ BX, "".~r1+120(FP)
0x0051 00081 (append.go:10) ADDQ $72, SP
0x0055 00085 (append.go:10) RET
0x0056 00086 (append.go:9) LEAQ type.[]uint8(SB), AX
0x005d 00093 (append.go:9) MOVQ AX, (SP)
0x0061 00097 (append.go:9) MOVQ "".b+80(FP), BP
0x0066 00102 (append.go:9) MOVQ BP, 8(SP)
0x006b 00107 (append.go:9) MOVQ CX, 16(SP)
0x0070 00112 (append.go:9) MOVQ BX, 24(SP)
0x0075 00117 (append.go:9) MOVQ DX, 32(SP)
0x007a 00122 (append.go:9) PCDATA $0, $0
0x007a 00122 (append.go:9) CALL runtime.growslice(SB)
0x007f 00127 (append.go:9) MOVQ 40(SP), AX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:8) MOVQ "".b+88(FP), CX
0x008e 00142 (append.go:9) MOVQ "".autotmp_0+64(SP), DX
0x0093 00147 (append.go:9) JMP 52
0x0095 00149 (append.go:9) NOP
0x0095 00149 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009a 00154 (append.go:8) JMP 0
After:
"".s t=1 size=176 args=0x30 locals=0x40
0x0000 00000 (append.go:8) TEXT "".s(SB), $64-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 151
0x0013 00019 (append.go:8) SUBQ $64, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+80(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ "".b+88(FP), BX
0x0025 00037 (append.go:9) CMPQ DX, BX
0x0028 00040 (append.go:9) JGT $0, 81
0x002a 00042 (append.go:8) MOVQ "".b+72(FP), AX
0x002f 00047 (append.go:9) MOVB $1, (AX)(CX*1)
0x0033 00051 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x003d 00061 (append.go:10) MOVQ AX, "".~r1+96(FP)
0x0042 00066 (append.go:10) MOVQ DX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ BX, "".~r1+112(FP)
0x004c 00076 (append.go:10) ADDQ $64, SP
0x0050 00080 (append.go:10) RET
0x0051 00081 (append.go:9) LEAQ type.[]uint8(SB), AX
0x0058 00088 (append.go:9) MOVQ AX, (SP)
0x005c 00092 (append.go:9) MOVQ "".b+72(FP), BP
0x0061 00097 (append.go:9) MOVQ BP, 8(SP)
0x0066 00102 (append.go:9) MOVQ CX, 16(SP)
0x006b 00107 (append.go:9) MOVQ BX, 24(SP)
0x0070 00112 (append.go:9) MOVQ DX, 32(SP)
0x0075 00117 (append.go:9) PCDATA $0, $0
0x0075 00117 (append.go:9) CALL runtime.growslice(SB)
0x007a 00122 (append.go:9) MOVQ 40(SP), AX
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
0x0095 00149 (append.go:9) JMP 47
0x0097 00151 (append.go:9) NOP
0x0097 00151 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009c 00156 (append.go:8) JMP 0
Observe that in the following sequence,
we should use DX directly instead of using
CX as a temporary register, which would make
the new code a strict improvement on the old:
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
Change-Id: I4ee50b18fa53865901d2d7f86c2cbb54c6fa6924
Reviewed-on: https://go-review.googlesource.com/21812
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:08:00 -07:00
|
|
|
ptrVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "ptr"}}
|
cmd/compile: re-enable in-place append optimization
CL 21891 was too clever in its attempts to avoid spills.
Storing newlen too early caused uses of append in the runtime
itself to receive an inconsistent view of a slice,
leading to corruption.
This CL makes the generate code much more similar to
the old backend. It spills more than before,
but those spills have been contained to the grow path.
It recalculates newlen unnecessarily on the fast path,
but that's measurably cheaper than spilling it.
CL 21891 caused runtime failures in 6 of 2000 runs
of net/http and crypto/x509 in my test setup.
This CL has gone 6000 runs without a failure.
Benchmarks going from master to this CL:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 439ns ± 2% 436ns ± 2% -0.72% (p=0.001 n=28+27)
AppendInPlace/NoGrow/1Ptr-8 901ns ± 0% 856ns ± 0% -4.95% (p=0.000 n=26+29)
AppendInPlace/NoGrow/2Ptr-8 2.15µs ± 1% 1.95µs ± 0% -9.07% (p=0.000 n=28+30)
AppendInPlace/NoGrow/3Ptr-8 2.66µs ± 0% 2.45µs ± 0% -7.93% (p=0.000 n=29+26)
AppendInPlace/NoGrow/4Ptr-8 3.24µs ± 1% 3.02µs ± 1% -6.75% (p=0.000 n=28+30)
AppendInPlace/Grow/Byte-8 269ns ± 1% 271ns ± 1% +0.84% (p=0.000 n=30+29)
AppendInPlace/Grow/1Ptr-8 275ns ± 1% 280ns ± 1% +1.75% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 384ns ± 0% 391ns ± 0% +1.94% (p=0.000 n=27+30)
AppendInPlace/Grow/3Ptr-8 455ns ± 0% 462ns ± 0% +1.43% (p=0.000 n=29+29)
AppendInPlace/Grow/4Ptr-8 478ns ± 0% 479ns ± 0% +0.23% (p=0.000 n=30+27)
However, for the large no-grow cases, there is still more work to be done.
Going from this CL to the non-SSA backend:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 436ns ± 2% 436ns ± 2% ~ (p=0.967 n=27+29)
AppendInPlace/NoGrow/1Ptr-8 856ns ± 0% 884ns ± 0% +3.28% (p=0.000 n=29+26)
AppendInPlace/NoGrow/2Ptr-8 1.95µs ± 0% 1.56µs ± 0% -20.28% (p=0.000 n=30+29)
AppendInPlace/NoGrow/3Ptr-8 2.45µs ± 0% 1.89µs ± 0% -22.88% (p=0.000 n=26+28)
AppendInPlace/NoGrow/4Ptr-8 3.02µs ± 1% 2.56µs ± 1% -15.35% (p=0.000 n=30+28)
AppendInPlace/Grow/Byte-8 271ns ± 1% 283ns ± 1% +4.56% (p=0.000 n=29+29)
AppendInPlace/Grow/1Ptr-8 280ns ± 1% 288ns ± 1% +2.99% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 391ns ± 0% 409ns ± 0% +4.66% (p=0.000 n=30+29)
AppendInPlace/Grow/3Ptr-8 462ns ± 0% 481ns ± 0% +4.13% (p=0.000 n=29+30)
AppendInPlace/Grow/4Ptr-8 479ns ± 0% 502ns ± 0% +4.81% (p=0.000 n=27+26)
New generated code:
var x []byte
func a() {
x = append(x, 1)
}
"".a t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (a.go:5) TEXT "".a(SB), $72-0
0x0000 00000 (a.go:5) MOVQ (TLS), CX
0x0009 00009 (a.go:5) CMPQ SP, 16(CX)
0x000d 00013 (a.go:5) JLS 190
0x0013 00019 (a.go:5) SUBQ $72, SP
0x0017 00023 (a.go:5) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:5) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:6) MOVQ "".x+16(SB), CX
0x001e 00030 (a.go:6) MOVQ "".x+8(SB), DX
0x0025 00037 (a.go:6) MOVQ "".x(SB), BX
0x002c 00044 (a.go:6) LEAQ 1(DX), BP
0x0030 00048 (a.go:6) CMPQ BP, CX
0x0033 00051 (a.go:6) JGT $0, 73
0x0035 00053 (a.go:6) LEAQ 1(DX), AX
0x0039 00057 (a.go:6) MOVQ AX, "".x+8(SB)
0x0040 00064 (a.go:6) MOVB $1, (BX)(DX*1)
0x0044 00068 (a.go:7) ADDQ $72, SP
0x0048 00072 (a.go:7) RET
0x0049 00073 (a.go:6) LEAQ type.[]uint8(SB), AX
0x0050 00080 (a.go:6) MOVQ AX, (SP)
0x0054 00084 (a.go:6) MOVQ BX, 8(SP)
0x0059 00089 (a.go:6) MOVQ DX, 16(SP)
0x005e 00094 (a.go:6) MOVQ CX, 24(SP)
0x0063 00099 (a.go:6) MOVQ BP, 32(SP)
0x0068 00104 (a.go:6) PCDATA $0, $0
0x0068 00104 (a.go:6) CALL runtime.growslice(SB)
0x006d 00109 (a.go:6) MOVQ 40(SP), CX
0x0072 00114 (a.go:6) MOVQ 48(SP), DX
0x0077 00119 (a.go:6) MOVQ DX, "".autotmp_0+64(SP)
0x007c 00124 (a.go:6) MOVQ 56(SP), BX
0x0081 00129 (a.go:6) MOVQ BX, "".x+16(SB)
0x0088 00136 (a.go:6) MOVL runtime.writeBarrier(SB), AX
0x008e 00142 (a.go:6) TESTB AL, AL
0x0090 00144 (a.go:6) JNE $0, 162
0x0092 00146 (a.go:6) MOVQ CX, "".x(SB)
0x0099 00153 (a.go:6) MOVQ "".x(SB), BX
0x00a0 00160 (a.go:6) JMP 53
0x00a2 00162 (a.go:6) LEAQ "".x(SB), BX
0x00a9 00169 (a.go:6) MOVQ BX, (SP)
0x00ad 00173 (a.go:6) MOVQ CX, 8(SP)
0x00b2 00178 (a.go:6) PCDATA $0, $0
0x00b2 00178 (a.go:6) CALL runtime.writebarrierptr(SB)
0x00b7 00183 (a.go:6) MOVQ "".autotmp_0+64(SP), DX
0x00bc 00188 (a.go:6) JMP 153
0x00be 00190 (a.go:6) NOP
0x00be 00190 (a.go:5) CALL runtime.morestack_noctxt(SB)
0x00c3 00195 (a.go:5) JMP 0
Fixes #14969 again
Change-Id: Ia50463b1f506011aad0718a4fef1d4738e43c32d
Reviewed-on: https://go-review.googlesource.com/22197
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-18 09:40:30 -07:00
|
|
|
lenVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "len"}}
|
cmd/compile: avoid a spill in append fast path
Instead of spilling newlen, recalculate it.
This removes a spill from the fast path,
at the cost of a cheap recalculation
on the (rare) growth path.
This uses 8 bytes less of stack space.
It generates two more bytes of code,
but that is due to suboptimal register allocation;
see far below.
Runtime append microbenchmarks are all over the map,
presumably due to incidental code movement.
Sample code:
func s(b []byte) []byte {
b = append(b, 1, 2, 3)
return b
}
Before:
"".s t=1 size=160 args=0x30 locals=0x48
0x0000 00000 (append.go:8) TEXT "".s(SB), $72-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 149
0x0013 00019 (append.go:8) SUBQ $72, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+88(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ DX, "".autotmp_0+64(SP)
0x0025 00037 (append.go:9) MOVQ "".b+96(FP), BX
0x002a 00042 (append.go:9) CMPQ DX, BX
0x002d 00045 (append.go:9) JGT $0, 86
0x002f 00047 (append.go:8) MOVQ "".b+80(FP), AX
0x0034 00052 (append.go:9) MOVB $1, (AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x003d 00061 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x0042 00066 (append.go:10) MOVQ AX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ DX, "".~r1+112(FP)
0x004c 00076 (append.go:10) MOVQ BX, "".~r1+120(FP)
0x0051 00081 (append.go:10) ADDQ $72, SP
0x0055 00085 (append.go:10) RET
0x0056 00086 (append.go:9) LEAQ type.[]uint8(SB), AX
0x005d 00093 (append.go:9) MOVQ AX, (SP)
0x0061 00097 (append.go:9) MOVQ "".b+80(FP), BP
0x0066 00102 (append.go:9) MOVQ BP, 8(SP)
0x006b 00107 (append.go:9) MOVQ CX, 16(SP)
0x0070 00112 (append.go:9) MOVQ BX, 24(SP)
0x0075 00117 (append.go:9) MOVQ DX, 32(SP)
0x007a 00122 (append.go:9) PCDATA $0, $0
0x007a 00122 (append.go:9) CALL runtime.growslice(SB)
0x007f 00127 (append.go:9) MOVQ 40(SP), AX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:8) MOVQ "".b+88(FP), CX
0x008e 00142 (append.go:9) MOVQ "".autotmp_0+64(SP), DX
0x0093 00147 (append.go:9) JMP 52
0x0095 00149 (append.go:9) NOP
0x0095 00149 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009a 00154 (append.go:8) JMP 0
After:
"".s t=1 size=176 args=0x30 locals=0x40
0x0000 00000 (append.go:8) TEXT "".s(SB), $64-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 151
0x0013 00019 (append.go:8) SUBQ $64, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+80(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ "".b+88(FP), BX
0x0025 00037 (append.go:9) CMPQ DX, BX
0x0028 00040 (append.go:9) JGT $0, 81
0x002a 00042 (append.go:8) MOVQ "".b+72(FP), AX
0x002f 00047 (append.go:9) MOVB $1, (AX)(CX*1)
0x0033 00051 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x003d 00061 (append.go:10) MOVQ AX, "".~r1+96(FP)
0x0042 00066 (append.go:10) MOVQ DX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ BX, "".~r1+112(FP)
0x004c 00076 (append.go:10) ADDQ $64, SP
0x0050 00080 (append.go:10) RET
0x0051 00081 (append.go:9) LEAQ type.[]uint8(SB), AX
0x0058 00088 (append.go:9) MOVQ AX, (SP)
0x005c 00092 (append.go:9) MOVQ "".b+72(FP), BP
0x0061 00097 (append.go:9) MOVQ BP, 8(SP)
0x0066 00102 (append.go:9) MOVQ CX, 16(SP)
0x006b 00107 (append.go:9) MOVQ BX, 24(SP)
0x0070 00112 (append.go:9) MOVQ DX, 32(SP)
0x0075 00117 (append.go:9) PCDATA $0, $0
0x0075 00117 (append.go:9) CALL runtime.growslice(SB)
0x007a 00122 (append.go:9) MOVQ 40(SP), AX
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
0x0095 00149 (append.go:9) JMP 47
0x0097 00151 (append.go:9) NOP
0x0097 00151 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009c 00156 (append.go:8) JMP 0
Observe that in the following sequence,
we should use DX directly instead of using
CX as a temporary register, which would make
the new code a strict improvement on the old:
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
Change-Id: I4ee50b18fa53865901d2d7f86c2cbb54c6fa6924
Reviewed-on: https://go-review.googlesource.com/21812
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:08:00 -07:00
|
|
|
newlenVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "newlen"}}
|
|
|
|
capVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "cap"}}
|
|
|
|
typVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "typ"}}
|
|
|
|
idataVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "idata"}}
|
|
|
|
okVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "ok"}}
|
|
|
|
deltaVar = Node{Op: ONAME, Class: Pxxx, Sym: &Sym{Name: "delta"}}
|
2015-09-17 10:31:16 -07:00
|
|
|
)
|
2015-09-12 23:27:26 -07:00
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
// startBlock sets the current block we're generating code in to b.
|
2015-05-18 16:44:20 -07:00
|
|
|
func (s *state) startBlock(b *ssa.Block) {
|
|
|
|
if s.curBlock != nil {
|
2015-06-24 14:03:39 -07:00
|
|
|
s.Fatalf("starting block %v when block %v has not ended", b, s.curBlock)
|
2015-05-18 16:44:20 -07:00
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
s.curBlock = b
|
2015-06-19 21:02:28 -07:00
|
|
|
s.vars = map[*Node]*ssa.Value{}
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
// endBlock marks the end of generating code for the current block.
|
2016-03-01 23:21:55 +00:00
|
|
|
// Returns the (former) current block. Returns nil if there is no current
|
2015-04-15 15:51:25 -07:00
|
|
|
// block, i.e. if no code flows to the current execution point.
|
2015-05-18 16:44:20 -07:00
|
|
|
func (s *state) endBlock() *ssa.Block {
|
2015-04-15 15:51:25 -07:00
|
|
|
b := s.curBlock
|
|
|
|
if b == nil {
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
for len(s.defvars) <= int(b.ID) {
|
|
|
|
s.defvars = append(s.defvars, nil)
|
|
|
|
}
|
|
|
|
s.defvars[b.ID] = s.vars
|
|
|
|
s.curBlock = nil
|
|
|
|
s.vars = nil
|
2015-05-30 01:03:06 -04:00
|
|
|
b.Line = s.peekLine()
|
2015-04-15 15:51:25 -07:00
|
|
|
return b
|
|
|
|
}
|
|
|
|
|
2015-05-30 01:03:06 -04:00
|
|
|
// pushLine pushes a line number on the line number stack.
|
|
|
|
func (s *state) pushLine(line int32) {
|
2016-06-30 06:36:31 -04:00
|
|
|
if line == 0 {
|
|
|
|
// the frontend may emit node with line number missing,
|
|
|
|
// use the parent line number in this case.
|
|
|
|
line = s.peekLine()
|
|
|
|
if Debug['K'] != 0 {
|
|
|
|
Warn("buildssa: line 0")
|
|
|
|
}
|
|
|
|
}
|
2015-05-30 01:03:06 -04:00
|
|
|
s.line = append(s.line, line)
|
|
|
|
}
|
|
|
|
|
|
|
|
// popLine pops the top of the line number stack.
|
|
|
|
func (s *state) popLine() {
|
|
|
|
s.line = s.line[:len(s.line)-1]
|
|
|
|
}
|
|
|
|
|
|
|
|
// peekLine peek the top of the line number stack.
|
|
|
|
func (s *state) peekLine() int32 {
|
|
|
|
return s.line[len(s.line)-1]
|
|
|
|
}
|
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
func (s *state) Error(msg string, args ...interface{}) {
|
2016-03-02 11:01:25 -08:00
|
|
|
yyerrorl(s.peekLine(), msg, args...)
|
2015-07-20 15:39:14 -07:00
|
|
|
}
|
|
|
|
|
2015-06-11 21:29:25 -07:00
|
|
|
// newValue0 adds a new value with no arguments to the current block.
|
|
|
|
func (s *state) newValue0(op ssa.Op, t ssa.Type) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue0(s.peekLine(), op, t)
|
|
|
|
}
|
|
|
|
|
|
|
|
// newValue0A adds a new value with no arguments and an aux value to the current block.
|
|
|
|
func (s *state) newValue0A(op ssa.Op, t ssa.Type, aux interface{}) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue0A(s.peekLine(), op, t, aux)
|
2015-05-30 01:03:06 -04:00
|
|
|
}
|
|
|
|
|
2015-09-03 18:24:22 -05:00
|
|
|
// newValue0I adds a new value with no arguments and an auxint value to the current block.
|
|
|
|
func (s *state) newValue0I(op ssa.Op, t ssa.Type, auxint int64) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue0I(s.peekLine(), op, t, auxint)
|
|
|
|
}
|
|
|
|
|
2015-05-30 01:03:06 -04:00
|
|
|
// newValue1 adds a new value with one argument to the current block.
|
2015-06-11 21:29:25 -07:00
|
|
|
func (s *state) newValue1(op ssa.Op, t ssa.Type, arg *ssa.Value) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue1(s.peekLine(), op, t, arg)
|
|
|
|
}
|
|
|
|
|
|
|
|
// newValue1A adds a new value with one argument and an aux value to the current block.
|
|
|
|
func (s *state) newValue1A(op ssa.Op, t ssa.Type, aux interface{}, arg *ssa.Value) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue1A(s.peekLine(), op, t, aux, arg)
|
2015-05-30 01:03:06 -04:00
|
|
|
}
|
|
|
|
|
2015-07-15 21:33:49 -07:00
|
|
|
// newValue1I adds a new value with one argument and an auxint value to the current block.
|
|
|
|
func (s *state) newValue1I(op ssa.Op, t ssa.Type, aux int64, arg *ssa.Value) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue1I(s.peekLine(), op, t, aux, arg)
|
|
|
|
}
|
|
|
|
|
2015-05-30 01:03:06 -04:00
|
|
|
// newValue2 adds a new value with two arguments to the current block.
|
2015-06-11 21:29:25 -07:00
|
|
|
func (s *state) newValue2(op ssa.Op, t ssa.Type, arg0, arg1 *ssa.Value) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue2(s.peekLine(), op, t, arg0, arg1)
|
2015-05-30 01:03:06 -04:00
|
|
|
}
|
|
|
|
|
2015-06-27 15:45:20 +01:00
|
|
|
// newValue2I adds a new value with two arguments and an auxint value to the current block.
|
|
|
|
func (s *state) newValue2I(op ssa.Op, t ssa.Type, aux int64, arg0, arg1 *ssa.Value) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue2I(s.peekLine(), op, t, aux, arg0, arg1)
|
|
|
|
}
|
|
|
|
|
2015-05-30 01:03:06 -04:00
|
|
|
// newValue3 adds a new value with three arguments to the current block.
|
2015-06-11 21:29:25 -07:00
|
|
|
func (s *state) newValue3(op ssa.Op, t ssa.Type, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue3(s.peekLine(), op, t, arg0, arg1, arg2)
|
2015-05-30 01:03:06 -04:00
|
|
|
}
|
|
|
|
|
2015-08-14 21:47:20 -07:00
|
|
|
// newValue3I adds a new value with three arguments and an auxint value to the current block.
|
|
|
|
func (s *state) newValue3I(op ssa.Op, t ssa.Type, aux int64, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue3I(s.peekLine(), op, t, aux, arg0, arg1, arg2)
|
|
|
|
}
|
|
|
|
|
2016-08-25 16:02:57 -07:00
|
|
|
// newValue4 adds a new value with four arguments to the current block.
|
|
|
|
func (s *state) newValue4(op ssa.Op, t ssa.Type, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
|
|
|
|
return s.curBlock.NewValue4(s.peekLine(), op, t, arg0, arg1, arg2, arg3)
|
|
|
|
}
|
|
|
|
|
2015-09-03 18:24:22 -05:00
|
|
|
// entryNewValue0 adds a new value with no arguments to the entry block.
|
2015-06-11 21:29:25 -07:00
|
|
|
func (s *state) entryNewValue0(op ssa.Op, t ssa.Type) *ssa.Value {
|
|
|
|
return s.f.Entry.NewValue0(s.peekLine(), op, t)
|
|
|
|
}
|
|
|
|
|
2015-09-03 18:24:22 -05:00
|
|
|
// entryNewValue0A adds a new value with no arguments and an aux value to the entry block.
|
2015-06-11 21:29:25 -07:00
|
|
|
func (s *state) entryNewValue0A(op ssa.Op, t ssa.Type, aux interface{}) *ssa.Value {
|
|
|
|
return s.f.Entry.NewValue0A(s.peekLine(), op, t, aux)
|
2015-05-30 01:03:06 -04:00
|
|
|
}
|
|
|
|
|
2015-09-03 18:24:22 -05:00
|
|
|
// entryNewValue0I adds a new value with no arguments and an auxint value to the entry block.
|
|
|
|
func (s *state) entryNewValue0I(op ssa.Op, t ssa.Type, auxint int64) *ssa.Value {
|
|
|
|
return s.f.Entry.NewValue0I(s.peekLine(), op, t, auxint)
|
|
|
|
}
|
|
|
|
|
2015-05-30 01:03:06 -04:00
|
|
|
// entryNewValue1 adds a new value with one argument to the entry block.
|
2015-06-11 21:29:25 -07:00
|
|
|
func (s *state) entryNewValue1(op ssa.Op, t ssa.Type, arg *ssa.Value) *ssa.Value {
|
|
|
|
return s.f.Entry.NewValue1(s.peekLine(), op, t, arg)
|
|
|
|
}
|
|
|
|
|
|
|
|
// entryNewValue1 adds a new value with one argument and an auxint value to the entry block.
|
|
|
|
func (s *state) entryNewValue1I(op ssa.Op, t ssa.Type, auxint int64, arg *ssa.Value) *ssa.Value {
|
|
|
|
return s.f.Entry.NewValue1I(s.peekLine(), op, t, auxint, arg)
|
2015-05-30 01:03:06 -04:00
|
|
|
}
|
|
|
|
|
2015-06-19 21:02:28 -07:00
|
|
|
// entryNewValue1A adds a new value with one argument and an aux value to the entry block.
|
|
|
|
func (s *state) entryNewValue1A(op ssa.Op, t ssa.Type, aux interface{}, arg *ssa.Value) *ssa.Value {
|
|
|
|
return s.f.Entry.NewValue1A(s.peekLine(), op, t, aux, arg)
|
|
|
|
}
|
|
|
|
|
2015-05-30 01:03:06 -04:00
|
|
|
// entryNewValue2 adds a new value with two arguments to the entry block.
|
2015-06-11 21:29:25 -07:00
|
|
|
func (s *state) entryNewValue2(op ssa.Op, t ssa.Type, arg0, arg1 *ssa.Value) *ssa.Value {
|
|
|
|
return s.f.Entry.NewValue2(s.peekLine(), op, t, arg0, arg1)
|
2015-05-30 01:03:06 -04:00
|
|
|
}
|
|
|
|
|
2015-09-08 16:52:25 -07:00
|
|
|
// const* routines add a new const value to the entry block.
|
2016-03-06 18:06:09 -08:00
|
|
|
func (s *state) constSlice(t ssa.Type) *ssa.Value { return s.f.ConstSlice(s.peekLine(), t) }
|
|
|
|
func (s *state) constInterface(t ssa.Type) *ssa.Value { return s.f.ConstInterface(s.peekLine(), t) }
|
|
|
|
func (s *state) constNil(t ssa.Type) *ssa.Value { return s.f.ConstNil(s.peekLine(), t) }
|
|
|
|
func (s *state) constEmptyString(t ssa.Type) *ssa.Value { return s.f.ConstEmptyString(s.peekLine(), t) }
|
2015-09-08 16:52:25 -07:00
|
|
|
func (s *state) constBool(c bool) *ssa.Value {
|
|
|
|
return s.f.ConstBool(s.peekLine(), Types[TBOOL], c)
|
|
|
|
}
|
2015-07-28 14:19:20 -07:00
|
|
|
func (s *state) constInt8(t ssa.Type, c int8) *ssa.Value {
|
|
|
|
return s.f.ConstInt8(s.peekLine(), t, c)
|
|
|
|
}
|
|
|
|
func (s *state) constInt16(t ssa.Type, c int16) *ssa.Value {
|
|
|
|
return s.f.ConstInt16(s.peekLine(), t, c)
|
|
|
|
}
|
|
|
|
func (s *state) constInt32(t ssa.Type, c int32) *ssa.Value {
|
|
|
|
return s.f.ConstInt32(s.peekLine(), t, c)
|
|
|
|
}
|
|
|
|
func (s *state) constInt64(t ssa.Type, c int64) *ssa.Value {
|
|
|
|
return s.f.ConstInt64(s.peekLine(), t, c)
|
|
|
|
}
|
2015-08-12 16:38:11 -04:00
|
|
|
func (s *state) constFloat32(t ssa.Type, c float64) *ssa.Value {
|
|
|
|
return s.f.ConstFloat32(s.peekLine(), t, c)
|
|
|
|
}
|
|
|
|
func (s *state) constFloat64(t ssa.Type, c float64) *ssa.Value {
|
|
|
|
return s.f.ConstFloat64(s.peekLine(), t, c)
|
|
|
|
}
|
2015-05-30 01:03:06 -04:00
|
|
|
func (s *state) constInt(t ssa.Type, c int64) *ssa.Value {
|
2015-07-28 14:19:20 -07:00
|
|
|
if s.config.IntSize == 8 {
|
|
|
|
return s.constInt64(t, c)
|
|
|
|
}
|
|
|
|
if int64(int32(c)) != c {
|
|
|
|
s.Fatalf("integer constant too big %d", c)
|
|
|
|
}
|
|
|
|
return s.constInt32(t, int32(c))
|
2015-05-30 01:03:06 -04:00
|
|
|
}
|
|
|
|
|
2016-02-29 13:31:48 -08:00
|
|
|
func (s *state) stmts(a Nodes) {
|
|
|
|
for _, x := range a.Slice() {
|
|
|
|
s.stmt(x)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
// ssaStmtList converts the statement n to SSA and adds it to s.
|
2016-03-08 10:26:20 -08:00
|
|
|
func (s *state) stmtList(l Nodes) {
|
2016-03-08 15:10:26 -08:00
|
|
|
for _, n := range l.Slice() {
|
|
|
|
s.stmt(n)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// ssaStmt converts the statement n to SSA and adds it to s.
|
2015-05-18 16:44:20 -07:00
|
|
|
func (s *state) stmt(n *Node) {
|
2015-05-30 01:03:06 -04:00
|
|
|
s.pushLine(n.Lineno)
|
|
|
|
defer s.popLine()
|
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
// If s.curBlock is nil, then we're about to generate dead code.
|
|
|
|
// We can't just short-circuit here, though,
|
|
|
|
// because we check labels and gotos as part of SSA generation.
|
|
|
|
// Provide a block for the dead code so that we don't have
|
|
|
|
// to add special cases everywhere else.
|
|
|
|
if s.curBlock == nil {
|
|
|
|
dead := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
s.startBlock(dead)
|
|
|
|
}
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
s.stmtList(n.Ninit)
|
|
|
|
switch n.Op {
|
|
|
|
|
|
|
|
case OBLOCK:
|
|
|
|
s.stmtList(n.List)
|
|
|
|
|
2015-07-16 13:25:36 -06:00
|
|
|
// No-ops
|
2015-08-28 21:19:40 -05:00
|
|
|
case OEMPTY, ODCLCONST, ODCLTYPE, OFALL:
|
2015-07-10 10:47:28 -06:00
|
|
|
|
2015-07-16 13:25:36 -06:00
|
|
|
// Expression statements
|
2016-08-23 16:49:28 -07:00
|
|
|
case OCALLFUNC:
|
|
|
|
if isIntrinsicCall(n) {
|
|
|
|
s.intrinsicCall(n)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
fallthrough
|
|
|
|
|
|
|
|
case OCALLMETH, OCALLINTER:
|
2015-09-09 23:56:59 -07:00
|
|
|
s.call(n, callNormal)
|
2016-02-24 16:19:20 -08:00
|
|
|
if n.Op == OCALLFUNC && n.Left.Op == ONAME && n.Left.Class == PFUNC &&
|
2016-04-13 18:37:18 -07:00
|
|
|
(compiling_runtime && n.Left.Sym.Name == "throw" ||
|
2016-02-27 17:49:31 -08:00
|
|
|
n.Left.Sym.Pkg == Runtimepkg && (n.Left.Sym.Name == "gopanic" || n.Left.Sym.Name == "selectgo" || n.Left.Sym.Name == "block")) {
|
2016-02-06 22:35:34 -08:00
|
|
|
m := s.mem()
|
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockExit
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(m)
|
2016-02-06 22:35:34 -08:00
|
|
|
// TODO: never rewrite OPANIC to OCALLFUNC in the
|
2016-03-01 23:21:55 +00:00
|
|
|
// first place. Need to wait until all backends
|
2016-02-06 22:35:34 -08:00
|
|
|
// go through SSA.
|
|
|
|
}
|
2015-09-09 23:56:59 -07:00
|
|
|
case ODEFER:
|
|
|
|
s.call(n.Left, callDefer)
|
|
|
|
case OPROC:
|
|
|
|
s.call(n.Left, callGo)
|
2015-07-16 13:25:36 -06:00
|
|
|
|
2015-09-17 10:31:16 -07:00
|
|
|
case OAS2DOTTYPE:
|
2016-03-08 15:10:26 -08:00
|
|
|
res, resok := s.dottype(n.Rlist.First(), true)
|
2016-05-27 14:07:37 -07:00
|
|
|
s.assign(n.List.First(), res, needwritebarrier(n.List.First(), n.Rlist.First()), false, n.Lineno, 0, false)
|
|
|
|
s.assign(n.List.Second(), resok, false, false, n.Lineno, 0, false)
|
2015-09-17 10:31:16 -07:00
|
|
|
return
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
case ODCL:
|
cmd/compile: fix liveness computation for heap-escaped parameters
The liveness computation of parameters generally was never
correct, but forcing all parameters to be live throughout the
function covered up that problem. The new SSA back end is
too clever: even though it currently keeps the parameter values live
throughout the function, it may find optimizations that mean
the current values are not written back to the original parameter
stack slots immediately or ever (for example if a parameter is set
to nil, SSA constant propagation may replace all later uses of the
parameter with a constant nil, eliminating the need to write the nil
value back to the stack slot), so the liveness code must now
track the actual operations on the stack slots, exposing these
problems.
One small problem in the handling of arguments is that nodarg
can return ONAME PPARAM nodes with adjusted offsets, so that
there are actually multiple *Node pointers for the same parameter
in the instruction stream. This might be possible to correct, but
not in this CL. For now, we fix this by using n.Orig instead of n
when considering PPARAM and PPARAMOUT nodes.
The major problem in the handling of arguments is general
confusion in the liveness code about the meaning of PPARAM|PHEAP
and PPARAMOUT|PHEAP nodes, especially as contrasted with PAUTO|PHEAP.
The difference between these two is that when a local variable "moves"
to the heap, it's really just allocated there to start with; in contrast,
when an argument moves to the heap, the actual data has to be copied
there from the stack at the beginning of the function, and when a
result "moves" to the heap the value in the heap has to be copied
back to the stack when the function returns
This general confusion is also present in the SSA back end.
The PHEAP bit worked decently when I first introduced it 7 years ago (!)
in 391425ae. The back end did nothing sophisticated, and in particular
there was no analysis at all: no escape analysis, no liveness analysis,
and certainly no SSA back end. But the complications caused in the
various downstream consumers suggest that this should be a detail
kept mainly in the front end.
This CL therefore eliminates both the PHEAP bit and even the idea of
"heap variables" from the back ends.
First, it replaces the PPARAM|PHEAP, PPARAMOUT|PHEAP, and PAUTO|PHEAP
variable classes with the single PAUTOHEAP, a pseudo-class indicating
a variable maintained on the heap and available by indirecting a
local variable kept on the stack (a plain PAUTO).
Second, walkexpr replaces all references to PAUTOHEAP variables
with indirections of the corresponding PAUTO variable.
The back ends and the liveness code now just see plain indirected
variables. This may actually produce better code, but the real goal
here is to eliminate these little-used and somewhat suspect code
paths in the back end analyses.
The OPARAM node type goes away too.
A followup CL will do the same to PPARAMREF. I'm not sure that
the back ends (SSA in particular) are handling those right either,
and with the framework established in this CL that change is trivial
and the result clearly more correct.
Fixes #15747.
Change-Id: I2770b1ce3cbc93981bfc7166be66a9da12013d74
Reviewed-on: https://go-review.googlesource.com/23393
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-05-25 01:33:24 -04:00
|
|
|
if n.Left.Class == PAUTOHEAP {
|
|
|
|
Fatalf("DCL %v", n)
|
2015-06-12 14:23:29 +01:00
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
case OLABEL:
|
|
|
|
sym := n.Left.Sym
|
|
|
|
|
|
|
|
if isblanksym(sym) {
|
2015-07-12 11:52:09 -07:00
|
|
|
// Empty identifier is valid but useless.
|
|
|
|
// See issues 11589, 11593.
|
|
|
|
return
|
|
|
|
}
|
2015-07-20 15:39:14 -07:00
|
|
|
|
|
|
|
lab := s.label(sym)
|
|
|
|
|
|
|
|
// Associate label with its control flow node, if any
|
|
|
|
if ctl := n.Name.Defn; ctl != nil {
|
|
|
|
switch ctl.Op {
|
|
|
|
case OFOR, OSWITCH, OSELECT:
|
|
|
|
s.labeledNodes[ctl] = lab
|
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
2015-07-20 15:39:14 -07:00
|
|
|
|
|
|
|
if !lab.defined() {
|
|
|
|
lab.defNode = n
|
|
|
|
} else {
|
2016-03-02 11:30:29 -08:00
|
|
|
s.Error("label %v already defined at %v", sym, linestr(lab.defNode.Lineno))
|
2015-07-20 15:39:14 -07:00
|
|
|
lab.reported = true
|
|
|
|
}
|
|
|
|
// The label might already have a target block via a goto.
|
|
|
|
if lab.target == nil {
|
|
|
|
lab.target = s.f.NewBlock(ssa.BlockPlain)
|
2015-06-12 16:24:33 -07:00
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
// go to that label (we pretend "label:" is preceded by "goto label")
|
|
|
|
b := s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(lab.target)
|
2015-07-20 15:39:14 -07:00
|
|
|
s.startBlock(lab.target)
|
|
|
|
|
|
|
|
case OGOTO:
|
|
|
|
sym := n.Left.Sym
|
|
|
|
|
|
|
|
lab := s.label(sym)
|
|
|
|
if lab.target == nil {
|
|
|
|
lab.target = s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
}
|
|
|
|
if !lab.used() {
|
|
|
|
lab.useNode = n
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
2015-07-20 15:39:14 -07:00
|
|
|
|
|
|
|
if lab.defined() {
|
|
|
|
s.checkgoto(n, lab.defNode)
|
|
|
|
} else {
|
|
|
|
s.fwdGotos = append(s.fwdGotos, n)
|
2015-06-12 11:01:13 -07:00
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
b := s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(lab.target)
|
2015-07-20 15:39:14 -07:00
|
|
|
|
2015-06-10 15:03:06 -07:00
|
|
|
case OAS, OASWB:
|
2015-07-28 10:56:39 -07:00
|
|
|
// Check whether we can generate static data rather than code.
|
|
|
|
// If so, ignore n and defer data generation until codegen.
|
|
|
|
// Failure to do this causes writes to readonly symbols.
|
|
|
|
if gen_as_init(n, true) {
|
|
|
|
var data []*Node
|
|
|
|
if s.f.StaticData != nil {
|
|
|
|
data = s.f.StaticData.([]*Node)
|
|
|
|
}
|
|
|
|
s.f.StaticData = append(data, n)
|
|
|
|
return
|
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
|
2016-04-01 11:05:30 -07:00
|
|
|
if n.Left == n.Right && n.Left.Op == ONAME {
|
|
|
|
// An x=x assignment. No point in doing anything
|
|
|
|
// here. In addition, skipping this assignment
|
|
|
|
// prevents generating:
|
|
|
|
// VARDEF x
|
|
|
|
// COPY x -> x
|
|
|
|
// which is bad because x is incorrectly considered
|
|
|
|
// dead before the vardef. See issue #14904.
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
2016-01-25 17:06:54 -08:00
|
|
|
var t *Type
|
2015-08-29 14:54:45 -07:00
|
|
|
if n.Right != nil {
|
2016-01-25 17:06:54 -08:00
|
|
|
t = n.Right.Type
|
|
|
|
} else {
|
|
|
|
t = n.Left.Type
|
|
|
|
}
|
|
|
|
|
|
|
|
// Evaluate RHS.
|
|
|
|
rhs := n.Right
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
if rhs != nil {
|
|
|
|
switch rhs.Op {
|
|
|
|
case OSTRUCTLIT, OARRAYLIT:
|
|
|
|
// All literals with nonzero fields have already been
|
|
|
|
// rewritten during walk. Any that remain are just T{}
|
|
|
|
// or equivalents. Use the zero value.
|
|
|
|
if !iszero(rhs) {
|
|
|
|
Fatalf("literal with nonzero value in SSA: %v", rhs)
|
|
|
|
}
|
|
|
|
rhs = nil
|
|
|
|
case OAPPEND:
|
|
|
|
// If we're writing the result of an append back to the same slice,
|
|
|
|
// handle it specially to avoid write barriers on the fast (non-growth) path.
|
|
|
|
// If the slice can be SSA'd, it'll be on the stack,
|
|
|
|
// so there will be no write barriers,
|
|
|
|
// so there's no need to attempt to prevent them.
|
cmd/compile: re-enable in-place append optimization
CL 21891 was too clever in its attempts to avoid spills.
Storing newlen too early caused uses of append in the runtime
itself to receive an inconsistent view of a slice,
leading to corruption.
This CL makes the generate code much more similar to
the old backend. It spills more than before,
but those spills have been contained to the grow path.
It recalculates newlen unnecessarily on the fast path,
but that's measurably cheaper than spilling it.
CL 21891 caused runtime failures in 6 of 2000 runs
of net/http and crypto/x509 in my test setup.
This CL has gone 6000 runs without a failure.
Benchmarks going from master to this CL:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 439ns ± 2% 436ns ± 2% -0.72% (p=0.001 n=28+27)
AppendInPlace/NoGrow/1Ptr-8 901ns ± 0% 856ns ± 0% -4.95% (p=0.000 n=26+29)
AppendInPlace/NoGrow/2Ptr-8 2.15µs ± 1% 1.95µs ± 0% -9.07% (p=0.000 n=28+30)
AppendInPlace/NoGrow/3Ptr-8 2.66µs ± 0% 2.45µs ± 0% -7.93% (p=0.000 n=29+26)
AppendInPlace/NoGrow/4Ptr-8 3.24µs ± 1% 3.02µs ± 1% -6.75% (p=0.000 n=28+30)
AppendInPlace/Grow/Byte-8 269ns ± 1% 271ns ± 1% +0.84% (p=0.000 n=30+29)
AppendInPlace/Grow/1Ptr-8 275ns ± 1% 280ns ± 1% +1.75% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 384ns ± 0% 391ns ± 0% +1.94% (p=0.000 n=27+30)
AppendInPlace/Grow/3Ptr-8 455ns ± 0% 462ns ± 0% +1.43% (p=0.000 n=29+29)
AppendInPlace/Grow/4Ptr-8 478ns ± 0% 479ns ± 0% +0.23% (p=0.000 n=30+27)
However, for the large no-grow cases, there is still more work to be done.
Going from this CL to the non-SSA backend:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 436ns ± 2% 436ns ± 2% ~ (p=0.967 n=27+29)
AppendInPlace/NoGrow/1Ptr-8 856ns ± 0% 884ns ± 0% +3.28% (p=0.000 n=29+26)
AppendInPlace/NoGrow/2Ptr-8 1.95µs ± 0% 1.56µs ± 0% -20.28% (p=0.000 n=30+29)
AppendInPlace/NoGrow/3Ptr-8 2.45µs ± 0% 1.89µs ± 0% -22.88% (p=0.000 n=26+28)
AppendInPlace/NoGrow/4Ptr-8 3.02µs ± 1% 2.56µs ± 1% -15.35% (p=0.000 n=30+28)
AppendInPlace/Grow/Byte-8 271ns ± 1% 283ns ± 1% +4.56% (p=0.000 n=29+29)
AppendInPlace/Grow/1Ptr-8 280ns ± 1% 288ns ± 1% +2.99% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 391ns ± 0% 409ns ± 0% +4.66% (p=0.000 n=30+29)
AppendInPlace/Grow/3Ptr-8 462ns ± 0% 481ns ± 0% +4.13% (p=0.000 n=29+30)
AppendInPlace/Grow/4Ptr-8 479ns ± 0% 502ns ± 0% +4.81% (p=0.000 n=27+26)
New generated code:
var x []byte
func a() {
x = append(x, 1)
}
"".a t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (a.go:5) TEXT "".a(SB), $72-0
0x0000 00000 (a.go:5) MOVQ (TLS), CX
0x0009 00009 (a.go:5) CMPQ SP, 16(CX)
0x000d 00013 (a.go:5) JLS 190
0x0013 00019 (a.go:5) SUBQ $72, SP
0x0017 00023 (a.go:5) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:5) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:6) MOVQ "".x+16(SB), CX
0x001e 00030 (a.go:6) MOVQ "".x+8(SB), DX
0x0025 00037 (a.go:6) MOVQ "".x(SB), BX
0x002c 00044 (a.go:6) LEAQ 1(DX), BP
0x0030 00048 (a.go:6) CMPQ BP, CX
0x0033 00051 (a.go:6) JGT $0, 73
0x0035 00053 (a.go:6) LEAQ 1(DX), AX
0x0039 00057 (a.go:6) MOVQ AX, "".x+8(SB)
0x0040 00064 (a.go:6) MOVB $1, (BX)(DX*1)
0x0044 00068 (a.go:7) ADDQ $72, SP
0x0048 00072 (a.go:7) RET
0x0049 00073 (a.go:6) LEAQ type.[]uint8(SB), AX
0x0050 00080 (a.go:6) MOVQ AX, (SP)
0x0054 00084 (a.go:6) MOVQ BX, 8(SP)
0x0059 00089 (a.go:6) MOVQ DX, 16(SP)
0x005e 00094 (a.go:6) MOVQ CX, 24(SP)
0x0063 00099 (a.go:6) MOVQ BP, 32(SP)
0x0068 00104 (a.go:6) PCDATA $0, $0
0x0068 00104 (a.go:6) CALL runtime.growslice(SB)
0x006d 00109 (a.go:6) MOVQ 40(SP), CX
0x0072 00114 (a.go:6) MOVQ 48(SP), DX
0x0077 00119 (a.go:6) MOVQ DX, "".autotmp_0+64(SP)
0x007c 00124 (a.go:6) MOVQ 56(SP), BX
0x0081 00129 (a.go:6) MOVQ BX, "".x+16(SB)
0x0088 00136 (a.go:6) MOVL runtime.writeBarrier(SB), AX
0x008e 00142 (a.go:6) TESTB AL, AL
0x0090 00144 (a.go:6) JNE $0, 162
0x0092 00146 (a.go:6) MOVQ CX, "".x(SB)
0x0099 00153 (a.go:6) MOVQ "".x(SB), BX
0x00a0 00160 (a.go:6) JMP 53
0x00a2 00162 (a.go:6) LEAQ "".x(SB), BX
0x00a9 00169 (a.go:6) MOVQ BX, (SP)
0x00ad 00173 (a.go:6) MOVQ CX, 8(SP)
0x00b2 00178 (a.go:6) PCDATA $0, $0
0x00b2 00178 (a.go:6) CALL runtime.writebarrierptr(SB)
0x00b7 00183 (a.go:6) MOVQ "".autotmp_0+64(SP), DX
0x00bc 00188 (a.go:6) JMP 153
0x00be 00190 (a.go:6) NOP
0x00be 00190 (a.go:5) CALL runtime.morestack_noctxt(SB)
0x00c3 00195 (a.go:5) JMP 0
Fixes #14969 again
Change-Id: Ia50463b1f506011aad0718a4fef1d4738e43c32d
Reviewed-on: https://go-review.googlesource.com/22197
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-18 09:40:30 -07:00
|
|
|
if samesafeexpr(n.Left, rhs.List.First()) && !s.canSSA(n.Left) {
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
s.append(rhs, true)
|
|
|
|
return
|
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
var r *ssa.Value
|
2016-05-27 14:07:37 -07:00
|
|
|
var isVolatile bool
|
2016-01-25 17:06:54 -08:00
|
|
|
needwb := n.Op == OASWB && rhs != nil
|
|
|
|
deref := !canSSAType(t)
|
|
|
|
if deref {
|
|
|
|
if rhs == nil {
|
|
|
|
r = nil // Signal assign to use OpZero.
|
|
|
|
} else {
|
2016-05-27 14:07:37 -07:00
|
|
|
r, isVolatile = s.addr(rhs, false)
|
2016-01-25 17:06:54 -08:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
if rhs == nil {
|
|
|
|
r = s.zeroVal(t)
|
2015-09-18 22:12:38 -07:00
|
|
|
} else {
|
2016-01-25 17:06:54 -08:00
|
|
|
r = s.expr(rhs)
|
2015-09-18 22:12:38 -07:00
|
|
|
}
|
2015-08-29 14:54:45 -07:00
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
if rhs != nil && rhs.Op == OAPPEND {
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
// The frontend gets rid of the write barrier to enable the special OAPPEND
|
|
|
|
// handling above, but since this is not a special case, we need it.
|
2015-09-11 11:02:57 -07:00
|
|
|
// TODO: just add a ptr graying to the end of growslice?
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
// TODO: check whether we need to provide special handling and a write barrier
|
|
|
|
// for ODOTTYPE and ORECV also.
|
2015-09-11 11:02:57 -07:00
|
|
|
// They get similar wb-removal treatment in walk.go:OAS.
|
2016-01-25 17:06:54 -08:00
|
|
|
needwb = true
|
2015-09-11 11:02:57 -07:00
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
|
2016-03-21 10:22:03 -07:00
|
|
|
var skip skipMask
|
|
|
|
if rhs != nil && (rhs.Op == OSLICE || rhs.Op == OSLICE3 || rhs.Op == OSLICESTR) && samesafeexpr(rhs.Left, n.Left) {
|
|
|
|
// We're assigning a slicing operation back to its source.
|
|
|
|
// Don't write back fields we aren't changing. See issue #14855.
|
2016-04-21 11:55:33 -07:00
|
|
|
i, j, k := rhs.SliceBounds()
|
2016-04-01 14:51:02 -07:00
|
|
|
if i != nil && (i.Op == OLITERAL && i.Val().Ctype() == CTINT && i.Int64() == 0) {
|
2016-03-21 10:22:03 -07:00
|
|
|
// [0:...] is the same as [:...]
|
|
|
|
i = nil
|
|
|
|
}
|
|
|
|
// TODO: detect defaults for len/cap also.
|
|
|
|
// Currently doesn't really work because (*p)[:len(*p)] appears here as:
|
|
|
|
// tmp = len(*p)
|
|
|
|
// (*p)[:tmp]
|
|
|
|
//if j != nil && (j.Op == OLEN && samesafeexpr(j.Left, n.Left)) {
|
|
|
|
// j = nil
|
|
|
|
//}
|
|
|
|
//if k != nil && (k.Op == OCAP && samesafeexpr(k.Left, n.Left)) {
|
|
|
|
// k = nil
|
|
|
|
//}
|
|
|
|
if i == nil {
|
|
|
|
skip |= skipPtr
|
|
|
|
if j == nil {
|
|
|
|
skip |= skipLen
|
|
|
|
}
|
|
|
|
if k == nil {
|
|
|
|
skip |= skipCap
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-05-27 14:07:37 -07:00
|
|
|
s.assign(n.Left, r, needwb, deref, n.Lineno, skip, isVolatile)
|
2015-06-12 14:23:29 +01:00
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
case OIF:
|
|
|
|
bThen := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bEnd := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
var bElse *ssa.Block
|
2016-03-08 15:10:26 -08:00
|
|
|
if n.Rlist.Len() != 0 {
|
2015-04-15 15:51:25 -07:00
|
|
|
bElse = s.f.NewBlock(ssa.BlockPlain)
|
2015-11-02 16:56:53 -08:00
|
|
|
s.condBranch(n.Left, bThen, bElse, n.Likely)
|
|
|
|
} else {
|
|
|
|
s.condBranch(n.Left, bThen, bEnd, n.Likely)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
s.startBlock(bThen)
|
2016-03-01 12:50:17 -08:00
|
|
|
s.stmts(n.Nbody)
|
2015-07-20 18:42:45 -07:00
|
|
|
if b := s.endBlock(); b != nil {
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bEnd)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
|
2016-03-08 15:10:26 -08:00
|
|
|
if n.Rlist.Len() != 0 {
|
2015-04-15 15:51:25 -07:00
|
|
|
s.startBlock(bElse)
|
2015-06-11 10:20:39 -07:00
|
|
|
s.stmtList(n.Rlist)
|
2015-07-20 18:42:45 -07:00
|
|
|
if b := s.endBlock(); b != nil {
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bEnd)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
}
|
|
|
|
s.startBlock(bEnd)
|
|
|
|
|
|
|
|
case ORETURN:
|
|
|
|
s.stmtList(n.List)
|
2016-02-27 17:49:31 -08:00
|
|
|
s.exit()
|
2015-09-08 21:28:44 -07:00
|
|
|
case ORETJMP:
|
|
|
|
s.stmtList(n.List)
|
2016-02-27 17:49:31 -08:00
|
|
|
b := s.exit()
|
|
|
|
b.Kind = ssa.BlockRetJmp // override BlockRet
|
2015-09-08 21:28:44 -07:00
|
|
|
b.Aux = n.Left.Sym
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
case OCONTINUE, OBREAK:
|
|
|
|
var op string
|
|
|
|
var to *ssa.Block
|
|
|
|
switch n.Op {
|
|
|
|
case OCONTINUE:
|
|
|
|
op = "continue"
|
|
|
|
to = s.continueTo
|
|
|
|
case OBREAK:
|
|
|
|
op = "break"
|
|
|
|
to = s.breakTo
|
|
|
|
}
|
|
|
|
if n.Left == nil {
|
|
|
|
// plain break/continue
|
|
|
|
if to == nil {
|
|
|
|
s.Error("%s is not in a loop", op)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
// nothing to do; "to" is already the correct target
|
|
|
|
} else {
|
|
|
|
// labeled break/continue; look up the target
|
|
|
|
sym := n.Left.Sym
|
|
|
|
lab := s.label(sym)
|
|
|
|
if !lab.used() {
|
|
|
|
lab.useNode = n.Left
|
|
|
|
}
|
|
|
|
if !lab.defined() {
|
|
|
|
s.Error("%s label not defined: %v", op, sym)
|
|
|
|
lab.reported = true
|
|
|
|
return
|
|
|
|
}
|
|
|
|
switch n.Op {
|
|
|
|
case OCONTINUE:
|
|
|
|
to = lab.continueTarget
|
|
|
|
case OBREAK:
|
|
|
|
to = lab.breakTarget
|
|
|
|
}
|
|
|
|
if to == nil {
|
|
|
|
// Valid label but not usable with a break/continue here, e.g.:
|
|
|
|
// for {
|
|
|
|
// continue abc
|
|
|
|
// }
|
|
|
|
// abc:
|
|
|
|
// for {}
|
|
|
|
s.Error("invalid %s label %v", op, sym)
|
|
|
|
lab.reported = true
|
|
|
|
return
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
b := s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(to)
|
2015-07-20 15:39:14 -07:00
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
case OFOR:
|
2015-07-06 15:29:39 -07:00
|
|
|
// OFOR: for Ninit; Left; Right { Nbody }
|
2015-04-15 15:51:25 -07:00
|
|
|
bCond := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bBody := s.f.NewBlock(ssa.BlockPlain)
|
2015-07-06 15:29:39 -07:00
|
|
|
bIncr := s.f.NewBlock(ssa.BlockPlain)
|
2015-04-15 15:51:25 -07:00
|
|
|
bEnd := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
|
|
|
|
// first, jump to condition test
|
|
|
|
b := s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bCond)
|
2015-04-15 15:51:25 -07:00
|
|
|
|
|
|
|
// generate code to test condition
|
|
|
|
s.startBlock(bCond)
|
2015-07-06 15:29:39 -07:00
|
|
|
if n.Left != nil {
|
2015-11-02 16:56:53 -08:00
|
|
|
s.condBranch(n.Left, bBody, bEnd, 1)
|
2015-07-06 15:29:39 -07:00
|
|
|
} else {
|
2015-11-02 16:56:53 -08:00
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockPlain
|
|
|
|
b.AddEdgeTo(bBody)
|
2015-07-06 15:29:39 -07:00
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
// set up for continue/break in body
|
|
|
|
prevContinue := s.continueTo
|
|
|
|
prevBreak := s.breakTo
|
|
|
|
s.continueTo = bIncr
|
|
|
|
s.breakTo = bEnd
|
|
|
|
lab := s.labeledNodes[n]
|
|
|
|
if lab != nil {
|
|
|
|
// labeled for loop
|
|
|
|
lab.continueTarget = bIncr
|
|
|
|
lab.breakTarget = bEnd
|
|
|
|
}
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
// generate body
|
|
|
|
s.startBlock(bBody)
|
2016-03-01 12:50:17 -08:00
|
|
|
s.stmts(n.Nbody)
|
2015-07-20 15:39:14 -07:00
|
|
|
|
|
|
|
// tear down continue/break
|
|
|
|
s.continueTo = prevContinue
|
|
|
|
s.breakTo = prevBreak
|
|
|
|
if lab != nil {
|
|
|
|
lab.continueTarget = nil
|
|
|
|
lab.breakTarget = nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// done with body, goto incr
|
2015-07-06 15:29:39 -07:00
|
|
|
if b := s.endBlock(); b != nil {
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bIncr)
|
2015-07-06 15:29:39 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
// generate incr
|
|
|
|
s.startBlock(bIncr)
|
2015-06-24 17:48:22 -07:00
|
|
|
if n.Right != nil {
|
|
|
|
s.stmt(n.Right)
|
|
|
|
}
|
2015-07-06 15:29:39 -07:00
|
|
|
if b := s.endBlock(); b != nil {
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bCond)
|
2015-07-04 09:07:54 -07:00
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
s.startBlock(bEnd)
|
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
case OSWITCH, OSELECT:
|
|
|
|
// These have been mostly rewritten by the front end into their Nbody fields.
|
|
|
|
// Our main task is to correctly hook up any break statements.
|
|
|
|
bEnd := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
|
|
|
|
prevBreak := s.breakTo
|
|
|
|
s.breakTo = bEnd
|
|
|
|
lab := s.labeledNodes[n]
|
|
|
|
if lab != nil {
|
|
|
|
// labeled
|
|
|
|
lab.breakTarget = bEnd
|
|
|
|
}
|
|
|
|
|
|
|
|
// generate body code
|
2016-03-01 12:50:17 -08:00
|
|
|
s.stmts(n.Nbody)
|
2015-07-20 15:39:14 -07:00
|
|
|
|
|
|
|
s.breakTo = prevBreak
|
|
|
|
if lab != nil {
|
|
|
|
lab.breakTarget = nil
|
|
|
|
}
|
|
|
|
|
2016-02-09 12:28:02 -08:00
|
|
|
// OSWITCH never falls through (s.curBlock == nil here).
|
|
|
|
// OSELECT does not fall through if we're calling selectgo.
|
|
|
|
// OSELECT does fall through if we're calling selectnb{send,recv}[2].
|
|
|
|
// In those latter cases, go to the code after the select.
|
2015-07-20 15:39:14 -07:00
|
|
|
if b := s.endBlock(); b != nil {
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bEnd)
|
2015-07-20 15:39:14 -07:00
|
|
|
}
|
|
|
|
s.startBlock(bEnd)
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
case OVARKILL:
|
2015-08-24 02:16:19 -07:00
|
|
|
// Insert a varkill op to record that a variable is no longer live.
|
|
|
|
// We only care about liveness info at call sites, so putting the
|
|
|
|
// varkill in the store chain is enough to keep it correctly ordered
|
|
|
|
// with respect to call ops.
|
2016-02-27 17:49:31 -08:00
|
|
|
if !s.canSSA(n.Left) {
|
2015-09-19 12:01:39 -07:00
|
|
|
s.vars[&memVar] = s.newValue1A(ssa.OpVarKill, ssa.TypeMem, n.Left, s.mem())
|
|
|
|
}
|
2015-08-28 22:51:01 -07:00
|
|
|
|
2016-01-19 09:59:21 -08:00
|
|
|
case OVARLIVE:
|
|
|
|
// Insert a varlive op to record that a variable is still live.
|
|
|
|
if !n.Left.Addrtaken {
|
|
|
|
s.Fatalf("VARLIVE variable %s must have Addrtaken set", n.Left)
|
|
|
|
}
|
|
|
|
s.vars[&memVar] = s.newValue1A(ssa.OpVarLive, ssa.TypeMem, n.Left, s.mem())
|
|
|
|
|
2015-09-12 14:06:44 -07:00
|
|
|
case OCHECKNIL:
|
|
|
|
p := s.expr(n.Left)
|
|
|
|
s.nilCheck(p)
|
|
|
|
|
2016-08-21 09:57:01 +02:00
|
|
|
case OSQRT:
|
|
|
|
s.expr(n.Left)
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
default:
|
2016-04-22 07:14:10 -07:00
|
|
|
s.Unimplementedf("unhandled stmt %s", n.Op)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-02-27 17:49:31 -08:00
|
|
|
// exit processes any code that needs to be generated just before returning.
|
2016-03-01 23:21:55 +00:00
|
|
|
// It returns a BlockRet block that ends the control flow. Its control value
|
2016-02-27 17:49:31 -08:00
|
|
|
// will be set to the final memory state.
|
|
|
|
func (s *state) exit() *ssa.Block {
|
2016-03-09 19:27:57 -08:00
|
|
|
if hasdefer {
|
|
|
|
s.rtcall(Deferreturn, true, nil)
|
|
|
|
}
|
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// Run exit code. Typically, this code copies heap-allocated PPARAMOUT
|
2016-02-27 17:49:31 -08:00
|
|
|
// variables back to the stack.
|
|
|
|
s.stmts(s.exitCode)
|
|
|
|
|
|
|
|
// Store SSAable PPARAMOUT variables back to stack locations.
|
|
|
|
for _, n := range s.returns {
|
2016-04-21 19:28:28 -07:00
|
|
|
addr := s.decladdrs[n]
|
2016-02-27 17:49:31 -08:00
|
|
|
val := s.variable(n, n.Type)
|
|
|
|
s.vars[&memVar] = s.newValue1A(ssa.OpVarDef, ssa.TypeMem, n, s.mem())
|
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, n.Type.Size(), addr, val, s.mem())
|
|
|
|
// TODO: if val is ever spilled, we'd like to use the
|
2016-03-01 23:21:55 +00:00
|
|
|
// PPARAMOUT slot for spilling it. That won't happen
|
2016-02-27 17:49:31 -08:00
|
|
|
// currently.
|
|
|
|
}
|
|
|
|
|
2016-04-21 19:28:28 -07:00
|
|
|
// Keep input pointer args live until the return. This is a bandaid
|
|
|
|
// fix for 1.7 for what will become in 1.8 explicit runtime.KeepAlive calls.
|
|
|
|
// For <= 1.7 we guarantee that pointer input arguments live to the end of
|
|
|
|
// the function to prevent premature (from the user's point of view)
|
|
|
|
// execution of finalizers. See issue 15277.
|
|
|
|
// TODO: remove for 1.8?
|
|
|
|
for _, n := range s.ptrargs {
|
|
|
|
s.vars[&memVar] = s.newValue2(ssa.OpKeepAlive, ssa.TypeMem, s.variable(n, n.Type), s.mem())
|
|
|
|
}
|
|
|
|
|
2016-02-27 17:49:31 -08:00
|
|
|
// Do actual return.
|
|
|
|
m := s.mem()
|
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockRet
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(m)
|
2016-02-27 17:49:31 -08:00
|
|
|
return b
|
|
|
|
}
|
|
|
|
|
2015-07-19 15:48:20 -07:00
|
|
|
type opAndType struct {
|
2015-11-16 13:20:16 -08:00
|
|
|
op Op
|
|
|
|
etype EType
|
2015-07-19 15:48:20 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
var opToSSA = map[opAndType]ssa.Op{
|
2015-08-12 16:38:11 -04:00
|
|
|
opAndType{OADD, TINT8}: ssa.OpAdd8,
|
|
|
|
opAndType{OADD, TUINT8}: ssa.OpAdd8,
|
|
|
|
opAndType{OADD, TINT16}: ssa.OpAdd16,
|
|
|
|
opAndType{OADD, TUINT16}: ssa.OpAdd16,
|
|
|
|
opAndType{OADD, TINT32}: ssa.OpAdd32,
|
|
|
|
opAndType{OADD, TUINT32}: ssa.OpAdd32,
|
|
|
|
opAndType{OADD, TPTR32}: ssa.OpAdd32,
|
|
|
|
opAndType{OADD, TINT64}: ssa.OpAdd64,
|
|
|
|
opAndType{OADD, TUINT64}: ssa.OpAdd64,
|
|
|
|
opAndType{OADD, TPTR64}: ssa.OpAdd64,
|
|
|
|
opAndType{OADD, TFLOAT32}: ssa.OpAdd32F,
|
|
|
|
opAndType{OADD, TFLOAT64}: ssa.OpAdd64F,
|
|
|
|
|
|
|
|
opAndType{OSUB, TINT8}: ssa.OpSub8,
|
|
|
|
opAndType{OSUB, TUINT8}: ssa.OpSub8,
|
|
|
|
opAndType{OSUB, TINT16}: ssa.OpSub16,
|
|
|
|
opAndType{OSUB, TUINT16}: ssa.OpSub16,
|
|
|
|
opAndType{OSUB, TINT32}: ssa.OpSub32,
|
|
|
|
opAndType{OSUB, TUINT32}: ssa.OpSub32,
|
|
|
|
opAndType{OSUB, TINT64}: ssa.OpSub64,
|
|
|
|
opAndType{OSUB, TUINT64}: ssa.OpSub64,
|
|
|
|
opAndType{OSUB, TFLOAT32}: ssa.OpSub32F,
|
|
|
|
opAndType{OSUB, TFLOAT64}: ssa.OpSub64F,
|
2015-07-19 15:48:20 -07:00
|
|
|
|
2015-07-22 19:19:40 -07:00
|
|
|
opAndType{ONOT, TBOOL}: ssa.OpNot,
|
|
|
|
|
2015-08-28 14:24:10 -04:00
|
|
|
opAndType{OMINUS, TINT8}: ssa.OpNeg8,
|
|
|
|
opAndType{OMINUS, TUINT8}: ssa.OpNeg8,
|
|
|
|
opAndType{OMINUS, TINT16}: ssa.OpNeg16,
|
|
|
|
opAndType{OMINUS, TUINT16}: ssa.OpNeg16,
|
|
|
|
opAndType{OMINUS, TINT32}: ssa.OpNeg32,
|
|
|
|
opAndType{OMINUS, TUINT32}: ssa.OpNeg32,
|
|
|
|
opAndType{OMINUS, TINT64}: ssa.OpNeg64,
|
|
|
|
opAndType{OMINUS, TUINT64}: ssa.OpNeg64,
|
|
|
|
opAndType{OMINUS, TFLOAT32}: ssa.OpNeg32F,
|
|
|
|
opAndType{OMINUS, TFLOAT64}: ssa.OpNeg64F,
|
2015-07-21 16:58:18 +02:00
|
|
|
|
2015-07-29 17:07:09 -07:00
|
|
|
opAndType{OCOM, TINT8}: ssa.OpCom8,
|
|
|
|
opAndType{OCOM, TUINT8}: ssa.OpCom8,
|
|
|
|
opAndType{OCOM, TINT16}: ssa.OpCom16,
|
|
|
|
opAndType{OCOM, TUINT16}: ssa.OpCom16,
|
|
|
|
opAndType{OCOM, TINT32}: ssa.OpCom32,
|
|
|
|
opAndType{OCOM, TUINT32}: ssa.OpCom32,
|
|
|
|
opAndType{OCOM, TINT64}: ssa.OpCom64,
|
|
|
|
opAndType{OCOM, TUINT64}: ssa.OpCom64,
|
|
|
|
|
2015-09-06 19:24:59 -07:00
|
|
|
opAndType{OIMAG, TCOMPLEX64}: ssa.OpComplexImag,
|
|
|
|
opAndType{OIMAG, TCOMPLEX128}: ssa.OpComplexImag,
|
|
|
|
opAndType{OREAL, TCOMPLEX64}: ssa.OpComplexReal,
|
|
|
|
opAndType{OREAL, TCOMPLEX128}: ssa.OpComplexReal,
|
|
|
|
|
2015-08-12 16:38:11 -04:00
|
|
|
opAndType{OMUL, TINT8}: ssa.OpMul8,
|
|
|
|
opAndType{OMUL, TUINT8}: ssa.OpMul8,
|
|
|
|
opAndType{OMUL, TINT16}: ssa.OpMul16,
|
|
|
|
opAndType{OMUL, TUINT16}: ssa.OpMul16,
|
|
|
|
opAndType{OMUL, TINT32}: ssa.OpMul32,
|
|
|
|
opAndType{OMUL, TUINT32}: ssa.OpMul32,
|
|
|
|
opAndType{OMUL, TINT64}: ssa.OpMul64,
|
|
|
|
opAndType{OMUL, TUINT64}: ssa.OpMul64,
|
|
|
|
opAndType{OMUL, TFLOAT32}: ssa.OpMul32F,
|
|
|
|
opAndType{OMUL, TFLOAT64}: ssa.OpMul64F,
|
|
|
|
|
|
|
|
opAndType{ODIV, TFLOAT32}: ssa.OpDiv32F,
|
|
|
|
opAndType{ODIV, TFLOAT64}: ssa.OpDiv64F,
|
2015-07-22 13:46:15 -07:00
|
|
|
|
2015-08-18 19:14:47 -05:00
|
|
|
opAndType{OHMUL, TINT8}: ssa.OpHmul8,
|
|
|
|
opAndType{OHMUL, TUINT8}: ssa.OpHmul8u,
|
|
|
|
opAndType{OHMUL, TINT16}: ssa.OpHmul16,
|
|
|
|
opAndType{OHMUL, TUINT16}: ssa.OpHmul16u,
|
|
|
|
opAndType{OHMUL, TINT32}: ssa.OpHmul32,
|
|
|
|
opAndType{OHMUL, TUINT32}: ssa.OpHmul32u,
|
|
|
|
|
2015-08-17 17:46:06 -05:00
|
|
|
opAndType{ODIV, TINT8}: ssa.OpDiv8,
|
|
|
|
opAndType{ODIV, TUINT8}: ssa.OpDiv8u,
|
|
|
|
opAndType{ODIV, TINT16}: ssa.OpDiv16,
|
|
|
|
opAndType{ODIV, TUINT16}: ssa.OpDiv16u,
|
|
|
|
opAndType{ODIV, TINT32}: ssa.OpDiv32,
|
|
|
|
opAndType{ODIV, TUINT32}: ssa.OpDiv32u,
|
|
|
|
opAndType{ODIV, TINT64}: ssa.OpDiv64,
|
|
|
|
opAndType{ODIV, TUINT64}: ssa.OpDiv64u,
|
|
|
|
|
2015-08-18 19:51:44 -05:00
|
|
|
opAndType{OMOD, TINT8}: ssa.OpMod8,
|
|
|
|
opAndType{OMOD, TUINT8}: ssa.OpMod8u,
|
|
|
|
opAndType{OMOD, TINT16}: ssa.OpMod16,
|
|
|
|
opAndType{OMOD, TUINT16}: ssa.OpMod16u,
|
|
|
|
opAndType{OMOD, TINT32}: ssa.OpMod32,
|
|
|
|
opAndType{OMOD, TUINT32}: ssa.OpMod32u,
|
|
|
|
opAndType{OMOD, TINT64}: ssa.OpMod64,
|
|
|
|
opAndType{OMOD, TUINT64}: ssa.OpMod64u,
|
|
|
|
|
2015-07-28 14:58:49 +02:00
|
|
|
opAndType{OAND, TINT8}: ssa.OpAnd8,
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
opAndType{OAND, TUINT8}: ssa.OpAnd8,
|
2015-07-28 14:58:49 +02:00
|
|
|
opAndType{OAND, TINT16}: ssa.OpAnd16,
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
opAndType{OAND, TUINT16}: ssa.OpAnd16,
|
2015-07-28 14:58:49 +02:00
|
|
|
opAndType{OAND, TINT32}: ssa.OpAnd32,
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
opAndType{OAND, TUINT32}: ssa.OpAnd32,
|
2015-07-28 14:58:49 +02:00
|
|
|
opAndType{OAND, TINT64}: ssa.OpAnd64,
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
opAndType{OAND, TUINT64}: ssa.OpAnd64,
|
2015-07-28 14:58:49 +02:00
|
|
|
|
2015-07-29 17:52:25 +02:00
|
|
|
opAndType{OOR, TINT8}: ssa.OpOr8,
|
|
|
|
opAndType{OOR, TUINT8}: ssa.OpOr8,
|
|
|
|
opAndType{OOR, TINT16}: ssa.OpOr16,
|
|
|
|
opAndType{OOR, TUINT16}: ssa.OpOr16,
|
|
|
|
opAndType{OOR, TINT32}: ssa.OpOr32,
|
|
|
|
opAndType{OOR, TUINT32}: ssa.OpOr32,
|
|
|
|
opAndType{OOR, TINT64}: ssa.OpOr64,
|
|
|
|
opAndType{OOR, TUINT64}: ssa.OpOr64,
|
|
|
|
|
2015-07-30 12:33:36 +02:00
|
|
|
opAndType{OXOR, TINT8}: ssa.OpXor8,
|
|
|
|
opAndType{OXOR, TUINT8}: ssa.OpXor8,
|
|
|
|
opAndType{OXOR, TINT16}: ssa.OpXor16,
|
|
|
|
opAndType{OXOR, TUINT16}: ssa.OpXor16,
|
|
|
|
opAndType{OXOR, TINT32}: ssa.OpXor32,
|
|
|
|
opAndType{OXOR, TUINT32}: ssa.OpXor32,
|
|
|
|
opAndType{OXOR, TINT64}: ssa.OpXor64,
|
|
|
|
opAndType{OXOR, TUINT64}: ssa.OpXor64,
|
|
|
|
|
2016-04-24 21:21:07 +02:00
|
|
|
opAndType{OEQ, TBOOL}: ssa.OpEqB,
|
2015-07-28 14:14:25 -07:00
|
|
|
opAndType{OEQ, TINT8}: ssa.OpEq8,
|
|
|
|
opAndType{OEQ, TUINT8}: ssa.OpEq8,
|
|
|
|
opAndType{OEQ, TINT16}: ssa.OpEq16,
|
|
|
|
opAndType{OEQ, TUINT16}: ssa.OpEq16,
|
|
|
|
opAndType{OEQ, TINT32}: ssa.OpEq32,
|
|
|
|
opAndType{OEQ, TUINT32}: ssa.OpEq32,
|
|
|
|
opAndType{OEQ, TINT64}: ssa.OpEq64,
|
|
|
|
opAndType{OEQ, TUINT64}: ssa.OpEq64,
|
2015-09-10 13:53:27 -07:00
|
|
|
opAndType{OEQ, TINTER}: ssa.OpEqInter,
|
2016-04-18 14:02:08 -07:00
|
|
|
opAndType{OEQ, TSLICE}: ssa.OpEqSlice,
|
2015-07-28 14:14:25 -07:00
|
|
|
opAndType{OEQ, TFUNC}: ssa.OpEqPtr,
|
|
|
|
opAndType{OEQ, TMAP}: ssa.OpEqPtr,
|
|
|
|
opAndType{OEQ, TCHAN}: ssa.OpEqPtr,
|
2016-05-13 11:25:07 -04:00
|
|
|
opAndType{OEQ, TPTR32}: ssa.OpEqPtr,
|
2015-08-30 20:47:26 -05:00
|
|
|
opAndType{OEQ, TPTR64}: ssa.OpEqPtr,
|
2015-07-28 14:14:25 -07:00
|
|
|
opAndType{OEQ, TUINTPTR}: ssa.OpEqPtr,
|
|
|
|
opAndType{OEQ, TUNSAFEPTR}: ssa.OpEqPtr,
|
2015-08-18 14:39:26 -04:00
|
|
|
opAndType{OEQ, TFLOAT64}: ssa.OpEq64F,
|
|
|
|
opAndType{OEQ, TFLOAT32}: ssa.OpEq32F,
|
2015-07-28 14:14:25 -07:00
|
|
|
|
2016-04-24 21:21:07 +02:00
|
|
|
opAndType{ONE, TBOOL}: ssa.OpNeqB,
|
2015-07-28 14:14:25 -07:00
|
|
|
opAndType{ONE, TINT8}: ssa.OpNeq8,
|
|
|
|
opAndType{ONE, TUINT8}: ssa.OpNeq8,
|
|
|
|
opAndType{ONE, TINT16}: ssa.OpNeq16,
|
|
|
|
opAndType{ONE, TUINT16}: ssa.OpNeq16,
|
|
|
|
opAndType{ONE, TINT32}: ssa.OpNeq32,
|
|
|
|
opAndType{ONE, TUINT32}: ssa.OpNeq32,
|
|
|
|
opAndType{ONE, TINT64}: ssa.OpNeq64,
|
|
|
|
opAndType{ONE, TUINT64}: ssa.OpNeq64,
|
2015-09-10 13:53:27 -07:00
|
|
|
opAndType{ONE, TINTER}: ssa.OpNeqInter,
|
2016-04-18 14:02:08 -07:00
|
|
|
opAndType{ONE, TSLICE}: ssa.OpNeqSlice,
|
2015-07-28 14:14:25 -07:00
|
|
|
opAndType{ONE, TFUNC}: ssa.OpNeqPtr,
|
|
|
|
opAndType{ONE, TMAP}: ssa.OpNeqPtr,
|
|
|
|
opAndType{ONE, TCHAN}: ssa.OpNeqPtr,
|
2016-05-13 11:25:07 -04:00
|
|
|
opAndType{ONE, TPTR32}: ssa.OpNeqPtr,
|
2015-08-30 20:47:26 -05:00
|
|
|
opAndType{ONE, TPTR64}: ssa.OpNeqPtr,
|
2015-07-28 14:14:25 -07:00
|
|
|
opAndType{ONE, TUINTPTR}: ssa.OpNeqPtr,
|
|
|
|
opAndType{ONE, TUNSAFEPTR}: ssa.OpNeqPtr,
|
2015-08-18 14:39:26 -04:00
|
|
|
opAndType{ONE, TFLOAT64}: ssa.OpNeq64F,
|
|
|
|
opAndType{ONE, TFLOAT32}: ssa.OpNeq32F,
|
|
|
|
|
|
|
|
opAndType{OLT, TINT8}: ssa.OpLess8,
|
|
|
|
opAndType{OLT, TUINT8}: ssa.OpLess8U,
|
|
|
|
opAndType{OLT, TINT16}: ssa.OpLess16,
|
|
|
|
opAndType{OLT, TUINT16}: ssa.OpLess16U,
|
|
|
|
opAndType{OLT, TINT32}: ssa.OpLess32,
|
|
|
|
opAndType{OLT, TUINT32}: ssa.OpLess32U,
|
|
|
|
opAndType{OLT, TINT64}: ssa.OpLess64,
|
|
|
|
opAndType{OLT, TUINT64}: ssa.OpLess64U,
|
|
|
|
opAndType{OLT, TFLOAT64}: ssa.OpLess64F,
|
|
|
|
opAndType{OLT, TFLOAT32}: ssa.OpLess32F,
|
|
|
|
|
|
|
|
opAndType{OGT, TINT8}: ssa.OpGreater8,
|
|
|
|
opAndType{OGT, TUINT8}: ssa.OpGreater8U,
|
|
|
|
opAndType{OGT, TINT16}: ssa.OpGreater16,
|
|
|
|
opAndType{OGT, TUINT16}: ssa.OpGreater16U,
|
|
|
|
opAndType{OGT, TINT32}: ssa.OpGreater32,
|
|
|
|
opAndType{OGT, TUINT32}: ssa.OpGreater32U,
|
|
|
|
opAndType{OGT, TINT64}: ssa.OpGreater64,
|
|
|
|
opAndType{OGT, TUINT64}: ssa.OpGreater64U,
|
|
|
|
opAndType{OGT, TFLOAT64}: ssa.OpGreater64F,
|
|
|
|
opAndType{OGT, TFLOAT32}: ssa.OpGreater32F,
|
|
|
|
|
|
|
|
opAndType{OLE, TINT8}: ssa.OpLeq8,
|
|
|
|
opAndType{OLE, TUINT8}: ssa.OpLeq8U,
|
|
|
|
opAndType{OLE, TINT16}: ssa.OpLeq16,
|
|
|
|
opAndType{OLE, TUINT16}: ssa.OpLeq16U,
|
|
|
|
opAndType{OLE, TINT32}: ssa.OpLeq32,
|
|
|
|
opAndType{OLE, TUINT32}: ssa.OpLeq32U,
|
|
|
|
opAndType{OLE, TINT64}: ssa.OpLeq64,
|
|
|
|
opAndType{OLE, TUINT64}: ssa.OpLeq64U,
|
|
|
|
opAndType{OLE, TFLOAT64}: ssa.OpLeq64F,
|
|
|
|
opAndType{OLE, TFLOAT32}: ssa.OpLeq32F,
|
|
|
|
|
|
|
|
opAndType{OGE, TINT8}: ssa.OpGeq8,
|
|
|
|
opAndType{OGE, TUINT8}: ssa.OpGeq8U,
|
|
|
|
opAndType{OGE, TINT16}: ssa.OpGeq16,
|
|
|
|
opAndType{OGE, TUINT16}: ssa.OpGeq16U,
|
|
|
|
opAndType{OGE, TINT32}: ssa.OpGeq32,
|
|
|
|
opAndType{OGE, TUINT32}: ssa.OpGeq32U,
|
|
|
|
opAndType{OGE, TINT64}: ssa.OpGeq64,
|
|
|
|
opAndType{OGE, TUINT64}: ssa.OpGeq64U,
|
|
|
|
opAndType{OGE, TFLOAT64}: ssa.OpGeq64F,
|
|
|
|
opAndType{OGE, TFLOAT32}: ssa.OpGeq32F,
|
2015-08-05 22:11:14 -04:00
|
|
|
|
|
|
|
opAndType{OLROT, TUINT8}: ssa.OpLrot8,
|
|
|
|
opAndType{OLROT, TUINT16}: ssa.OpLrot16,
|
|
|
|
opAndType{OLROT, TUINT32}: ssa.OpLrot32,
|
|
|
|
opAndType{OLROT, TUINT64}: ssa.OpLrot64,
|
2015-09-12 13:26:57 -07:00
|
|
|
|
|
|
|
opAndType{OSQRT, TFLOAT64}: ssa.OpSqrt,
|
2015-07-19 15:48:20 -07:00
|
|
|
}
|
|
|
|
|
2015-11-16 13:20:16 -08:00
|
|
|
func (s *state) concreteEtype(t *Type) EType {
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
e := t.Etype
|
|
|
|
switch e {
|
|
|
|
default:
|
|
|
|
return e
|
2015-07-19 15:48:20 -07:00
|
|
|
case TINT:
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
if s.config.IntSize == 8 {
|
|
|
|
return TINT64
|
2015-07-19 15:48:20 -07:00
|
|
|
}
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
return TINT32
|
2015-07-19 15:48:20 -07:00
|
|
|
case TUINT:
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
if s.config.IntSize == 8 {
|
|
|
|
return TUINT64
|
|
|
|
}
|
|
|
|
return TUINT32
|
|
|
|
case TUINTPTR:
|
2015-07-19 15:48:20 -07:00
|
|
|
if s.config.PtrSize == 8 {
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
return TUINT64
|
2015-07-19 15:48:20 -07:00
|
|
|
}
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
return TUINT32
|
2015-07-19 15:48:20 -07:00
|
|
|
}
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
}
|
|
|
|
|
2015-11-16 13:20:16 -08:00
|
|
|
func (s *state) ssaOp(op Op, t *Type) ssa.Op {
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
etype := s.concreteEtype(t)
|
2015-07-19 15:48:20 -07:00
|
|
|
x, ok := opToSSA[opAndType{op, etype}]
|
|
|
|
if !ok {
|
2016-04-22 08:39:56 -07:00
|
|
|
s.Unimplementedf("unhandled binary op %s %s", op, etype)
|
2015-07-19 15:48:20 -07:00
|
|
|
}
|
|
|
|
return x
|
2015-06-24 17:48:22 -07:00
|
|
|
}
|
|
|
|
|
2015-08-28 14:24:10 -04:00
|
|
|
func floatForComplex(t *Type) *Type {
|
|
|
|
if t.Size() == 8 {
|
|
|
|
return Types[TFLOAT32]
|
|
|
|
} else {
|
|
|
|
return Types[TFLOAT64]
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-07-29 17:07:09 -07:00
|
|
|
type opAndTwoTypes struct {
|
2015-11-16 13:20:16 -08:00
|
|
|
op Op
|
|
|
|
etype1 EType
|
|
|
|
etype2 EType
|
2015-07-29 17:07:09 -07:00
|
|
|
}
|
|
|
|
|
2015-09-01 17:09:00 -04:00
|
|
|
type twoTypes struct {
|
2015-11-16 13:20:16 -08:00
|
|
|
etype1 EType
|
|
|
|
etype2 EType
|
2015-09-01 17:09:00 -04:00
|
|
|
}
|
|
|
|
|
|
|
|
type twoOpsAndType struct {
|
|
|
|
op1 ssa.Op
|
|
|
|
op2 ssa.Op
|
2015-11-16 13:20:16 -08:00
|
|
|
intermediateType EType
|
2015-09-01 17:09:00 -04:00
|
|
|
}
|
|
|
|
|
|
|
|
var fpConvOpToSSA = map[twoTypes]twoOpsAndType{
|
|
|
|
|
|
|
|
twoTypes{TINT8, TFLOAT32}: twoOpsAndType{ssa.OpSignExt8to32, ssa.OpCvt32to32F, TINT32},
|
|
|
|
twoTypes{TINT16, TFLOAT32}: twoOpsAndType{ssa.OpSignExt16to32, ssa.OpCvt32to32F, TINT32},
|
|
|
|
twoTypes{TINT32, TFLOAT32}: twoOpsAndType{ssa.OpCopy, ssa.OpCvt32to32F, TINT32},
|
|
|
|
twoTypes{TINT64, TFLOAT32}: twoOpsAndType{ssa.OpCopy, ssa.OpCvt64to32F, TINT64},
|
|
|
|
|
|
|
|
twoTypes{TINT8, TFLOAT64}: twoOpsAndType{ssa.OpSignExt8to32, ssa.OpCvt32to64F, TINT32},
|
|
|
|
twoTypes{TINT16, TFLOAT64}: twoOpsAndType{ssa.OpSignExt16to32, ssa.OpCvt32to64F, TINT32},
|
|
|
|
twoTypes{TINT32, TFLOAT64}: twoOpsAndType{ssa.OpCopy, ssa.OpCvt32to64F, TINT32},
|
|
|
|
twoTypes{TINT64, TFLOAT64}: twoOpsAndType{ssa.OpCopy, ssa.OpCvt64to64F, TINT64},
|
|
|
|
|
|
|
|
twoTypes{TFLOAT32, TINT8}: twoOpsAndType{ssa.OpCvt32Fto32, ssa.OpTrunc32to8, TINT32},
|
|
|
|
twoTypes{TFLOAT32, TINT16}: twoOpsAndType{ssa.OpCvt32Fto32, ssa.OpTrunc32to16, TINT32},
|
|
|
|
twoTypes{TFLOAT32, TINT32}: twoOpsAndType{ssa.OpCvt32Fto32, ssa.OpCopy, TINT32},
|
|
|
|
twoTypes{TFLOAT32, TINT64}: twoOpsAndType{ssa.OpCvt32Fto64, ssa.OpCopy, TINT64},
|
|
|
|
|
|
|
|
twoTypes{TFLOAT64, TINT8}: twoOpsAndType{ssa.OpCvt64Fto32, ssa.OpTrunc32to8, TINT32},
|
|
|
|
twoTypes{TFLOAT64, TINT16}: twoOpsAndType{ssa.OpCvt64Fto32, ssa.OpTrunc32to16, TINT32},
|
|
|
|
twoTypes{TFLOAT64, TINT32}: twoOpsAndType{ssa.OpCvt64Fto32, ssa.OpCopy, TINT32},
|
|
|
|
twoTypes{TFLOAT64, TINT64}: twoOpsAndType{ssa.OpCvt64Fto64, ssa.OpCopy, TINT64},
|
|
|
|
// unsigned
|
|
|
|
twoTypes{TUINT8, TFLOAT32}: twoOpsAndType{ssa.OpZeroExt8to32, ssa.OpCvt32to32F, TINT32},
|
|
|
|
twoTypes{TUINT16, TFLOAT32}: twoOpsAndType{ssa.OpZeroExt16to32, ssa.OpCvt32to32F, TINT32},
|
|
|
|
twoTypes{TUINT32, TFLOAT32}: twoOpsAndType{ssa.OpZeroExt32to64, ssa.OpCvt64to32F, TINT64}, // go wide to dodge unsigned
|
|
|
|
twoTypes{TUINT64, TFLOAT32}: twoOpsAndType{ssa.OpCopy, ssa.OpInvalid, TUINT64}, // Cvt64Uto32F, branchy code expansion instead
|
|
|
|
|
|
|
|
twoTypes{TUINT8, TFLOAT64}: twoOpsAndType{ssa.OpZeroExt8to32, ssa.OpCvt32to64F, TINT32},
|
|
|
|
twoTypes{TUINT16, TFLOAT64}: twoOpsAndType{ssa.OpZeroExt16to32, ssa.OpCvt32to64F, TINT32},
|
|
|
|
twoTypes{TUINT32, TFLOAT64}: twoOpsAndType{ssa.OpZeroExt32to64, ssa.OpCvt64to64F, TINT64}, // go wide to dodge unsigned
|
|
|
|
twoTypes{TUINT64, TFLOAT64}: twoOpsAndType{ssa.OpCopy, ssa.OpInvalid, TUINT64}, // Cvt64Uto64F, branchy code expansion instead
|
|
|
|
|
|
|
|
twoTypes{TFLOAT32, TUINT8}: twoOpsAndType{ssa.OpCvt32Fto32, ssa.OpTrunc32to8, TINT32},
|
|
|
|
twoTypes{TFLOAT32, TUINT16}: twoOpsAndType{ssa.OpCvt32Fto32, ssa.OpTrunc32to16, TINT32},
|
|
|
|
twoTypes{TFLOAT32, TUINT32}: twoOpsAndType{ssa.OpCvt32Fto64, ssa.OpTrunc64to32, TINT64}, // go wide to dodge unsigned
|
|
|
|
twoTypes{TFLOAT32, TUINT64}: twoOpsAndType{ssa.OpInvalid, ssa.OpCopy, TUINT64}, // Cvt32Fto64U, branchy code expansion instead
|
|
|
|
|
|
|
|
twoTypes{TFLOAT64, TUINT8}: twoOpsAndType{ssa.OpCvt64Fto32, ssa.OpTrunc32to8, TINT32},
|
|
|
|
twoTypes{TFLOAT64, TUINT16}: twoOpsAndType{ssa.OpCvt64Fto32, ssa.OpTrunc32to16, TINT32},
|
|
|
|
twoTypes{TFLOAT64, TUINT32}: twoOpsAndType{ssa.OpCvt64Fto64, ssa.OpTrunc64to32, TINT64}, // go wide to dodge unsigned
|
|
|
|
twoTypes{TFLOAT64, TUINT64}: twoOpsAndType{ssa.OpInvalid, ssa.OpCopy, TUINT64}, // Cvt64Fto64U, branchy code expansion instead
|
|
|
|
|
|
|
|
// float
|
|
|
|
twoTypes{TFLOAT64, TFLOAT32}: twoOpsAndType{ssa.OpCvt64Fto32F, ssa.OpCopy, TFLOAT32},
|
|
|
|
twoTypes{TFLOAT64, TFLOAT64}: twoOpsAndType{ssa.OpCopy, ssa.OpCopy, TFLOAT64},
|
|
|
|
twoTypes{TFLOAT32, TFLOAT32}: twoOpsAndType{ssa.OpCopy, ssa.OpCopy, TFLOAT32},
|
|
|
|
twoTypes{TFLOAT32, TFLOAT64}: twoOpsAndType{ssa.OpCvt32Fto64F, ssa.OpCopy, TFLOAT64},
|
|
|
|
}
|
|
|
|
|
2016-05-31 11:27:16 -04:00
|
|
|
// this map is used only for 32-bit arch, and only includes the difference
|
|
|
|
// on 32-bit arch, don't use int64<->float conversion for uint32
|
|
|
|
var fpConvOpToSSA32 = map[twoTypes]twoOpsAndType{
|
|
|
|
twoTypes{TUINT32, TFLOAT32}: twoOpsAndType{ssa.OpCopy, ssa.OpCvt32Uto32F, TUINT32},
|
|
|
|
twoTypes{TUINT32, TFLOAT64}: twoOpsAndType{ssa.OpCopy, ssa.OpCvt32Uto64F, TUINT32},
|
|
|
|
twoTypes{TFLOAT32, TUINT32}: twoOpsAndType{ssa.OpCvt32Fto32U, ssa.OpCopy, TUINT32},
|
|
|
|
twoTypes{TFLOAT64, TUINT32}: twoOpsAndType{ssa.OpCvt64Fto32U, ssa.OpCopy, TUINT32},
|
|
|
|
}
|
|
|
|
|
2016-08-16 14:17:33 -04:00
|
|
|
// uint64<->float conversions, only on machines that have intructions for that
|
|
|
|
var uint64fpConvOpToSSA = map[twoTypes]twoOpsAndType{
|
|
|
|
twoTypes{TUINT64, TFLOAT32}: twoOpsAndType{ssa.OpCopy, ssa.OpCvt64Uto32F, TUINT64},
|
|
|
|
twoTypes{TUINT64, TFLOAT64}: twoOpsAndType{ssa.OpCopy, ssa.OpCvt64Uto64F, TUINT64},
|
|
|
|
twoTypes{TFLOAT32, TUINT64}: twoOpsAndType{ssa.OpCvt32Fto64U, ssa.OpCopy, TUINT64},
|
|
|
|
twoTypes{TFLOAT64, TUINT64}: twoOpsAndType{ssa.OpCvt64Fto64U, ssa.OpCopy, TUINT64},
|
|
|
|
}
|
|
|
|
|
2015-07-29 17:07:09 -07:00
|
|
|
var shiftOpToSSA = map[opAndTwoTypes]ssa.Op{
|
|
|
|
opAndTwoTypes{OLSH, TINT8, TUINT8}: ssa.OpLsh8x8,
|
|
|
|
opAndTwoTypes{OLSH, TUINT8, TUINT8}: ssa.OpLsh8x8,
|
|
|
|
opAndTwoTypes{OLSH, TINT8, TUINT16}: ssa.OpLsh8x16,
|
|
|
|
opAndTwoTypes{OLSH, TUINT8, TUINT16}: ssa.OpLsh8x16,
|
|
|
|
opAndTwoTypes{OLSH, TINT8, TUINT32}: ssa.OpLsh8x32,
|
|
|
|
opAndTwoTypes{OLSH, TUINT8, TUINT32}: ssa.OpLsh8x32,
|
|
|
|
opAndTwoTypes{OLSH, TINT8, TUINT64}: ssa.OpLsh8x64,
|
|
|
|
opAndTwoTypes{OLSH, TUINT8, TUINT64}: ssa.OpLsh8x64,
|
|
|
|
|
|
|
|
opAndTwoTypes{OLSH, TINT16, TUINT8}: ssa.OpLsh16x8,
|
|
|
|
opAndTwoTypes{OLSH, TUINT16, TUINT8}: ssa.OpLsh16x8,
|
|
|
|
opAndTwoTypes{OLSH, TINT16, TUINT16}: ssa.OpLsh16x16,
|
|
|
|
opAndTwoTypes{OLSH, TUINT16, TUINT16}: ssa.OpLsh16x16,
|
|
|
|
opAndTwoTypes{OLSH, TINT16, TUINT32}: ssa.OpLsh16x32,
|
|
|
|
opAndTwoTypes{OLSH, TUINT16, TUINT32}: ssa.OpLsh16x32,
|
|
|
|
opAndTwoTypes{OLSH, TINT16, TUINT64}: ssa.OpLsh16x64,
|
|
|
|
opAndTwoTypes{OLSH, TUINT16, TUINT64}: ssa.OpLsh16x64,
|
|
|
|
|
|
|
|
opAndTwoTypes{OLSH, TINT32, TUINT8}: ssa.OpLsh32x8,
|
|
|
|
opAndTwoTypes{OLSH, TUINT32, TUINT8}: ssa.OpLsh32x8,
|
|
|
|
opAndTwoTypes{OLSH, TINT32, TUINT16}: ssa.OpLsh32x16,
|
|
|
|
opAndTwoTypes{OLSH, TUINT32, TUINT16}: ssa.OpLsh32x16,
|
|
|
|
opAndTwoTypes{OLSH, TINT32, TUINT32}: ssa.OpLsh32x32,
|
|
|
|
opAndTwoTypes{OLSH, TUINT32, TUINT32}: ssa.OpLsh32x32,
|
|
|
|
opAndTwoTypes{OLSH, TINT32, TUINT64}: ssa.OpLsh32x64,
|
|
|
|
opAndTwoTypes{OLSH, TUINT32, TUINT64}: ssa.OpLsh32x64,
|
|
|
|
|
|
|
|
opAndTwoTypes{OLSH, TINT64, TUINT8}: ssa.OpLsh64x8,
|
|
|
|
opAndTwoTypes{OLSH, TUINT64, TUINT8}: ssa.OpLsh64x8,
|
|
|
|
opAndTwoTypes{OLSH, TINT64, TUINT16}: ssa.OpLsh64x16,
|
|
|
|
opAndTwoTypes{OLSH, TUINT64, TUINT16}: ssa.OpLsh64x16,
|
|
|
|
opAndTwoTypes{OLSH, TINT64, TUINT32}: ssa.OpLsh64x32,
|
|
|
|
opAndTwoTypes{OLSH, TUINT64, TUINT32}: ssa.OpLsh64x32,
|
|
|
|
opAndTwoTypes{OLSH, TINT64, TUINT64}: ssa.OpLsh64x64,
|
|
|
|
opAndTwoTypes{OLSH, TUINT64, TUINT64}: ssa.OpLsh64x64,
|
|
|
|
|
|
|
|
opAndTwoTypes{ORSH, TINT8, TUINT8}: ssa.OpRsh8x8,
|
|
|
|
opAndTwoTypes{ORSH, TUINT8, TUINT8}: ssa.OpRsh8Ux8,
|
|
|
|
opAndTwoTypes{ORSH, TINT8, TUINT16}: ssa.OpRsh8x16,
|
|
|
|
opAndTwoTypes{ORSH, TUINT8, TUINT16}: ssa.OpRsh8Ux16,
|
|
|
|
opAndTwoTypes{ORSH, TINT8, TUINT32}: ssa.OpRsh8x32,
|
|
|
|
opAndTwoTypes{ORSH, TUINT8, TUINT32}: ssa.OpRsh8Ux32,
|
|
|
|
opAndTwoTypes{ORSH, TINT8, TUINT64}: ssa.OpRsh8x64,
|
|
|
|
opAndTwoTypes{ORSH, TUINT8, TUINT64}: ssa.OpRsh8Ux64,
|
|
|
|
|
|
|
|
opAndTwoTypes{ORSH, TINT16, TUINT8}: ssa.OpRsh16x8,
|
|
|
|
opAndTwoTypes{ORSH, TUINT16, TUINT8}: ssa.OpRsh16Ux8,
|
|
|
|
opAndTwoTypes{ORSH, TINT16, TUINT16}: ssa.OpRsh16x16,
|
|
|
|
opAndTwoTypes{ORSH, TUINT16, TUINT16}: ssa.OpRsh16Ux16,
|
|
|
|
opAndTwoTypes{ORSH, TINT16, TUINT32}: ssa.OpRsh16x32,
|
|
|
|
opAndTwoTypes{ORSH, TUINT16, TUINT32}: ssa.OpRsh16Ux32,
|
|
|
|
opAndTwoTypes{ORSH, TINT16, TUINT64}: ssa.OpRsh16x64,
|
|
|
|
opAndTwoTypes{ORSH, TUINT16, TUINT64}: ssa.OpRsh16Ux64,
|
|
|
|
|
|
|
|
opAndTwoTypes{ORSH, TINT32, TUINT8}: ssa.OpRsh32x8,
|
|
|
|
opAndTwoTypes{ORSH, TUINT32, TUINT8}: ssa.OpRsh32Ux8,
|
|
|
|
opAndTwoTypes{ORSH, TINT32, TUINT16}: ssa.OpRsh32x16,
|
|
|
|
opAndTwoTypes{ORSH, TUINT32, TUINT16}: ssa.OpRsh32Ux16,
|
|
|
|
opAndTwoTypes{ORSH, TINT32, TUINT32}: ssa.OpRsh32x32,
|
|
|
|
opAndTwoTypes{ORSH, TUINT32, TUINT32}: ssa.OpRsh32Ux32,
|
|
|
|
opAndTwoTypes{ORSH, TINT32, TUINT64}: ssa.OpRsh32x64,
|
|
|
|
opAndTwoTypes{ORSH, TUINT32, TUINT64}: ssa.OpRsh32Ux64,
|
|
|
|
|
|
|
|
opAndTwoTypes{ORSH, TINT64, TUINT8}: ssa.OpRsh64x8,
|
|
|
|
opAndTwoTypes{ORSH, TUINT64, TUINT8}: ssa.OpRsh64Ux8,
|
|
|
|
opAndTwoTypes{ORSH, TINT64, TUINT16}: ssa.OpRsh64x16,
|
|
|
|
opAndTwoTypes{ORSH, TUINT64, TUINT16}: ssa.OpRsh64Ux16,
|
|
|
|
opAndTwoTypes{ORSH, TINT64, TUINT32}: ssa.OpRsh64x32,
|
|
|
|
opAndTwoTypes{ORSH, TUINT64, TUINT32}: ssa.OpRsh64Ux32,
|
|
|
|
opAndTwoTypes{ORSH, TINT64, TUINT64}: ssa.OpRsh64x64,
|
|
|
|
opAndTwoTypes{ORSH, TUINT64, TUINT64}: ssa.OpRsh64Ux64,
|
|
|
|
}
|
|
|
|
|
2015-11-16 13:20:16 -08:00
|
|
|
func (s *state) ssaShiftOp(op Op, t *Type, u *Type) ssa.Op {
|
2015-07-29 17:07:09 -07:00
|
|
|
etype1 := s.concreteEtype(t)
|
|
|
|
etype2 := s.concreteEtype(u)
|
|
|
|
x, ok := shiftOpToSSA[opAndTwoTypes{op, etype1, etype2}]
|
|
|
|
if !ok {
|
2016-04-22 08:39:56 -07:00
|
|
|
s.Unimplementedf("unhandled shift op %s etype=%s/%s", op, etype1, etype2)
|
2015-07-29 17:07:09 -07:00
|
|
|
}
|
|
|
|
return x
|
|
|
|
}
|
|
|
|
|
2015-11-16 13:20:16 -08:00
|
|
|
func (s *state) ssaRotateOp(op Op, t *Type) ssa.Op {
|
2015-08-05 22:11:14 -04:00
|
|
|
etype1 := s.concreteEtype(t)
|
|
|
|
x, ok := opToSSA[opAndType{op, etype1}]
|
|
|
|
if !ok {
|
2016-04-22 08:39:56 -07:00
|
|
|
s.Unimplementedf("unhandled rotate op %s etype=%s", op, etype1)
|
2015-08-05 22:11:14 -04:00
|
|
|
}
|
|
|
|
return x
|
|
|
|
}
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
// expr converts the expression n to ssa, adds it to s and returns the ssa result.
|
2015-05-18 16:44:20 -07:00
|
|
|
func (s *state) expr(n *Node) *ssa.Value {
|
2016-04-26 15:22:33 -07:00
|
|
|
if !(n.Op == ONAME || n.Op == OLITERAL && n.Sym != nil) {
|
|
|
|
// ONAMEs and named OLITERALs have the line number
|
|
|
|
// of the decl, not the use. See issue 14742.
|
|
|
|
s.pushLine(n.Lineno)
|
|
|
|
defer s.popLine()
|
|
|
|
}
|
2015-05-30 01:03:06 -04:00
|
|
|
|
2015-07-11 11:39:12 -07:00
|
|
|
s.stmtList(n.Ninit)
|
2015-04-15 15:51:25 -07:00
|
|
|
switch n.Op {
|
2015-09-07 19:07:02 -05:00
|
|
|
case OCFUNC:
|
2016-04-29 09:02:27 -07:00
|
|
|
aux := s.lookupSymbol(n, &ssa.ExternSymbol{Typ: n.Type, Sym: n.Left.Sym})
|
2015-09-07 19:07:02 -05:00
|
|
|
return s.entryNewValue1A(ssa.OpAddr, n.Type, aux, s.sb)
|
2015-04-15 15:51:25 -07:00
|
|
|
case ONAME:
|
2015-06-10 15:03:06 -07:00
|
|
|
if n.Class == PFUNC {
|
|
|
|
// "value" of a function is the address of the function's closure
|
2015-06-19 21:02:28 -07:00
|
|
|
sym := funcsym(n.Sym)
|
2016-04-29 09:02:27 -07:00
|
|
|
aux := &ssa.ExternSymbol{Typ: n.Type, Sym: sym}
|
2015-06-19 21:02:28 -07:00
|
|
|
return s.entryNewValue1A(ssa.OpAddr, Ptrto(n.Type), aux, s.sb)
|
2015-06-10 15:03:06 -07:00
|
|
|
}
|
2016-02-27 17:49:31 -08:00
|
|
|
if s.canSSA(n) {
|
2015-06-19 21:02:28 -07:00
|
|
|
return s.variable(n, n.Type)
|
2015-05-12 15:16:52 -07:00
|
|
|
}
|
2016-05-27 14:07:37 -07:00
|
|
|
addr, _ := s.addr(n, false)
|
2015-06-11 21:29:25 -07:00
|
|
|
return s.newValue2(ssa.OpLoad, n.Type, addr, s.mem())
|
2015-09-11 16:40:05 -04:00
|
|
|
case OCLOSUREVAR:
|
2016-05-27 14:07:37 -07:00
|
|
|
addr, _ := s.addr(n, false)
|
2015-09-11 16:40:05 -04:00
|
|
|
return s.newValue2(ssa.OpLoad, n.Type, addr, s.mem())
|
2015-04-15 15:51:25 -07:00
|
|
|
case OLITERAL:
|
2016-04-22 12:27:29 -07:00
|
|
|
switch u := n.Val().U.(type) {
|
|
|
|
case *Mpint:
|
|
|
|
i := u.Int64()
|
2015-07-28 14:19:20 -07:00
|
|
|
switch n.Type.Size() {
|
|
|
|
case 1:
|
|
|
|
return s.constInt8(n.Type, int8(i))
|
|
|
|
case 2:
|
|
|
|
return s.constInt16(n.Type, int16(i))
|
|
|
|
case 4:
|
|
|
|
return s.constInt32(n.Type, int32(i))
|
|
|
|
case 8:
|
|
|
|
return s.constInt64(n.Type, i)
|
|
|
|
default:
|
|
|
|
s.Fatalf("bad integer size %d", n.Type.Size())
|
|
|
|
return nil
|
|
|
|
}
|
2016-04-22 12:27:29 -07:00
|
|
|
case string:
|
|
|
|
if u == "" {
|
2016-03-06 18:06:09 -08:00
|
|
|
return s.constEmptyString(n.Type)
|
|
|
|
}
|
2016-04-22 12:27:29 -07:00
|
|
|
return s.entryNewValue0A(ssa.OpConstString, n.Type, u)
|
|
|
|
case bool:
|
2016-04-26 15:22:33 -07:00
|
|
|
return s.constBool(u)
|
2016-04-22 12:27:29 -07:00
|
|
|
case *NilVal:
|
2015-08-18 10:26:28 -07:00
|
|
|
t := n.Type
|
|
|
|
switch {
|
|
|
|
case t.IsSlice():
|
2016-03-06 18:06:09 -08:00
|
|
|
return s.constSlice(t)
|
2015-08-18 10:26:28 -07:00
|
|
|
case t.IsInterface():
|
2016-03-06 18:06:09 -08:00
|
|
|
return s.constInterface(t)
|
2015-08-18 10:26:28 -07:00
|
|
|
default:
|
2016-03-06 18:06:09 -08:00
|
|
|
return s.constNil(t)
|
2015-08-18 10:26:28 -07:00
|
|
|
}
|
2016-04-22 12:27:29 -07:00
|
|
|
case *Mpflt:
|
2015-08-12 16:38:11 -04:00
|
|
|
switch n.Type.Size() {
|
|
|
|
case 4:
|
2016-04-22 12:27:29 -07:00
|
|
|
return s.constFloat32(n.Type, u.Float32())
|
2015-08-12 16:38:11 -04:00
|
|
|
case 8:
|
2016-04-22 12:27:29 -07:00
|
|
|
return s.constFloat64(n.Type, u.Float64())
|
2015-08-12 16:38:11 -04:00
|
|
|
default:
|
|
|
|
s.Fatalf("bad float size %d", n.Type.Size())
|
|
|
|
return nil
|
|
|
|
}
|
2016-04-22 12:27:29 -07:00
|
|
|
case *Mpcplx:
|
|
|
|
r := &u.Real
|
|
|
|
i := &u.Imag
|
2015-08-28 14:24:10 -04:00
|
|
|
switch n.Type.Size() {
|
|
|
|
case 8:
|
2016-04-22 12:27:29 -07:00
|
|
|
pt := Types[TFLOAT32]
|
|
|
|
return s.newValue2(ssa.OpComplexMake, n.Type,
|
|
|
|
s.constFloat32(pt, r.Float32()),
|
|
|
|
s.constFloat32(pt, i.Float32()))
|
2015-08-28 14:24:10 -04:00
|
|
|
case 16:
|
2016-04-22 12:27:29 -07:00
|
|
|
pt := Types[TFLOAT64]
|
|
|
|
return s.newValue2(ssa.OpComplexMake, n.Type,
|
|
|
|
s.constFloat64(pt, r.Float64()),
|
|
|
|
s.constFloat64(pt, i.Float64()))
|
2015-08-28 14:24:10 -04:00
|
|
|
default:
|
|
|
|
s.Fatalf("bad float size %d", n.Type.Size())
|
|
|
|
return nil
|
|
|
|
}
|
2015-08-12 16:38:11 -04:00
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
default:
|
2015-06-24 14:03:39 -07:00
|
|
|
s.Unimplementedf("unhandled OLITERAL %v", n.Val().Ctype())
|
2015-04-15 15:51:25 -07:00
|
|
|
return nil
|
|
|
|
}
|
2015-06-12 16:24:33 -07:00
|
|
|
case OCONVNOP:
|
2015-07-28 14:31:25 -07:00
|
|
|
to := n.Type
|
|
|
|
from := n.Left.Type
|
|
|
|
|
|
|
|
// Assume everything will work out, so set up our return value.
|
|
|
|
// Anything interesting that happens from here is a fatal.
|
2015-06-12 16:24:33 -07:00
|
|
|
x := s.expr(n.Left)
|
2015-10-19 11:36:07 -04:00
|
|
|
|
|
|
|
// Special case for not confusing GC and liveness.
|
|
|
|
// We don't want pointers accidentally classified
|
|
|
|
// as not-pointers or vice-versa because of copy
|
|
|
|
// elision.
|
2016-03-28 10:55:44 -07:00
|
|
|
if to.IsPtrShaped() != from.IsPtrShaped() {
|
2015-11-10 15:35:36 -08:00
|
|
|
return s.newValue2(ssa.OpConvert, to, x, s.mem())
|
2015-10-19 11:36:07 -04:00
|
|
|
}
|
|
|
|
|
2015-07-28 14:31:25 -07:00
|
|
|
v := s.newValue1(ssa.OpCopy, to, x) // ensure that v has the right type
|
|
|
|
|
2015-09-07 19:07:02 -05:00
|
|
|
// CONVNOP closure
|
2016-03-28 10:55:44 -07:00
|
|
|
if to.Etype == TFUNC && from.IsPtrShaped() {
|
2015-09-07 19:07:02 -05:00
|
|
|
return v
|
|
|
|
}
|
|
|
|
|
2015-07-28 14:31:25 -07:00
|
|
|
// named <--> unnamed type or typed <--> untyped const
|
|
|
|
if from.Etype == to.Etype {
|
|
|
|
return v
|
|
|
|
}
|
2015-10-19 11:36:07 -04:00
|
|
|
|
2015-07-28 14:31:25 -07:00
|
|
|
// unsafe.Pointer <--> *T
|
|
|
|
if to.Etype == TUNSAFEPTR && from.IsPtr() || from.Etype == TUNSAFEPTR && to.IsPtr() {
|
|
|
|
return v
|
|
|
|
}
|
|
|
|
|
|
|
|
dowidth(from)
|
|
|
|
dowidth(to)
|
|
|
|
if from.Width != to.Width {
|
|
|
|
s.Fatalf("CONVNOP width mismatch %v (%d) -> %v (%d)\n", from, from.Width, to, to.Width)
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
if etypesign(from.Etype) != etypesign(to.Etype) {
|
2016-04-22 08:39:56 -07:00
|
|
|
s.Fatalf("CONVNOP sign mismatch %v (%s) -> %v (%s)\n", from, from.Etype, to, to.Etype)
|
2015-07-28 14:31:25 -07:00
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
2016-03-01 15:17:34 -08:00
|
|
|
if instrumenting {
|
2015-10-09 16:48:30 -04:00
|
|
|
// These appear to be fine, but they fail the
|
|
|
|
// integer constraint below, so okay them here.
|
|
|
|
// Sample non-integer conversion: map[string]string -> *uint8
|
|
|
|
return v
|
2015-07-28 14:31:25 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
if etypesign(from.Etype) == 0 {
|
|
|
|
s.Fatalf("CONVNOP unrecognized non-integer %v -> %v\n", from, to)
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// integer, same width, same sign
|
|
|
|
return v
|
|
|
|
|
2015-06-14 11:38:46 -07:00
|
|
|
case OCONV:
|
|
|
|
x := s.expr(n.Left)
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
ft := n.Left.Type // from type
|
|
|
|
tt := n.Type // to type
|
|
|
|
if ft.IsInteger() && tt.IsInteger() {
|
|
|
|
var op ssa.Op
|
|
|
|
if tt.Size() == ft.Size() {
|
2015-07-28 14:31:25 -07:00
|
|
|
op = ssa.OpCopy
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
} else if tt.Size() < ft.Size() {
|
|
|
|
// truncation
|
|
|
|
switch 10*ft.Size() + tt.Size() {
|
|
|
|
case 21:
|
|
|
|
op = ssa.OpTrunc16to8
|
|
|
|
case 41:
|
|
|
|
op = ssa.OpTrunc32to8
|
|
|
|
case 42:
|
|
|
|
op = ssa.OpTrunc32to16
|
|
|
|
case 81:
|
|
|
|
op = ssa.OpTrunc64to8
|
|
|
|
case 82:
|
|
|
|
op = ssa.OpTrunc64to16
|
|
|
|
case 84:
|
|
|
|
op = ssa.OpTrunc64to32
|
|
|
|
default:
|
|
|
|
s.Fatalf("weird integer truncation %s -> %s", ft, tt)
|
|
|
|
}
|
|
|
|
} else if ft.IsSigned() {
|
|
|
|
// sign extension
|
|
|
|
switch 10*ft.Size() + tt.Size() {
|
|
|
|
case 12:
|
|
|
|
op = ssa.OpSignExt8to16
|
|
|
|
case 14:
|
|
|
|
op = ssa.OpSignExt8to32
|
|
|
|
case 18:
|
|
|
|
op = ssa.OpSignExt8to64
|
|
|
|
case 24:
|
|
|
|
op = ssa.OpSignExt16to32
|
|
|
|
case 28:
|
|
|
|
op = ssa.OpSignExt16to64
|
|
|
|
case 48:
|
|
|
|
op = ssa.OpSignExt32to64
|
|
|
|
default:
|
|
|
|
s.Fatalf("bad integer sign extension %s -> %s", ft, tt)
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// zero extension
|
|
|
|
switch 10*ft.Size() + tt.Size() {
|
|
|
|
case 12:
|
|
|
|
op = ssa.OpZeroExt8to16
|
|
|
|
case 14:
|
|
|
|
op = ssa.OpZeroExt8to32
|
|
|
|
case 18:
|
|
|
|
op = ssa.OpZeroExt8to64
|
|
|
|
case 24:
|
|
|
|
op = ssa.OpZeroExt16to32
|
|
|
|
case 28:
|
|
|
|
op = ssa.OpZeroExt16to64
|
|
|
|
case 48:
|
|
|
|
op = ssa.OpZeroExt32to64
|
|
|
|
default:
|
|
|
|
s.Fatalf("weird integer sign extension %s -> %s", ft, tt)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return s.newValue1(op, n.Type, x)
|
|
|
|
}
|
2015-08-20 15:14:20 -04:00
|
|
|
|
2015-09-01 17:09:00 -04:00
|
|
|
if ft.IsFloat() || tt.IsFloat() {
|
|
|
|
conv, ok := fpConvOpToSSA[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]
|
2016-08-08 11:26:25 -07:00
|
|
|
if s.config.IntSize == 4 && Thearch.LinkArch.Name != "amd64p32" {
|
2016-05-31 11:27:16 -04:00
|
|
|
if conv1, ok1 := fpConvOpToSSA32[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]; ok1 {
|
|
|
|
conv = conv1
|
|
|
|
}
|
|
|
|
}
|
2016-08-16 14:17:33 -04:00
|
|
|
if Thearch.LinkArch.Name == "arm64" {
|
|
|
|
if conv1, ok1 := uint64fpConvOpToSSA[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]; ok1 {
|
|
|
|
conv = conv1
|
|
|
|
}
|
|
|
|
}
|
2015-09-01 17:09:00 -04:00
|
|
|
if !ok {
|
|
|
|
s.Fatalf("weird float conversion %s -> %s", ft, tt)
|
2015-08-20 15:14:20 -04:00
|
|
|
}
|
2015-09-01 17:09:00 -04:00
|
|
|
op1, op2, it := conv.op1, conv.op2, conv.intermediateType
|
|
|
|
|
|
|
|
if op1 != ssa.OpInvalid && op2 != ssa.OpInvalid {
|
|
|
|
// normal case, not tripping over unsigned 64
|
|
|
|
if op1 == ssa.OpCopy {
|
|
|
|
if op2 == ssa.OpCopy {
|
|
|
|
return x
|
|
|
|
}
|
|
|
|
return s.newValue1(op2, n.Type, x)
|
|
|
|
}
|
|
|
|
if op2 == ssa.OpCopy {
|
|
|
|
return s.newValue1(op1, n.Type, x)
|
|
|
|
}
|
|
|
|
return s.newValue1(op2, n.Type, s.newValue1(op1, Types[it], x))
|
|
|
|
}
|
|
|
|
// Tricky 64-bit unsigned cases.
|
|
|
|
if ft.IsInteger() {
|
|
|
|
// therefore tt is float32 or float64, and ft is also unsigned
|
2015-08-20 15:14:20 -04:00
|
|
|
if tt.Size() == 4 {
|
|
|
|
return s.uint64Tofloat32(n, x, ft, tt)
|
|
|
|
}
|
|
|
|
if tt.Size() == 8 {
|
|
|
|
return s.uint64Tofloat64(n, x, ft, tt)
|
|
|
|
}
|
2015-09-01 17:09:00 -04:00
|
|
|
s.Fatalf("weird unsigned integer to float conversion %s -> %s", ft, tt)
|
2015-08-26 14:25:40 -04:00
|
|
|
}
|
2015-09-01 17:09:00 -04:00
|
|
|
// therefore ft is float32 or float64, and tt is unsigned integer
|
2015-08-26 14:25:40 -04:00
|
|
|
if ft.Size() == 4 {
|
2015-09-01 17:09:00 -04:00
|
|
|
return s.float32ToUint64(n, x, ft, tt)
|
2015-08-26 14:25:40 -04:00
|
|
|
}
|
2015-09-01 17:09:00 -04:00
|
|
|
if ft.Size() == 8 {
|
|
|
|
return s.float64ToUint64(n, x, ft, tt)
|
2015-08-26 14:25:40 -04:00
|
|
|
}
|
2015-09-01 17:09:00 -04:00
|
|
|
s.Fatalf("weird float to unsigned integer conversion %s -> %s", ft, tt)
|
|
|
|
return nil
|
2015-08-20 15:14:20 -04:00
|
|
|
}
|
2015-08-28 14:24:10 -04:00
|
|
|
|
|
|
|
if ft.IsComplex() && tt.IsComplex() {
|
|
|
|
var op ssa.Op
|
|
|
|
if ft.Size() == tt.Size() {
|
|
|
|
op = ssa.OpCopy
|
|
|
|
} else if ft.Size() == 8 && tt.Size() == 16 {
|
|
|
|
op = ssa.OpCvt32Fto64F
|
|
|
|
} else if ft.Size() == 16 && tt.Size() == 8 {
|
|
|
|
op = ssa.OpCvt64Fto32F
|
|
|
|
} else {
|
|
|
|
s.Fatalf("weird complex conversion %s -> %s", ft, tt)
|
|
|
|
}
|
|
|
|
ftp := floatForComplex(ft)
|
|
|
|
ttp := floatForComplex(tt)
|
|
|
|
return s.newValue2(ssa.OpComplexMake, tt,
|
|
|
|
s.newValue1(op, ttp, s.newValue1(ssa.OpComplexReal, ftp, x)),
|
|
|
|
s.newValue1(op, ttp, s.newValue1(ssa.OpComplexImag, ftp, x)))
|
|
|
|
}
|
2015-08-20 15:14:20 -04:00
|
|
|
|
2016-04-22 08:39:56 -07:00
|
|
|
s.Unimplementedf("unhandled OCONV %s -> %s", n.Left.Type.Etype, n.Type.Etype)
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
return nil
|
2015-05-18 16:44:20 -07:00
|
|
|
|
2015-09-17 10:31:16 -07:00
|
|
|
case ODOTTYPE:
|
|
|
|
res, _ := s.dottype(n, false)
|
|
|
|
return res
|
|
|
|
|
2015-06-24 17:48:22 -07:00
|
|
|
// binary ops
|
|
|
|
case OLT, OEQ, ONE, OLE, OGE, OGT:
|
2015-05-28 10:47:24 -07:00
|
|
|
a := s.expr(n.Left)
|
|
|
|
b := s.expr(n.Right)
|
2015-09-10 11:05:42 -07:00
|
|
|
if n.Left.Type.IsComplex() {
|
2015-09-10 14:59:00 -07:00
|
|
|
pt := floatForComplex(n.Left.Type)
|
2015-09-10 11:05:42 -07:00
|
|
|
op := s.ssaOp(OEQ, pt)
|
|
|
|
r := s.newValue2(op, Types[TBOOL], s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b))
|
|
|
|
i := s.newValue2(op, Types[TBOOL], s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b))
|
|
|
|
c := s.newValue2(ssa.OpAnd8, Types[TBOOL], r, i)
|
|
|
|
switch n.Op {
|
|
|
|
case OEQ:
|
|
|
|
return c
|
|
|
|
case ONE:
|
|
|
|
return s.newValue1(ssa.OpNot, Types[TBOOL], c)
|
|
|
|
default:
|
2016-04-22 07:14:10 -07:00
|
|
|
s.Fatalf("ordered complex compare %s", n.Op)
|
2015-09-10 11:05:42 -07:00
|
|
|
}
|
|
|
|
}
|
2015-07-30 11:03:05 -07:00
|
|
|
return s.newValue2(s.ssaOp(n.Op, n.Left.Type), Types[TBOOL], a, b)
|
2015-08-28 14:24:10 -04:00
|
|
|
case OMUL:
|
|
|
|
a := s.expr(n.Left)
|
|
|
|
b := s.expr(n.Right)
|
|
|
|
if n.Type.IsComplex() {
|
|
|
|
mulop := ssa.OpMul64F
|
|
|
|
addop := ssa.OpAdd64F
|
|
|
|
subop := ssa.OpSub64F
|
|
|
|
pt := floatForComplex(n.Type) // Could be Float32 or Float64
|
2016-06-30 19:24:06 +00:00
|
|
|
wt := Types[TFLOAT64] // Compute in Float64 to minimize cancelation error
|
2015-08-28 14:24:10 -04:00
|
|
|
|
|
|
|
areal := s.newValue1(ssa.OpComplexReal, pt, a)
|
|
|
|
breal := s.newValue1(ssa.OpComplexReal, pt, b)
|
|
|
|
aimag := s.newValue1(ssa.OpComplexImag, pt, a)
|
|
|
|
bimag := s.newValue1(ssa.OpComplexImag, pt, b)
|
|
|
|
|
|
|
|
if pt != wt { // Widen for calculation
|
|
|
|
areal = s.newValue1(ssa.OpCvt32Fto64F, wt, areal)
|
|
|
|
breal = s.newValue1(ssa.OpCvt32Fto64F, wt, breal)
|
|
|
|
aimag = s.newValue1(ssa.OpCvt32Fto64F, wt, aimag)
|
|
|
|
bimag = s.newValue1(ssa.OpCvt32Fto64F, wt, bimag)
|
|
|
|
}
|
|
|
|
|
|
|
|
xreal := s.newValue2(subop, wt, s.newValue2(mulop, wt, areal, breal), s.newValue2(mulop, wt, aimag, bimag))
|
|
|
|
ximag := s.newValue2(addop, wt, s.newValue2(mulop, wt, areal, bimag), s.newValue2(mulop, wt, aimag, breal))
|
|
|
|
|
|
|
|
if pt != wt { // Narrow to store back
|
|
|
|
xreal = s.newValue1(ssa.OpCvt64Fto32F, pt, xreal)
|
|
|
|
ximag = s.newValue1(ssa.OpCvt64Fto32F, pt, ximag)
|
|
|
|
}
|
|
|
|
|
|
|
|
return s.newValue2(ssa.OpComplexMake, n.Type, xreal, ximag)
|
|
|
|
}
|
|
|
|
return s.newValue2(s.ssaOp(n.Op, n.Type), a.Type, a, b)
|
|
|
|
|
|
|
|
case ODIV:
|
|
|
|
a := s.expr(n.Left)
|
|
|
|
b := s.expr(n.Right)
|
|
|
|
if n.Type.IsComplex() {
|
|
|
|
// TODO this is not executed because the front-end substitutes a runtime call.
|
|
|
|
// That probably ought to change; with modest optimization the widen/narrow
|
|
|
|
// conversions could all be elided in larger expression trees.
|
|
|
|
mulop := ssa.OpMul64F
|
|
|
|
addop := ssa.OpAdd64F
|
|
|
|
subop := ssa.OpSub64F
|
|
|
|
divop := ssa.OpDiv64F
|
|
|
|
pt := floatForComplex(n.Type) // Could be Float32 or Float64
|
2016-06-30 19:24:06 +00:00
|
|
|
wt := Types[TFLOAT64] // Compute in Float64 to minimize cancelation error
|
2015-08-28 14:24:10 -04:00
|
|
|
|
|
|
|
areal := s.newValue1(ssa.OpComplexReal, pt, a)
|
|
|
|
breal := s.newValue1(ssa.OpComplexReal, pt, b)
|
|
|
|
aimag := s.newValue1(ssa.OpComplexImag, pt, a)
|
|
|
|
bimag := s.newValue1(ssa.OpComplexImag, pt, b)
|
|
|
|
|
|
|
|
if pt != wt { // Widen for calculation
|
|
|
|
areal = s.newValue1(ssa.OpCvt32Fto64F, wt, areal)
|
|
|
|
breal = s.newValue1(ssa.OpCvt32Fto64F, wt, breal)
|
|
|
|
aimag = s.newValue1(ssa.OpCvt32Fto64F, wt, aimag)
|
|
|
|
bimag = s.newValue1(ssa.OpCvt32Fto64F, wt, bimag)
|
|
|
|
}
|
|
|
|
|
|
|
|
denom := s.newValue2(addop, wt, s.newValue2(mulop, wt, breal, breal), s.newValue2(mulop, wt, bimag, bimag))
|
|
|
|
xreal := s.newValue2(addop, wt, s.newValue2(mulop, wt, areal, breal), s.newValue2(mulop, wt, aimag, bimag))
|
|
|
|
ximag := s.newValue2(subop, wt, s.newValue2(mulop, wt, aimag, breal), s.newValue2(mulop, wt, areal, bimag))
|
|
|
|
|
|
|
|
// TODO not sure if this is best done in wide precision or narrow
|
|
|
|
// Double-rounding might be an issue.
|
|
|
|
// Note that the pre-SSA implementation does the entire calculation
|
|
|
|
// in wide format, so wide is compatible.
|
|
|
|
xreal = s.newValue2(divop, wt, xreal, denom)
|
|
|
|
ximag = s.newValue2(divop, wt, ximag, denom)
|
|
|
|
|
|
|
|
if pt != wt { // Narrow to store back
|
|
|
|
xreal = s.newValue1(ssa.OpCvt64Fto32F, pt, xreal)
|
|
|
|
ximag = s.newValue1(ssa.OpCvt64Fto32F, pt, ximag)
|
|
|
|
}
|
|
|
|
return s.newValue2(ssa.OpComplexMake, n.Type, xreal, ximag)
|
|
|
|
}
|
2015-10-28 13:55:46 -04:00
|
|
|
if n.Type.IsFloat() {
|
|
|
|
return s.newValue2(s.ssaOp(n.Op, n.Type), a.Type, a, b)
|
|
|
|
} else {
|
|
|
|
// do a size-appropriate check for zero
|
|
|
|
cmp := s.newValue2(s.ssaOp(ONE, n.Type), Types[TBOOL], b, s.zeroVal(n.Type))
|
|
|
|
s.check(cmp, panicdivide)
|
|
|
|
return s.newValue2(s.ssaOp(n.Op, n.Type), a.Type, a, b)
|
|
|
|
}
|
|
|
|
case OMOD:
|
|
|
|
a := s.expr(n.Left)
|
|
|
|
b := s.expr(n.Right)
|
|
|
|
// do a size-appropriate check for zero
|
|
|
|
cmp := s.newValue2(s.ssaOp(ONE, n.Type), Types[TBOOL], b, s.zeroVal(n.Type))
|
|
|
|
s.check(cmp, panicdivide)
|
2015-08-28 14:24:10 -04:00
|
|
|
return s.newValue2(s.ssaOp(n.Op, n.Type), a.Type, a, b)
|
|
|
|
case OADD, OSUB:
|
|
|
|
a := s.expr(n.Left)
|
|
|
|
b := s.expr(n.Right)
|
|
|
|
if n.Type.IsComplex() {
|
|
|
|
pt := floatForComplex(n.Type)
|
|
|
|
op := s.ssaOp(n.Op, pt)
|
|
|
|
return s.newValue2(ssa.OpComplexMake, n.Type,
|
|
|
|
s.newValue2(op, pt, s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b)),
|
|
|
|
s.newValue2(op, pt, s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b)))
|
|
|
|
}
|
|
|
|
return s.newValue2(s.ssaOp(n.Op, n.Type), a.Type, a, b)
|
2015-10-28 13:55:46 -04:00
|
|
|
case OAND, OOR, OHMUL, OXOR:
|
2015-05-28 10:47:24 -07:00
|
|
|
a := s.expr(n.Left)
|
|
|
|
b := s.expr(n.Right)
|
2015-07-19 15:48:20 -07:00
|
|
|
return s.newValue2(s.ssaOp(n.Op, n.Type), a.Type, a, b)
|
2015-07-29 17:07:09 -07:00
|
|
|
case OLSH, ORSH:
|
|
|
|
a := s.expr(n.Left)
|
|
|
|
b := s.expr(n.Right)
|
|
|
|
return s.newValue2(s.ssaShiftOp(n.Op, n.Type, n.Right.Type), a.Type, a, b)
|
2015-08-05 22:11:14 -04:00
|
|
|
case OLROT:
|
|
|
|
a := s.expr(n.Left)
|
2016-04-01 14:51:02 -07:00
|
|
|
i := n.Right.Int64()
|
2015-08-05 22:11:14 -04:00
|
|
|
if i <= 0 || i >= n.Type.Size()*8 {
|
|
|
|
s.Fatalf("Wrong rotate distance for LROT, expected 1 through %d, saw %d", n.Type.Size()*8-1, i)
|
|
|
|
}
|
|
|
|
return s.newValue1I(s.ssaRotateOp(n.Op, n.Type), a.Type, i, a)
|
2015-07-10 12:58:53 -06:00
|
|
|
case OANDAND, OOROR:
|
|
|
|
// To implement OANDAND (and OOROR), we introduce a
|
|
|
|
// new temporary variable to hold the result. The
|
|
|
|
// variable is associated with the OANDAND node in the
|
|
|
|
// s.vars table (normally variables are only
|
|
|
|
// associated with ONAME nodes). We convert
|
|
|
|
// A && B
|
|
|
|
// to
|
|
|
|
// var = A
|
|
|
|
// if var {
|
|
|
|
// var = B
|
|
|
|
// }
|
|
|
|
// Using var in the subsequent block introduces the
|
|
|
|
// necessary phi variable.
|
|
|
|
el := s.expr(n.Left)
|
|
|
|
s.vars[n] = el
|
|
|
|
|
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(el)
|
2015-08-11 17:28:56 -07:00
|
|
|
// In theory, we should set b.Likely here based on context.
|
|
|
|
// However, gc only gives us likeliness hints
|
|
|
|
// in a single place, for plain OIF statements,
|
|
|
|
// and passing around context is finnicky, so don't bother for now.
|
2015-07-10 12:58:53 -06:00
|
|
|
|
|
|
|
bRight := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bResult := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
if n.Op == OANDAND {
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bRight)
|
|
|
|
b.AddEdgeTo(bResult)
|
2015-07-10 12:58:53 -06:00
|
|
|
} else if n.Op == OOROR {
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bResult)
|
|
|
|
b.AddEdgeTo(bRight)
|
2015-07-10 12:58:53 -06:00
|
|
|
}
|
|
|
|
|
|
|
|
s.startBlock(bRight)
|
|
|
|
er := s.expr(n.Right)
|
|
|
|
s.vars[n] = er
|
|
|
|
|
|
|
|
b = s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bResult)
|
2015-07-10 12:58:53 -06:00
|
|
|
|
|
|
|
s.startBlock(bResult)
|
2015-08-27 10:11:08 -07:00
|
|
|
return s.variable(n, Types[TBOOL])
|
2015-09-12 14:14:02 -07:00
|
|
|
case OCOMPLEX:
|
|
|
|
r := s.expr(n.Left)
|
|
|
|
i := s.expr(n.Right)
|
|
|
|
return s.newValue2(ssa.OpComplexMake, n.Type, r, i)
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-09-05 19:28:00 -07:00
|
|
|
// unary ops
|
2015-08-28 14:24:10 -04:00
|
|
|
case OMINUS:
|
|
|
|
a := s.expr(n.Left)
|
|
|
|
if n.Type.IsComplex() {
|
|
|
|
tp := floatForComplex(n.Type)
|
|
|
|
negop := s.ssaOp(n.Op, tp)
|
|
|
|
return s.newValue2(ssa.OpComplexMake, n.Type,
|
|
|
|
s.newValue1(negop, tp, s.newValue1(ssa.OpComplexReal, tp, a)),
|
|
|
|
s.newValue1(negop, tp, s.newValue1(ssa.OpComplexImag, tp, a)))
|
|
|
|
}
|
|
|
|
return s.newValue1(s.ssaOp(n.Op, n.Type), a.Type, a)
|
2015-09-12 13:26:57 -07:00
|
|
|
case ONOT, OCOM, OSQRT:
|
2015-07-10 11:25:48 -06:00
|
|
|
a := s.expr(n.Left)
|
2015-07-21 16:58:18 +02:00
|
|
|
return s.newValue1(s.ssaOp(n.Op, n.Type), a.Type, a)
|
2015-09-10 11:37:09 -07:00
|
|
|
case OIMAG, OREAL:
|
|
|
|
a := s.expr(n.Left)
|
|
|
|
return s.newValue1(s.ssaOp(n.Op, n.Left.Type), n.Type, a)
|
2015-09-05 19:28:00 -07:00
|
|
|
case OPLUS:
|
|
|
|
return s.expr(n.Left)
|
2015-07-10 11:25:48 -06:00
|
|
|
|
2015-05-18 16:44:20 -07:00
|
|
|
case OADDR:
|
2016-05-27 14:07:37 -07:00
|
|
|
a, _ := s.addr(n.Left, n.Bounded)
|
|
|
|
// Note we know the volatile result is false because you can't write &f() in Go.
|
|
|
|
return a
|
2015-05-18 16:44:20 -07:00
|
|
|
|
2015-07-28 12:37:46 -07:00
|
|
|
case OINDREG:
|
|
|
|
if int(n.Reg) != Thearch.REGSP {
|
|
|
|
s.Unimplementedf("OINDREG of non-SP register %s in expr: %v", obj.Rconv(int(n.Reg)), n)
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
addr := s.entryNewValue1I(ssa.OpOffPtr, Ptrto(n.Type), n.Xoffset, s.sp)
|
|
|
|
return s.newValue2(ssa.OpLoad, n.Type, addr, s.mem())
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
case OIND:
|
2016-04-19 21:06:53 -07:00
|
|
|
p := s.exprPtr(n.Left, false, n.Lineno)
|
2015-06-11 21:29:25 -07:00
|
|
|
return s.newValue2(ssa.OpLoad, n.Type, p, s.mem())
|
2015-05-18 16:44:20 -07:00
|
|
|
|
2015-07-15 21:33:49 -07:00
|
|
|
case ODOT:
|
2016-01-11 21:05:33 -08:00
|
|
|
t := n.Left.Type
|
|
|
|
if canSSAType(t) {
|
|
|
|
v := s.expr(n.Left)
|
2016-03-14 12:45:18 -07:00
|
|
|
return s.newValue1I(ssa.OpStructSelect, n.Type, int64(fieldIdx(n)), v)
|
2016-01-11 21:05:33 -08:00
|
|
|
}
|
2016-05-27 14:07:37 -07:00
|
|
|
p, _ := s.addr(n, false)
|
2015-10-09 09:33:29 -07:00
|
|
|
return s.newValue2(ssa.OpLoad, n.Type, p, s.mem())
|
2015-07-15 21:33:49 -07:00
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
case ODOTPTR:
|
2016-04-19 21:06:53 -07:00
|
|
|
p := s.exprPtr(n.Left, false, n.Lineno)
|
2016-03-04 12:34:43 -08:00
|
|
|
p = s.newValue1I(ssa.OpOffPtr, p.Type, n.Xoffset, p)
|
2015-06-11 21:29:25 -07:00
|
|
|
return s.newValue2(ssa.OpLoad, n.Type, p, s.mem())
|
2015-04-15 15:51:25 -07:00
|
|
|
|
|
|
|
case OINDEX:
|
2015-10-09 09:33:29 -07:00
|
|
|
switch {
|
|
|
|
case n.Left.Type.IsString():
|
2015-05-18 16:44:20 -07:00
|
|
|
a := s.expr(n.Left)
|
|
|
|
i := s.expr(n.Right)
|
2016-05-25 09:49:28 -04:00
|
|
|
i = s.extendIndex(i, Panicindex)
|
2015-10-09 09:33:29 -07:00
|
|
|
if !n.Bounded {
|
|
|
|
len := s.newValue1(ssa.OpStringLen, Types[TINT], a)
|
|
|
|
s.boundsCheck(i, len)
|
2015-08-18 14:17:30 -07:00
|
|
|
}
|
2015-10-09 09:33:29 -07:00
|
|
|
ptrtyp := Ptrto(Types[TUINT8])
|
|
|
|
ptr := s.newValue1(ssa.OpStringPtr, ptrtyp, a)
|
2016-03-04 12:34:43 -08:00
|
|
|
if Isconst(n.Right, CTINT) {
|
2016-04-01 14:51:02 -07:00
|
|
|
ptr = s.newValue1I(ssa.OpOffPtr, ptrtyp, n.Right.Int64(), ptr)
|
2016-03-04 12:34:43 -08:00
|
|
|
} else {
|
|
|
|
ptr = s.newValue2(ssa.OpAddPtr, ptrtyp, ptr, i)
|
|
|
|
}
|
2015-10-09 09:33:29 -07:00
|
|
|
return s.newValue2(ssa.OpLoad, Types[TUINT8], ptr, s.mem())
|
|
|
|
case n.Left.Type.IsSlice():
|
2016-05-27 14:07:37 -07:00
|
|
|
p, _ := s.addr(n, false)
|
2016-03-30 10:57:47 -07:00
|
|
|
return s.newValue2(ssa.OpLoad, n.Left.Type.Elem(), p, s.mem())
|
2015-10-09 09:33:29 -07:00
|
|
|
case n.Left.Type.IsArray():
|
|
|
|
// TODO: fix when we can SSA arrays of length 1.
|
2016-05-27 14:07:37 -07:00
|
|
|
p, _ := s.addr(n, false)
|
2016-03-30 10:57:47 -07:00
|
|
|
return s.newValue2(ssa.OpLoad, n.Left.Type.Elem(), p, s.mem())
|
2015-10-09 09:33:29 -07:00
|
|
|
default:
|
|
|
|
s.Fatalf("bad type for index %v", n.Left.Type)
|
|
|
|
return nil
|
2015-05-18 16:44:20 -07:00
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
|
2015-07-10 10:47:28 -06:00
|
|
|
case OLEN, OCAP:
|
2015-07-03 18:41:28 -07:00
|
|
|
switch {
|
2015-07-10 10:47:28 -06:00
|
|
|
case n.Left.Type.IsSlice():
|
|
|
|
op := ssa.OpSliceLen
|
|
|
|
if n.Op == OCAP {
|
|
|
|
op = ssa.OpSliceCap
|
|
|
|
}
|
2015-07-30 11:03:05 -07:00
|
|
|
return s.newValue1(op, Types[TINT], s.expr(n.Left))
|
2015-07-10 10:47:28 -06:00
|
|
|
case n.Left.Type.IsString(): // string; not reachable for OCAP
|
2015-07-30 11:03:05 -07:00
|
|
|
return s.newValue1(ssa.OpStringLen, Types[TINT], s.expr(n.Left))
|
2015-08-28 15:56:43 -05:00
|
|
|
case n.Left.Type.IsMap(), n.Left.Type.IsChan():
|
|
|
|
return s.referenceTypeBuiltin(n, s.expr(n.Left))
|
2015-07-03 18:41:28 -07:00
|
|
|
default: // array
|
2016-03-31 14:46:04 -07:00
|
|
|
return s.constInt(Types[TINT], n.Left.Type.NumElem())
|
2015-07-03 18:41:28 -07:00
|
|
|
}
|
|
|
|
|
2015-08-12 10:12:14 -07:00
|
|
|
case OSPTR:
|
|
|
|
a := s.expr(n.Left)
|
|
|
|
if n.Left.Type.IsSlice() {
|
|
|
|
return s.newValue1(ssa.OpSlicePtr, n.Type, a)
|
|
|
|
} else {
|
|
|
|
return s.newValue1(ssa.OpStringPtr, n.Type, a)
|
|
|
|
}
|
|
|
|
|
2015-08-04 15:47:22 -07:00
|
|
|
case OITAB:
|
|
|
|
a := s.expr(n.Left)
|
|
|
|
return s.newValue1(ssa.OpITab, n.Type, a)
|
|
|
|
|
2016-06-06 12:38:19 -07:00
|
|
|
case OIDATA:
|
|
|
|
a := s.expr(n.Left)
|
|
|
|
return s.newValue1(ssa.OpIData, n.Type, a)
|
|
|
|
|
2015-09-05 19:28:27 -07:00
|
|
|
case OEFACE:
|
|
|
|
tab := s.expr(n.Left)
|
|
|
|
data := s.expr(n.Right)
|
2015-10-07 14:35:25 -07:00
|
|
|
// The frontend allows putting things like struct{*byte} in
|
2016-03-01 23:21:55 +00:00
|
|
|
// the data portion of an eface. But we don't want struct{*byte}
|
2015-10-07 14:35:25 -07:00
|
|
|
// as a register type because (among other reasons) the liveness
|
|
|
|
// analysis is confused by the "fat" variables that result from
|
|
|
|
// such types being spilled.
|
|
|
|
// So here we ensure that we are selecting the underlying pointer
|
|
|
|
// when we build an eface.
|
2016-01-11 21:05:33 -08:00
|
|
|
// TODO: get rid of this now that structs can be SSA'd?
|
2016-03-28 10:55:44 -07:00
|
|
|
for !data.Type.IsPtrShaped() {
|
2015-10-07 14:35:25 -07:00
|
|
|
switch {
|
|
|
|
case data.Type.IsArray():
|
2016-03-10 14:35:39 -08:00
|
|
|
data = s.newValue1I(ssa.OpArrayIndex, data.Type.ElemType(), 0, data)
|
2015-10-07 14:35:25 -07:00
|
|
|
case data.Type.IsStruct():
|
|
|
|
for i := data.Type.NumFields() - 1; i >= 0; i-- {
|
|
|
|
f := data.Type.FieldType(i)
|
|
|
|
if f.Size() == 0 {
|
|
|
|
// eface type could also be struct{p *byte; q [0]int}
|
|
|
|
continue
|
|
|
|
}
|
2016-03-14 12:45:18 -07:00
|
|
|
data = s.newValue1I(ssa.OpStructSelect, f, int64(i), data)
|
2015-10-07 14:35:25 -07:00
|
|
|
break
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
s.Fatalf("type being put into an eface isn't a pointer")
|
|
|
|
}
|
|
|
|
}
|
2015-09-05 19:28:27 -07:00
|
|
|
return s.newValue2(ssa.OpIMake, n.Type, tab, data)
|
|
|
|
|
2016-04-21 11:55:33 -07:00
|
|
|
case OSLICE, OSLICEARR, OSLICE3, OSLICE3ARR:
|
2015-09-12 23:27:26 -07:00
|
|
|
v := s.expr(n.Left)
|
2016-04-21 11:55:33 -07:00
|
|
|
var i, j, k *ssa.Value
|
|
|
|
low, high, max := n.SliceBounds()
|
|
|
|
if low != nil {
|
2016-05-25 09:49:28 -04:00
|
|
|
i = s.extendIndex(s.expr(low), panicslice)
|
2015-09-12 23:27:26 -07:00
|
|
|
}
|
2016-04-21 11:55:33 -07:00
|
|
|
if high != nil {
|
2016-05-25 09:49:28 -04:00
|
|
|
j = s.extendIndex(s.expr(high), panicslice)
|
2015-09-12 23:27:26 -07:00
|
|
|
}
|
2016-04-21 11:55:33 -07:00
|
|
|
if max != nil {
|
2016-05-25 09:49:28 -04:00
|
|
|
k = s.extendIndex(s.expr(max), panicslice)
|
2016-04-21 11:55:33 -07:00
|
|
|
}
|
|
|
|
p, l, c := s.slice(n.Left.Type, v, i, j, k)
|
2015-09-12 23:27:26 -07:00
|
|
|
return s.newValue3(ssa.OpSliceMake, n.Type, p, l, c)
|
2016-04-21 11:55:33 -07:00
|
|
|
|
2015-08-24 23:52:03 -07:00
|
|
|
case OSLICESTR:
|
2015-09-12 23:27:26 -07:00
|
|
|
v := s.expr(n.Left)
|
|
|
|
var i, j *ssa.Value
|
2016-04-21 11:55:33 -07:00
|
|
|
low, high, _ := n.SliceBounds()
|
|
|
|
if low != nil {
|
2016-05-25 09:49:28 -04:00
|
|
|
i = s.extendIndex(s.expr(low), panicslice)
|
2015-08-24 23:52:03 -07:00
|
|
|
}
|
2016-04-21 11:55:33 -07:00
|
|
|
if high != nil {
|
2016-05-25 09:49:28 -04:00
|
|
|
j = s.extendIndex(s.expr(high), panicslice)
|
2015-08-24 23:52:03 -07:00
|
|
|
}
|
2015-09-12 23:27:26 -07:00
|
|
|
p, l, _ := s.slice(n.Left.Type, v, i, j, nil)
|
|
|
|
return s.newValue2(ssa.OpStringMake, n.Type, p, l)
|
2015-08-24 23:52:03 -07:00
|
|
|
|
2016-03-11 00:10:52 -05:00
|
|
|
case OCALLFUNC:
|
2016-08-23 16:49:28 -07:00
|
|
|
if isIntrinsicCall(n) {
|
|
|
|
return s.intrinsicCall(n)
|
2016-03-11 00:10:52 -05:00
|
|
|
}
|
|
|
|
fallthrough
|
|
|
|
|
|
|
|
case OCALLINTER, OCALLMETH:
|
2016-01-25 17:06:54 -08:00
|
|
|
a := s.call(n, callNormal)
|
|
|
|
return s.newValue2(ssa.OpLoad, n.Type, a, s.mem())
|
2015-08-12 11:22:16 -07:00
|
|
|
|
|
|
|
case OGETG:
|
2015-10-19 18:54:40 -07:00
|
|
|
return s.newValue1(ssa.OpGetG, n.Type, s.mem())
|
2015-08-12 11:22:16 -07:00
|
|
|
|
2015-09-11 11:02:57 -07:00
|
|
|
case OAPPEND:
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
return s.append(n, false)
|
2015-09-11 11:02:57 -07:00
|
|
|
|
2016-04-04 10:58:21 -07:00
|
|
|
default:
|
2016-04-22 07:14:10 -07:00
|
|
|
s.Unimplementedf("unhandled expr %s", n.Op)
|
2016-04-04 10:58:21 -07:00
|
|
|
return nil
|
|
|
|
}
|
|
|
|
}
|
2015-09-11 11:02:57 -07:00
|
|
|
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
// append converts an OAPPEND node to SSA.
|
|
|
|
// If inplace is false, it converts the OAPPEND expression n to an ssa.Value,
|
|
|
|
// adds it to s, and returns the Value.
|
|
|
|
// If inplace is true, it writes the result of the OAPPEND expression n
|
|
|
|
// back to the slice being appended to, and returns nil.
|
|
|
|
// inplace MUST be set to false if the slice can be SSA'd.
|
|
|
|
func (s *state) append(n *Node, inplace bool) *ssa.Value {
|
|
|
|
// If inplace is false, process as expression "append(s, e1, e2, e3)":
|
|
|
|
//
|
cmd/compile: avoid a spill in append fast path
Instead of spilling newlen, recalculate it.
This removes a spill from the fast path,
at the cost of a cheap recalculation
on the (rare) growth path.
This uses 8 bytes less of stack space.
It generates two more bytes of code,
but that is due to suboptimal register allocation;
see far below.
Runtime append microbenchmarks are all over the map,
presumably due to incidental code movement.
Sample code:
func s(b []byte) []byte {
b = append(b, 1, 2, 3)
return b
}
Before:
"".s t=1 size=160 args=0x30 locals=0x48
0x0000 00000 (append.go:8) TEXT "".s(SB), $72-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 149
0x0013 00019 (append.go:8) SUBQ $72, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+88(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ DX, "".autotmp_0+64(SP)
0x0025 00037 (append.go:9) MOVQ "".b+96(FP), BX
0x002a 00042 (append.go:9) CMPQ DX, BX
0x002d 00045 (append.go:9) JGT $0, 86
0x002f 00047 (append.go:8) MOVQ "".b+80(FP), AX
0x0034 00052 (append.go:9) MOVB $1, (AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x003d 00061 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x0042 00066 (append.go:10) MOVQ AX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ DX, "".~r1+112(FP)
0x004c 00076 (append.go:10) MOVQ BX, "".~r1+120(FP)
0x0051 00081 (append.go:10) ADDQ $72, SP
0x0055 00085 (append.go:10) RET
0x0056 00086 (append.go:9) LEAQ type.[]uint8(SB), AX
0x005d 00093 (append.go:9) MOVQ AX, (SP)
0x0061 00097 (append.go:9) MOVQ "".b+80(FP), BP
0x0066 00102 (append.go:9) MOVQ BP, 8(SP)
0x006b 00107 (append.go:9) MOVQ CX, 16(SP)
0x0070 00112 (append.go:9) MOVQ BX, 24(SP)
0x0075 00117 (append.go:9) MOVQ DX, 32(SP)
0x007a 00122 (append.go:9) PCDATA $0, $0
0x007a 00122 (append.go:9) CALL runtime.growslice(SB)
0x007f 00127 (append.go:9) MOVQ 40(SP), AX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:8) MOVQ "".b+88(FP), CX
0x008e 00142 (append.go:9) MOVQ "".autotmp_0+64(SP), DX
0x0093 00147 (append.go:9) JMP 52
0x0095 00149 (append.go:9) NOP
0x0095 00149 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009a 00154 (append.go:8) JMP 0
After:
"".s t=1 size=176 args=0x30 locals=0x40
0x0000 00000 (append.go:8) TEXT "".s(SB), $64-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 151
0x0013 00019 (append.go:8) SUBQ $64, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+80(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ "".b+88(FP), BX
0x0025 00037 (append.go:9) CMPQ DX, BX
0x0028 00040 (append.go:9) JGT $0, 81
0x002a 00042 (append.go:8) MOVQ "".b+72(FP), AX
0x002f 00047 (append.go:9) MOVB $1, (AX)(CX*1)
0x0033 00051 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x003d 00061 (append.go:10) MOVQ AX, "".~r1+96(FP)
0x0042 00066 (append.go:10) MOVQ DX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ BX, "".~r1+112(FP)
0x004c 00076 (append.go:10) ADDQ $64, SP
0x0050 00080 (append.go:10) RET
0x0051 00081 (append.go:9) LEAQ type.[]uint8(SB), AX
0x0058 00088 (append.go:9) MOVQ AX, (SP)
0x005c 00092 (append.go:9) MOVQ "".b+72(FP), BP
0x0061 00097 (append.go:9) MOVQ BP, 8(SP)
0x0066 00102 (append.go:9) MOVQ CX, 16(SP)
0x006b 00107 (append.go:9) MOVQ BX, 24(SP)
0x0070 00112 (append.go:9) MOVQ DX, 32(SP)
0x0075 00117 (append.go:9) PCDATA $0, $0
0x0075 00117 (append.go:9) CALL runtime.growslice(SB)
0x007a 00122 (append.go:9) MOVQ 40(SP), AX
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
0x0095 00149 (append.go:9) JMP 47
0x0097 00151 (append.go:9) NOP
0x0097 00151 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009c 00156 (append.go:8) JMP 0
Observe that in the following sequence,
we should use DX directly instead of using
CX as a temporary register, which would make
the new code a strict improvement on the old:
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
Change-Id: I4ee50b18fa53865901d2d7f86c2cbb54c6fa6924
Reviewed-on: https://go-review.googlesource.com/21812
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:08:00 -07:00
|
|
|
// ptr, len, cap := s
|
2016-04-04 10:58:21 -07:00
|
|
|
// newlen := len + 3
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
// if newlen > cap {
|
cmd/compile: avoid a spill in append fast path
Instead of spilling newlen, recalculate it.
This removes a spill from the fast path,
at the cost of a cheap recalculation
on the (rare) growth path.
This uses 8 bytes less of stack space.
It generates two more bytes of code,
but that is due to suboptimal register allocation;
see far below.
Runtime append microbenchmarks are all over the map,
presumably due to incidental code movement.
Sample code:
func s(b []byte) []byte {
b = append(b, 1, 2, 3)
return b
}
Before:
"".s t=1 size=160 args=0x30 locals=0x48
0x0000 00000 (append.go:8) TEXT "".s(SB), $72-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 149
0x0013 00019 (append.go:8) SUBQ $72, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+88(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ DX, "".autotmp_0+64(SP)
0x0025 00037 (append.go:9) MOVQ "".b+96(FP), BX
0x002a 00042 (append.go:9) CMPQ DX, BX
0x002d 00045 (append.go:9) JGT $0, 86
0x002f 00047 (append.go:8) MOVQ "".b+80(FP), AX
0x0034 00052 (append.go:9) MOVB $1, (AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x003d 00061 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x0042 00066 (append.go:10) MOVQ AX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ DX, "".~r1+112(FP)
0x004c 00076 (append.go:10) MOVQ BX, "".~r1+120(FP)
0x0051 00081 (append.go:10) ADDQ $72, SP
0x0055 00085 (append.go:10) RET
0x0056 00086 (append.go:9) LEAQ type.[]uint8(SB), AX
0x005d 00093 (append.go:9) MOVQ AX, (SP)
0x0061 00097 (append.go:9) MOVQ "".b+80(FP), BP
0x0066 00102 (append.go:9) MOVQ BP, 8(SP)
0x006b 00107 (append.go:9) MOVQ CX, 16(SP)
0x0070 00112 (append.go:9) MOVQ BX, 24(SP)
0x0075 00117 (append.go:9) MOVQ DX, 32(SP)
0x007a 00122 (append.go:9) PCDATA $0, $0
0x007a 00122 (append.go:9) CALL runtime.growslice(SB)
0x007f 00127 (append.go:9) MOVQ 40(SP), AX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:8) MOVQ "".b+88(FP), CX
0x008e 00142 (append.go:9) MOVQ "".autotmp_0+64(SP), DX
0x0093 00147 (append.go:9) JMP 52
0x0095 00149 (append.go:9) NOP
0x0095 00149 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009a 00154 (append.go:8) JMP 0
After:
"".s t=1 size=176 args=0x30 locals=0x40
0x0000 00000 (append.go:8) TEXT "".s(SB), $64-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 151
0x0013 00019 (append.go:8) SUBQ $64, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+80(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ "".b+88(FP), BX
0x0025 00037 (append.go:9) CMPQ DX, BX
0x0028 00040 (append.go:9) JGT $0, 81
0x002a 00042 (append.go:8) MOVQ "".b+72(FP), AX
0x002f 00047 (append.go:9) MOVB $1, (AX)(CX*1)
0x0033 00051 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x003d 00061 (append.go:10) MOVQ AX, "".~r1+96(FP)
0x0042 00066 (append.go:10) MOVQ DX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ BX, "".~r1+112(FP)
0x004c 00076 (append.go:10) ADDQ $64, SP
0x0050 00080 (append.go:10) RET
0x0051 00081 (append.go:9) LEAQ type.[]uint8(SB), AX
0x0058 00088 (append.go:9) MOVQ AX, (SP)
0x005c 00092 (append.go:9) MOVQ "".b+72(FP), BP
0x0061 00097 (append.go:9) MOVQ BP, 8(SP)
0x0066 00102 (append.go:9) MOVQ CX, 16(SP)
0x006b 00107 (append.go:9) MOVQ BX, 24(SP)
0x0070 00112 (append.go:9) MOVQ DX, 32(SP)
0x0075 00117 (append.go:9) PCDATA $0, $0
0x0075 00117 (append.go:9) CALL runtime.growslice(SB)
0x007a 00122 (append.go:9) MOVQ 40(SP), AX
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
0x0095 00149 (append.go:9) JMP 47
0x0097 00151 (append.go:9) NOP
0x0097 00151 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009c 00156 (append.go:8) JMP 0
Observe that in the following sequence,
we should use DX directly instead of using
CX as a temporary register, which would make
the new code a strict improvement on the old:
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
Change-Id: I4ee50b18fa53865901d2d7f86c2cbb54c6fa6924
Reviewed-on: https://go-review.googlesource.com/21812
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:08:00 -07:00
|
|
|
// ptr, len, cap = growslice(s, newlen)
|
|
|
|
// newlen = len + 3 // recalculate to avoid a spill
|
2016-04-04 10:58:21 -07:00
|
|
|
// }
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
// // with write barriers, if needed:
|
|
|
|
// *(ptr+len) = e1
|
|
|
|
// *(ptr+len+1) = e2
|
|
|
|
// *(ptr+len+2) = e3
|
|
|
|
// return makeslice(ptr, newlen, cap)
|
|
|
|
//
|
|
|
|
//
|
|
|
|
// If inplace is true, process as statement "s = append(s, e1, e2, e3)":
|
|
|
|
//
|
|
|
|
// a := &s
|
|
|
|
// ptr, len, cap := s
|
|
|
|
// newlen := len + 3
|
|
|
|
// if newlen > cap {
|
cmd/compile: re-enable in-place append optimization
CL 21891 was too clever in its attempts to avoid spills.
Storing newlen too early caused uses of append in the runtime
itself to receive an inconsistent view of a slice,
leading to corruption.
This CL makes the generate code much more similar to
the old backend. It spills more than before,
but those spills have been contained to the grow path.
It recalculates newlen unnecessarily on the fast path,
but that's measurably cheaper than spilling it.
CL 21891 caused runtime failures in 6 of 2000 runs
of net/http and crypto/x509 in my test setup.
This CL has gone 6000 runs without a failure.
Benchmarks going from master to this CL:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 439ns ± 2% 436ns ± 2% -0.72% (p=0.001 n=28+27)
AppendInPlace/NoGrow/1Ptr-8 901ns ± 0% 856ns ± 0% -4.95% (p=0.000 n=26+29)
AppendInPlace/NoGrow/2Ptr-8 2.15µs ± 1% 1.95µs ± 0% -9.07% (p=0.000 n=28+30)
AppendInPlace/NoGrow/3Ptr-8 2.66µs ± 0% 2.45µs ± 0% -7.93% (p=0.000 n=29+26)
AppendInPlace/NoGrow/4Ptr-8 3.24µs ± 1% 3.02µs ± 1% -6.75% (p=0.000 n=28+30)
AppendInPlace/Grow/Byte-8 269ns ± 1% 271ns ± 1% +0.84% (p=0.000 n=30+29)
AppendInPlace/Grow/1Ptr-8 275ns ± 1% 280ns ± 1% +1.75% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 384ns ± 0% 391ns ± 0% +1.94% (p=0.000 n=27+30)
AppendInPlace/Grow/3Ptr-8 455ns ± 0% 462ns ± 0% +1.43% (p=0.000 n=29+29)
AppendInPlace/Grow/4Ptr-8 478ns ± 0% 479ns ± 0% +0.23% (p=0.000 n=30+27)
However, for the large no-grow cases, there is still more work to be done.
Going from this CL to the non-SSA backend:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 436ns ± 2% 436ns ± 2% ~ (p=0.967 n=27+29)
AppendInPlace/NoGrow/1Ptr-8 856ns ± 0% 884ns ± 0% +3.28% (p=0.000 n=29+26)
AppendInPlace/NoGrow/2Ptr-8 1.95µs ± 0% 1.56µs ± 0% -20.28% (p=0.000 n=30+29)
AppendInPlace/NoGrow/3Ptr-8 2.45µs ± 0% 1.89µs ± 0% -22.88% (p=0.000 n=26+28)
AppendInPlace/NoGrow/4Ptr-8 3.02µs ± 1% 2.56µs ± 1% -15.35% (p=0.000 n=30+28)
AppendInPlace/Grow/Byte-8 271ns ± 1% 283ns ± 1% +4.56% (p=0.000 n=29+29)
AppendInPlace/Grow/1Ptr-8 280ns ± 1% 288ns ± 1% +2.99% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 391ns ± 0% 409ns ± 0% +4.66% (p=0.000 n=30+29)
AppendInPlace/Grow/3Ptr-8 462ns ± 0% 481ns ± 0% +4.13% (p=0.000 n=29+30)
AppendInPlace/Grow/4Ptr-8 479ns ± 0% 502ns ± 0% +4.81% (p=0.000 n=27+26)
New generated code:
var x []byte
func a() {
x = append(x, 1)
}
"".a t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (a.go:5) TEXT "".a(SB), $72-0
0x0000 00000 (a.go:5) MOVQ (TLS), CX
0x0009 00009 (a.go:5) CMPQ SP, 16(CX)
0x000d 00013 (a.go:5) JLS 190
0x0013 00019 (a.go:5) SUBQ $72, SP
0x0017 00023 (a.go:5) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:5) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:6) MOVQ "".x+16(SB), CX
0x001e 00030 (a.go:6) MOVQ "".x+8(SB), DX
0x0025 00037 (a.go:6) MOVQ "".x(SB), BX
0x002c 00044 (a.go:6) LEAQ 1(DX), BP
0x0030 00048 (a.go:6) CMPQ BP, CX
0x0033 00051 (a.go:6) JGT $0, 73
0x0035 00053 (a.go:6) LEAQ 1(DX), AX
0x0039 00057 (a.go:6) MOVQ AX, "".x+8(SB)
0x0040 00064 (a.go:6) MOVB $1, (BX)(DX*1)
0x0044 00068 (a.go:7) ADDQ $72, SP
0x0048 00072 (a.go:7) RET
0x0049 00073 (a.go:6) LEAQ type.[]uint8(SB), AX
0x0050 00080 (a.go:6) MOVQ AX, (SP)
0x0054 00084 (a.go:6) MOVQ BX, 8(SP)
0x0059 00089 (a.go:6) MOVQ DX, 16(SP)
0x005e 00094 (a.go:6) MOVQ CX, 24(SP)
0x0063 00099 (a.go:6) MOVQ BP, 32(SP)
0x0068 00104 (a.go:6) PCDATA $0, $0
0x0068 00104 (a.go:6) CALL runtime.growslice(SB)
0x006d 00109 (a.go:6) MOVQ 40(SP), CX
0x0072 00114 (a.go:6) MOVQ 48(SP), DX
0x0077 00119 (a.go:6) MOVQ DX, "".autotmp_0+64(SP)
0x007c 00124 (a.go:6) MOVQ 56(SP), BX
0x0081 00129 (a.go:6) MOVQ BX, "".x+16(SB)
0x0088 00136 (a.go:6) MOVL runtime.writeBarrier(SB), AX
0x008e 00142 (a.go:6) TESTB AL, AL
0x0090 00144 (a.go:6) JNE $0, 162
0x0092 00146 (a.go:6) MOVQ CX, "".x(SB)
0x0099 00153 (a.go:6) MOVQ "".x(SB), BX
0x00a0 00160 (a.go:6) JMP 53
0x00a2 00162 (a.go:6) LEAQ "".x(SB), BX
0x00a9 00169 (a.go:6) MOVQ BX, (SP)
0x00ad 00173 (a.go:6) MOVQ CX, 8(SP)
0x00b2 00178 (a.go:6) PCDATA $0, $0
0x00b2 00178 (a.go:6) CALL runtime.writebarrierptr(SB)
0x00b7 00183 (a.go:6) MOVQ "".autotmp_0+64(SP), DX
0x00bc 00188 (a.go:6) JMP 153
0x00be 00190 (a.go:6) NOP
0x00be 00190 (a.go:5) CALL runtime.morestack_noctxt(SB)
0x00c3 00195 (a.go:5) JMP 0
Fixes #14969 again
Change-Id: Ia50463b1f506011aad0718a4fef1d4738e43c32d
Reviewed-on: https://go-review.googlesource.com/22197
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-18 09:40:30 -07:00
|
|
|
// newptr, len, newcap = growslice(ptr, len, cap, newlen)
|
|
|
|
// vardef(a) // if necessary, advise liveness we are writing a new a
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
// *a.cap = newcap // write before ptr to avoid a spill
|
|
|
|
// *a.ptr = newptr // with write barrier
|
|
|
|
// }
|
cmd/compile: re-enable in-place append optimization
CL 21891 was too clever in its attempts to avoid spills.
Storing newlen too early caused uses of append in the runtime
itself to receive an inconsistent view of a slice,
leading to corruption.
This CL makes the generate code much more similar to
the old backend. It spills more than before,
but those spills have been contained to the grow path.
It recalculates newlen unnecessarily on the fast path,
but that's measurably cheaper than spilling it.
CL 21891 caused runtime failures in 6 of 2000 runs
of net/http and crypto/x509 in my test setup.
This CL has gone 6000 runs without a failure.
Benchmarks going from master to this CL:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 439ns ± 2% 436ns ± 2% -0.72% (p=0.001 n=28+27)
AppendInPlace/NoGrow/1Ptr-8 901ns ± 0% 856ns ± 0% -4.95% (p=0.000 n=26+29)
AppendInPlace/NoGrow/2Ptr-8 2.15µs ± 1% 1.95µs ± 0% -9.07% (p=0.000 n=28+30)
AppendInPlace/NoGrow/3Ptr-8 2.66µs ± 0% 2.45µs ± 0% -7.93% (p=0.000 n=29+26)
AppendInPlace/NoGrow/4Ptr-8 3.24µs ± 1% 3.02µs ± 1% -6.75% (p=0.000 n=28+30)
AppendInPlace/Grow/Byte-8 269ns ± 1% 271ns ± 1% +0.84% (p=0.000 n=30+29)
AppendInPlace/Grow/1Ptr-8 275ns ± 1% 280ns ± 1% +1.75% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 384ns ± 0% 391ns ± 0% +1.94% (p=0.000 n=27+30)
AppendInPlace/Grow/3Ptr-8 455ns ± 0% 462ns ± 0% +1.43% (p=0.000 n=29+29)
AppendInPlace/Grow/4Ptr-8 478ns ± 0% 479ns ± 0% +0.23% (p=0.000 n=30+27)
However, for the large no-grow cases, there is still more work to be done.
Going from this CL to the non-SSA backend:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 436ns ± 2% 436ns ± 2% ~ (p=0.967 n=27+29)
AppendInPlace/NoGrow/1Ptr-8 856ns ± 0% 884ns ± 0% +3.28% (p=0.000 n=29+26)
AppendInPlace/NoGrow/2Ptr-8 1.95µs ± 0% 1.56µs ± 0% -20.28% (p=0.000 n=30+29)
AppendInPlace/NoGrow/3Ptr-8 2.45µs ± 0% 1.89µs ± 0% -22.88% (p=0.000 n=26+28)
AppendInPlace/NoGrow/4Ptr-8 3.02µs ± 1% 2.56µs ± 1% -15.35% (p=0.000 n=30+28)
AppendInPlace/Grow/Byte-8 271ns ± 1% 283ns ± 1% +4.56% (p=0.000 n=29+29)
AppendInPlace/Grow/1Ptr-8 280ns ± 1% 288ns ± 1% +2.99% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 391ns ± 0% 409ns ± 0% +4.66% (p=0.000 n=30+29)
AppendInPlace/Grow/3Ptr-8 462ns ± 0% 481ns ± 0% +4.13% (p=0.000 n=29+30)
AppendInPlace/Grow/4Ptr-8 479ns ± 0% 502ns ± 0% +4.81% (p=0.000 n=27+26)
New generated code:
var x []byte
func a() {
x = append(x, 1)
}
"".a t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (a.go:5) TEXT "".a(SB), $72-0
0x0000 00000 (a.go:5) MOVQ (TLS), CX
0x0009 00009 (a.go:5) CMPQ SP, 16(CX)
0x000d 00013 (a.go:5) JLS 190
0x0013 00019 (a.go:5) SUBQ $72, SP
0x0017 00023 (a.go:5) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:5) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:6) MOVQ "".x+16(SB), CX
0x001e 00030 (a.go:6) MOVQ "".x+8(SB), DX
0x0025 00037 (a.go:6) MOVQ "".x(SB), BX
0x002c 00044 (a.go:6) LEAQ 1(DX), BP
0x0030 00048 (a.go:6) CMPQ BP, CX
0x0033 00051 (a.go:6) JGT $0, 73
0x0035 00053 (a.go:6) LEAQ 1(DX), AX
0x0039 00057 (a.go:6) MOVQ AX, "".x+8(SB)
0x0040 00064 (a.go:6) MOVB $1, (BX)(DX*1)
0x0044 00068 (a.go:7) ADDQ $72, SP
0x0048 00072 (a.go:7) RET
0x0049 00073 (a.go:6) LEAQ type.[]uint8(SB), AX
0x0050 00080 (a.go:6) MOVQ AX, (SP)
0x0054 00084 (a.go:6) MOVQ BX, 8(SP)
0x0059 00089 (a.go:6) MOVQ DX, 16(SP)
0x005e 00094 (a.go:6) MOVQ CX, 24(SP)
0x0063 00099 (a.go:6) MOVQ BP, 32(SP)
0x0068 00104 (a.go:6) PCDATA $0, $0
0x0068 00104 (a.go:6) CALL runtime.growslice(SB)
0x006d 00109 (a.go:6) MOVQ 40(SP), CX
0x0072 00114 (a.go:6) MOVQ 48(SP), DX
0x0077 00119 (a.go:6) MOVQ DX, "".autotmp_0+64(SP)
0x007c 00124 (a.go:6) MOVQ 56(SP), BX
0x0081 00129 (a.go:6) MOVQ BX, "".x+16(SB)
0x0088 00136 (a.go:6) MOVL runtime.writeBarrier(SB), AX
0x008e 00142 (a.go:6) TESTB AL, AL
0x0090 00144 (a.go:6) JNE $0, 162
0x0092 00146 (a.go:6) MOVQ CX, "".x(SB)
0x0099 00153 (a.go:6) MOVQ "".x(SB), BX
0x00a0 00160 (a.go:6) JMP 53
0x00a2 00162 (a.go:6) LEAQ "".x(SB), BX
0x00a9 00169 (a.go:6) MOVQ BX, (SP)
0x00ad 00173 (a.go:6) MOVQ CX, 8(SP)
0x00b2 00178 (a.go:6) PCDATA $0, $0
0x00b2 00178 (a.go:6) CALL runtime.writebarrierptr(SB)
0x00b7 00183 (a.go:6) MOVQ "".autotmp_0+64(SP), DX
0x00bc 00188 (a.go:6) JMP 153
0x00be 00190 (a.go:6) NOP
0x00be 00190 (a.go:5) CALL runtime.morestack_noctxt(SB)
0x00c3 00195 (a.go:5) JMP 0
Fixes #14969 again
Change-Id: Ia50463b1f506011aad0718a4fef1d4738e43c32d
Reviewed-on: https://go-review.googlesource.com/22197
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-18 09:40:30 -07:00
|
|
|
// newlen = len + 3 // recalculate to avoid a spill
|
|
|
|
// *a.len = newlen
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
// // with write barriers, if needed:
|
2016-04-04 10:58:21 -07:00
|
|
|
// *(ptr+len) = e1
|
|
|
|
// *(ptr+len+1) = e2
|
|
|
|
// *(ptr+len+2) = e3
|
|
|
|
|
|
|
|
et := n.Type.Elem()
|
|
|
|
pt := Ptrto(et)
|
|
|
|
|
|
|
|
// Evaluate slice
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
sn := n.List.First() // the slice node is the first in the list
|
|
|
|
|
|
|
|
var slice, addr *ssa.Value
|
|
|
|
if inplace {
|
2016-05-27 14:07:37 -07:00
|
|
|
addr, _ = s.addr(sn, false)
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
slice = s.newValue2(ssa.OpLoad, n.Type, addr, s.mem())
|
|
|
|
} else {
|
|
|
|
slice = s.expr(sn)
|
|
|
|
}
|
2016-04-04 10:58:21 -07:00
|
|
|
|
|
|
|
// Allocate new blocks
|
|
|
|
grow := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
assign := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
|
|
|
|
// Decide if we need to grow
|
|
|
|
nargs := int64(n.List.Len() - 1)
|
|
|
|
p := s.newValue1(ssa.OpSlicePtr, pt, slice)
|
|
|
|
l := s.newValue1(ssa.OpSliceLen, Types[TINT], slice)
|
|
|
|
c := s.newValue1(ssa.OpSliceCap, Types[TINT], slice)
|
|
|
|
nl := s.newValue2(s.ssaOp(OADD, Types[TINT]), Types[TINT], l, s.constInt(Types[TINT], nargs))
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
|
2016-04-04 10:58:21 -07:00
|
|
|
cmp := s.newValue2(s.ssaOp(OGT, Types[TINT]), Types[TBOOL], nl, c)
|
|
|
|
s.vars[&ptrVar] = p
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
|
|
|
|
if !inplace {
|
|
|
|
s.vars[&newlenVar] = nl
|
|
|
|
s.vars[&capVar] = c
|
cmd/compile: re-enable in-place append optimization
CL 21891 was too clever in its attempts to avoid spills.
Storing newlen too early caused uses of append in the runtime
itself to receive an inconsistent view of a slice,
leading to corruption.
This CL makes the generate code much more similar to
the old backend. It spills more than before,
but those spills have been contained to the grow path.
It recalculates newlen unnecessarily on the fast path,
but that's measurably cheaper than spilling it.
CL 21891 caused runtime failures in 6 of 2000 runs
of net/http and crypto/x509 in my test setup.
This CL has gone 6000 runs without a failure.
Benchmarks going from master to this CL:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 439ns ± 2% 436ns ± 2% -0.72% (p=0.001 n=28+27)
AppendInPlace/NoGrow/1Ptr-8 901ns ± 0% 856ns ± 0% -4.95% (p=0.000 n=26+29)
AppendInPlace/NoGrow/2Ptr-8 2.15µs ± 1% 1.95µs ± 0% -9.07% (p=0.000 n=28+30)
AppendInPlace/NoGrow/3Ptr-8 2.66µs ± 0% 2.45µs ± 0% -7.93% (p=0.000 n=29+26)
AppendInPlace/NoGrow/4Ptr-8 3.24µs ± 1% 3.02µs ± 1% -6.75% (p=0.000 n=28+30)
AppendInPlace/Grow/Byte-8 269ns ± 1% 271ns ± 1% +0.84% (p=0.000 n=30+29)
AppendInPlace/Grow/1Ptr-8 275ns ± 1% 280ns ± 1% +1.75% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 384ns ± 0% 391ns ± 0% +1.94% (p=0.000 n=27+30)
AppendInPlace/Grow/3Ptr-8 455ns ± 0% 462ns ± 0% +1.43% (p=0.000 n=29+29)
AppendInPlace/Grow/4Ptr-8 478ns ± 0% 479ns ± 0% +0.23% (p=0.000 n=30+27)
However, for the large no-grow cases, there is still more work to be done.
Going from this CL to the non-SSA backend:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 436ns ± 2% 436ns ± 2% ~ (p=0.967 n=27+29)
AppendInPlace/NoGrow/1Ptr-8 856ns ± 0% 884ns ± 0% +3.28% (p=0.000 n=29+26)
AppendInPlace/NoGrow/2Ptr-8 1.95µs ± 0% 1.56µs ± 0% -20.28% (p=0.000 n=30+29)
AppendInPlace/NoGrow/3Ptr-8 2.45µs ± 0% 1.89µs ± 0% -22.88% (p=0.000 n=26+28)
AppendInPlace/NoGrow/4Ptr-8 3.02µs ± 1% 2.56µs ± 1% -15.35% (p=0.000 n=30+28)
AppendInPlace/Grow/Byte-8 271ns ± 1% 283ns ± 1% +4.56% (p=0.000 n=29+29)
AppendInPlace/Grow/1Ptr-8 280ns ± 1% 288ns ± 1% +2.99% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 391ns ± 0% 409ns ± 0% +4.66% (p=0.000 n=30+29)
AppendInPlace/Grow/3Ptr-8 462ns ± 0% 481ns ± 0% +4.13% (p=0.000 n=29+30)
AppendInPlace/Grow/4Ptr-8 479ns ± 0% 502ns ± 0% +4.81% (p=0.000 n=27+26)
New generated code:
var x []byte
func a() {
x = append(x, 1)
}
"".a t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (a.go:5) TEXT "".a(SB), $72-0
0x0000 00000 (a.go:5) MOVQ (TLS), CX
0x0009 00009 (a.go:5) CMPQ SP, 16(CX)
0x000d 00013 (a.go:5) JLS 190
0x0013 00019 (a.go:5) SUBQ $72, SP
0x0017 00023 (a.go:5) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:5) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:6) MOVQ "".x+16(SB), CX
0x001e 00030 (a.go:6) MOVQ "".x+8(SB), DX
0x0025 00037 (a.go:6) MOVQ "".x(SB), BX
0x002c 00044 (a.go:6) LEAQ 1(DX), BP
0x0030 00048 (a.go:6) CMPQ BP, CX
0x0033 00051 (a.go:6) JGT $0, 73
0x0035 00053 (a.go:6) LEAQ 1(DX), AX
0x0039 00057 (a.go:6) MOVQ AX, "".x+8(SB)
0x0040 00064 (a.go:6) MOVB $1, (BX)(DX*1)
0x0044 00068 (a.go:7) ADDQ $72, SP
0x0048 00072 (a.go:7) RET
0x0049 00073 (a.go:6) LEAQ type.[]uint8(SB), AX
0x0050 00080 (a.go:6) MOVQ AX, (SP)
0x0054 00084 (a.go:6) MOVQ BX, 8(SP)
0x0059 00089 (a.go:6) MOVQ DX, 16(SP)
0x005e 00094 (a.go:6) MOVQ CX, 24(SP)
0x0063 00099 (a.go:6) MOVQ BP, 32(SP)
0x0068 00104 (a.go:6) PCDATA $0, $0
0x0068 00104 (a.go:6) CALL runtime.growslice(SB)
0x006d 00109 (a.go:6) MOVQ 40(SP), CX
0x0072 00114 (a.go:6) MOVQ 48(SP), DX
0x0077 00119 (a.go:6) MOVQ DX, "".autotmp_0+64(SP)
0x007c 00124 (a.go:6) MOVQ 56(SP), BX
0x0081 00129 (a.go:6) MOVQ BX, "".x+16(SB)
0x0088 00136 (a.go:6) MOVL runtime.writeBarrier(SB), AX
0x008e 00142 (a.go:6) TESTB AL, AL
0x0090 00144 (a.go:6) JNE $0, 162
0x0092 00146 (a.go:6) MOVQ CX, "".x(SB)
0x0099 00153 (a.go:6) MOVQ "".x(SB), BX
0x00a0 00160 (a.go:6) JMP 53
0x00a2 00162 (a.go:6) LEAQ "".x(SB), BX
0x00a9 00169 (a.go:6) MOVQ BX, (SP)
0x00ad 00173 (a.go:6) MOVQ CX, 8(SP)
0x00b2 00178 (a.go:6) PCDATA $0, $0
0x00b2 00178 (a.go:6) CALL runtime.writebarrierptr(SB)
0x00b7 00183 (a.go:6) MOVQ "".autotmp_0+64(SP), DX
0x00bc 00188 (a.go:6) JMP 153
0x00be 00190 (a.go:6) NOP
0x00be 00190 (a.go:5) CALL runtime.morestack_noctxt(SB)
0x00c3 00195 (a.go:5) JMP 0
Fixes #14969 again
Change-Id: Ia50463b1f506011aad0718a4fef1d4738e43c32d
Reviewed-on: https://go-review.googlesource.com/22197
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-18 09:40:30 -07:00
|
|
|
} else {
|
|
|
|
s.vars[&lenVar] = l
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
}
|
|
|
|
|
2016-04-04 10:58:21 -07:00
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
|
|
|
b.Likely = ssa.BranchUnlikely
|
|
|
|
b.SetControl(cmp)
|
|
|
|
b.AddEdgeTo(grow)
|
|
|
|
b.AddEdgeTo(assign)
|
|
|
|
|
|
|
|
// Call growslice
|
|
|
|
s.startBlock(grow)
|
2016-04-29 09:02:27 -07:00
|
|
|
taddr := s.newValue1A(ssa.OpAddr, Types[TUINTPTR], &ssa.ExternSymbol{Typ: Types[TUINTPTR], Sym: typenamesym(n.Type.Elem())}, s.sb)
|
2016-04-04 10:58:21 -07:00
|
|
|
|
|
|
|
r := s.rtcall(growslice, true, []*Type{pt, Types[TINT], Types[TINT]}, taddr, p, l, c, nl)
|
|
|
|
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
if inplace {
|
cmd/compile: re-enable in-place append optimization
CL 21891 was too clever in its attempts to avoid spills.
Storing newlen too early caused uses of append in the runtime
itself to receive an inconsistent view of a slice,
leading to corruption.
This CL makes the generate code much more similar to
the old backend. It spills more than before,
but those spills have been contained to the grow path.
It recalculates newlen unnecessarily on the fast path,
but that's measurably cheaper than spilling it.
CL 21891 caused runtime failures in 6 of 2000 runs
of net/http and crypto/x509 in my test setup.
This CL has gone 6000 runs without a failure.
Benchmarks going from master to this CL:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 439ns ± 2% 436ns ± 2% -0.72% (p=0.001 n=28+27)
AppendInPlace/NoGrow/1Ptr-8 901ns ± 0% 856ns ± 0% -4.95% (p=0.000 n=26+29)
AppendInPlace/NoGrow/2Ptr-8 2.15µs ± 1% 1.95µs ± 0% -9.07% (p=0.000 n=28+30)
AppendInPlace/NoGrow/3Ptr-8 2.66µs ± 0% 2.45µs ± 0% -7.93% (p=0.000 n=29+26)
AppendInPlace/NoGrow/4Ptr-8 3.24µs ± 1% 3.02µs ± 1% -6.75% (p=0.000 n=28+30)
AppendInPlace/Grow/Byte-8 269ns ± 1% 271ns ± 1% +0.84% (p=0.000 n=30+29)
AppendInPlace/Grow/1Ptr-8 275ns ± 1% 280ns ± 1% +1.75% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 384ns ± 0% 391ns ± 0% +1.94% (p=0.000 n=27+30)
AppendInPlace/Grow/3Ptr-8 455ns ± 0% 462ns ± 0% +1.43% (p=0.000 n=29+29)
AppendInPlace/Grow/4Ptr-8 478ns ± 0% 479ns ± 0% +0.23% (p=0.000 n=30+27)
However, for the large no-grow cases, there is still more work to be done.
Going from this CL to the non-SSA backend:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 436ns ± 2% 436ns ± 2% ~ (p=0.967 n=27+29)
AppendInPlace/NoGrow/1Ptr-8 856ns ± 0% 884ns ± 0% +3.28% (p=0.000 n=29+26)
AppendInPlace/NoGrow/2Ptr-8 1.95µs ± 0% 1.56µs ± 0% -20.28% (p=0.000 n=30+29)
AppendInPlace/NoGrow/3Ptr-8 2.45µs ± 0% 1.89µs ± 0% -22.88% (p=0.000 n=26+28)
AppendInPlace/NoGrow/4Ptr-8 3.02µs ± 1% 2.56µs ± 1% -15.35% (p=0.000 n=30+28)
AppendInPlace/Grow/Byte-8 271ns ± 1% 283ns ± 1% +4.56% (p=0.000 n=29+29)
AppendInPlace/Grow/1Ptr-8 280ns ± 1% 288ns ± 1% +2.99% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 391ns ± 0% 409ns ± 0% +4.66% (p=0.000 n=30+29)
AppendInPlace/Grow/3Ptr-8 462ns ± 0% 481ns ± 0% +4.13% (p=0.000 n=29+30)
AppendInPlace/Grow/4Ptr-8 479ns ± 0% 502ns ± 0% +4.81% (p=0.000 n=27+26)
New generated code:
var x []byte
func a() {
x = append(x, 1)
}
"".a t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (a.go:5) TEXT "".a(SB), $72-0
0x0000 00000 (a.go:5) MOVQ (TLS), CX
0x0009 00009 (a.go:5) CMPQ SP, 16(CX)
0x000d 00013 (a.go:5) JLS 190
0x0013 00019 (a.go:5) SUBQ $72, SP
0x0017 00023 (a.go:5) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:5) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:6) MOVQ "".x+16(SB), CX
0x001e 00030 (a.go:6) MOVQ "".x+8(SB), DX
0x0025 00037 (a.go:6) MOVQ "".x(SB), BX
0x002c 00044 (a.go:6) LEAQ 1(DX), BP
0x0030 00048 (a.go:6) CMPQ BP, CX
0x0033 00051 (a.go:6) JGT $0, 73
0x0035 00053 (a.go:6) LEAQ 1(DX), AX
0x0039 00057 (a.go:6) MOVQ AX, "".x+8(SB)
0x0040 00064 (a.go:6) MOVB $1, (BX)(DX*1)
0x0044 00068 (a.go:7) ADDQ $72, SP
0x0048 00072 (a.go:7) RET
0x0049 00073 (a.go:6) LEAQ type.[]uint8(SB), AX
0x0050 00080 (a.go:6) MOVQ AX, (SP)
0x0054 00084 (a.go:6) MOVQ BX, 8(SP)
0x0059 00089 (a.go:6) MOVQ DX, 16(SP)
0x005e 00094 (a.go:6) MOVQ CX, 24(SP)
0x0063 00099 (a.go:6) MOVQ BP, 32(SP)
0x0068 00104 (a.go:6) PCDATA $0, $0
0x0068 00104 (a.go:6) CALL runtime.growslice(SB)
0x006d 00109 (a.go:6) MOVQ 40(SP), CX
0x0072 00114 (a.go:6) MOVQ 48(SP), DX
0x0077 00119 (a.go:6) MOVQ DX, "".autotmp_0+64(SP)
0x007c 00124 (a.go:6) MOVQ 56(SP), BX
0x0081 00129 (a.go:6) MOVQ BX, "".x+16(SB)
0x0088 00136 (a.go:6) MOVL runtime.writeBarrier(SB), AX
0x008e 00142 (a.go:6) TESTB AL, AL
0x0090 00144 (a.go:6) JNE $0, 162
0x0092 00146 (a.go:6) MOVQ CX, "".x(SB)
0x0099 00153 (a.go:6) MOVQ "".x(SB), BX
0x00a0 00160 (a.go:6) JMP 53
0x00a2 00162 (a.go:6) LEAQ "".x(SB), BX
0x00a9 00169 (a.go:6) MOVQ BX, (SP)
0x00ad 00173 (a.go:6) MOVQ CX, 8(SP)
0x00b2 00178 (a.go:6) PCDATA $0, $0
0x00b2 00178 (a.go:6) CALL runtime.writebarrierptr(SB)
0x00b7 00183 (a.go:6) MOVQ "".autotmp_0+64(SP), DX
0x00bc 00188 (a.go:6) JMP 153
0x00be 00190 (a.go:6) NOP
0x00be 00190 (a.go:5) CALL runtime.morestack_noctxt(SB)
0x00c3 00195 (a.go:5) JMP 0
Fixes #14969 again
Change-Id: Ia50463b1f506011aad0718a4fef1d4738e43c32d
Reviewed-on: https://go-review.googlesource.com/22197
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-18 09:40:30 -07:00
|
|
|
if sn.Op == ONAME {
|
|
|
|
// Tell liveness we're about to build a new slice
|
|
|
|
s.vars[&memVar] = s.newValue1A(ssa.OpVarDef, ssa.TypeMem, sn, s.mem())
|
|
|
|
}
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
capaddr := s.newValue1I(ssa.OpOffPtr, pt, int64(Array_cap), addr)
|
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, s.config.IntSize, capaddr, r[2], s.mem())
|
|
|
|
s.insertWBstore(pt, addr, r[0], n.Lineno, 0)
|
|
|
|
// load the value we just stored to avoid having to spill it
|
|
|
|
s.vars[&ptrVar] = s.newValue2(ssa.OpLoad, pt, addr, s.mem())
|
cmd/compile: re-enable in-place append optimization
CL 21891 was too clever in its attempts to avoid spills.
Storing newlen too early caused uses of append in the runtime
itself to receive an inconsistent view of a slice,
leading to corruption.
This CL makes the generate code much more similar to
the old backend. It spills more than before,
but those spills have been contained to the grow path.
It recalculates newlen unnecessarily on the fast path,
but that's measurably cheaper than spilling it.
CL 21891 caused runtime failures in 6 of 2000 runs
of net/http and crypto/x509 in my test setup.
This CL has gone 6000 runs without a failure.
Benchmarks going from master to this CL:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 439ns ± 2% 436ns ± 2% -0.72% (p=0.001 n=28+27)
AppendInPlace/NoGrow/1Ptr-8 901ns ± 0% 856ns ± 0% -4.95% (p=0.000 n=26+29)
AppendInPlace/NoGrow/2Ptr-8 2.15µs ± 1% 1.95µs ± 0% -9.07% (p=0.000 n=28+30)
AppendInPlace/NoGrow/3Ptr-8 2.66µs ± 0% 2.45µs ± 0% -7.93% (p=0.000 n=29+26)
AppendInPlace/NoGrow/4Ptr-8 3.24µs ± 1% 3.02µs ± 1% -6.75% (p=0.000 n=28+30)
AppendInPlace/Grow/Byte-8 269ns ± 1% 271ns ± 1% +0.84% (p=0.000 n=30+29)
AppendInPlace/Grow/1Ptr-8 275ns ± 1% 280ns ± 1% +1.75% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 384ns ± 0% 391ns ± 0% +1.94% (p=0.000 n=27+30)
AppendInPlace/Grow/3Ptr-8 455ns ± 0% 462ns ± 0% +1.43% (p=0.000 n=29+29)
AppendInPlace/Grow/4Ptr-8 478ns ± 0% 479ns ± 0% +0.23% (p=0.000 n=30+27)
However, for the large no-grow cases, there is still more work to be done.
Going from this CL to the non-SSA backend:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 436ns ± 2% 436ns ± 2% ~ (p=0.967 n=27+29)
AppendInPlace/NoGrow/1Ptr-8 856ns ± 0% 884ns ± 0% +3.28% (p=0.000 n=29+26)
AppendInPlace/NoGrow/2Ptr-8 1.95µs ± 0% 1.56µs ± 0% -20.28% (p=0.000 n=30+29)
AppendInPlace/NoGrow/3Ptr-8 2.45µs ± 0% 1.89µs ± 0% -22.88% (p=0.000 n=26+28)
AppendInPlace/NoGrow/4Ptr-8 3.02µs ± 1% 2.56µs ± 1% -15.35% (p=0.000 n=30+28)
AppendInPlace/Grow/Byte-8 271ns ± 1% 283ns ± 1% +4.56% (p=0.000 n=29+29)
AppendInPlace/Grow/1Ptr-8 280ns ± 1% 288ns ± 1% +2.99% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 391ns ± 0% 409ns ± 0% +4.66% (p=0.000 n=30+29)
AppendInPlace/Grow/3Ptr-8 462ns ± 0% 481ns ± 0% +4.13% (p=0.000 n=29+30)
AppendInPlace/Grow/4Ptr-8 479ns ± 0% 502ns ± 0% +4.81% (p=0.000 n=27+26)
New generated code:
var x []byte
func a() {
x = append(x, 1)
}
"".a t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (a.go:5) TEXT "".a(SB), $72-0
0x0000 00000 (a.go:5) MOVQ (TLS), CX
0x0009 00009 (a.go:5) CMPQ SP, 16(CX)
0x000d 00013 (a.go:5) JLS 190
0x0013 00019 (a.go:5) SUBQ $72, SP
0x0017 00023 (a.go:5) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:5) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:6) MOVQ "".x+16(SB), CX
0x001e 00030 (a.go:6) MOVQ "".x+8(SB), DX
0x0025 00037 (a.go:6) MOVQ "".x(SB), BX
0x002c 00044 (a.go:6) LEAQ 1(DX), BP
0x0030 00048 (a.go:6) CMPQ BP, CX
0x0033 00051 (a.go:6) JGT $0, 73
0x0035 00053 (a.go:6) LEAQ 1(DX), AX
0x0039 00057 (a.go:6) MOVQ AX, "".x+8(SB)
0x0040 00064 (a.go:6) MOVB $1, (BX)(DX*1)
0x0044 00068 (a.go:7) ADDQ $72, SP
0x0048 00072 (a.go:7) RET
0x0049 00073 (a.go:6) LEAQ type.[]uint8(SB), AX
0x0050 00080 (a.go:6) MOVQ AX, (SP)
0x0054 00084 (a.go:6) MOVQ BX, 8(SP)
0x0059 00089 (a.go:6) MOVQ DX, 16(SP)
0x005e 00094 (a.go:6) MOVQ CX, 24(SP)
0x0063 00099 (a.go:6) MOVQ BP, 32(SP)
0x0068 00104 (a.go:6) PCDATA $0, $0
0x0068 00104 (a.go:6) CALL runtime.growslice(SB)
0x006d 00109 (a.go:6) MOVQ 40(SP), CX
0x0072 00114 (a.go:6) MOVQ 48(SP), DX
0x0077 00119 (a.go:6) MOVQ DX, "".autotmp_0+64(SP)
0x007c 00124 (a.go:6) MOVQ 56(SP), BX
0x0081 00129 (a.go:6) MOVQ BX, "".x+16(SB)
0x0088 00136 (a.go:6) MOVL runtime.writeBarrier(SB), AX
0x008e 00142 (a.go:6) TESTB AL, AL
0x0090 00144 (a.go:6) JNE $0, 162
0x0092 00146 (a.go:6) MOVQ CX, "".x(SB)
0x0099 00153 (a.go:6) MOVQ "".x(SB), BX
0x00a0 00160 (a.go:6) JMP 53
0x00a2 00162 (a.go:6) LEAQ "".x(SB), BX
0x00a9 00169 (a.go:6) MOVQ BX, (SP)
0x00ad 00173 (a.go:6) MOVQ CX, 8(SP)
0x00b2 00178 (a.go:6) PCDATA $0, $0
0x00b2 00178 (a.go:6) CALL runtime.writebarrierptr(SB)
0x00b7 00183 (a.go:6) MOVQ "".autotmp_0+64(SP), DX
0x00bc 00188 (a.go:6) JMP 153
0x00be 00190 (a.go:6) NOP
0x00be 00190 (a.go:5) CALL runtime.morestack_noctxt(SB)
0x00c3 00195 (a.go:5) JMP 0
Fixes #14969 again
Change-Id: Ia50463b1f506011aad0718a4fef1d4738e43c32d
Reviewed-on: https://go-review.googlesource.com/22197
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-18 09:40:30 -07:00
|
|
|
s.vars[&lenVar] = r[1] // avoid a spill in the fast path
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
} else {
|
|
|
|
s.vars[&ptrVar] = r[0]
|
|
|
|
s.vars[&newlenVar] = s.newValue2(s.ssaOp(OADD, Types[TINT]), Types[TINT], r[1], s.constInt(Types[TINT], nargs))
|
|
|
|
s.vars[&capVar] = r[2]
|
|
|
|
}
|
|
|
|
|
2016-04-04 10:58:21 -07:00
|
|
|
b = s.endBlock()
|
|
|
|
b.AddEdgeTo(assign)
|
|
|
|
|
|
|
|
// assign new elements to slots
|
|
|
|
s.startBlock(assign)
|
|
|
|
|
cmd/compile: re-enable in-place append optimization
CL 21891 was too clever in its attempts to avoid spills.
Storing newlen too early caused uses of append in the runtime
itself to receive an inconsistent view of a slice,
leading to corruption.
This CL makes the generate code much more similar to
the old backend. It spills more than before,
but those spills have been contained to the grow path.
It recalculates newlen unnecessarily on the fast path,
but that's measurably cheaper than spilling it.
CL 21891 caused runtime failures in 6 of 2000 runs
of net/http and crypto/x509 in my test setup.
This CL has gone 6000 runs without a failure.
Benchmarks going from master to this CL:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 439ns ± 2% 436ns ± 2% -0.72% (p=0.001 n=28+27)
AppendInPlace/NoGrow/1Ptr-8 901ns ± 0% 856ns ± 0% -4.95% (p=0.000 n=26+29)
AppendInPlace/NoGrow/2Ptr-8 2.15µs ± 1% 1.95µs ± 0% -9.07% (p=0.000 n=28+30)
AppendInPlace/NoGrow/3Ptr-8 2.66µs ± 0% 2.45µs ± 0% -7.93% (p=0.000 n=29+26)
AppendInPlace/NoGrow/4Ptr-8 3.24µs ± 1% 3.02µs ± 1% -6.75% (p=0.000 n=28+30)
AppendInPlace/Grow/Byte-8 269ns ± 1% 271ns ± 1% +0.84% (p=0.000 n=30+29)
AppendInPlace/Grow/1Ptr-8 275ns ± 1% 280ns ± 1% +1.75% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 384ns ± 0% 391ns ± 0% +1.94% (p=0.000 n=27+30)
AppendInPlace/Grow/3Ptr-8 455ns ± 0% 462ns ± 0% +1.43% (p=0.000 n=29+29)
AppendInPlace/Grow/4Ptr-8 478ns ± 0% 479ns ± 0% +0.23% (p=0.000 n=30+27)
However, for the large no-grow cases, there is still more work to be done.
Going from this CL to the non-SSA backend:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 436ns ± 2% 436ns ± 2% ~ (p=0.967 n=27+29)
AppendInPlace/NoGrow/1Ptr-8 856ns ± 0% 884ns ± 0% +3.28% (p=0.000 n=29+26)
AppendInPlace/NoGrow/2Ptr-8 1.95µs ± 0% 1.56µs ± 0% -20.28% (p=0.000 n=30+29)
AppendInPlace/NoGrow/3Ptr-8 2.45µs ± 0% 1.89µs ± 0% -22.88% (p=0.000 n=26+28)
AppendInPlace/NoGrow/4Ptr-8 3.02µs ± 1% 2.56µs ± 1% -15.35% (p=0.000 n=30+28)
AppendInPlace/Grow/Byte-8 271ns ± 1% 283ns ± 1% +4.56% (p=0.000 n=29+29)
AppendInPlace/Grow/1Ptr-8 280ns ± 1% 288ns ± 1% +2.99% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 391ns ± 0% 409ns ± 0% +4.66% (p=0.000 n=30+29)
AppendInPlace/Grow/3Ptr-8 462ns ± 0% 481ns ± 0% +4.13% (p=0.000 n=29+30)
AppendInPlace/Grow/4Ptr-8 479ns ± 0% 502ns ± 0% +4.81% (p=0.000 n=27+26)
New generated code:
var x []byte
func a() {
x = append(x, 1)
}
"".a t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (a.go:5) TEXT "".a(SB), $72-0
0x0000 00000 (a.go:5) MOVQ (TLS), CX
0x0009 00009 (a.go:5) CMPQ SP, 16(CX)
0x000d 00013 (a.go:5) JLS 190
0x0013 00019 (a.go:5) SUBQ $72, SP
0x0017 00023 (a.go:5) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:5) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:6) MOVQ "".x+16(SB), CX
0x001e 00030 (a.go:6) MOVQ "".x+8(SB), DX
0x0025 00037 (a.go:6) MOVQ "".x(SB), BX
0x002c 00044 (a.go:6) LEAQ 1(DX), BP
0x0030 00048 (a.go:6) CMPQ BP, CX
0x0033 00051 (a.go:6) JGT $0, 73
0x0035 00053 (a.go:6) LEAQ 1(DX), AX
0x0039 00057 (a.go:6) MOVQ AX, "".x+8(SB)
0x0040 00064 (a.go:6) MOVB $1, (BX)(DX*1)
0x0044 00068 (a.go:7) ADDQ $72, SP
0x0048 00072 (a.go:7) RET
0x0049 00073 (a.go:6) LEAQ type.[]uint8(SB), AX
0x0050 00080 (a.go:6) MOVQ AX, (SP)
0x0054 00084 (a.go:6) MOVQ BX, 8(SP)
0x0059 00089 (a.go:6) MOVQ DX, 16(SP)
0x005e 00094 (a.go:6) MOVQ CX, 24(SP)
0x0063 00099 (a.go:6) MOVQ BP, 32(SP)
0x0068 00104 (a.go:6) PCDATA $0, $0
0x0068 00104 (a.go:6) CALL runtime.growslice(SB)
0x006d 00109 (a.go:6) MOVQ 40(SP), CX
0x0072 00114 (a.go:6) MOVQ 48(SP), DX
0x0077 00119 (a.go:6) MOVQ DX, "".autotmp_0+64(SP)
0x007c 00124 (a.go:6) MOVQ 56(SP), BX
0x0081 00129 (a.go:6) MOVQ BX, "".x+16(SB)
0x0088 00136 (a.go:6) MOVL runtime.writeBarrier(SB), AX
0x008e 00142 (a.go:6) TESTB AL, AL
0x0090 00144 (a.go:6) JNE $0, 162
0x0092 00146 (a.go:6) MOVQ CX, "".x(SB)
0x0099 00153 (a.go:6) MOVQ "".x(SB), BX
0x00a0 00160 (a.go:6) JMP 53
0x00a2 00162 (a.go:6) LEAQ "".x(SB), BX
0x00a9 00169 (a.go:6) MOVQ BX, (SP)
0x00ad 00173 (a.go:6) MOVQ CX, 8(SP)
0x00b2 00178 (a.go:6) PCDATA $0, $0
0x00b2 00178 (a.go:6) CALL runtime.writebarrierptr(SB)
0x00b7 00183 (a.go:6) MOVQ "".autotmp_0+64(SP), DX
0x00bc 00188 (a.go:6) JMP 153
0x00be 00190 (a.go:6) NOP
0x00be 00190 (a.go:5) CALL runtime.morestack_noctxt(SB)
0x00c3 00195 (a.go:5) JMP 0
Fixes #14969 again
Change-Id: Ia50463b1f506011aad0718a4fef1d4738e43c32d
Reviewed-on: https://go-review.googlesource.com/22197
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-18 09:40:30 -07:00
|
|
|
if inplace {
|
|
|
|
l = s.variable(&lenVar, Types[TINT]) // generates phi for len
|
|
|
|
nl = s.newValue2(s.ssaOp(OADD, Types[TINT]), Types[TINT], l, s.constInt(Types[TINT], nargs))
|
|
|
|
lenaddr := s.newValue1I(ssa.OpOffPtr, pt, int64(Array_nel), addr)
|
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, s.config.IntSize, lenaddr, nl, s.mem())
|
|
|
|
}
|
|
|
|
|
2016-04-04 10:58:21 -07:00
|
|
|
// Evaluate args
|
2016-05-27 14:07:37 -07:00
|
|
|
type argRec struct {
|
|
|
|
// if store is true, we're appending the value v. If false, we're appending the
|
|
|
|
// value at *v. If store==false, isVolatile reports whether the source
|
|
|
|
// is in the outargs section of the stack frame.
|
|
|
|
v *ssa.Value
|
|
|
|
store bool
|
|
|
|
isVolatile bool
|
|
|
|
}
|
|
|
|
args := make([]argRec, 0, nargs)
|
2016-04-04 10:58:21 -07:00
|
|
|
for _, n := range n.List.Slice()[1:] {
|
|
|
|
if canSSAType(n.Type) {
|
2016-05-27 14:07:37 -07:00
|
|
|
args = append(args, argRec{v: s.expr(n), store: true})
|
2016-04-04 10:58:21 -07:00
|
|
|
} else {
|
2016-05-27 14:07:37 -07:00
|
|
|
v, isVolatile := s.addr(n, false)
|
|
|
|
args = append(args, argRec{v: v, isVolatile: isVolatile})
|
2016-04-04 10:58:21 -07:00
|
|
|
}
|
|
|
|
}
|
2015-09-18 15:11:30 -07:00
|
|
|
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
p = s.variable(&ptrVar, pt) // generates phi for ptr
|
|
|
|
if !inplace {
|
|
|
|
nl = s.variable(&newlenVar, Types[TINT]) // generates phi for nl
|
|
|
|
c = s.variable(&capVar, Types[TINT]) // generates phi for cap
|
|
|
|
}
|
2016-04-04 10:58:21 -07:00
|
|
|
p2 := s.newValue2(ssa.OpPtrIndex, pt, p, l)
|
|
|
|
// TODO: just one write barrier call for all of these writes?
|
|
|
|
// TODO: maybe just one writeBarrier.enabled check?
|
|
|
|
for i, arg := range args {
|
|
|
|
addr := s.newValue2(ssa.OpPtrIndex, pt, p2, s.constInt(Types[TINT], int64(i)))
|
2016-05-27 14:07:37 -07:00
|
|
|
if arg.store {
|
2016-04-04 10:58:21 -07:00
|
|
|
if haspointers(et) {
|
2016-05-27 14:07:37 -07:00
|
|
|
s.insertWBstore(et, addr, arg.v, n.Lineno, 0)
|
2015-10-07 14:35:25 -07:00
|
|
|
} else {
|
2016-05-27 14:07:37 -07:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, et.Size(), addr, arg.v, s.mem())
|
2015-10-07 14:35:25 -07:00
|
|
|
}
|
2016-04-04 10:58:21 -07:00
|
|
|
} else {
|
|
|
|
if haspointers(et) {
|
2016-05-27 14:07:37 -07:00
|
|
|
s.insertWBmove(et, addr, arg.v, n.Lineno, arg.isVolatile)
|
2015-10-07 14:35:25 -07:00
|
|
|
} else {
|
2016-06-27 16:54:57 -04:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpMove, ssa.TypeMem, SizeAlignAuxInt(et), addr, arg.v, s.mem())
|
2015-09-11 11:02:57 -07:00
|
|
|
}
|
|
|
|
}
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
2016-04-04 10:58:21 -07:00
|
|
|
|
|
|
|
delete(s.vars, &ptrVar)
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
if inplace {
|
cmd/compile: re-enable in-place append optimization
CL 21891 was too clever in its attempts to avoid spills.
Storing newlen too early caused uses of append in the runtime
itself to receive an inconsistent view of a slice,
leading to corruption.
This CL makes the generate code much more similar to
the old backend. It spills more than before,
but those spills have been contained to the grow path.
It recalculates newlen unnecessarily on the fast path,
but that's measurably cheaper than spilling it.
CL 21891 caused runtime failures in 6 of 2000 runs
of net/http and crypto/x509 in my test setup.
This CL has gone 6000 runs without a failure.
Benchmarks going from master to this CL:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 439ns ± 2% 436ns ± 2% -0.72% (p=0.001 n=28+27)
AppendInPlace/NoGrow/1Ptr-8 901ns ± 0% 856ns ± 0% -4.95% (p=0.000 n=26+29)
AppendInPlace/NoGrow/2Ptr-8 2.15µs ± 1% 1.95µs ± 0% -9.07% (p=0.000 n=28+30)
AppendInPlace/NoGrow/3Ptr-8 2.66µs ± 0% 2.45µs ± 0% -7.93% (p=0.000 n=29+26)
AppendInPlace/NoGrow/4Ptr-8 3.24µs ± 1% 3.02µs ± 1% -6.75% (p=0.000 n=28+30)
AppendInPlace/Grow/Byte-8 269ns ± 1% 271ns ± 1% +0.84% (p=0.000 n=30+29)
AppendInPlace/Grow/1Ptr-8 275ns ± 1% 280ns ± 1% +1.75% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 384ns ± 0% 391ns ± 0% +1.94% (p=0.000 n=27+30)
AppendInPlace/Grow/3Ptr-8 455ns ± 0% 462ns ± 0% +1.43% (p=0.000 n=29+29)
AppendInPlace/Grow/4Ptr-8 478ns ± 0% 479ns ± 0% +0.23% (p=0.000 n=30+27)
However, for the large no-grow cases, there is still more work to be done.
Going from this CL to the non-SSA backend:
name old time/op new time/op delta
AppendInPlace/NoGrow/Byte-8 436ns ± 2% 436ns ± 2% ~ (p=0.967 n=27+29)
AppendInPlace/NoGrow/1Ptr-8 856ns ± 0% 884ns ± 0% +3.28% (p=0.000 n=29+26)
AppendInPlace/NoGrow/2Ptr-8 1.95µs ± 0% 1.56µs ± 0% -20.28% (p=0.000 n=30+29)
AppendInPlace/NoGrow/3Ptr-8 2.45µs ± 0% 1.89µs ± 0% -22.88% (p=0.000 n=26+28)
AppendInPlace/NoGrow/4Ptr-8 3.02µs ± 1% 2.56µs ± 1% -15.35% (p=0.000 n=30+28)
AppendInPlace/Grow/Byte-8 271ns ± 1% 283ns ± 1% +4.56% (p=0.000 n=29+29)
AppendInPlace/Grow/1Ptr-8 280ns ± 1% 288ns ± 1% +2.99% (p=0.000 n=30+30)
AppendInPlace/Grow/2Ptr-8 391ns ± 0% 409ns ± 0% +4.66% (p=0.000 n=30+29)
AppendInPlace/Grow/3Ptr-8 462ns ± 0% 481ns ± 0% +4.13% (p=0.000 n=29+30)
AppendInPlace/Grow/4Ptr-8 479ns ± 0% 502ns ± 0% +4.81% (p=0.000 n=27+26)
New generated code:
var x []byte
func a() {
x = append(x, 1)
}
"".a t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (a.go:5) TEXT "".a(SB), $72-0
0x0000 00000 (a.go:5) MOVQ (TLS), CX
0x0009 00009 (a.go:5) CMPQ SP, 16(CX)
0x000d 00013 (a.go:5) JLS 190
0x0013 00019 (a.go:5) SUBQ $72, SP
0x0017 00023 (a.go:5) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:5) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (a.go:6) MOVQ "".x+16(SB), CX
0x001e 00030 (a.go:6) MOVQ "".x+8(SB), DX
0x0025 00037 (a.go:6) MOVQ "".x(SB), BX
0x002c 00044 (a.go:6) LEAQ 1(DX), BP
0x0030 00048 (a.go:6) CMPQ BP, CX
0x0033 00051 (a.go:6) JGT $0, 73
0x0035 00053 (a.go:6) LEAQ 1(DX), AX
0x0039 00057 (a.go:6) MOVQ AX, "".x+8(SB)
0x0040 00064 (a.go:6) MOVB $1, (BX)(DX*1)
0x0044 00068 (a.go:7) ADDQ $72, SP
0x0048 00072 (a.go:7) RET
0x0049 00073 (a.go:6) LEAQ type.[]uint8(SB), AX
0x0050 00080 (a.go:6) MOVQ AX, (SP)
0x0054 00084 (a.go:6) MOVQ BX, 8(SP)
0x0059 00089 (a.go:6) MOVQ DX, 16(SP)
0x005e 00094 (a.go:6) MOVQ CX, 24(SP)
0x0063 00099 (a.go:6) MOVQ BP, 32(SP)
0x0068 00104 (a.go:6) PCDATA $0, $0
0x0068 00104 (a.go:6) CALL runtime.growslice(SB)
0x006d 00109 (a.go:6) MOVQ 40(SP), CX
0x0072 00114 (a.go:6) MOVQ 48(SP), DX
0x0077 00119 (a.go:6) MOVQ DX, "".autotmp_0+64(SP)
0x007c 00124 (a.go:6) MOVQ 56(SP), BX
0x0081 00129 (a.go:6) MOVQ BX, "".x+16(SB)
0x0088 00136 (a.go:6) MOVL runtime.writeBarrier(SB), AX
0x008e 00142 (a.go:6) TESTB AL, AL
0x0090 00144 (a.go:6) JNE $0, 162
0x0092 00146 (a.go:6) MOVQ CX, "".x(SB)
0x0099 00153 (a.go:6) MOVQ "".x(SB), BX
0x00a0 00160 (a.go:6) JMP 53
0x00a2 00162 (a.go:6) LEAQ "".x(SB), BX
0x00a9 00169 (a.go:6) MOVQ BX, (SP)
0x00ad 00173 (a.go:6) MOVQ CX, 8(SP)
0x00b2 00178 (a.go:6) PCDATA $0, $0
0x00b2 00178 (a.go:6) CALL runtime.writebarrierptr(SB)
0x00b7 00183 (a.go:6) MOVQ "".autotmp_0+64(SP), DX
0x00bc 00188 (a.go:6) JMP 153
0x00be 00190 (a.go:6) NOP
0x00be 00190 (a.go:5) CALL runtime.morestack_noctxt(SB)
0x00c3 00195 (a.go:5) JMP 0
Fixes #14969 again
Change-Id: Ia50463b1f506011aad0718a4fef1d4738e43c32d
Reviewed-on: https://go-review.googlesource.com/22197
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-18 09:40:30 -07:00
|
|
|
delete(s.vars, &lenVar)
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
return nil
|
|
|
|
}
|
cmd/compile: avoid a spill in append fast path
Instead of spilling newlen, recalculate it.
This removes a spill from the fast path,
at the cost of a cheap recalculation
on the (rare) growth path.
This uses 8 bytes less of stack space.
It generates two more bytes of code,
but that is due to suboptimal register allocation;
see far below.
Runtime append microbenchmarks are all over the map,
presumably due to incidental code movement.
Sample code:
func s(b []byte) []byte {
b = append(b, 1, 2, 3)
return b
}
Before:
"".s t=1 size=160 args=0x30 locals=0x48
0x0000 00000 (append.go:8) TEXT "".s(SB), $72-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 149
0x0013 00019 (append.go:8) SUBQ $72, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+88(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ DX, "".autotmp_0+64(SP)
0x0025 00037 (append.go:9) MOVQ "".b+96(FP), BX
0x002a 00042 (append.go:9) CMPQ DX, BX
0x002d 00045 (append.go:9) JGT $0, 86
0x002f 00047 (append.go:8) MOVQ "".b+80(FP), AX
0x0034 00052 (append.go:9) MOVB $1, (AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x003d 00061 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x0042 00066 (append.go:10) MOVQ AX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ DX, "".~r1+112(FP)
0x004c 00076 (append.go:10) MOVQ BX, "".~r1+120(FP)
0x0051 00081 (append.go:10) ADDQ $72, SP
0x0055 00085 (append.go:10) RET
0x0056 00086 (append.go:9) LEAQ type.[]uint8(SB), AX
0x005d 00093 (append.go:9) MOVQ AX, (SP)
0x0061 00097 (append.go:9) MOVQ "".b+80(FP), BP
0x0066 00102 (append.go:9) MOVQ BP, 8(SP)
0x006b 00107 (append.go:9) MOVQ CX, 16(SP)
0x0070 00112 (append.go:9) MOVQ BX, 24(SP)
0x0075 00117 (append.go:9) MOVQ DX, 32(SP)
0x007a 00122 (append.go:9) PCDATA $0, $0
0x007a 00122 (append.go:9) CALL runtime.growslice(SB)
0x007f 00127 (append.go:9) MOVQ 40(SP), AX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:8) MOVQ "".b+88(FP), CX
0x008e 00142 (append.go:9) MOVQ "".autotmp_0+64(SP), DX
0x0093 00147 (append.go:9) JMP 52
0x0095 00149 (append.go:9) NOP
0x0095 00149 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009a 00154 (append.go:8) JMP 0
After:
"".s t=1 size=176 args=0x30 locals=0x40
0x0000 00000 (append.go:8) TEXT "".s(SB), $64-48
0x0000 00000 (append.go:8) MOVQ (TLS), CX
0x0009 00009 (append.go:8) CMPQ SP, 16(CX)
0x000d 00013 (append.go:8) JLS 151
0x0013 00019 (append.go:8) SUBQ $64, SP
0x0017 00023 (append.go:8) FUNCDATA $0, gclocals·6432f8c6a0d23fa7bee6c5d96f21a92a(SB)
0x0017 00023 (append.go:8) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:9) MOVQ "".b+80(FP), CX
0x001c 00028 (append.go:9) LEAQ 3(CX), DX
0x0020 00032 (append.go:9) MOVQ "".b+88(FP), BX
0x0025 00037 (append.go:9) CMPQ DX, BX
0x0028 00040 (append.go:9) JGT $0, 81
0x002a 00042 (append.go:8) MOVQ "".b+72(FP), AX
0x002f 00047 (append.go:9) MOVB $1, (AX)(CX*1)
0x0033 00051 (append.go:9) MOVB $2, 1(AX)(CX*1)
0x0038 00056 (append.go:9) MOVB $3, 2(AX)(CX*1)
0x003d 00061 (append.go:10) MOVQ AX, "".~r1+96(FP)
0x0042 00066 (append.go:10) MOVQ DX, "".~r1+104(FP)
0x0047 00071 (append.go:10) MOVQ BX, "".~r1+112(FP)
0x004c 00076 (append.go:10) ADDQ $64, SP
0x0050 00080 (append.go:10) RET
0x0051 00081 (append.go:9) LEAQ type.[]uint8(SB), AX
0x0058 00088 (append.go:9) MOVQ AX, (SP)
0x005c 00092 (append.go:9) MOVQ "".b+72(FP), BP
0x0061 00097 (append.go:9) MOVQ BP, 8(SP)
0x0066 00102 (append.go:9) MOVQ CX, 16(SP)
0x006b 00107 (append.go:9) MOVQ BX, 24(SP)
0x0070 00112 (append.go:9) MOVQ DX, 32(SP)
0x0075 00117 (append.go:9) PCDATA $0, $0
0x0075 00117 (append.go:9) CALL runtime.growslice(SB)
0x007a 00122 (append.go:9) MOVQ 40(SP), AX
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
0x0095 00149 (append.go:9) JMP 47
0x0097 00151 (append.go:9) NOP
0x0097 00151 (append.go:8) CALL runtime.morestack_noctxt(SB)
0x009c 00156 (append.go:8) JMP 0
Observe that in the following sequence,
we should use DX directly instead of using
CX as a temporary register, which would make
the new code a strict improvement on the old:
0x007f 00127 (append.go:9) MOVQ 48(SP), CX
0x0084 00132 (append.go:9) MOVQ 56(SP), BX
0x0089 00137 (append.go:9) ADDQ $3, CX
0x008d 00141 (append.go:9) MOVQ CX, DX
0x0090 00144 (append.go:8) MOVQ "".b+80(FP), CX
Change-Id: I4ee50b18fa53865901d2d7f86c2cbb54c6fa6924
Reviewed-on: https://go-review.googlesource.com/21812
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:08:00 -07:00
|
|
|
delete(s.vars, &newlenVar)
|
2016-04-04 10:58:21 -07:00
|
|
|
delete(s.vars, &capVar)
|
cmd/compile: avoid write barrier in append fast path
When we are writing the result of an append back
to the same slice, we don’t need a write barrier
on the fast path.
This re-implements an optimization that was present
in the old backend.
Updates #14921
Fixes #14969
Sample code:
var x []byte
func p() {
x = append(x, 1, 2, 3)
}
Before:
"".p t=1 size=224 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 199
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x(SB), DX
0x0025 00037 (append.go:19) MOVQ "".x+8(SB), BX
0x002c 00044 (append.go:19) MOVQ BX, "".autotmp_0+64(SP)
0x0031 00049 (append.go:22) LEAQ 3(BX), BP
0x0035 00053 (append.go:22) CMPQ BP, CX
0x0038 00056 (append.go:22) JGT $0, 131
0x003a 00058 (append.go:22) MOVB $1, (DX)(BX*1)
0x003e 00062 (append.go:22) MOVB $2, 1(DX)(BX*1)
0x0043 00067 (append.go:22) MOVB $3, 2(DX)(BX*1)
0x0048 00072 (append.go:22) MOVQ BP, "".x+8(SB)
0x004f 00079 (append.go:22) MOVQ CX, "".x+16(SB)
0x0056 00086 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x005c 00092 (append.go:22) TESTB AL, AL
0x005e 00094 (append.go:22) JNE $0, 108
0x0060 00096 (append.go:22) MOVQ DX, "".x(SB)
0x0067 00103 (append.go:23) ADDQ $72, SP
0x006b 00107 (append.go:23) RET
0x006c 00108 (append.go:22) LEAQ "".x(SB), CX
0x0073 00115 (append.go:22) MOVQ CX, (SP)
0x0077 00119 (append.go:22) MOVQ DX, 8(SP)
0x007c 00124 (append.go:22) PCDATA $0, $0
0x007c 00124 (append.go:22) CALL runtime.writebarrierptr(SB)
0x0081 00129 (append.go:23) JMP 103
0x0083 00131 (append.go:22) LEAQ type.[]uint8(SB), AX
0x008a 00138 (append.go:22) MOVQ AX, (SP)
0x008e 00142 (append.go:22) MOVQ DX, 8(SP)
0x0093 00147 (append.go:22) MOVQ BX, 16(SP)
0x0098 00152 (append.go:22) MOVQ CX, 24(SP)
0x009d 00157 (append.go:22) MOVQ BP, 32(SP)
0x00a2 00162 (append.go:22) PCDATA $0, $0
0x00a2 00162 (append.go:22) CALL runtime.growslice(SB)
0x00a7 00167 (append.go:22) MOVQ 40(SP), DX
0x00ac 00172 (append.go:22) MOVQ 48(SP), AX
0x00b1 00177 (append.go:22) MOVQ 56(SP), CX
0x00b6 00182 (append.go:22) ADDQ $3, AX
0x00ba 00186 (append.go:19) MOVQ "".autotmp_0+64(SP), BX
0x00bf 00191 (append.go:22) MOVQ AX, BP
0x00c2 00194 (append.go:22) JMP 58
0x00c7 00199 (append.go:22) NOP
0x00c7 00199 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00cc 00204 (append.go:21) JMP 0
After:
"".p t=1 size=208 args=0x0 locals=0x48
0x0000 00000 (append.go:21) TEXT "".p(SB), $72-0
0x0000 00000 (append.go:21) MOVQ (TLS), CX
0x0009 00009 (append.go:21) CMPQ SP, 16(CX)
0x000d 00013 (append.go:21) JLS 191
0x0013 00019 (append.go:21) SUBQ $72, SP
0x0017 00023 (append.go:21) FUNCDATA $0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:21) FUNCDATA $1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
0x0017 00023 (append.go:19) MOVQ "".x+16(SB), CX
0x001e 00030 (append.go:19) MOVQ "".x+8(SB), DX
0x0025 00037 (append.go:19) MOVQ DX, "".autotmp_0+64(SP)
0x002a 00042 (append.go:19) MOVQ "".x(SB), BX
0x0031 00049 (append.go:22) LEAQ 3(DX), BP
0x0035 00053 (append.go:22) MOVQ BP, "".x+8(SB)
0x003c 00060 (append.go:22) CMPQ BP, CX
0x003f 00063 (append.go:22) JGT $0, 84
0x0041 00065 (append.go:22) MOVB $1, (BX)(DX*1)
0x0045 00069 (append.go:22) MOVB $2, 1(BX)(DX*1)
0x004a 00074 (append.go:22) MOVB $3, 2(BX)(DX*1)
0x004f 00079 (append.go:23) ADDQ $72, SP
0x0053 00083 (append.go:23) RET
0x0054 00084 (append.go:22) LEAQ type.[]uint8(SB), AX
0x005b 00091 (append.go:22) MOVQ AX, (SP)
0x005f 00095 (append.go:22) MOVQ BX, 8(SP)
0x0064 00100 (append.go:22) MOVQ DX, 16(SP)
0x0069 00105 (append.go:22) MOVQ CX, 24(SP)
0x006e 00110 (append.go:22) MOVQ BP, 32(SP)
0x0073 00115 (append.go:22) PCDATA $0, $0
0x0073 00115 (append.go:22) CALL runtime.growslice(SB)
0x0078 00120 (append.go:22) MOVQ 40(SP), CX
0x007d 00125 (append.go:22) MOVQ 56(SP), AX
0x0082 00130 (append.go:22) MOVQ AX, "".x+16(SB)
0x0089 00137 (append.go:22) MOVL runtime.writeBarrier(SB), AX
0x008f 00143 (append.go:22) TESTB AL, AL
0x0091 00145 (append.go:22) JNE $0, 168
0x0093 00147 (append.go:22) MOVQ CX, "".x(SB)
0x009a 00154 (append.go:22) MOVQ "".x(SB), BX
0x00a1 00161 (append.go:19) MOVQ "".autotmp_0+64(SP), DX
0x00a6 00166 (append.go:22) JMP 65
0x00a8 00168 (append.go:22) LEAQ "".x(SB), DX
0x00af 00175 (append.go:22) MOVQ DX, (SP)
0x00b3 00179 (append.go:22) MOVQ CX, 8(SP)
0x00b8 00184 (append.go:22) PCDATA $0, $0
0x00b8 00184 (append.go:22) CALL runtime.writebarrierptr(SB)
0x00bd 00189 (append.go:22) JMP 154
0x00bf 00191 (append.go:22) NOP
0x00bf 00191 (append.go:21) CALL runtime.morestack_noctxt(SB)
0x00c4 00196 (append.go:21) JMP 0
Change-Id: I77a41ad3a22557a4bb4654de7d6d24a029efe34a
Reviewed-on: https://go-review.googlesource.com/21813
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
2016-04-10 09:44:17 -07:00
|
|
|
// make result
|
2016-04-04 10:58:21 -07:00
|
|
|
return s.newValue3(ssa.OpSliceMake, n.Type, p, nl, c)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
|
2015-11-02 16:56:53 -08:00
|
|
|
// condBranch evaluates the boolean expression cond and branches to yes
|
|
|
|
// if cond is true and no if cond is false.
|
|
|
|
// This function is intended to handle && and || better than just calling
|
|
|
|
// s.expr(cond) and branching on the result.
|
|
|
|
func (s *state) condBranch(cond *Node, yes, no *ssa.Block, likely int8) {
|
|
|
|
if cond.Op == OANDAND {
|
|
|
|
mid := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
s.stmtList(cond.Ninit)
|
|
|
|
s.condBranch(cond.Left, mid, no, max8(likely, 0))
|
|
|
|
s.startBlock(mid)
|
|
|
|
s.condBranch(cond.Right, yes, no, likely)
|
|
|
|
return
|
|
|
|
// Note: if likely==1, then both recursive calls pass 1.
|
|
|
|
// If likely==-1, then we don't have enough information to decide
|
2016-03-01 23:21:55 +00:00
|
|
|
// whether the first branch is likely or not. So we pass 0 for
|
2015-11-02 16:56:53 -08:00
|
|
|
// the likeliness of the first branch.
|
|
|
|
// TODO: have the frontend give us branch prediction hints for
|
|
|
|
// OANDAND and OOROR nodes (if it ever has such info).
|
|
|
|
}
|
|
|
|
if cond.Op == OOROR {
|
|
|
|
mid := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
s.stmtList(cond.Ninit)
|
|
|
|
s.condBranch(cond.Left, yes, mid, min8(likely, 0))
|
|
|
|
s.startBlock(mid)
|
|
|
|
s.condBranch(cond.Right, yes, no, likely)
|
|
|
|
return
|
|
|
|
// Note: if likely==-1, then both recursive calls pass -1.
|
|
|
|
// If likely==1, then we don't have enough info to decide
|
|
|
|
// the likelihood of the first branch.
|
|
|
|
}
|
2015-11-03 09:30:17 -08:00
|
|
|
if cond.Op == ONOT {
|
|
|
|
s.stmtList(cond.Ninit)
|
|
|
|
s.condBranch(cond.Left, no, yes, -likely)
|
|
|
|
return
|
|
|
|
}
|
2015-11-02 16:56:53 -08:00
|
|
|
c := s.expr(cond)
|
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(c)
|
2015-11-02 16:56:53 -08:00
|
|
|
b.Likely = ssa.BranchPrediction(likely) // gc and ssa both use -1/0/+1 for likeliness
|
|
|
|
b.AddEdgeTo(yes)
|
|
|
|
b.AddEdgeTo(no)
|
|
|
|
}
|
|
|
|
|
2016-03-21 10:22:03 -07:00
|
|
|
type skipMask uint8
|
|
|
|
|
|
|
|
const (
|
|
|
|
skipPtr skipMask = 1 << iota
|
|
|
|
skipLen
|
|
|
|
skipCap
|
|
|
|
)
|
|
|
|
|
2016-01-25 17:06:54 -08:00
|
|
|
// assign does left = right.
|
|
|
|
// Right has already been evaluated to ssa, left has not.
|
|
|
|
// If deref is true, then we do left = *right instead (and right has already been nil-checked).
|
|
|
|
// If deref is true and right == nil, just do left = 0.
|
2016-05-27 14:07:37 -07:00
|
|
|
// If deref is true, rightIsVolatile reports whether right points to volatile (clobbered by a call) storage.
|
2016-01-25 17:06:54 -08:00
|
|
|
// Include a write barrier if wb is true.
|
2016-03-21 10:22:03 -07:00
|
|
|
// skip indicates assignments (at the top level) that can be avoided.
|
2016-05-27 14:07:37 -07:00
|
|
|
func (s *state) assign(left *Node, right *ssa.Value, wb, deref bool, line int32, skip skipMask, rightIsVolatile bool) {
|
2015-08-14 21:47:20 -07:00
|
|
|
if left.Op == ONAME && isblank(left) {
|
|
|
|
return
|
|
|
|
}
|
|
|
|
t := left.Type
|
|
|
|
dowidth(t)
|
2016-02-27 17:49:31 -08:00
|
|
|
if s.canSSA(left) {
|
2016-01-25 17:06:54 -08:00
|
|
|
if deref {
|
|
|
|
s.Fatalf("can SSA LHS %s but not RHS %s", left, right)
|
|
|
|
}
|
2016-01-11 21:05:33 -08:00
|
|
|
if left.Op == ODOT {
|
|
|
|
// We're assigning to a field of an ssa-able value.
|
|
|
|
// We need to build a new structure with the new value for the
|
|
|
|
// field we're assigning and the old values for the other fields.
|
|
|
|
// For instance:
|
|
|
|
// type T struct {a, b, c int}
|
|
|
|
// var T x
|
|
|
|
// x.b = 5
|
|
|
|
// For the x.b = 5 assignment we want to generate x = T{x.a, 5, x.c}
|
|
|
|
|
|
|
|
// Grab information about the structure type.
|
|
|
|
t := left.Left.Type
|
|
|
|
nf := t.NumFields()
|
|
|
|
idx := fieldIdx(left)
|
|
|
|
|
|
|
|
// Grab old value of structure.
|
|
|
|
old := s.expr(left.Left)
|
|
|
|
|
|
|
|
// Make new structure.
|
|
|
|
new := s.newValue0(ssa.StructMakeOp(t.NumFields()), t)
|
|
|
|
|
|
|
|
// Add fields as args.
|
2016-03-14 12:45:18 -07:00
|
|
|
for i := 0; i < nf; i++ {
|
2016-01-11 21:05:33 -08:00
|
|
|
if i == idx {
|
|
|
|
new.AddArg(right)
|
|
|
|
} else {
|
2016-03-14 12:45:18 -07:00
|
|
|
new.AddArg(s.newValue1I(ssa.OpStructSelect, t.FieldType(i), int64(i), old))
|
2016-01-11 21:05:33 -08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Recursively assign the new value we've made to the base of the dot op.
|
2016-05-27 14:07:37 -07:00
|
|
|
s.assign(left.Left, new, false, false, line, 0, rightIsVolatile)
|
2016-01-11 21:05:33 -08:00
|
|
|
// TODO: do we need to update named values here?
|
|
|
|
return
|
|
|
|
}
|
2015-06-12 14:23:29 +01:00
|
|
|
// Update variable assignment.
|
2015-08-29 14:54:45 -07:00
|
|
|
s.vars[left] = right
|
2015-10-22 14:22:38 -07:00
|
|
|
s.addNamedValue(left, right)
|
2015-06-12 14:23:29 +01:00
|
|
|
return
|
|
|
|
}
|
2016-03-01 23:21:55 +00:00
|
|
|
// Left is not ssa-able. Compute its address.
|
2016-05-27 14:07:37 -07:00
|
|
|
addr, _ := s.addr(left, false)
|
2016-04-29 12:09:32 -07:00
|
|
|
if left.Op == ONAME && skip == 0 {
|
2015-09-17 16:45:10 -07:00
|
|
|
s.vars[&memVar] = s.newValue1A(ssa.OpVarDef, ssa.TypeMem, left, s.mem())
|
2015-08-24 02:16:19 -07:00
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
if deref {
|
|
|
|
// Treat as a mem->mem move.
|
|
|
|
if right == nil {
|
2016-06-27 16:54:57 -04:00
|
|
|
s.vars[&memVar] = s.newValue2I(ssa.OpZero, ssa.TypeMem, SizeAlignAuxInt(t), addr, s.mem())
|
2016-01-25 17:06:54 -08:00
|
|
|
return
|
|
|
|
}
|
|
|
|
if wb {
|
2016-05-27 14:07:37 -07:00
|
|
|
s.insertWBmove(t, addr, right, line, rightIsVolatile)
|
2016-01-25 17:06:54 -08:00
|
|
|
return
|
|
|
|
}
|
2016-06-27 16:54:57 -04:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpMove, ssa.TypeMem, SizeAlignAuxInt(t), addr, right, s.mem())
|
2016-01-25 17:06:54 -08:00
|
|
|
return
|
|
|
|
}
|
|
|
|
// Treat as a store.
|
2015-09-07 23:18:02 -07:00
|
|
|
if wb {
|
2016-03-21 10:22:03 -07:00
|
|
|
if skip&skipPtr != 0 {
|
|
|
|
// Special case: if we don't write back the pointers, don't bother
|
|
|
|
// doing the write barrier check.
|
|
|
|
s.storeTypeScalars(t, addr, right, skip)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
s.insertWBstore(t, addr, right, line, skip)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
if skip != 0 {
|
|
|
|
if skip&skipPtr == 0 {
|
|
|
|
s.storeTypePtrs(t, addr, right)
|
|
|
|
}
|
|
|
|
s.storeTypeScalars(t, addr, right, skip)
|
2016-01-25 17:06:54 -08:00
|
|
|
return
|
2015-09-07 23:18:02 -07:00
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, t.Size(), addr, right, s.mem())
|
2015-06-12 14:23:29 +01:00
|
|
|
}
|
|
|
|
|
2015-07-20 15:30:52 -07:00
|
|
|
// zeroVal returns the zero value for type t.
|
|
|
|
func (s *state) zeroVal(t *Type) *ssa.Value {
|
|
|
|
switch {
|
2015-07-28 14:19:20 -07:00
|
|
|
case t.IsInteger():
|
|
|
|
switch t.Size() {
|
|
|
|
case 1:
|
|
|
|
return s.constInt8(t, 0)
|
|
|
|
case 2:
|
|
|
|
return s.constInt16(t, 0)
|
|
|
|
case 4:
|
|
|
|
return s.constInt32(t, 0)
|
|
|
|
case 8:
|
|
|
|
return s.constInt64(t, 0)
|
|
|
|
default:
|
|
|
|
s.Fatalf("bad sized integer type %s", t)
|
|
|
|
}
|
2015-08-25 19:21:45 -05:00
|
|
|
case t.IsFloat():
|
|
|
|
switch t.Size() {
|
|
|
|
case 4:
|
|
|
|
return s.constFloat32(t, 0)
|
|
|
|
case 8:
|
|
|
|
return s.constFloat64(t, 0)
|
|
|
|
default:
|
|
|
|
s.Fatalf("bad sized float type %s", t)
|
|
|
|
}
|
2015-08-28 14:24:10 -04:00
|
|
|
case t.IsComplex():
|
|
|
|
switch t.Size() {
|
|
|
|
case 8:
|
|
|
|
z := s.constFloat32(Types[TFLOAT32], 0)
|
2015-08-28 13:52:26 -07:00
|
|
|
return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
|
2015-08-28 14:24:10 -04:00
|
|
|
case 16:
|
|
|
|
z := s.constFloat64(Types[TFLOAT64], 0)
|
2015-08-28 13:52:26 -07:00
|
|
|
return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
|
2015-08-28 14:24:10 -04:00
|
|
|
default:
|
|
|
|
s.Fatalf("bad sized complex type %s", t)
|
|
|
|
}
|
|
|
|
|
2015-07-20 15:30:52 -07:00
|
|
|
case t.IsString():
|
2016-03-06 18:06:09 -08:00
|
|
|
return s.constEmptyString(t)
|
2016-03-28 10:55:44 -07:00
|
|
|
case t.IsPtrShaped():
|
2016-03-06 18:06:09 -08:00
|
|
|
return s.constNil(t)
|
2015-07-20 15:30:52 -07:00
|
|
|
case t.IsBoolean():
|
2015-09-08 16:52:25 -07:00
|
|
|
return s.constBool(false)
|
2015-08-18 10:26:28 -07:00
|
|
|
case t.IsInterface():
|
2016-03-06 18:06:09 -08:00
|
|
|
return s.constInterface(t)
|
2015-08-18 10:26:28 -07:00
|
|
|
case t.IsSlice():
|
2016-03-06 18:06:09 -08:00
|
|
|
return s.constSlice(t)
|
2016-01-11 21:05:33 -08:00
|
|
|
case t.IsStruct():
|
|
|
|
n := t.NumFields()
|
|
|
|
v := s.entryNewValue0(ssa.StructMakeOp(t.NumFields()), t)
|
2016-03-14 12:45:18 -07:00
|
|
|
for i := 0; i < n; i++ {
|
2016-01-11 21:05:33 -08:00
|
|
|
v.AddArg(s.zeroVal(t.FieldType(i).(*Type)))
|
|
|
|
}
|
|
|
|
return v
|
2015-07-20 15:30:52 -07:00
|
|
|
}
|
|
|
|
s.Unimplementedf("zero for type %v not implemented", t)
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
2015-09-09 23:56:59 -07:00
|
|
|
type callKind int8
|
|
|
|
|
|
|
|
const (
|
|
|
|
callNormal callKind = iota
|
|
|
|
callDefer
|
|
|
|
callGo
|
|
|
|
)
|
|
|
|
|
2016-08-23 16:49:28 -07:00
|
|
|
// isSSAIntrinsic returns true if n is a call to a recognized intrinsic
|
2016-03-11 00:10:52 -05:00
|
|
|
// that can be handled by the SSA backend.
|
|
|
|
// SSA uses this, but so does the front end to see if should not
|
|
|
|
// inline a function because it is a candidate for intrinsic
|
|
|
|
// substitution.
|
2016-08-23 16:49:28 -07:00
|
|
|
func isSSAIntrinsic(s *Sym) bool {
|
2016-03-11 00:10:52 -05:00
|
|
|
// The test below is not quite accurate -- in the event that
|
|
|
|
// a function is disabled on a per-function basis, for example
|
|
|
|
// because of hash-keyed binary failure search, SSA might be
|
|
|
|
// disabled for that function but it would not be noted here,
|
|
|
|
// and thus an inlining would not occur (in practice, inlining
|
|
|
|
// so far has only been noticed for Bswap32 and the 16-bit count
|
|
|
|
// leading/trailing instructions, but heuristics might change
|
|
|
|
// in the future or on different architectures).
|
2016-04-06 12:01:40 -07:00
|
|
|
if !ssaEnabled || ssa.IntrinsicsDisable || Thearch.LinkArch.Family != sys.AMD64 {
|
2016-03-11 00:10:52 -05:00
|
|
|
return false
|
|
|
|
}
|
|
|
|
if s != nil && s.Pkg != nil && s.Pkg.Path == "runtime/internal/sys" {
|
|
|
|
switch s.Name {
|
|
|
|
case
|
2016-08-23 10:43:47 -07:00
|
|
|
"Ctz64", "Ctz32",
|
2016-03-11 00:10:52 -05:00
|
|
|
"Bswap64", "Bswap32":
|
|
|
|
return true
|
|
|
|
}
|
|
|
|
}
|
2016-08-23 16:49:28 -07:00
|
|
|
if s != nil && s.Pkg != nil && s.Pkg.Path == "runtime/internal/atomic" {
|
|
|
|
switch s.Name {
|
|
|
|
case "Load", "Load64", "Loadint64", "Loadp", "Loaduint", "Loaduintptr":
|
|
|
|
return true
|
|
|
|
case "Store", "Store64", "StorepNoWB", "Storeuintptr":
|
|
|
|
return true
|
2016-08-25 16:02:57 -07:00
|
|
|
case "Xchg", "Xchg64", "Xchguintptr":
|
|
|
|
return true
|
|
|
|
case "Xadd", "Xadd64", "Xaddint64", "Xadduintptr":
|
|
|
|
return true
|
|
|
|
case "Cas", "Cas64", "Casp1", "Casuintptr":
|
|
|
|
return true
|
|
|
|
case "And8", "Or8":
|
|
|
|
return true
|
2016-08-23 16:49:28 -07:00
|
|
|
}
|
|
|
|
}
|
2016-03-11 00:10:52 -05:00
|
|
|
return false
|
|
|
|
}
|
|
|
|
|
2016-08-23 16:49:28 -07:00
|
|
|
func isIntrinsicCall(n *Node) bool {
|
2016-03-11 00:10:52 -05:00
|
|
|
if n == nil || n.Left == nil {
|
|
|
|
return false
|
|
|
|
}
|
2016-08-23 16:49:28 -07:00
|
|
|
return isSSAIntrinsic(n.Left.Sym)
|
2016-03-11 00:10:52 -05:00
|
|
|
}
|
|
|
|
|
2016-08-23 16:49:28 -07:00
|
|
|
// intrinsicArg extracts the ith arg from n.List and returns its value.
|
|
|
|
func (s *state) intrinsicArg(n *Node, i int) *ssa.Value {
|
|
|
|
x := n.List.Slice()[i]
|
2016-03-11 00:10:52 -05:00
|
|
|
if x.Op == OAS {
|
|
|
|
x = x.Right
|
|
|
|
}
|
|
|
|
return s.expr(x)
|
|
|
|
}
|
2016-08-23 16:49:28 -07:00
|
|
|
func (s *state) intrinsicFirstArg(n *Node) *ssa.Value {
|
|
|
|
return s.intrinsicArg(n, 0)
|
|
|
|
}
|
2016-03-11 00:10:52 -05:00
|
|
|
|
2016-08-23 16:49:28 -07:00
|
|
|
// intrinsicCall converts a call to a recognized intrinsic function into the intrinsic SSA operation.
|
|
|
|
func (s *state) intrinsicCall(n *Node) (ret *ssa.Value) {
|
2016-03-11 00:10:52 -05:00
|
|
|
var result *ssa.Value
|
2016-08-23 16:49:28 -07:00
|
|
|
name := n.Left.Sym.Name
|
|
|
|
switch {
|
|
|
|
case name == "Ctz64":
|
2016-03-11 00:10:52 -05:00
|
|
|
result = s.newValue1(ssa.OpCtz64, Types[TUINT64], s.intrinsicFirstArg(n))
|
2016-08-23 16:49:28 -07:00
|
|
|
ret = result
|
|
|
|
case name == "Ctz32":
|
2016-03-11 00:10:52 -05:00
|
|
|
result = s.newValue1(ssa.OpCtz32, Types[TUINT32], s.intrinsicFirstArg(n))
|
2016-08-23 16:49:28 -07:00
|
|
|
ret = result
|
|
|
|
case name == "Bswap64":
|
2016-03-11 00:10:52 -05:00
|
|
|
result = s.newValue1(ssa.OpBswap64, Types[TUINT64], s.intrinsicFirstArg(n))
|
2016-08-23 16:49:28 -07:00
|
|
|
ret = result
|
|
|
|
case name == "Bswap32":
|
2016-03-11 00:10:52 -05:00
|
|
|
result = s.newValue1(ssa.OpBswap32, Types[TUINT32], s.intrinsicFirstArg(n))
|
2016-08-23 16:49:28 -07:00
|
|
|
ret = result
|
|
|
|
case name == "Load" || name == "Loaduint" && s.config.IntSize == 4 || name == "Loaduintptr" && s.config.PtrSize == 4:
|
|
|
|
result = s.newValue2(ssa.OpAtomicLoad32, ssa.MakeTuple(Types[TUINT32], ssa.TypeMem), s.intrinsicArg(n, 0), s.mem())
|
|
|
|
s.vars[&memVar] = s.newValue1(ssa.OpSelect1, ssa.TypeMem, result)
|
|
|
|
ret = s.newValue1(ssa.OpSelect0, Types[TUINT32], result)
|
|
|
|
case name == "Load64" || name == "Loadint64" || name == "Loaduint" && s.config.IntSize == 8 || name == "Loaduintptr" && s.config.PtrSize == 8:
|
|
|
|
result = s.newValue2(ssa.OpAtomicLoad64, ssa.MakeTuple(Types[TUINT64], ssa.TypeMem), s.intrinsicArg(n, 0), s.mem())
|
|
|
|
s.vars[&memVar] = s.newValue1(ssa.OpSelect1, ssa.TypeMem, result)
|
|
|
|
ret = s.newValue1(ssa.OpSelect0, Types[TUINT64], result)
|
|
|
|
case name == "Loadp":
|
|
|
|
result = s.newValue2(ssa.OpAtomicLoadPtr, ssa.MakeTuple(Ptrto(Types[TUINT8]), ssa.TypeMem), s.intrinsicArg(n, 0), s.mem())
|
|
|
|
s.vars[&memVar] = s.newValue1(ssa.OpSelect1, ssa.TypeMem, result)
|
|
|
|
ret = s.newValue1(ssa.OpSelect0, Ptrto(Types[TUINT8]), result)
|
|
|
|
case name == "Store" || name == "Storeuintptr" && s.config.PtrSize == 4:
|
|
|
|
result = s.newValue3(ssa.OpAtomicStore32, ssa.TypeMem, s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.mem())
|
|
|
|
s.vars[&memVar] = result
|
|
|
|
case name == "Store64" || name == "Storeuintptr" && s.config.PtrSize == 8:
|
|
|
|
result = s.newValue3(ssa.OpAtomicStore64, ssa.TypeMem, s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.mem())
|
|
|
|
s.vars[&memVar] = result
|
|
|
|
case name == "StorepNoWB":
|
|
|
|
result = s.newValue3(ssa.OpAtomicStorePtrNoWB, ssa.TypeMem, s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.mem())
|
|
|
|
s.vars[&memVar] = result
|
2016-08-25 16:02:57 -07:00
|
|
|
case name == "Xchg" || name == "Xchguintptr" && s.config.PtrSize == 4:
|
|
|
|
result = s.newValue3(ssa.OpAtomicExchange32, ssa.MakeTuple(Types[TUINT32], ssa.TypeMem), s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.mem())
|
|
|
|
s.vars[&memVar] = s.newValue1(ssa.OpSelect1, ssa.TypeMem, result)
|
|
|
|
ret = s.newValue1(ssa.OpSelect0, Types[TUINT32], result)
|
|
|
|
case name == "Xchg64" || name == "Xchguintptr" && s.config.PtrSize == 8:
|
|
|
|
result = s.newValue3(ssa.OpAtomicExchange64, ssa.MakeTuple(Types[TUINT64], ssa.TypeMem), s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.mem())
|
|
|
|
s.vars[&memVar] = s.newValue1(ssa.OpSelect1, ssa.TypeMem, result)
|
|
|
|
ret = s.newValue1(ssa.OpSelect0, Types[TUINT64], result)
|
|
|
|
case name == "Xadd" || name == "Xadduintptr" && s.config.PtrSize == 4:
|
|
|
|
result = s.newValue3(ssa.OpAtomicAdd32, ssa.MakeTuple(Types[TUINT32], ssa.TypeMem), s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.mem())
|
|
|
|
s.vars[&memVar] = s.newValue1(ssa.OpSelect1, ssa.TypeMem, result)
|
|
|
|
ret = s.newValue1(ssa.OpSelect0, Types[TUINT32], result)
|
|
|
|
case name == "Xadd64" || name == "Xaddint64" || name == "Xadduintptr" && s.config.PtrSize == 8:
|
|
|
|
result = s.newValue3(ssa.OpAtomicAdd64, ssa.MakeTuple(Types[TUINT64], ssa.TypeMem), s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.mem())
|
|
|
|
s.vars[&memVar] = s.newValue1(ssa.OpSelect1, ssa.TypeMem, result)
|
|
|
|
ret = s.newValue1(ssa.OpSelect0, Types[TUINT64], result)
|
|
|
|
case name == "Cas" || (name == "Casp1" || name == "Casuintptr") && s.config.PtrSize == 4:
|
|
|
|
result = s.newValue4(ssa.OpAtomicCompareAndSwap32, ssa.MakeTuple(Types[TBOOL], ssa.TypeMem), s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.intrinsicArg(n, 2), s.mem())
|
|
|
|
s.vars[&memVar] = s.newValue1(ssa.OpSelect1, ssa.TypeMem, result)
|
|
|
|
ret = s.newValue1(ssa.OpSelect0, Types[TBOOL], result)
|
|
|
|
case name == "Cas64" || (name == "Casp1" || name == "Casuintptr") && s.config.PtrSize == 8:
|
|
|
|
result = s.newValue4(ssa.OpAtomicCompareAndSwap64, ssa.MakeTuple(Types[TBOOL], ssa.TypeMem), s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.intrinsicArg(n, 2), s.mem())
|
|
|
|
s.vars[&memVar] = s.newValue1(ssa.OpSelect1, ssa.TypeMem, result)
|
|
|
|
ret = s.newValue1(ssa.OpSelect0, Types[TBOOL], result)
|
|
|
|
case name == "And8":
|
|
|
|
result = s.newValue3(ssa.OpAtomicAnd8, ssa.TypeMem, s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.mem())
|
|
|
|
s.vars[&memVar] = result
|
|
|
|
case name == "Or8":
|
|
|
|
result = s.newValue3(ssa.OpAtomicOr8, ssa.TypeMem, s.intrinsicArg(n, 0), s.intrinsicArg(n, 1), s.mem())
|
|
|
|
s.vars[&memVar] = result
|
2016-03-11 00:10:52 -05:00
|
|
|
}
|
|
|
|
if result == nil {
|
|
|
|
Fatalf("Unknown special call: %v", n.Left.Sym)
|
|
|
|
}
|
|
|
|
if ssa.IntrinsicsDebug > 0 {
|
|
|
|
Warnl(n.Lineno, "intrinsic substitution for %v with %s", n.Left.Sym.Name, result.LongString())
|
|
|
|
}
|
2016-08-23 16:49:28 -07:00
|
|
|
return
|
2016-03-11 00:10:52 -05:00
|
|
|
}
|
|
|
|
|
2016-01-25 17:06:54 -08:00
|
|
|
// Calls the function n using the specified call type.
|
|
|
|
// Returns the address of the return value (or nil if none).
|
2015-09-09 23:56:59 -07:00
|
|
|
func (s *state) call(n *Node, k callKind) *ssa.Value {
|
|
|
|
var sym *Sym // target symbol (if static)
|
|
|
|
var closure *ssa.Value // ptr to closure to run (if dynamic)
|
|
|
|
var codeptr *ssa.Value // ptr to target code (if dynamic)
|
|
|
|
var rcvr *ssa.Value // receiver to set
|
|
|
|
fn := n.Left
|
|
|
|
switch n.Op {
|
|
|
|
case OCALLFUNC:
|
|
|
|
if k == callNormal && fn.Op == ONAME && fn.Class == PFUNC {
|
|
|
|
sym = fn.Sym
|
|
|
|
break
|
|
|
|
}
|
|
|
|
closure = s.expr(fn)
|
|
|
|
case OCALLMETH:
|
|
|
|
if fn.Op != ODOTMETH {
|
|
|
|
Fatalf("OCALLMETH: n.Left not an ODOTMETH: %v", fn)
|
|
|
|
}
|
|
|
|
if k == callNormal {
|
cmd/compile: change ODOT and friends to use Sym, not Right
The Node type ODOT and its variants all represent a selector, with a
simple name to the right of the dot. Before this change this was
represented by using an ONAME Node in the Right field. This ONAME node
served no useful purpose. This CL changes these Node types to store the
symbol in the Sym field instead, thus not requiring allocating a Node
for each selector.
When compiling x/tools/go/types this CL eliminates nearly 5000 calls to
newname and reduces the total number of Nodes allocated by about 6.6%.
It seems to cut compilation time by 1 to 2 percent.
Getting this right was somewhat subtle, and I added two dubious changes
to produce the exact same output as before. One is to ishairy in
inl.go: the ONAME node increased the cost of ODOT and friends by 1, and
I retained that, although really ODOT is not more expensive than any
other node. The other is to varexpr in walk.go: because the ONAME in
the Right field of an ODOT has no class, varexpr would always return
false for an ODOT, although in fact for some ODOT's it seemingly ought
to return true; I added an && false for now. I will send separate CLs,
that will break toolstash -cmp, to clean these up.
This CL passes toolstash -cmp.
Change-Id: I4af8a10cc59078c436130ce472f25abc3a9b2f80
Reviewed-on: https://go-review.googlesource.com/20890
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
2016-03-18 16:52:30 -07:00
|
|
|
sym = fn.Sym
|
2015-09-09 23:56:59 -07:00
|
|
|
break
|
|
|
|
}
|
cmd/compile: change ODOT and friends to use Sym, not Right
The Node type ODOT and its variants all represent a selector, with a
simple name to the right of the dot. Before this change this was
represented by using an ONAME Node in the Right field. This ONAME node
served no useful purpose. This CL changes these Node types to store the
symbol in the Sym field instead, thus not requiring allocating a Node
for each selector.
When compiling x/tools/go/types this CL eliminates nearly 5000 calls to
newname and reduces the total number of Nodes allocated by about 6.6%.
It seems to cut compilation time by 1 to 2 percent.
Getting this right was somewhat subtle, and I added two dubious changes
to produce the exact same output as before. One is to ishairy in
inl.go: the ONAME node increased the cost of ODOT and friends by 1, and
I retained that, although really ODOT is not more expensive than any
other node. The other is to varexpr in walk.go: because the ONAME in
the Right field of an ODOT has no class, varexpr would always return
false for an ODOT, although in fact for some ODOT's it seemingly ought
to return true; I added an && false for now. I will send separate CLs,
that will break toolstash -cmp, to clean these up.
This CL passes toolstash -cmp.
Change-Id: I4af8a10cc59078c436130ce472f25abc3a9b2f80
Reviewed-on: https://go-review.googlesource.com/20890
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
2016-03-18 16:52:30 -07:00
|
|
|
n2 := newname(fn.Sym)
|
2015-09-09 23:56:59 -07:00
|
|
|
n2.Class = PFUNC
|
cmd/compile: change ODOT and friends to use Sym, not Right
The Node type ODOT and its variants all represent a selector, with a
simple name to the right of the dot. Before this change this was
represented by using an ONAME Node in the Right field. This ONAME node
served no useful purpose. This CL changes these Node types to store the
symbol in the Sym field instead, thus not requiring allocating a Node
for each selector.
When compiling x/tools/go/types this CL eliminates nearly 5000 calls to
newname and reduces the total number of Nodes allocated by about 6.6%.
It seems to cut compilation time by 1 to 2 percent.
Getting this right was somewhat subtle, and I added two dubious changes
to produce the exact same output as before. One is to ishairy in
inl.go: the ONAME node increased the cost of ODOT and friends by 1, and
I retained that, although really ODOT is not more expensive than any
other node. The other is to varexpr in walk.go: because the ONAME in
the Right field of an ODOT has no class, varexpr would always return
false for an ODOT, although in fact for some ODOT's it seemingly ought
to return true; I added an && false for now. I will send separate CLs,
that will break toolstash -cmp, to clean these up.
This CL passes toolstash -cmp.
Change-Id: I4af8a10cc59078c436130ce472f25abc3a9b2f80
Reviewed-on: https://go-review.googlesource.com/20890
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
2016-03-18 16:52:30 -07:00
|
|
|
n2.Lineno = fn.Lineno
|
|
|
|
closure = s.expr(n2)
|
2015-09-09 23:56:59 -07:00
|
|
|
// Note: receiver is already assigned in n.List, so we don't
|
|
|
|
// want to set it here.
|
|
|
|
case OCALLINTER:
|
|
|
|
if fn.Op != ODOTINTER {
|
2016-04-27 15:10:10 +10:00
|
|
|
Fatalf("OCALLINTER: n.Left not an ODOTINTER: %v", fn.Op)
|
2015-09-09 23:56:59 -07:00
|
|
|
}
|
|
|
|
i := s.expr(fn.Left)
|
|
|
|
itab := s.newValue1(ssa.OpITab, Types[TUINTPTR], i)
|
2016-06-06 16:00:33 -04:00
|
|
|
if k != callNormal {
|
|
|
|
s.nilCheck(itab)
|
|
|
|
}
|
2015-09-09 23:56:59 -07:00
|
|
|
itabidx := fn.Xoffset + 3*int64(Widthptr) + 8 // offset of fun field in runtime.itab
|
2016-06-08 10:12:30 -04:00
|
|
|
itab = s.newValue1I(ssa.OpOffPtr, Ptrto(Types[TUINTPTR]), itabidx, itab)
|
2015-09-09 23:56:59 -07:00
|
|
|
if k == callNormal {
|
|
|
|
codeptr = s.newValue2(ssa.OpLoad, Types[TUINTPTR], itab, s.mem())
|
|
|
|
} else {
|
|
|
|
closure = itab
|
|
|
|
}
|
|
|
|
rcvr = s.newValue1(ssa.OpIData, Types[TUINTPTR], i)
|
|
|
|
}
|
|
|
|
dowidth(fn.Type)
|
2016-03-28 14:31:57 -07:00
|
|
|
stksize := fn.Type.ArgWidth() // includes receiver
|
2015-09-09 23:56:59 -07:00
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// Run all argument assignments. The arg slots have already
|
2015-09-09 23:56:59 -07:00
|
|
|
// been offset by the appropriate amount (+2*widthptr for go/defer,
|
|
|
|
// +widthptr for interface calls).
|
|
|
|
// For OCALLMETH, the receiver is set in these statements.
|
|
|
|
s.stmtList(n.List)
|
|
|
|
|
|
|
|
// Set receiver (for interface calls)
|
|
|
|
if rcvr != nil {
|
2015-10-19 13:56:55 -07:00
|
|
|
argStart := Ctxt.FixedFrameSize()
|
2015-09-09 23:56:59 -07:00
|
|
|
if k != callNormal {
|
|
|
|
argStart += int64(2 * Widthptr)
|
|
|
|
}
|
2016-06-08 10:12:30 -04:00
|
|
|
addr := s.entryNewValue1I(ssa.OpOffPtr, Ptrto(Types[TUINTPTR]), argStart, s.sp)
|
2015-09-17 16:45:10 -07:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, int64(Widthptr), addr, rcvr, s.mem())
|
2015-09-09 23:56:59 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
// Defer/go args
|
|
|
|
if k != callNormal {
|
|
|
|
// Write argsize and closure (args to Newproc/Deferproc).
|
2016-05-15 00:12:56 -04:00
|
|
|
argStart := Ctxt.FixedFrameSize()
|
2015-09-09 23:56:59 -07:00
|
|
|
argsize := s.constInt32(Types[TUINT32], int32(stksize))
|
2016-06-08 10:12:30 -04:00
|
|
|
addr := s.entryNewValue1I(ssa.OpOffPtr, Ptrto(Types[TUINT32]), argStart, s.sp)
|
2016-05-15 00:12:56 -04:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, 4, addr, argsize, s.mem())
|
|
|
|
addr = s.entryNewValue1I(ssa.OpOffPtr, Ptrto(Types[TUINTPTR]), argStart+int64(Widthptr), s.sp)
|
2015-09-17 16:45:10 -07:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, int64(Widthptr), addr, closure, s.mem())
|
2015-09-09 23:56:59 -07:00
|
|
|
stksize += 2 * int64(Widthptr)
|
|
|
|
}
|
|
|
|
|
|
|
|
// call target
|
|
|
|
bNext := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
var call *ssa.Value
|
|
|
|
switch {
|
|
|
|
case k == callDefer:
|
|
|
|
call = s.newValue1(ssa.OpDeferCall, ssa.TypeMem, s.mem())
|
|
|
|
case k == callGo:
|
|
|
|
call = s.newValue1(ssa.OpGoCall, ssa.TypeMem, s.mem())
|
|
|
|
case closure != nil:
|
|
|
|
codeptr = s.newValue2(ssa.OpLoad, Types[TUINTPTR], closure, s.mem())
|
|
|
|
call = s.newValue3(ssa.OpClosureCall, ssa.TypeMem, codeptr, closure, s.mem())
|
|
|
|
case codeptr != nil:
|
|
|
|
call = s.newValue2(ssa.OpInterCall, ssa.TypeMem, codeptr, s.mem())
|
|
|
|
case sym != nil:
|
|
|
|
call = s.newValue1A(ssa.OpStaticCall, ssa.TypeMem, sym, s.mem())
|
|
|
|
default:
|
2016-04-22 07:14:10 -07:00
|
|
|
Fatalf("bad call type %s %v", n.Op, n)
|
2015-09-09 23:56:59 -07:00
|
|
|
}
|
|
|
|
call.AuxInt = stksize // Call operations carry the argsize of the callee along with them
|
|
|
|
|
|
|
|
// Finish call block
|
2015-09-17 16:45:10 -07:00
|
|
|
s.vars[&memVar] = call
|
2015-09-09 23:56:59 -07:00
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockCall
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(call)
|
2015-09-09 23:56:59 -07:00
|
|
|
b.AddEdgeTo(bNext)
|
2016-03-09 19:27:57 -08:00
|
|
|
if k == callDefer {
|
|
|
|
// Add recover edge to exit code.
|
|
|
|
b.Kind = ssa.BlockDefer
|
|
|
|
r := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
s.startBlock(r)
|
|
|
|
s.exit()
|
|
|
|
b.AddEdgeTo(r)
|
|
|
|
b.Likely = ssa.BranchLikely
|
|
|
|
}
|
2015-09-09 23:56:59 -07:00
|
|
|
|
2016-01-25 17:06:54 -08:00
|
|
|
// Start exit block, find address of result.
|
2015-09-09 23:56:59 -07:00
|
|
|
s.startBlock(bNext)
|
2016-04-21 19:28:28 -07:00
|
|
|
// Keep input pointer args live across calls. This is a bandaid until 1.8.
|
|
|
|
for _, n := range s.ptrargs {
|
|
|
|
s.vars[&memVar] = s.newValue2(ssa.OpKeepAlive, ssa.TypeMem, s.variable(n, n.Type), s.mem())
|
|
|
|
}
|
2016-03-15 11:06:03 -07:00
|
|
|
res := n.Left.Type.Results()
|
|
|
|
if res.NumFields() == 0 || k != callNormal {
|
2015-09-09 23:56:59 -07:00
|
|
|
// call has no return value. Continue with the next statement.
|
|
|
|
return nil
|
|
|
|
}
|
2016-03-15 11:06:03 -07:00
|
|
|
fp := res.Field(0)
|
2016-04-18 10:30:20 -04:00
|
|
|
return s.entryNewValue1I(ssa.OpOffPtr, Ptrto(fp.Type), fp.Offset+Ctxt.FixedFrameSize(), s.sp)
|
2015-09-09 23:56:59 -07:00
|
|
|
}
|
|
|
|
|
2015-07-28 14:31:25 -07:00
|
|
|
// etypesign returns the signed-ness of e, for integer/pointer etypes.
|
|
|
|
// -1 means signed, +1 means unsigned, 0 means non-integer/non-pointer.
|
2015-11-16 13:20:16 -08:00
|
|
|
func etypesign(e EType) int8 {
|
2015-07-28 14:31:25 -07:00
|
|
|
switch e {
|
|
|
|
case TINT8, TINT16, TINT32, TINT64, TINT:
|
|
|
|
return -1
|
|
|
|
case TUINT8, TUINT16, TUINT32, TUINT64, TUINT, TUINTPTR, TUNSAFEPTR:
|
|
|
|
return +1
|
|
|
|
}
|
|
|
|
return 0
|
|
|
|
}
|
|
|
|
|
2015-10-15 20:25:32 -05:00
|
|
|
// lookupSymbol is used to retrieve the symbol (Extern, Arg or Auto) used for a particular node.
|
|
|
|
// This improves the effectiveness of cse by using the same Aux values for the
|
|
|
|
// same symbols.
|
|
|
|
func (s *state) lookupSymbol(n *Node, sym interface{}) interface{} {
|
|
|
|
switch sym.(type) {
|
|
|
|
default:
|
|
|
|
s.Fatalf("sym %v is of uknown type %T", sym, sym)
|
|
|
|
case *ssa.ExternSymbol, *ssa.ArgSymbol, *ssa.AutoSymbol:
|
|
|
|
// these are the only valid types
|
|
|
|
}
|
|
|
|
|
|
|
|
if lsym, ok := s.varsyms[n]; ok {
|
|
|
|
return lsym
|
|
|
|
} else {
|
|
|
|
s.varsyms[n] = sym
|
|
|
|
return sym
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-06-02 09:16:22 -07:00
|
|
|
// addr converts the address of the expression n to SSA, adds it to s and returns the SSA result.
|
2016-05-27 14:07:37 -07:00
|
|
|
// Also returns a bool reporting whether the returned value is "volatile", that is it
|
|
|
|
// points to the outargs section and thus the referent will be clobbered by any call.
|
2015-07-13 15:55:37 -07:00
|
|
|
// The value that the returned Value represents is guaranteed to be non-nil.
|
2015-10-09 16:48:30 -04:00
|
|
|
// If bounded is true then this address does not require a nil check for its operand
|
|
|
|
// even if that would otherwise be implied.
|
2016-05-27 14:07:37 -07:00
|
|
|
func (s *state) addr(n *Node, bounded bool) (*ssa.Value, bool) {
|
2015-10-22 14:22:38 -07:00
|
|
|
t := Ptrto(n.Type)
|
2015-05-18 16:44:20 -07:00
|
|
|
switch n.Op {
|
|
|
|
case ONAME:
|
2015-06-10 15:03:06 -07:00
|
|
|
switch n.Class {
|
|
|
|
case PEXTERN:
|
2015-05-18 16:44:20 -07:00
|
|
|
// global variable
|
2016-04-29 09:02:27 -07:00
|
|
|
aux := s.lookupSymbol(n, &ssa.ExternSymbol{Typ: n.Type, Sym: n.Sym})
|
2015-10-22 14:22:38 -07:00
|
|
|
v := s.entryNewValue1A(ssa.OpAddr, t, aux, s.sb)
|
2015-07-28 11:08:44 -07:00
|
|
|
// TODO: Make OpAddr use AuxInt as well as Aux.
|
|
|
|
if n.Xoffset != 0 {
|
|
|
|
v = s.entryNewValue1I(ssa.OpOffPtr, v.Type, n.Xoffset, v)
|
|
|
|
}
|
2016-05-27 14:07:37 -07:00
|
|
|
return v, false
|
2015-09-11 16:40:05 -04:00
|
|
|
case PPARAM:
|
|
|
|
// parameter slot
|
2015-06-29 11:56:28 -07:00
|
|
|
v := s.decladdrs[n]
|
2015-11-16 13:20:16 -08:00
|
|
|
if v != nil {
|
2016-05-27 14:07:37 -07:00
|
|
|
return v, false
|
2015-06-29 11:56:28 -07:00
|
|
|
}
|
2015-11-16 13:20:16 -08:00
|
|
|
if n.String() == ".fp" {
|
|
|
|
// Special arg that points to the frame pointer.
|
|
|
|
// (Used by the race detector, others?)
|
|
|
|
aux := s.lookupSymbol(n, &ssa.ArgSymbol{Typ: n.Type, Node: n})
|
2016-05-27 14:07:37 -07:00
|
|
|
return s.entryNewValue1A(ssa.OpAddr, t, aux, s.sp), false
|
2015-11-16 13:20:16 -08:00
|
|
|
}
|
|
|
|
s.Fatalf("addr of undeclared ONAME %v. declared: %v", n, s.decladdrs)
|
2016-05-27 14:07:37 -07:00
|
|
|
return nil, false
|
2015-08-24 02:16:19 -07:00
|
|
|
case PAUTO:
|
2016-03-11 20:03:17 -06:00
|
|
|
aux := s.lookupSymbol(n, &ssa.AutoSymbol{Typ: n.Type, Node: n})
|
2016-05-27 14:07:37 -07:00
|
|
|
return s.newValue1A(ssa.OpAddr, t, aux, s.sp), false
|
2015-09-11 16:40:05 -04:00
|
|
|
case PPARAMOUT: // Same as PAUTO -- cannot generate LEA early.
|
2015-10-15 20:25:32 -05:00
|
|
|
// ensure that we reuse symbols for out parameters so
|
|
|
|
// that cse works on their addresses
|
|
|
|
aux := s.lookupSymbol(n, &ssa.ArgSymbol{Typ: n.Type, Node: n})
|
2016-05-27 14:07:37 -07:00
|
|
|
return s.newValue1A(ssa.OpAddr, t, aux, s.sp), false
|
2015-06-10 15:03:06 -07:00
|
|
|
default:
|
cmd/compile: fix liveness computation for heap-escaped parameters
The liveness computation of parameters generally was never
correct, but forcing all parameters to be live throughout the
function covered up that problem. The new SSA back end is
too clever: even though it currently keeps the parameter values live
throughout the function, it may find optimizations that mean
the current values are not written back to the original parameter
stack slots immediately or ever (for example if a parameter is set
to nil, SSA constant propagation may replace all later uses of the
parameter with a constant nil, eliminating the need to write the nil
value back to the stack slot), so the liveness code must now
track the actual operations on the stack slots, exposing these
problems.
One small problem in the handling of arguments is that nodarg
can return ONAME PPARAM nodes with adjusted offsets, so that
there are actually multiple *Node pointers for the same parameter
in the instruction stream. This might be possible to correct, but
not in this CL. For now, we fix this by using n.Orig instead of n
when considering PPARAM and PPARAMOUT nodes.
The major problem in the handling of arguments is general
confusion in the liveness code about the meaning of PPARAM|PHEAP
and PPARAMOUT|PHEAP nodes, especially as contrasted with PAUTO|PHEAP.
The difference between these two is that when a local variable "moves"
to the heap, it's really just allocated there to start with; in contrast,
when an argument moves to the heap, the actual data has to be copied
there from the stack at the beginning of the function, and when a
result "moves" to the heap the value in the heap has to be copied
back to the stack when the function returns
This general confusion is also present in the SSA back end.
The PHEAP bit worked decently when I first introduced it 7 years ago (!)
in 391425ae. The back end did nothing sophisticated, and in particular
there was no analysis at all: no escape analysis, no liveness analysis,
and certainly no SSA back end. But the complications caused in the
various downstream consumers suggest that this should be a detail
kept mainly in the front end.
This CL therefore eliminates both the PHEAP bit and even the idea of
"heap variables" from the back ends.
First, it replaces the PPARAM|PHEAP, PPARAMOUT|PHEAP, and PAUTO|PHEAP
variable classes with the single PAUTOHEAP, a pseudo-class indicating
a variable maintained on the heap and available by indirecting a
local variable kept on the stack (a plain PAUTO).
Second, walkexpr replaces all references to PAUTOHEAP variables
with indirections of the corresponding PAUTO variable.
The back ends and the liveness code now just see plain indirected
variables. This may actually produce better code, but the real goal
here is to eliminate these little-used and somewhat suspect code
paths in the back end analyses.
The OPARAM node type goes away too.
A followup CL will do the same to PPARAMREF. I'm not sure that
the back ends (SSA in particular) are handling those right either,
and with the framework established in this CL that change is trivial
and the result clearly more correct.
Fixes #15747.
Change-Id: I2770b1ce3cbc93981bfc7166be66a9da12013d74
Reviewed-on: https://go-review.googlesource.com/23393
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-05-25 01:33:24 -04:00
|
|
|
s.Unimplementedf("variable address class %v not implemented", classnames[n.Class])
|
2016-05-27 14:07:37 -07:00
|
|
|
return nil, false
|
2015-05-18 16:44:20 -07:00
|
|
|
}
|
|
|
|
case OINDREG:
|
2015-07-28 12:37:46 -07:00
|
|
|
// indirect off a register
|
2015-05-18 16:44:20 -07:00
|
|
|
// used for storing/loading arguments/returns to/from callees
|
2015-07-28 12:37:46 -07:00
|
|
|
if int(n.Reg) != Thearch.REGSP {
|
|
|
|
s.Unimplementedf("OINDREG of non-SP register %s in addr: %v", obj.Rconv(int(n.Reg)), n)
|
2016-05-27 14:07:37 -07:00
|
|
|
return nil, false
|
2015-07-28 12:37:46 -07:00
|
|
|
}
|
2016-05-27 14:07:37 -07:00
|
|
|
return s.entryNewValue1I(ssa.OpOffPtr, t, n.Xoffset, s.sp), true
|
2015-05-18 16:44:20 -07:00
|
|
|
case OINDEX:
|
2015-07-10 10:47:28 -06:00
|
|
|
if n.Left.Type.IsSlice() {
|
2015-05-18 16:44:20 -07:00
|
|
|
a := s.expr(n.Left)
|
|
|
|
i := s.expr(n.Right)
|
2016-05-25 09:49:28 -04:00
|
|
|
i = s.extendIndex(i, Panicindex)
|
2015-10-22 14:22:38 -07:00
|
|
|
len := s.newValue1(ssa.OpSliceLen, Types[TINT], a)
|
2015-08-18 14:17:30 -07:00
|
|
|
if !n.Bounded {
|
|
|
|
s.boundsCheck(i, len)
|
|
|
|
}
|
2015-10-22 14:22:38 -07:00
|
|
|
p := s.newValue1(ssa.OpSlicePtr, t, a)
|
2016-05-27 14:07:37 -07:00
|
|
|
return s.newValue2(ssa.OpPtrIndex, t, p, i), false
|
2015-07-10 10:47:28 -06:00
|
|
|
} else { // array
|
2016-05-27 14:07:37 -07:00
|
|
|
a, isVolatile := s.addr(n.Left, bounded)
|
2015-07-10 10:47:28 -06:00
|
|
|
i := s.expr(n.Right)
|
2016-05-25 09:49:28 -04:00
|
|
|
i = s.extendIndex(i, Panicindex)
|
2016-03-31 14:46:04 -07:00
|
|
|
len := s.constInt(Types[TINT], n.Left.Type.NumElem())
|
2015-08-18 14:17:30 -07:00
|
|
|
if !n.Bounded {
|
|
|
|
s.boundsCheck(i, len)
|
|
|
|
}
|
2016-05-27 14:07:37 -07:00
|
|
|
return s.newValue2(ssa.OpPtrIndex, Ptrto(n.Left.Type.Elem()), a, i), isVolatile
|
2015-05-18 16:44:20 -07:00
|
|
|
}
|
2015-07-13 21:22:16 -05:00
|
|
|
case OIND:
|
2016-05-27 14:07:37 -07:00
|
|
|
return s.exprPtr(n.Left, bounded, n.Lineno), false
|
2015-07-13 15:55:37 -07:00
|
|
|
case ODOT:
|
2016-05-27 14:07:37 -07:00
|
|
|
p, isVolatile := s.addr(n.Left, bounded)
|
|
|
|
return s.newValue1I(ssa.OpOffPtr, t, n.Xoffset, p), isVolatile
|
2015-07-13 15:55:37 -07:00
|
|
|
case ODOTPTR:
|
2016-04-19 21:06:53 -07:00
|
|
|
p := s.exprPtr(n.Left, bounded, n.Lineno)
|
2016-05-27 14:07:37 -07:00
|
|
|
return s.newValue1I(ssa.OpOffPtr, t, n.Xoffset, p), false
|
2015-09-11 16:40:05 -04:00
|
|
|
case OCLOSUREVAR:
|
2016-03-04 12:34:43 -08:00
|
|
|
return s.newValue1I(ssa.OpOffPtr, t, n.Xoffset,
|
2016-05-27 14:07:37 -07:00
|
|
|
s.entryNewValue0(ssa.OpGetClosurePtr, Ptrto(Types[TUINT8]))), false
|
2015-10-09 16:48:30 -04:00
|
|
|
case OCONVNOP:
|
2016-05-27 14:07:37 -07:00
|
|
|
addr, isVolatile := s.addr(n.Left, bounded)
|
|
|
|
return s.newValue1(ssa.OpCopy, t, addr), isVolatile // ensure that addr has the right type
|
2016-01-25 17:06:54 -08:00
|
|
|
case OCALLFUNC, OCALLINTER, OCALLMETH:
|
2016-05-27 14:07:37 -07:00
|
|
|
return s.call(n, callNormal), true
|
2015-10-09 16:48:30 -04:00
|
|
|
|
2015-05-18 16:44:20 -07:00
|
|
|
default:
|
2016-04-27 15:10:10 +10:00
|
|
|
s.Unimplementedf("unhandled addr %v", n.Op)
|
2016-05-27 14:07:37 -07:00
|
|
|
return nil, false
|
2015-05-18 16:44:20 -07:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-06-10 15:03:06 -07:00
|
|
|
// canSSA reports whether n is SSA-able.
|
2016-01-11 21:05:33 -08:00
|
|
|
// n must be an ONAME (or an ODOT sequence with an ONAME base).
|
2016-02-27 17:49:31 -08:00
|
|
|
func (s *state) canSSA(n *Node) bool {
|
2016-04-29 12:09:32 -07:00
|
|
|
if Debug['N'] != 0 {
|
|
|
|
return false
|
|
|
|
}
|
2016-01-11 21:05:33 -08:00
|
|
|
for n.Op == ODOT {
|
|
|
|
n = n.Left
|
|
|
|
}
|
2015-06-10 15:03:06 -07:00
|
|
|
if n.Op != ONAME {
|
2015-06-27 15:45:20 +01:00
|
|
|
return false
|
2015-06-10 15:03:06 -07:00
|
|
|
}
|
|
|
|
if n.Addrtaken {
|
|
|
|
return false
|
|
|
|
}
|
cmd/compile: fix liveness computation for heap-escaped parameters
The liveness computation of parameters generally was never
correct, but forcing all parameters to be live throughout the
function covered up that problem. The new SSA back end is
too clever: even though it currently keeps the parameter values live
throughout the function, it may find optimizations that mean
the current values are not written back to the original parameter
stack slots immediately or ever (for example if a parameter is set
to nil, SSA constant propagation may replace all later uses of the
parameter with a constant nil, eliminating the need to write the nil
value back to the stack slot), so the liveness code must now
track the actual operations on the stack slots, exposing these
problems.
One small problem in the handling of arguments is that nodarg
can return ONAME PPARAM nodes with adjusted offsets, so that
there are actually multiple *Node pointers for the same parameter
in the instruction stream. This might be possible to correct, but
not in this CL. For now, we fix this by using n.Orig instead of n
when considering PPARAM and PPARAMOUT nodes.
The major problem in the handling of arguments is general
confusion in the liveness code about the meaning of PPARAM|PHEAP
and PPARAMOUT|PHEAP nodes, especially as contrasted with PAUTO|PHEAP.
The difference between these two is that when a local variable "moves"
to the heap, it's really just allocated there to start with; in contrast,
when an argument moves to the heap, the actual data has to be copied
there from the stack at the beginning of the function, and when a
result "moves" to the heap the value in the heap has to be copied
back to the stack when the function returns
This general confusion is also present in the SSA back end.
The PHEAP bit worked decently when I first introduced it 7 years ago (!)
in 391425ae. The back end did nothing sophisticated, and in particular
there was no analysis at all: no escape analysis, no liveness analysis,
and certainly no SSA back end. But the complications caused in the
various downstream consumers suggest that this should be a detail
kept mainly in the front end.
This CL therefore eliminates both the PHEAP bit and even the idea of
"heap variables" from the back ends.
First, it replaces the PPARAM|PHEAP, PPARAMOUT|PHEAP, and PAUTO|PHEAP
variable classes with the single PAUTOHEAP, a pseudo-class indicating
a variable maintained on the heap and available by indirecting a
local variable kept on the stack (a plain PAUTO).
Second, walkexpr replaces all references to PAUTOHEAP variables
with indirections of the corresponding PAUTO variable.
The back ends and the liveness code now just see plain indirected
variables. This may actually produce better code, but the real goal
here is to eliminate these little-used and somewhat suspect code
paths in the back end analyses.
The OPARAM node type goes away too.
A followup CL will do the same to PPARAMREF. I'm not sure that
the back ends (SSA in particular) are handling those right either,
and with the framework established in this CL that change is trivial
and the result clearly more correct.
Fixes #15747.
Change-Id: I2770b1ce3cbc93981bfc7166be66a9da12013d74
Reviewed-on: https://go-review.googlesource.com/23393
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-05-25 01:33:24 -04:00
|
|
|
if n.isParamHeapCopy() {
|
2015-06-10 15:03:06 -07:00
|
|
|
return false
|
|
|
|
}
|
cmd/compile: fix liveness computation for heap-escaped parameters
The liveness computation of parameters generally was never
correct, but forcing all parameters to be live throughout the
function covered up that problem. The new SSA back end is
too clever: even though it currently keeps the parameter values live
throughout the function, it may find optimizations that mean
the current values are not written back to the original parameter
stack slots immediately or ever (for example if a parameter is set
to nil, SSA constant propagation may replace all later uses of the
parameter with a constant nil, eliminating the need to write the nil
value back to the stack slot), so the liveness code must now
track the actual operations on the stack slots, exposing these
problems.
One small problem in the handling of arguments is that nodarg
can return ONAME PPARAM nodes with adjusted offsets, so that
there are actually multiple *Node pointers for the same parameter
in the instruction stream. This might be possible to correct, but
not in this CL. For now, we fix this by using n.Orig instead of n
when considering PPARAM and PPARAMOUT nodes.
The major problem in the handling of arguments is general
confusion in the liveness code about the meaning of PPARAM|PHEAP
and PPARAMOUT|PHEAP nodes, especially as contrasted with PAUTO|PHEAP.
The difference between these two is that when a local variable "moves"
to the heap, it's really just allocated there to start with; in contrast,
when an argument moves to the heap, the actual data has to be copied
there from the stack at the beginning of the function, and when a
result "moves" to the heap the value in the heap has to be copied
back to the stack when the function returns
This general confusion is also present in the SSA back end.
The PHEAP bit worked decently when I first introduced it 7 years ago (!)
in 391425ae. The back end did nothing sophisticated, and in particular
there was no analysis at all: no escape analysis, no liveness analysis,
and certainly no SSA back end. But the complications caused in the
various downstream consumers suggest that this should be a detail
kept mainly in the front end.
This CL therefore eliminates both the PHEAP bit and even the idea of
"heap variables" from the back ends.
First, it replaces the PPARAM|PHEAP, PPARAMOUT|PHEAP, and PAUTO|PHEAP
variable classes with the single PAUTOHEAP, a pseudo-class indicating
a variable maintained on the heap and available by indirecting a
local variable kept on the stack (a plain PAUTO).
Second, walkexpr replaces all references to PAUTOHEAP variables
with indirections of the corresponding PAUTO variable.
The back ends and the liveness code now just see plain indirected
variables. This may actually produce better code, but the real goal
here is to eliminate these little-used and somewhat suspect code
paths in the back end analyses.
The OPARAM node type goes away too.
A followup CL will do the same to PPARAMREF. I'm not sure that
the back ends (SSA in particular) are handling those right either,
and with the framework established in this CL that change is trivial
and the result clearly more correct.
Fixes #15747.
Change-Id: I2770b1ce3cbc93981bfc7166be66a9da12013d74
Reviewed-on: https://go-review.googlesource.com/23393
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-05-25 01:33:24 -04:00
|
|
|
if n.Class == PAUTOHEAP {
|
|
|
|
Fatalf("canSSA of PAUTOHEAP %v", n)
|
|
|
|
}
|
2015-08-28 13:35:32 -07:00
|
|
|
switch n.Class {
|
2016-05-25 10:29:50 -04:00
|
|
|
case PEXTERN:
|
2015-06-10 15:03:06 -07:00
|
|
|
return false
|
2016-02-27 17:49:31 -08:00
|
|
|
case PPARAMOUT:
|
|
|
|
if hasdefer {
|
|
|
|
// TODO: handle this case? Named return values must be
|
|
|
|
// in memory so that the deferred function can see them.
|
|
|
|
// Maybe do: if !strings.HasPrefix(n.String(), "~") { return false }
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
if s.cgoUnsafeArgs {
|
|
|
|
// Cgo effectively takes the address of all result args,
|
|
|
|
// but the compiler can't see that.
|
|
|
|
return false
|
|
|
|
}
|
2015-06-10 15:03:06 -07:00
|
|
|
}
|
2015-09-08 21:28:44 -07:00
|
|
|
if n.Class == PPARAM && n.String() == ".this" {
|
|
|
|
// wrappers generated by genwrapper need to update
|
|
|
|
// the .this pointer in place.
|
2016-02-27 17:49:31 -08:00
|
|
|
// TODO: treat as a PPARMOUT?
|
2015-09-08 21:28:44 -07:00
|
|
|
return false
|
|
|
|
}
|
2015-08-18 10:26:28 -07:00
|
|
|
return canSSAType(n.Type)
|
|
|
|
// TODO: try to make more variables SSAable?
|
|
|
|
}
|
|
|
|
|
|
|
|
// canSSA reports whether variables of type t are SSA-able.
|
|
|
|
func canSSAType(t *Type) bool {
|
|
|
|
dowidth(t)
|
|
|
|
if t.Width > int64(4*Widthptr) {
|
2016-03-01 23:21:55 +00:00
|
|
|
// 4*Widthptr is an arbitrary constant. We want it
|
2015-08-18 10:26:28 -07:00
|
|
|
// to be at least 3*Widthptr so slices can be registerized.
|
|
|
|
// Too big and we'll introduce too much register pressure.
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
switch t.Etype {
|
|
|
|
case TARRAY:
|
|
|
|
// We can't do arrays because dynamic indexing is
|
|
|
|
// not supported on SSA variables.
|
|
|
|
// TODO: maybe allow if length is <=1? All indexes
|
|
|
|
// are constant? Might be good for the arrays
|
|
|
|
// introduced by the compiler for variadic functions.
|
2015-06-27 15:45:20 +01:00
|
|
|
return false
|
2015-08-18 10:26:28 -07:00
|
|
|
case TSTRUCT:
|
2016-03-17 13:26:08 -07:00
|
|
|
if t.NumFields() > ssa.MaxStruct {
|
2015-08-18 10:26:28 -07:00
|
|
|
return false
|
|
|
|
}
|
2016-03-17 01:32:18 -07:00
|
|
|
for _, t1 := range t.Fields().Slice() {
|
2015-08-18 10:26:28 -07:00
|
|
|
if !canSSAType(t1.Type) {
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
}
|
2016-01-11 21:05:33 -08:00
|
|
|
return true
|
2015-08-18 10:26:28 -07:00
|
|
|
default:
|
|
|
|
return true
|
2015-06-27 15:45:20 +01:00
|
|
|
}
|
2015-06-10 15:03:06 -07:00
|
|
|
}
|
|
|
|
|
2016-04-19 21:06:53 -07:00
|
|
|
// exprPtr evaluates n to a pointer and nil-checks it.
|
|
|
|
func (s *state) exprPtr(n *Node, bounded bool, lineno int32) *ssa.Value {
|
|
|
|
p := s.expr(n)
|
|
|
|
if bounded || n.NonNil {
|
|
|
|
if s.f.Config.Debug_checknil() && lineno > 1 {
|
|
|
|
s.f.Config.Warnl(lineno, "removed nil check")
|
|
|
|
}
|
|
|
|
return p
|
|
|
|
}
|
|
|
|
s.nilCheck(p)
|
|
|
|
return p
|
|
|
|
}
|
|
|
|
|
2015-05-18 16:44:20 -07:00
|
|
|
// nilCheck generates nil pointer checking code.
|
2015-08-11 09:47:45 -07:00
|
|
|
// Starts a new block on return, unless nil checks are disabled.
|
2015-07-24 11:55:52 -07:00
|
|
|
// Used only for automatically inserted nil checks,
|
|
|
|
// not for user code like 'x != nil'.
|
2015-05-18 16:44:20 -07:00
|
|
|
func (s *state) nilCheck(ptr *ssa.Value) {
|
2015-08-11 09:47:45 -07:00
|
|
|
if Disable_checknil != 0 {
|
|
|
|
return
|
|
|
|
}
|
2015-10-23 19:12:49 -07:00
|
|
|
chk := s.newValue2(ssa.OpNilCheck, ssa.TypeVoid, ptr, s.mem())
|
2015-05-18 16:44:20 -07:00
|
|
|
b := s.endBlock()
|
2015-10-23 19:12:49 -07:00
|
|
|
b.Kind = ssa.BlockCheck
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(chk)
|
2015-05-18 16:44:20 -07:00
|
|
|
bNext := s.f.NewBlock(ssa.BlockPlain)
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bNext)
|
2015-08-11 09:47:45 -07:00
|
|
|
s.startBlock(bNext)
|
2015-05-18 16:44:20 -07:00
|
|
|
}
|
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// boundsCheck generates bounds checking code. Checks if 0 <= idx < len, branches to exit if not.
|
2015-05-18 16:44:20 -07:00
|
|
|
// Starts a new block on return.
|
2016-05-25 09:49:28 -04:00
|
|
|
// idx is already converted to full int width.
|
2015-05-18 16:44:20 -07:00
|
|
|
func (s *state) boundsCheck(idx, len *ssa.Value) {
|
2015-08-18 15:25:40 -07:00
|
|
|
if Debug['B'] != 0 {
|
|
|
|
return
|
|
|
|
}
|
2015-05-18 16:44:20 -07:00
|
|
|
|
|
|
|
// bounds check
|
2015-07-30 11:03:05 -07:00
|
|
|
cmp := s.newValue2(ssa.OpIsInBounds, Types[TBOOL], idx, len)
|
2015-09-17 16:54:15 -07:00
|
|
|
s.check(cmp, Panicindex)
|
2015-08-24 23:52:03 -07:00
|
|
|
}
|
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// sliceBoundsCheck generates slice bounds checking code. Checks if 0 <= idx <= len, branches to exit if not.
|
2015-08-24 23:52:03 -07:00
|
|
|
// Starts a new block on return.
|
2016-05-25 09:49:28 -04:00
|
|
|
// idx and len are already converted to full int width.
|
2015-08-24 23:52:03 -07:00
|
|
|
func (s *state) sliceBoundsCheck(idx, len *ssa.Value) {
|
|
|
|
if Debug['B'] != 0 {
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
|
|
|
// bounds check
|
|
|
|
cmp := s.newValue2(ssa.OpIsSliceInBounds, Types[TBOOL], idx, len)
|
2015-09-17 16:54:15 -07:00
|
|
|
s.check(cmp, panicslice)
|
2015-08-24 23:52:03 -07:00
|
|
|
}
|
|
|
|
|
2016-05-25 09:49:28 -04:00
|
|
|
// If cmp (a bool) is false, panic using the given function.
|
2015-09-17 16:54:15 -07:00
|
|
|
func (s *state) check(cmp *ssa.Value, fn *Node) {
|
2015-05-18 16:44:20 -07:00
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(cmp)
|
2015-08-11 17:28:56 -07:00
|
|
|
b.Likely = ssa.BranchLikely
|
2015-05-18 16:44:20 -07:00
|
|
|
bNext := s.f.NewBlock(ssa.BlockPlain)
|
2015-11-09 21:35:40 -08:00
|
|
|
line := s.peekLine()
|
|
|
|
bPanic := s.panics[funcLine{fn, line}]
|
|
|
|
if bPanic == nil {
|
|
|
|
bPanic = s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
s.panics[funcLine{fn, line}] = bPanic
|
|
|
|
s.startBlock(bPanic)
|
|
|
|
// The panic call takes/returns memory to ensure that the right
|
|
|
|
// memory state is observed if the panic happens.
|
|
|
|
s.rtcall(fn, false, nil)
|
|
|
|
}
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bNext)
|
|
|
|
b.AddEdgeTo(bPanic)
|
2015-05-18 16:44:20 -07:00
|
|
|
s.startBlock(bNext)
|
|
|
|
}
|
|
|
|
|
2015-09-18 15:11:30 -07:00
|
|
|
// rtcall issues a call to the given runtime function fn with the listed args.
|
|
|
|
// Returns a slice of results of the given result types.
|
|
|
|
// The call is added to the end of the current block.
|
|
|
|
// If returns is false, the block is marked as an exit block.
|
2016-03-01 23:21:55 +00:00
|
|
|
// If returns is true, the block is marked as a call block. A new block
|
2015-09-18 15:11:30 -07:00
|
|
|
// is started to load the return values.
|
|
|
|
func (s *state) rtcall(fn *Node, returns bool, results []*Type, args ...*ssa.Value) []*ssa.Value {
|
|
|
|
// Write args to the stack
|
2016-05-15 00:12:56 -04:00
|
|
|
off := Ctxt.FixedFrameSize()
|
2015-09-18 15:11:30 -07:00
|
|
|
for _, arg := range args {
|
|
|
|
t := arg.Type
|
|
|
|
off = Rnd(off, t.Alignment())
|
|
|
|
ptr := s.sp
|
|
|
|
if off != 0 {
|
2016-06-08 10:12:30 -04:00
|
|
|
ptr = s.newValue1I(ssa.OpOffPtr, t.PtrTo(), off, s.sp)
|
2015-09-18 15:11:30 -07:00
|
|
|
}
|
|
|
|
size := t.Size()
|
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, size, ptr, arg, s.mem())
|
|
|
|
off += size
|
|
|
|
}
|
|
|
|
off = Rnd(off, int64(Widthptr))
|
2016-08-08 11:26:25 -07:00
|
|
|
if Thearch.LinkArch.Name == "amd64p32" {
|
|
|
|
// amd64p32 wants 8-byte alignment of the start of the return values.
|
|
|
|
off = Rnd(off, 8)
|
|
|
|
}
|
2015-09-18 15:11:30 -07:00
|
|
|
|
|
|
|
// Issue call
|
|
|
|
call := s.newValue1A(ssa.OpStaticCall, ssa.TypeMem, fn.Sym, s.mem())
|
|
|
|
s.vars[&memVar] = call
|
|
|
|
|
|
|
|
// Finish block
|
|
|
|
b := s.endBlock()
|
|
|
|
if !returns {
|
|
|
|
b.Kind = ssa.BlockExit
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(call)
|
[dev.ssa] cmd/compile: fix argument size of runtime call in SSA for ARM
The argument size for runtime call was incorrectly includes the size
of LR (FixedFrameSize in general). This makes the stack frame
sometimes unnecessarily 4 bytes larger on ARM.
For example,
func f(b []byte) byte { return b[0] }
compiles to
0x0000 00000 (h.go:6) TEXT "".f(SB), $4-16 // <-- framesize = 4
0x0000 00000 (h.go:6) MOVW 8(g), R1
0x0004 00004 (h.go:6) CMP R1, R13
0x0008 00008 (h.go:6) BLS 52
0x000c 00012 (h.go:6) MOVW.W R14, -8(R13)
0x0010 00016 (h.go:6) FUNCDATA $0, gclocals·8355ad952265fec823c17fcf739bd009(SB)
0x0010 00016 (h.go:6) FUNCDATA $1, gclocals·69c1753bd5f81501d95132d08af04464(SB)
0x0010 00016 (h.go:6) MOVW "".b+4(FP), R0
0x0014 00020 (h.go:6) CMP $0, R0
0x0018 00024 (h.go:6) BLS 44
0x001c 00028 (h.go:6) MOVW "".b(FP), R0
0x0020 00032 (h.go:6) MOVBU (R0), R0
0x0024 00036 (h.go:6) MOVB R0, "".~r1+12(FP)
0x0028 00040 (h.go:6) MOVW.P 8(R13), R15
0x002c 00044 (h.go:6) PCDATA $0, $1
0x002c 00044 (h.go:6) CALL runtime.panicindex(SB)
0x0030 00048 (h.go:6) UNDEF
0x0034 00052 (h.go:6) NOP
0x0034 00052 (h.go:6) MOVW R14, R3
0x0038 00056 (h.go:6) CALL runtime.morestack_noctxt(SB)
0x003c 00060 (h.go:6) JMP 0
Note that the frame size is 4, but there is actually no local. It
incorrectly thinks call to runtime.panicindex needs 4 bytes space
for argument.
This CL fixes it.
Updates #15365.
Change-Id: Ic65d55283a6aa8a7861d7a3fbc7b63c35785eeec
Reviewed-on: https://go-review.googlesource.com/24909
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-07-13 09:22:48 -06:00
|
|
|
call.AuxInt = off - Ctxt.FixedFrameSize()
|
2015-09-18 15:11:30 -07:00
|
|
|
if len(results) > 0 {
|
|
|
|
Fatalf("panic call can't have results")
|
|
|
|
}
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
b.Kind = ssa.BlockCall
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(call)
|
2015-09-18 15:11:30 -07:00
|
|
|
bNext := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
b.AddEdgeTo(bNext)
|
|
|
|
s.startBlock(bNext)
|
|
|
|
|
2016-04-21 19:28:28 -07:00
|
|
|
// Keep input pointer args live across calls. This is a bandaid until 1.8.
|
|
|
|
for _, n := range s.ptrargs {
|
|
|
|
s.vars[&memVar] = s.newValue2(ssa.OpKeepAlive, ssa.TypeMem, s.variable(n, n.Type), s.mem())
|
|
|
|
}
|
|
|
|
|
2015-09-18 15:11:30 -07:00
|
|
|
// Load results
|
|
|
|
res := make([]*ssa.Value, len(results))
|
|
|
|
for i, t := range results {
|
|
|
|
off = Rnd(off, t.Alignment())
|
|
|
|
ptr := s.sp
|
|
|
|
if off != 0 {
|
2016-06-08 10:12:30 -04:00
|
|
|
ptr = s.newValue1I(ssa.OpOffPtr, Ptrto(t), off, s.sp)
|
2015-09-18 15:11:30 -07:00
|
|
|
}
|
|
|
|
res[i] = s.newValue2(ssa.OpLoad, t, ptr, s.mem())
|
|
|
|
off += t.Size()
|
|
|
|
}
|
|
|
|
off = Rnd(off, int64(Widthptr))
|
|
|
|
|
|
|
|
// Remember how much callee stack space we needed.
|
|
|
|
call.AuxInt = off
|
|
|
|
|
|
|
|
return res
|
|
|
|
}
|
|
|
|
|
2016-01-25 17:06:54 -08:00
|
|
|
// insertWBmove inserts the assignment *left = *right including a write barrier.
|
|
|
|
// t is the type being assigned.
|
2016-05-27 14:07:37 -07:00
|
|
|
func (s *state) insertWBmove(t *Type, left, right *ssa.Value, line int32, rightIsVolatile bool) {
|
2015-11-16 13:20:16 -08:00
|
|
|
// if writeBarrier.enabled {
|
2016-01-25 17:06:54 -08:00
|
|
|
// typedmemmove(&t, left, right)
|
|
|
|
// } else {
|
|
|
|
// *left = *right
|
2015-09-11 11:02:57 -07:00
|
|
|
// }
|
2016-03-16 21:51:17 -07:00
|
|
|
|
|
|
|
if s.noWB {
|
|
|
|
s.Fatalf("write barrier prohibited")
|
|
|
|
}
|
|
|
|
if s.WBLineno == 0 {
|
|
|
|
s.WBLineno = left.Line
|
|
|
|
}
|
2015-09-11 11:02:57 -07:00
|
|
|
bThen := s.f.NewBlock(ssa.BlockPlain)
|
2016-01-25 17:06:54 -08:00
|
|
|
bElse := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bEnd := s.f.NewBlock(ssa.BlockPlain)
|
2015-09-11 11:02:57 -07:00
|
|
|
|
2016-04-29 09:02:27 -07:00
|
|
|
aux := &ssa.ExternSymbol{Typ: Types[TBOOL], Sym: syslook("writeBarrier").Sym}
|
2016-02-28 11:15:22 -05:00
|
|
|
flagaddr := s.newValue1A(ssa.OpAddr, Ptrto(Types[TUINT32]), aux, s.sb)
|
2016-05-06 10:12:57 -07:00
|
|
|
// Load word, test word, avoiding partial register write from load byte.
|
2016-02-28 11:15:22 -05:00
|
|
|
flag := s.newValue2(ssa.OpLoad, Types[TUINT32], flagaddr, s.mem())
|
2016-05-06 10:12:57 -07:00
|
|
|
flag = s.newValue2(ssa.OpNeq32, Types[TBOOL], flag, s.constInt32(Types[TUINT32], 0))
|
2015-09-11 11:02:57 -07:00
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
|
|
|
b.Likely = ssa.BranchUnlikely
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(flag)
|
2015-09-11 11:02:57 -07:00
|
|
|
b.AddEdgeTo(bThen)
|
2016-01-25 17:06:54 -08:00
|
|
|
b.AddEdgeTo(bElse)
|
2015-09-11 11:02:57 -07:00
|
|
|
|
|
|
|
s.startBlock(bThen)
|
2016-05-27 14:07:37 -07:00
|
|
|
|
|
|
|
if !rightIsVolatile {
|
|
|
|
// Issue typedmemmove call.
|
|
|
|
taddr := s.newValue1A(ssa.OpAddr, Types[TUINTPTR], &ssa.ExternSymbol{Typ: Types[TUINTPTR], Sym: typenamesym(t)}, s.sb)
|
|
|
|
s.rtcall(typedmemmove, true, nil, taddr, left, right)
|
|
|
|
} else {
|
|
|
|
// Copy to temp location if the source is volatile (will be clobbered by
|
|
|
|
// a function call). Marshaling the args to typedmemmove might clobber the
|
|
|
|
// value we're trying to move.
|
|
|
|
tmp := temp(t)
|
|
|
|
s.vars[&memVar] = s.newValue1A(ssa.OpVarDef, ssa.TypeMem, tmp, s.mem())
|
|
|
|
tmpaddr, _ := s.addr(tmp, true)
|
2016-06-27 16:54:57 -04:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpMove, ssa.TypeMem, SizeAlignAuxInt(t), tmpaddr, right, s.mem())
|
2016-05-27 14:07:37 -07:00
|
|
|
// Issue typedmemmove call.
|
|
|
|
taddr := s.newValue1A(ssa.OpAddr, Types[TUINTPTR], &ssa.ExternSymbol{Typ: Types[TUINTPTR], Sym: typenamesym(t)}, s.sb)
|
|
|
|
s.rtcall(typedmemmove, true, nil, taddr, left, tmpaddr)
|
|
|
|
// Mark temp as dead.
|
|
|
|
s.vars[&memVar] = s.newValue1A(ssa.OpVarKill, ssa.TypeMem, tmp, s.mem())
|
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
s.endBlock().AddEdgeTo(bEnd)
|
|
|
|
|
|
|
|
s.startBlock(bElse)
|
2016-06-27 16:54:57 -04:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpMove, ssa.TypeMem, SizeAlignAuxInt(t), left, right, s.mem())
|
2016-01-25 17:06:54 -08:00
|
|
|
s.endBlock().AddEdgeTo(bEnd)
|
|
|
|
|
|
|
|
s.startBlock(bEnd)
|
2015-09-11 11:02:57 -07:00
|
|
|
|
2015-10-26 17:34:06 -04:00
|
|
|
if Debug_wb > 0 {
|
2016-03-02 11:01:25 -08:00
|
|
|
Warnl(line, "write barrier")
|
2015-10-26 17:34:06 -04:00
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
}
|
2015-10-26 17:34:06 -04:00
|
|
|
|
2016-01-25 17:06:54 -08:00
|
|
|
// insertWBstore inserts the assignment *left = right including a write barrier.
|
|
|
|
// t is the type being assigned.
|
2016-03-21 10:22:03 -07:00
|
|
|
func (s *state) insertWBstore(t *Type, left, right *ssa.Value, line int32, skip skipMask) {
|
2016-01-25 17:06:54 -08:00
|
|
|
// store scalar fields
|
|
|
|
// if writeBarrier.enabled {
|
|
|
|
// writebarrierptr for pointer fields
|
|
|
|
// } else {
|
|
|
|
// store pointer fields
|
|
|
|
// }
|
|
|
|
|
2016-03-16 21:51:17 -07:00
|
|
|
if s.noWB {
|
|
|
|
s.Fatalf("write barrier prohibited")
|
|
|
|
}
|
|
|
|
if s.WBLineno == 0 {
|
|
|
|
s.WBLineno = left.Line
|
|
|
|
}
|
2016-03-21 10:22:03 -07:00
|
|
|
s.storeTypeScalars(t, left, right, skip)
|
2016-01-29 21:57:57 -08:00
|
|
|
|
|
|
|
bThen := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bElse := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bEnd := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
|
2016-04-29 09:02:27 -07:00
|
|
|
aux := &ssa.ExternSymbol{Typ: Types[TBOOL], Sym: syslook("writeBarrier").Sym}
|
2016-02-28 11:15:22 -05:00
|
|
|
flagaddr := s.newValue1A(ssa.OpAddr, Ptrto(Types[TUINT32]), aux, s.sb)
|
2016-05-06 10:12:57 -07:00
|
|
|
// Load word, test word, avoiding partial register write from load byte.
|
2016-02-28 11:15:22 -05:00
|
|
|
flag := s.newValue2(ssa.OpLoad, Types[TUINT32], flagaddr, s.mem())
|
2016-05-06 10:12:57 -07:00
|
|
|
flag = s.newValue2(ssa.OpNeq32, Types[TBOOL], flag, s.constInt32(Types[TUINT32], 0))
|
2016-01-29 21:57:57 -08:00
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
|
|
|
b.Likely = ssa.BranchUnlikely
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(flag)
|
2016-01-29 21:57:57 -08:00
|
|
|
b.AddEdgeTo(bThen)
|
|
|
|
b.AddEdgeTo(bElse)
|
|
|
|
|
|
|
|
// Issue write barriers for pointer writes.
|
|
|
|
s.startBlock(bThen)
|
|
|
|
s.storeTypePtrsWB(t, left, right)
|
|
|
|
s.endBlock().AddEdgeTo(bEnd)
|
|
|
|
|
|
|
|
// Issue regular stores for pointer writes.
|
|
|
|
s.startBlock(bElse)
|
|
|
|
s.storeTypePtrs(t, left, right)
|
|
|
|
s.endBlock().AddEdgeTo(bEnd)
|
|
|
|
|
|
|
|
s.startBlock(bEnd)
|
|
|
|
|
|
|
|
if Debug_wb > 0 {
|
2016-03-02 11:01:25 -08:00
|
|
|
Warnl(line, "write barrier")
|
2016-01-25 17:06:54 -08:00
|
|
|
}
|
2016-01-29 21:57:57 -08:00
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
|
2016-01-29 21:57:57 -08:00
|
|
|
// do *left = right for all scalar (non-pointer) parts of t.
|
2016-03-21 10:22:03 -07:00
|
|
|
func (s *state) storeTypeScalars(t *Type, left, right *ssa.Value, skip skipMask) {
|
2016-01-25 17:06:54 -08:00
|
|
|
switch {
|
2016-01-29 21:57:57 -08:00
|
|
|
case t.IsBoolean() || t.IsInteger() || t.IsFloat() || t.IsComplex():
|
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, t.Size(), left, right, s.mem())
|
2016-03-28 10:55:44 -07:00
|
|
|
case t.IsPtrShaped():
|
2016-01-25 17:06:54 -08:00
|
|
|
// no scalar fields.
|
|
|
|
case t.IsString():
|
2016-03-21 10:22:03 -07:00
|
|
|
if skip&skipLen != 0 {
|
|
|
|
return
|
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
len := s.newValue1(ssa.OpStringLen, Types[TINT], right)
|
|
|
|
lenAddr := s.newValue1I(ssa.OpOffPtr, Ptrto(Types[TINT]), s.config.IntSize, left)
|
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, s.config.IntSize, lenAddr, len, s.mem())
|
|
|
|
case t.IsSlice():
|
2016-03-21 10:22:03 -07:00
|
|
|
if skip&skipLen == 0 {
|
|
|
|
len := s.newValue1(ssa.OpSliceLen, Types[TINT], right)
|
|
|
|
lenAddr := s.newValue1I(ssa.OpOffPtr, Ptrto(Types[TINT]), s.config.IntSize, left)
|
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, s.config.IntSize, lenAddr, len, s.mem())
|
|
|
|
}
|
|
|
|
if skip&skipCap == 0 {
|
|
|
|
cap := s.newValue1(ssa.OpSliceCap, Types[TINT], right)
|
|
|
|
capAddr := s.newValue1I(ssa.OpOffPtr, Ptrto(Types[TINT]), 2*s.config.IntSize, left)
|
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, s.config.IntSize, capAddr, cap, s.mem())
|
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
case t.IsInterface():
|
|
|
|
// itab field doesn't need a write barrier (even though it is a pointer).
|
|
|
|
itab := s.newValue1(ssa.OpITab, Ptrto(Types[TUINT8]), right)
|
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, s.config.IntSize, left, itab, s.mem())
|
2016-01-29 21:57:57 -08:00
|
|
|
case t.IsStruct():
|
|
|
|
n := t.NumFields()
|
2016-03-14 12:45:18 -07:00
|
|
|
for i := 0; i < n; i++ {
|
2016-01-29 21:57:57 -08:00
|
|
|
ft := t.FieldType(i)
|
|
|
|
addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
|
2016-03-14 12:45:18 -07:00
|
|
|
val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
|
2016-03-21 10:22:03 -07:00
|
|
|
s.storeTypeScalars(ft.(*Type), addr, val, 0)
|
2016-01-29 21:57:57 -08:00
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
default:
|
|
|
|
s.Fatalf("bad write barrier type %s", t)
|
|
|
|
}
|
2016-01-29 21:57:57 -08:00
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
|
2016-01-29 21:57:57 -08:00
|
|
|
// do *left = right for all pointer parts of t.
|
|
|
|
func (s *state) storeTypePtrs(t *Type, left, right *ssa.Value) {
|
2016-01-25 17:06:54 -08:00
|
|
|
switch {
|
2016-03-28 10:55:44 -07:00
|
|
|
case t.IsPtrShaped():
|
2016-01-29 21:57:57 -08:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, s.config.PtrSize, left, right, s.mem())
|
2016-01-25 17:06:54 -08:00
|
|
|
case t.IsString():
|
|
|
|
ptr := s.newValue1(ssa.OpStringPtr, Ptrto(Types[TUINT8]), right)
|
2016-01-29 21:57:57 -08:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, s.config.PtrSize, left, ptr, s.mem())
|
2016-01-25 17:06:54 -08:00
|
|
|
case t.IsSlice():
|
|
|
|
ptr := s.newValue1(ssa.OpSlicePtr, Ptrto(Types[TUINT8]), right)
|
2016-01-29 21:57:57 -08:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, s.config.PtrSize, left, ptr, s.mem())
|
2016-01-25 17:06:54 -08:00
|
|
|
case t.IsInterface():
|
2016-01-29 21:57:57 -08:00
|
|
|
// itab field is treated as a scalar.
|
2016-01-25 17:06:54 -08:00
|
|
|
idata := s.newValue1(ssa.OpIData, Ptrto(Types[TUINT8]), right)
|
|
|
|
idataAddr := s.newValue1I(ssa.OpOffPtr, Ptrto(Types[TUINT8]), s.config.PtrSize, left)
|
2016-01-29 21:57:57 -08:00
|
|
|
s.vars[&memVar] = s.newValue3I(ssa.OpStore, ssa.TypeMem, s.config.PtrSize, idataAddr, idata, s.mem())
|
|
|
|
case t.IsStruct():
|
|
|
|
n := t.NumFields()
|
2016-03-14 12:45:18 -07:00
|
|
|
for i := 0; i < n; i++ {
|
2016-01-29 21:57:57 -08:00
|
|
|
ft := t.FieldType(i)
|
|
|
|
if !haspointers(ft.(*Type)) {
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
|
2016-03-14 12:45:18 -07:00
|
|
|
val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
|
2016-01-29 21:57:57 -08:00
|
|
|
s.storeTypePtrs(ft.(*Type), addr, val)
|
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
default:
|
|
|
|
s.Fatalf("bad write barrier type %s", t)
|
|
|
|
}
|
2016-01-29 21:57:57 -08:00
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
|
2016-01-29 21:57:57 -08:00
|
|
|
// do *left = right with a write barrier for all pointer parts of t.
|
|
|
|
func (s *state) storeTypePtrsWB(t *Type, left, right *ssa.Value) {
|
2016-01-25 17:06:54 -08:00
|
|
|
switch {
|
2016-03-28 10:55:44 -07:00
|
|
|
case t.IsPtrShaped():
|
2016-01-29 21:57:57 -08:00
|
|
|
s.rtcall(writebarrierptr, true, nil, left, right)
|
2016-01-25 17:06:54 -08:00
|
|
|
case t.IsString():
|
|
|
|
ptr := s.newValue1(ssa.OpStringPtr, Ptrto(Types[TUINT8]), right)
|
2016-01-29 21:57:57 -08:00
|
|
|
s.rtcall(writebarrierptr, true, nil, left, ptr)
|
2016-01-25 17:06:54 -08:00
|
|
|
case t.IsSlice():
|
|
|
|
ptr := s.newValue1(ssa.OpSlicePtr, Ptrto(Types[TUINT8]), right)
|
2016-01-29 21:57:57 -08:00
|
|
|
s.rtcall(writebarrierptr, true, nil, left, ptr)
|
2016-01-25 17:06:54 -08:00
|
|
|
case t.IsInterface():
|
|
|
|
idata := s.newValue1(ssa.OpIData, Ptrto(Types[TUINT8]), right)
|
|
|
|
idataAddr := s.newValue1I(ssa.OpOffPtr, Ptrto(Types[TUINT8]), s.config.PtrSize, left)
|
2016-01-29 21:57:57 -08:00
|
|
|
s.rtcall(writebarrierptr, true, nil, idataAddr, idata)
|
|
|
|
case t.IsStruct():
|
|
|
|
n := t.NumFields()
|
2016-03-14 12:45:18 -07:00
|
|
|
for i := 0; i < n; i++ {
|
2016-01-29 21:57:57 -08:00
|
|
|
ft := t.FieldType(i)
|
|
|
|
if !haspointers(ft.(*Type)) {
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
|
2016-03-14 12:45:18 -07:00
|
|
|
val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
|
2016-01-29 21:57:57 -08:00
|
|
|
s.storeTypePtrsWB(ft.(*Type), addr, val)
|
|
|
|
}
|
2016-01-25 17:06:54 -08:00
|
|
|
default:
|
|
|
|
s.Fatalf("bad write barrier type %s", t)
|
|
|
|
}
|
2015-09-11 11:02:57 -07:00
|
|
|
}
|
|
|
|
|
2015-09-12 23:27:26 -07:00
|
|
|
// slice computes the slice v[i:j:k] and returns ptr, len, and cap of result.
|
|
|
|
// i,j,k may be nil, in which case they are set to their default value.
|
|
|
|
// t is a slice, ptr to array, or string type.
|
|
|
|
func (s *state) slice(t *Type, v, i, j, k *ssa.Value) (p, l, c *ssa.Value) {
|
|
|
|
var elemtype *Type
|
|
|
|
var ptrtype *Type
|
|
|
|
var ptr *ssa.Value
|
|
|
|
var len *ssa.Value
|
|
|
|
var cap *ssa.Value
|
|
|
|
zero := s.constInt(Types[TINT], 0)
|
|
|
|
switch {
|
|
|
|
case t.IsSlice():
|
2016-03-30 10:57:47 -07:00
|
|
|
elemtype = t.Elem()
|
2015-09-12 23:27:26 -07:00
|
|
|
ptrtype = Ptrto(elemtype)
|
|
|
|
ptr = s.newValue1(ssa.OpSlicePtr, ptrtype, v)
|
|
|
|
len = s.newValue1(ssa.OpSliceLen, Types[TINT], v)
|
|
|
|
cap = s.newValue1(ssa.OpSliceCap, Types[TINT], v)
|
|
|
|
case t.IsString():
|
|
|
|
elemtype = Types[TUINT8]
|
|
|
|
ptrtype = Ptrto(elemtype)
|
|
|
|
ptr = s.newValue1(ssa.OpStringPtr, ptrtype, v)
|
|
|
|
len = s.newValue1(ssa.OpStringLen, Types[TINT], v)
|
|
|
|
cap = len
|
|
|
|
case t.IsPtr():
|
2016-03-30 10:57:47 -07:00
|
|
|
if !t.Elem().IsArray() {
|
2015-09-12 23:27:26 -07:00
|
|
|
s.Fatalf("bad ptr to array in slice %v\n", t)
|
|
|
|
}
|
2016-03-30 10:57:47 -07:00
|
|
|
elemtype = t.Elem().Elem()
|
2015-09-12 23:27:26 -07:00
|
|
|
ptrtype = Ptrto(elemtype)
|
|
|
|
s.nilCheck(v)
|
|
|
|
ptr = v
|
2016-03-31 14:46:04 -07:00
|
|
|
len = s.constInt(Types[TINT], t.Elem().NumElem())
|
2015-09-12 23:27:26 -07:00
|
|
|
cap = len
|
|
|
|
default:
|
|
|
|
s.Fatalf("bad type in slice %v\n", t)
|
|
|
|
}
|
|
|
|
|
|
|
|
// Set default values
|
|
|
|
if i == nil {
|
|
|
|
i = zero
|
|
|
|
}
|
|
|
|
if j == nil {
|
|
|
|
j = len
|
|
|
|
}
|
|
|
|
if k == nil {
|
|
|
|
k = cap
|
|
|
|
}
|
|
|
|
|
|
|
|
// Panic if slice indices are not in bounds.
|
|
|
|
s.sliceBoundsCheck(i, j)
|
|
|
|
if j != k {
|
|
|
|
s.sliceBoundsCheck(j, k)
|
|
|
|
}
|
|
|
|
if k != cap {
|
|
|
|
s.sliceBoundsCheck(k, cap)
|
|
|
|
}
|
|
|
|
|
|
|
|
// Generate the following code assuming that indexes are in bounds.
|
|
|
|
// The conditional is to make sure that we don't generate a slice
|
|
|
|
// that points to the next object in memory.
|
cmd/compile: change the way SSA does slice zero-cap detection
There is a special case for slicing s[i:j] when the resulting
slice has zero capacity, to prevent pointing to the next object
in memory.
Change this special case code from:
rptr := rcap == 0 ? ptr : ptr+i*elemsize
to
rptr := ptr + (rcap == 0 ? 0 : i) * elemsize
This change leads to slightly smaller generated code, replacing
a load with a register zero.
old:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function ...
0x004e 00078 (slice.go:7) MOVQ "".a+8(FP), AX
0x0053 00083 (slice.go:8) JMP 62
new:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function...
0x004e 00078 (slice.go:8) MOVQ $0, CX
0x0050 00080 (slice.go:8) JMP 52
Change-Id: I2a396616b0d7b090c226a47c92a7ba14b128401f
Reviewed-on: https://go-review.googlesource.com/20994
Reviewed-by: David Chase <drchase@google.com>
2016-03-21 15:24:08 -07:00
|
|
|
// rlen = j-i
|
|
|
|
// rcap = k-i
|
|
|
|
// delta = i*elemsize
|
|
|
|
// if rcap == 0 {
|
|
|
|
// delta = 0
|
2015-09-12 23:27:26 -07:00
|
|
|
// }
|
cmd/compile: change the way SSA does slice zero-cap detection
There is a special case for slicing s[i:j] when the resulting
slice has zero capacity, to prevent pointing to the next object
in memory.
Change this special case code from:
rptr := rcap == 0 ? ptr : ptr+i*elemsize
to
rptr := ptr + (rcap == 0 ? 0 : i) * elemsize
This change leads to slightly smaller generated code, replacing
a load with a register zero.
old:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function ...
0x004e 00078 (slice.go:7) MOVQ "".a+8(FP), AX
0x0053 00083 (slice.go:8) JMP 62
new:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function...
0x004e 00078 (slice.go:8) MOVQ $0, CX
0x0050 00080 (slice.go:8) JMP 52
Change-Id: I2a396616b0d7b090c226a47c92a7ba14b128401f
Reviewed-on: https://go-review.googlesource.com/20994
Reviewed-by: David Chase <drchase@google.com>
2016-03-21 15:24:08 -07:00
|
|
|
// rptr = p+delta
|
|
|
|
// result = (SliceMake rptr rlen rcap)
|
2015-11-02 21:28:13 -08:00
|
|
|
subOp := s.ssaOp(OSUB, Types[TINT])
|
cmd/compile: change the way SSA does slice zero-cap detection
There is a special case for slicing s[i:j] when the resulting
slice has zero capacity, to prevent pointing to the next object
in memory.
Change this special case code from:
rptr := rcap == 0 ? ptr : ptr+i*elemsize
to
rptr := ptr + (rcap == 0 ? 0 : i) * elemsize
This change leads to slightly smaller generated code, replacing
a load with a register zero.
old:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function ...
0x004e 00078 (slice.go:7) MOVQ "".a+8(FP), AX
0x0053 00083 (slice.go:8) JMP 62
new:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function...
0x004e 00078 (slice.go:8) MOVQ $0, CX
0x0050 00080 (slice.go:8) JMP 52
Change-Id: I2a396616b0d7b090c226a47c92a7ba14b128401f
Reviewed-on: https://go-review.googlesource.com/20994
Reviewed-by: David Chase <drchase@google.com>
2016-03-21 15:24:08 -07:00
|
|
|
eqOp := s.ssaOp(OEQ, Types[TINT])
|
2015-11-02 21:28:13 -08:00
|
|
|
mulOp := s.ssaOp(OMUL, Types[TINT])
|
|
|
|
rlen := s.newValue2(subOp, Types[TINT], j, i)
|
2015-09-12 23:27:26 -07:00
|
|
|
var rcap *ssa.Value
|
|
|
|
switch {
|
|
|
|
case t.IsString():
|
2016-03-01 23:21:55 +00:00
|
|
|
// Capacity of the result is unimportant. However, we use
|
2015-09-12 23:27:26 -07:00
|
|
|
// rcap to test if we've generated a zero-length slice.
|
|
|
|
// Use length of strings for that.
|
|
|
|
rcap = rlen
|
|
|
|
case j == k:
|
|
|
|
rcap = rlen
|
|
|
|
default:
|
2015-11-02 21:28:13 -08:00
|
|
|
rcap = s.newValue2(subOp, Types[TINT], k, i)
|
2015-09-12 23:27:26 -07:00
|
|
|
}
|
|
|
|
|
cmd/compile: change the way SSA does slice zero-cap detection
There is a special case for slicing s[i:j] when the resulting
slice has zero capacity, to prevent pointing to the next object
in memory.
Change this special case code from:
rptr := rcap == 0 ? ptr : ptr+i*elemsize
to
rptr := ptr + (rcap == 0 ? 0 : i) * elemsize
This change leads to slightly smaller generated code, replacing
a load with a register zero.
old:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function ...
0x004e 00078 (slice.go:7) MOVQ "".a+8(FP), AX
0x0053 00083 (slice.go:8) JMP 62
new:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function...
0x004e 00078 (slice.go:8) MOVQ $0, CX
0x0050 00080 (slice.go:8) JMP 52
Change-Id: I2a396616b0d7b090c226a47c92a7ba14b128401f
Reviewed-on: https://go-review.googlesource.com/20994
Reviewed-by: David Chase <drchase@google.com>
2016-03-21 15:24:08 -07:00
|
|
|
// delta = # of elements to offset pointer by.
|
|
|
|
s.vars[&deltaVar] = i
|
2015-09-12 23:27:26 -07:00
|
|
|
|
cmd/compile: change the way SSA does slice zero-cap detection
There is a special case for slicing s[i:j] when the resulting
slice has zero capacity, to prevent pointing to the next object
in memory.
Change this special case code from:
rptr := rcap == 0 ? ptr : ptr+i*elemsize
to
rptr := ptr + (rcap == 0 ? 0 : i) * elemsize
This change leads to slightly smaller generated code, replacing
a load with a register zero.
old:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function ...
0x004e 00078 (slice.go:7) MOVQ "".a+8(FP), AX
0x0053 00083 (slice.go:8) JMP 62
new:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function...
0x004e 00078 (slice.go:8) MOVQ $0, CX
0x0050 00080 (slice.go:8) JMP 52
Change-Id: I2a396616b0d7b090c226a47c92a7ba14b128401f
Reviewed-on: https://go-review.googlesource.com/20994
Reviewed-by: David Chase <drchase@google.com>
2016-03-21 15:24:08 -07:00
|
|
|
// Generate code to set delta=0 if the resulting capacity is zero.
|
|
|
|
if !((i.Op == ssa.OpConst64 && i.AuxInt == 0) ||
|
|
|
|
(i.Op == ssa.OpConst32 && int32(i.AuxInt) == 0)) {
|
|
|
|
cmp := s.newValue2(eqOp, Types[TBOOL], rcap, zero)
|
2015-09-12 23:27:26 -07:00
|
|
|
|
cmd/compile: change the way SSA does slice zero-cap detection
There is a special case for slicing s[i:j] when the resulting
slice has zero capacity, to prevent pointing to the next object
in memory.
Change this special case code from:
rptr := rcap == 0 ? ptr : ptr+i*elemsize
to
rptr := ptr + (rcap == 0 ? 0 : i) * elemsize
This change leads to slightly smaller generated code, replacing
a load with a register zero.
old:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function ...
0x004e 00078 (slice.go:7) MOVQ "".a+8(FP), AX
0x0053 00083 (slice.go:8) JMP 62
new:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function...
0x004e 00078 (slice.go:8) MOVQ $0, CX
0x0050 00080 (slice.go:8) JMP 52
Change-Id: I2a396616b0d7b090c226a47c92a7ba14b128401f
Reviewed-on: https://go-review.googlesource.com/20994
Reviewed-by: David Chase <drchase@google.com>
2016-03-21 15:24:08 -07:00
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
|
|
|
b.Likely = ssa.BranchUnlikely
|
|
|
|
b.SetControl(cmp)
|
2015-09-12 23:27:26 -07:00
|
|
|
|
cmd/compile: change the way SSA does slice zero-cap detection
There is a special case for slicing s[i:j] when the resulting
slice has zero capacity, to prevent pointing to the next object
in memory.
Change this special case code from:
rptr := rcap == 0 ? ptr : ptr+i*elemsize
to
rptr := ptr + (rcap == 0 ? 0 : i) * elemsize
This change leads to slightly smaller generated code, replacing
a load with a register zero.
old:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function ...
0x004e 00078 (slice.go:7) MOVQ "".a+8(FP), AX
0x0053 00083 (slice.go:8) JMP 62
new:
0x002e 00046 (slice.go:8) CMPQ BX, $0
0x0032 00050 (slice.go:8) JEQ $0, 78
0x0034 00052 (slice.go:8) MOVQ "".a+8(FP), BP
0x0039 00057 (slice.go:8) LEAQ (BP)(CX*8), AX
0x003e 00062 ... rest of function...
0x004e 00078 (slice.go:8) MOVQ $0, CX
0x0050 00080 (slice.go:8) JMP 52
Change-Id: I2a396616b0d7b090c226a47c92a7ba14b128401f
Reviewed-on: https://go-review.googlesource.com/20994
Reviewed-by: David Chase <drchase@google.com>
2016-03-21 15:24:08 -07:00
|
|
|
// Generate block which zeros the delta variable.
|
|
|
|
nz := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
b.AddEdgeTo(nz)
|
|
|
|
s.startBlock(nz)
|
|
|
|
s.vars[&deltaVar] = zero
|
|
|
|
s.endBlock()
|
|
|
|
|
|
|
|
// All done.
|
|
|
|
merge := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
b.AddEdgeTo(merge)
|
|
|
|
nz.AddEdgeTo(merge)
|
|
|
|
s.startBlock(merge)
|
|
|
|
|
|
|
|
// TODO: use conditional moves somehow?
|
|
|
|
}
|
|
|
|
|
|
|
|
// Compute rptr = ptr + delta * elemsize
|
|
|
|
rptr := s.newValue2(ssa.OpAddPtr, ptrtype, ptr, s.newValue2(mulOp, Types[TINT], s.variable(&deltaVar, Types[TINT]), s.constInt(Types[TINT], elemtype.Width)))
|
|
|
|
delete(s.vars, &deltaVar)
|
2015-09-12 23:27:26 -07:00
|
|
|
return rptr, rlen, rcap
|
|
|
|
}
|
|
|
|
|
2015-08-20 15:14:20 -04:00
|
|
|
type u2fcvtTab struct {
|
|
|
|
geq, cvt2F, and, rsh, or, add ssa.Op
|
|
|
|
one func(*state, ssa.Type, int64) *ssa.Value
|
|
|
|
}
|
|
|
|
|
|
|
|
var u64_f64 u2fcvtTab = u2fcvtTab{
|
|
|
|
geq: ssa.OpGeq64,
|
|
|
|
cvt2F: ssa.OpCvt64to64F,
|
|
|
|
and: ssa.OpAnd64,
|
|
|
|
rsh: ssa.OpRsh64Ux64,
|
|
|
|
or: ssa.OpOr64,
|
|
|
|
add: ssa.OpAdd64F,
|
|
|
|
one: (*state).constInt64,
|
|
|
|
}
|
|
|
|
|
|
|
|
var u64_f32 u2fcvtTab = u2fcvtTab{
|
|
|
|
geq: ssa.OpGeq64,
|
|
|
|
cvt2F: ssa.OpCvt64to32F,
|
|
|
|
and: ssa.OpAnd64,
|
|
|
|
rsh: ssa.OpRsh64Ux64,
|
|
|
|
or: ssa.OpOr64,
|
|
|
|
add: ssa.OpAdd32F,
|
|
|
|
one: (*state).constInt64,
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *state) uint64Tofloat64(n *Node, x *ssa.Value, ft, tt *Type) *ssa.Value {
|
|
|
|
return s.uintTofloat(&u64_f64, n, x, ft, tt)
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *state) uint64Tofloat32(n *Node, x *ssa.Value, ft, tt *Type) *ssa.Value {
|
|
|
|
return s.uintTofloat(&u64_f32, n, x, ft, tt)
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *state) uintTofloat(cvttab *u2fcvtTab, n *Node, x *ssa.Value, ft, tt *Type) *ssa.Value {
|
|
|
|
// if x >= 0 {
|
|
|
|
// result = (floatY) x
|
|
|
|
// } else {
|
|
|
|
// y = uintX(x) ; y = x & 1
|
|
|
|
// z = uintX(x) ; z = z >> 1
|
|
|
|
// z = z >> 1
|
|
|
|
// z = z | y
|
2015-08-26 14:25:40 -04:00
|
|
|
// result = floatY(z)
|
|
|
|
// result = result + result
|
2015-08-20 15:14:20 -04:00
|
|
|
// }
|
|
|
|
//
|
|
|
|
// Code borrowed from old code generator.
|
|
|
|
// What's going on: large 64-bit "unsigned" looks like
|
|
|
|
// negative number to hardware's integer-to-float
|
2016-03-01 23:21:55 +00:00
|
|
|
// conversion. However, because the mantissa is only
|
2015-08-20 15:14:20 -04:00
|
|
|
// 63 bits, we don't need the LSB, so instead we do an
|
|
|
|
// unsigned right shift (divide by two), convert, and
|
2016-03-01 23:21:55 +00:00
|
|
|
// double. However, before we do that, we need to be
|
2015-08-20 15:14:20 -04:00
|
|
|
// sure that we do not lose a "1" if that made the
|
2016-03-01 23:21:55 +00:00
|
|
|
// difference in the resulting rounding. Therefore, we
|
|
|
|
// preserve it, and OR (not ADD) it back in. The case
|
2015-08-20 15:14:20 -04:00
|
|
|
// that matters is when the eleven discarded bits are
|
|
|
|
// equal to 10000000001; that rounds up, and the 1 cannot
|
|
|
|
// be lost else it would round down if the LSB of the
|
|
|
|
// candidate mantissa is 0.
|
|
|
|
cmp := s.newValue2(cvttab.geq, Types[TBOOL], x, s.zeroVal(ft))
|
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(cmp)
|
2015-08-20 15:14:20 -04:00
|
|
|
b.Likely = ssa.BranchLikely
|
|
|
|
|
|
|
|
bThen := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bElse := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bAfter := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bThen)
|
2015-08-20 15:14:20 -04:00
|
|
|
s.startBlock(bThen)
|
|
|
|
a0 := s.newValue1(cvttab.cvt2F, tt, x)
|
|
|
|
s.vars[n] = a0
|
|
|
|
s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
bThen.AddEdgeTo(bAfter)
|
2015-08-20 15:14:20 -04:00
|
|
|
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bElse)
|
2015-08-20 15:14:20 -04:00
|
|
|
s.startBlock(bElse)
|
|
|
|
one := cvttab.one(s, ft, 1)
|
|
|
|
y := s.newValue2(cvttab.and, ft, x, one)
|
|
|
|
z := s.newValue2(cvttab.rsh, ft, x, one)
|
|
|
|
z = s.newValue2(cvttab.or, ft, z, y)
|
|
|
|
a := s.newValue1(cvttab.cvt2F, tt, z)
|
|
|
|
a1 := s.newValue2(cvttab.add, tt, a, a)
|
|
|
|
s.vars[n] = a1
|
|
|
|
s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
bElse.AddEdgeTo(bAfter)
|
2015-08-20 15:14:20 -04:00
|
|
|
|
|
|
|
s.startBlock(bAfter)
|
|
|
|
return s.variable(n, n.Type)
|
|
|
|
}
|
|
|
|
|
2015-08-28 15:56:43 -05:00
|
|
|
// referenceTypeBuiltin generates code for the len/cap builtins for maps and channels.
|
|
|
|
func (s *state) referenceTypeBuiltin(n *Node, x *ssa.Value) *ssa.Value {
|
|
|
|
if !n.Left.Type.IsMap() && !n.Left.Type.IsChan() {
|
|
|
|
s.Fatalf("node must be a map or a channel")
|
|
|
|
}
|
2015-08-26 18:40:52 -05:00
|
|
|
// if n == nil {
|
|
|
|
// return 0
|
|
|
|
// } else {
|
2015-08-28 15:56:43 -05:00
|
|
|
// // len
|
2015-08-26 18:40:52 -05:00
|
|
|
// return *((*int)n)
|
2015-08-28 15:56:43 -05:00
|
|
|
// // cap
|
|
|
|
// return *(((*int)n)+1)
|
2015-08-26 18:40:52 -05:00
|
|
|
// }
|
|
|
|
lenType := n.Type
|
2016-03-06 18:06:09 -08:00
|
|
|
nilValue := s.constNil(Types[TUINTPTR])
|
2015-08-28 15:20:54 -05:00
|
|
|
cmp := s.newValue2(ssa.OpEqPtr, Types[TBOOL], x, nilValue)
|
2015-08-26 18:40:52 -05:00
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(cmp)
|
2015-08-26 18:40:52 -05:00
|
|
|
b.Likely = ssa.BranchUnlikely
|
|
|
|
|
|
|
|
bThen := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bElse := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bAfter := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
|
2015-08-28 15:56:43 -05:00
|
|
|
// length/capacity of a nil map/chan is zero
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bThen)
|
2015-08-26 18:40:52 -05:00
|
|
|
s.startBlock(bThen)
|
|
|
|
s.vars[n] = s.zeroVal(lenType)
|
|
|
|
s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
bThen.AddEdgeTo(bAfter)
|
2015-08-26 18:40:52 -05:00
|
|
|
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bElse)
|
2015-08-26 18:40:52 -05:00
|
|
|
s.startBlock(bElse)
|
2015-08-28 15:56:43 -05:00
|
|
|
if n.Op == OLEN {
|
|
|
|
// length is stored in the first word for map/chan
|
|
|
|
s.vars[n] = s.newValue2(ssa.OpLoad, lenType, x, s.mem())
|
|
|
|
} else if n.Op == OCAP {
|
|
|
|
// capacity is stored in the second word for chan
|
|
|
|
sw := s.newValue1I(ssa.OpOffPtr, lenType.PtrTo(), lenType.Width, x)
|
|
|
|
s.vars[n] = s.newValue2(ssa.OpLoad, lenType, sw, s.mem())
|
|
|
|
} else {
|
|
|
|
s.Fatalf("op must be OLEN or OCAP")
|
|
|
|
}
|
2015-08-26 18:40:52 -05:00
|
|
|
s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
bElse.AddEdgeTo(bAfter)
|
2015-08-26 18:40:52 -05:00
|
|
|
|
|
|
|
s.startBlock(bAfter)
|
|
|
|
return s.variable(n, lenType)
|
|
|
|
}
|
|
|
|
|
2015-08-26 14:25:40 -04:00
|
|
|
type f2uCvtTab struct {
|
|
|
|
ltf, cvt2U, subf ssa.Op
|
|
|
|
value func(*state, ssa.Type, float64) *ssa.Value
|
|
|
|
}
|
|
|
|
|
|
|
|
var f32_u64 f2uCvtTab = f2uCvtTab{
|
|
|
|
ltf: ssa.OpLess32F,
|
|
|
|
cvt2U: ssa.OpCvt32Fto64,
|
|
|
|
subf: ssa.OpSub32F,
|
|
|
|
value: (*state).constFloat32,
|
|
|
|
}
|
|
|
|
|
|
|
|
var f64_u64 f2uCvtTab = f2uCvtTab{
|
|
|
|
ltf: ssa.OpLess64F,
|
|
|
|
cvt2U: ssa.OpCvt64Fto64,
|
|
|
|
subf: ssa.OpSub64F,
|
|
|
|
value: (*state).constFloat64,
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *state) float32ToUint64(n *Node, x *ssa.Value, ft, tt *Type) *ssa.Value {
|
|
|
|
return s.floatToUint(&f32_u64, n, x, ft, tt)
|
|
|
|
}
|
|
|
|
func (s *state) float64ToUint64(n *Node, x *ssa.Value, ft, tt *Type) *ssa.Value {
|
|
|
|
return s.floatToUint(&f64_u64, n, x, ft, tt)
|
|
|
|
}
|
|
|
|
|
|
|
|
func (s *state) floatToUint(cvttab *f2uCvtTab, n *Node, x *ssa.Value, ft, tt *Type) *ssa.Value {
|
|
|
|
// if x < 9223372036854775808.0 {
|
|
|
|
// result = uintY(x)
|
|
|
|
// } else {
|
|
|
|
// y = x - 9223372036854775808.0
|
|
|
|
// z = uintY(y)
|
|
|
|
// result = z | -9223372036854775808
|
|
|
|
// }
|
|
|
|
twoToThe63 := cvttab.value(s, ft, 9223372036854775808.0)
|
|
|
|
cmp := s.newValue2(cvttab.ltf, Types[TBOOL], x, twoToThe63)
|
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(cmp)
|
2015-08-26 14:25:40 -04:00
|
|
|
b.Likely = ssa.BranchLikely
|
|
|
|
|
|
|
|
bThen := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bElse := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bAfter := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bThen)
|
2015-08-26 14:25:40 -04:00
|
|
|
s.startBlock(bThen)
|
|
|
|
a0 := s.newValue1(cvttab.cvt2U, tt, x)
|
|
|
|
s.vars[n] = a0
|
|
|
|
s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
bThen.AddEdgeTo(bAfter)
|
2015-08-26 14:25:40 -04:00
|
|
|
|
2015-08-28 21:36:29 -05:00
|
|
|
b.AddEdgeTo(bElse)
|
2015-08-26 14:25:40 -04:00
|
|
|
s.startBlock(bElse)
|
|
|
|
y := s.newValue2(cvttab.subf, ft, x, twoToThe63)
|
|
|
|
y = s.newValue1(cvttab.cvt2U, tt, y)
|
|
|
|
z := s.constInt64(tt, -9223372036854775808)
|
|
|
|
a1 := s.newValue2(ssa.OpOr64, tt, y, z)
|
|
|
|
s.vars[n] = a1
|
|
|
|
s.endBlock()
|
2015-08-28 21:36:29 -05:00
|
|
|
bElse.AddEdgeTo(bAfter)
|
2015-08-26 14:25:40 -04:00
|
|
|
|
|
|
|
s.startBlock(bAfter)
|
|
|
|
return s.variable(n, n.Type)
|
|
|
|
}
|
|
|
|
|
2015-09-17 10:31:16 -07:00
|
|
|
// ifaceType returns the value for the word containing the type.
|
|
|
|
// n is the node for the interface expression.
|
|
|
|
// v is the corresponding value.
|
|
|
|
func (s *state) ifaceType(n *Node, v *ssa.Value) *ssa.Value {
|
|
|
|
byteptr := Ptrto(Types[TUINT8]) // type used in runtime prototypes for runtime type (*byte)
|
|
|
|
|
2016-04-01 13:36:24 -07:00
|
|
|
if n.Type.IsEmptyInterface() {
|
2015-09-17 10:31:16 -07:00
|
|
|
// Have *eface. The type is the first word in the struct.
|
|
|
|
return s.newValue1(ssa.OpITab, byteptr, v)
|
|
|
|
}
|
|
|
|
|
|
|
|
// Have *iface.
|
|
|
|
// The first word in the struct is the *itab.
|
|
|
|
// If the *itab is nil, return 0.
|
|
|
|
// Otherwise, the second word in the *itab is the type.
|
|
|
|
|
|
|
|
tab := s.newValue1(ssa.OpITab, byteptr, v)
|
|
|
|
s.vars[&typVar] = tab
|
2016-03-06 18:06:09 -08:00
|
|
|
isnonnil := s.newValue2(ssa.OpNeqPtr, Types[TBOOL], tab, s.constNil(byteptr))
|
2015-09-17 10:31:16 -07:00
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(isnonnil)
|
2015-09-17 10:31:16 -07:00
|
|
|
b.Likely = ssa.BranchLikely
|
|
|
|
|
|
|
|
bLoad := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bEnd := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
|
|
|
|
b.AddEdgeTo(bLoad)
|
|
|
|
b.AddEdgeTo(bEnd)
|
|
|
|
bLoad.AddEdgeTo(bEnd)
|
|
|
|
|
|
|
|
s.startBlock(bLoad)
|
|
|
|
off := s.newValue1I(ssa.OpOffPtr, byteptr, int64(Widthptr), tab)
|
|
|
|
s.vars[&typVar] = s.newValue2(ssa.OpLoad, byteptr, off, s.mem())
|
|
|
|
s.endBlock()
|
|
|
|
|
|
|
|
s.startBlock(bEnd)
|
|
|
|
typ := s.variable(&typVar, byteptr)
|
|
|
|
delete(s.vars, &typVar)
|
|
|
|
return typ
|
|
|
|
}
|
|
|
|
|
|
|
|
// dottype generates SSA for a type assertion node.
|
|
|
|
// commaok indicates whether to panic or return a bool.
|
|
|
|
// If commaok is false, resok will be nil.
|
|
|
|
func (s *state) dottype(n *Node, commaok bool) (res, resok *ssa.Value) {
|
|
|
|
iface := s.expr(n.Left)
|
|
|
|
typ := s.ifaceType(n.Left, iface) // actual concrete type
|
|
|
|
target := s.expr(typename(n.Type)) // target type
|
|
|
|
if !isdirectiface(n.Type) {
|
|
|
|
// walk rewrites ODOTTYPE/OAS2DOTTYPE into runtime calls except for this case.
|
|
|
|
Fatalf("dottype needs a direct iface type %s", n.Type)
|
|
|
|
}
|
|
|
|
|
2015-10-26 17:34:06 -04:00
|
|
|
if Debug_typeassert > 0 {
|
2016-03-02 11:01:25 -08:00
|
|
|
Warnl(n.Lineno, "type assertion inlined")
|
2015-10-26 17:34:06 -04:00
|
|
|
}
|
|
|
|
|
2015-09-17 10:31:16 -07:00
|
|
|
// TODO: If we have a nonempty interface and its itab field is nil,
|
|
|
|
// then this test is redundant and ifaceType should just branch directly to bFail.
|
|
|
|
cond := s.newValue2(ssa.OpEqPtr, Types[TBOOL], typ, target)
|
|
|
|
b := s.endBlock()
|
|
|
|
b.Kind = ssa.BlockIf
|
2016-03-15 20:45:50 -07:00
|
|
|
b.SetControl(cond)
|
2015-09-17 10:31:16 -07:00
|
|
|
b.Likely = ssa.BranchLikely
|
|
|
|
|
|
|
|
byteptr := Ptrto(Types[TUINT8])
|
|
|
|
|
|
|
|
bOk := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
bFail := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
b.AddEdgeTo(bOk)
|
|
|
|
b.AddEdgeTo(bFail)
|
|
|
|
|
|
|
|
if !commaok {
|
|
|
|
// on failure, panic by calling panicdottype
|
|
|
|
s.startBlock(bFail)
|
2016-04-29 09:02:27 -07:00
|
|
|
taddr := s.newValue1A(ssa.OpAddr, byteptr, &ssa.ExternSymbol{Typ: byteptr, Sym: typenamesym(n.Left.Type)}, s.sb)
|
2015-09-18 15:11:30 -07:00
|
|
|
s.rtcall(panicdottype, false, nil, typ, target, taddr)
|
2015-09-17 10:31:16 -07:00
|
|
|
|
|
|
|
// on success, return idata field
|
|
|
|
s.startBlock(bOk)
|
|
|
|
return s.newValue1(ssa.OpIData, n.Type, iface), nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// commaok is the more complicated case because we have
|
|
|
|
// a control flow merge point.
|
|
|
|
bEnd := s.f.NewBlock(ssa.BlockPlain)
|
|
|
|
|
|
|
|
// type assertion succeeded
|
|
|
|
s.startBlock(bOk)
|
|
|
|
s.vars[&idataVar] = s.newValue1(ssa.OpIData, n.Type, iface)
|
|
|
|
s.vars[&okVar] = s.constBool(true)
|
|
|
|
s.endBlock()
|
|
|
|
bOk.AddEdgeTo(bEnd)
|
|
|
|
|
|
|
|
// type assertion failed
|
|
|
|
s.startBlock(bFail)
|
2016-03-06 18:06:09 -08:00
|
|
|
s.vars[&idataVar] = s.constNil(byteptr)
|
2015-09-17 10:31:16 -07:00
|
|
|
s.vars[&okVar] = s.constBool(false)
|
|
|
|
s.endBlock()
|
|
|
|
bFail.AddEdgeTo(bEnd)
|
|
|
|
|
|
|
|
// merge point
|
|
|
|
s.startBlock(bEnd)
|
|
|
|
res = s.variable(&idataVar, byteptr)
|
|
|
|
resok = s.variable(&okVar, Types[TBOOL])
|
|
|
|
delete(s.vars, &idataVar)
|
|
|
|
delete(s.vars, &okVar)
|
|
|
|
return res, resok
|
|
|
|
}
|
|
|
|
|
2015-07-20 15:39:14 -07:00
|
|
|
// checkgoto checks that a goto from from to to does not
|
|
|
|
// jump into a block or jump over variable declarations.
|
|
|
|
// It is a copy of checkgoto in the pre-SSA backend,
|
|
|
|
// modified only for line number handling.
|
|
|
|
// TODO: document how this works and why it is designed the way it is.
|
|
|
|
func (s *state) checkgoto(from *Node, to *Node) {
|
|
|
|
if from.Sym == to.Sym {
|
|
|
|
return
|
|
|
|
}
|
|
|
|
|
|
|
|
nf := 0
|
|
|
|
for fs := from.Sym; fs != nil; fs = fs.Link {
|
|
|
|
nf++
|
|
|
|
}
|
|
|
|
nt := 0
|
|
|
|
for fs := to.Sym; fs != nil; fs = fs.Link {
|
|
|
|
nt++
|
|
|
|
}
|
|
|
|
fs := from.Sym
|
|
|
|
for ; nf > nt; nf-- {
|
|
|
|
fs = fs.Link
|
|
|
|
}
|
|
|
|
if fs != to.Sym {
|
|
|
|
// decide what to complain about.
|
|
|
|
// prefer to complain about 'into block' over declarations,
|
|
|
|
// so scan backward to find most recent block or else dcl.
|
|
|
|
var block *Sym
|
|
|
|
|
|
|
|
var dcl *Sym
|
|
|
|
ts := to.Sym
|
|
|
|
for ; nt > nf; nt-- {
|
|
|
|
if ts.Pkg == nil {
|
|
|
|
block = ts
|
|
|
|
} else {
|
|
|
|
dcl = ts
|
|
|
|
}
|
|
|
|
ts = ts.Link
|
|
|
|
}
|
|
|
|
|
|
|
|
for ts != fs {
|
|
|
|
if ts.Pkg == nil {
|
|
|
|
block = ts
|
|
|
|
} else {
|
|
|
|
dcl = ts
|
|
|
|
}
|
|
|
|
ts = ts.Link
|
|
|
|
fs = fs.Link
|
|
|
|
}
|
|
|
|
|
2016-03-02 11:01:25 -08:00
|
|
|
lno := from.Left.Lineno
|
2015-07-20 15:39:14 -07:00
|
|
|
if block != nil {
|
2016-03-02 11:30:29 -08:00
|
|
|
yyerrorl(lno, "goto %v jumps into block starting at %v", from.Left.Sym, linestr(block.Lastlineno))
|
2015-07-20 15:39:14 -07:00
|
|
|
} else {
|
2016-03-02 11:30:29 -08:00
|
|
|
yyerrorl(lno, "goto %v jumps over declaration of %v at %v", from.Left.Sym, dcl, linestr(dcl.Lastlineno))
|
2015-07-20 15:39:14 -07:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-04-15 15:51:25 -07:00
|
|
|
// variable returns the value of a variable at the current location.
|
2015-06-19 21:02:28 -07:00
|
|
|
func (s *state) variable(name *Node, t ssa.Type) *ssa.Value {
|
2015-04-15 15:51:25 -07:00
|
|
|
v := s.vars[name]
|
|
|
|
if v == nil {
|
2015-06-11 21:29:25 -07:00
|
|
|
v = s.newValue0A(ssa.OpFwdRef, t, name)
|
2016-01-14 16:02:23 -08:00
|
|
|
s.fwdRefs = append(s.fwdRefs, v)
|
2015-04-15 15:51:25 -07:00
|
|
|
s.vars[name] = v
|
2016-01-14 16:02:23 -08:00
|
|
|
s.addNamedValue(name, v)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
return v
|
|
|
|
}
|
|
|
|
|
2015-05-18 16:44:20 -07:00
|
|
|
func (s *state) mem() *ssa.Value {
|
2015-09-17 16:45:10 -07:00
|
|
|
return s.variable(&memVar, ssa.TypeMem)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
|
cmd/compile: use sparse algorithm for phis in large program
This adds a sparse method for locating nearest ancestors
in a dominator tree, and checks blocks with more than one
predecessor for differences and inserts phi functions where
there are.
Uses reversed post order to cut number of passes, running
it from first def to last use ("last use" for paramout and
mem is end-of-program; last use for a phi input from a
backedge is the source of the back edge)
Includes a cutover from old algorithm to new to avoid paying
large constant factor for small programs. This keeps normal
builds running at about the same time, while not running
over-long on large machine-generated inputs.
Add "phase" flags for ssa/build -- ssa/build/stats prints
number of blocks, values (before and after linking references
and inserting phis, so expansion can be measured), and their
product; the product governs the cutover, where a good value
seems to be somewhere between 1 and 5 million.
Among the files compiled by make.bash, this is the shape of
the tail of the distribution for #blocks, #vars, and their
product:
#blocks #vars product
max 6171 28180 173,898,780
99.9% 1641 6548 10,401,878
99% 463 1909 873,721
95% 152 639 95,235
90% 84 359 30,021
The old algorithm is indeed usually fastest, for 99%ile
values of usually.
The fix to LookupVarOutgoing
( https://go-review.googlesource.com/#/c/22790/ )
deals with some of the same problems addressed by this CL,
but on at least one bug ( #15537 ) this change is still
a significant help.
With this CL:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 4m35.200s
user 13m16.644s
sys 0m36.712s
and pprof reports 3.4GB allocated in one of the larger profiles
With tip:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 10m36.569s
user 25m52.286s
sys 4m3.696s
and pprof reports 8.3GB allocated in the same larger profile
With this CL, most of the compilation time on the benchmarked
input is spent in register/stack allocation (cumulative 53%)
and in the sparse lookup algorithm itself (cumulative 20%).
Fixes #15537.
Change-Id: Ia0299dda6a291534d8b08e5f9883216ded677a00
Reviewed-on: https://go-review.googlesource.com/22342
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-04-21 13:24:58 -04:00
|
|
|
func (s *state) linkForwardReferences(dm *sparseDefState) {
|
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// Build SSA graph. Each variable on its first use in a basic block
|
2015-04-15 15:51:25 -07:00
|
|
|
// leaves a FwdRef in that block representing the incoming value
|
2016-03-01 23:21:55 +00:00
|
|
|
// of that variable. This function links that ref up with possible definitions,
|
|
|
|
// inserting Phi values as needed. This is essentially the algorithm
|
2016-01-14 16:02:23 -08:00
|
|
|
// described by Braun, Buchwald, Hack, Leißa, Mallon, and Zwinkau:
|
2015-04-15 15:51:25 -07:00
|
|
|
// http://pp.info.uni-karlsruhe.de/uploads/publikationen/braun13cc.pdf
|
2016-01-14 16:02:23 -08:00
|
|
|
// Differences:
|
|
|
|
// - We use FwdRef nodes to postpone phi building until the CFG is
|
2016-03-01 23:21:55 +00:00
|
|
|
// completely built. That way we can avoid the notion of "sealed"
|
2016-01-14 16:02:23 -08:00
|
|
|
// blocks.
|
|
|
|
// - Phi optimization is a separate pass (in ../ssa/phielim.go).
|
|
|
|
for len(s.fwdRefs) > 0 {
|
|
|
|
v := s.fwdRefs[len(s.fwdRefs)-1]
|
|
|
|
s.fwdRefs = s.fwdRefs[:len(s.fwdRefs)-1]
|
cmd/compile: use sparse algorithm for phis in large program
This adds a sparse method for locating nearest ancestors
in a dominator tree, and checks blocks with more than one
predecessor for differences and inserts phi functions where
there are.
Uses reversed post order to cut number of passes, running
it from first def to last use ("last use" for paramout and
mem is end-of-program; last use for a phi input from a
backedge is the source of the back edge)
Includes a cutover from old algorithm to new to avoid paying
large constant factor for small programs. This keeps normal
builds running at about the same time, while not running
over-long on large machine-generated inputs.
Add "phase" flags for ssa/build -- ssa/build/stats prints
number of blocks, values (before and after linking references
and inserting phis, so expansion can be measured), and their
product; the product governs the cutover, where a good value
seems to be somewhere between 1 and 5 million.
Among the files compiled by make.bash, this is the shape of
the tail of the distribution for #blocks, #vars, and their
product:
#blocks #vars product
max 6171 28180 173,898,780
99.9% 1641 6548 10,401,878
99% 463 1909 873,721
95% 152 639 95,235
90% 84 359 30,021
The old algorithm is indeed usually fastest, for 99%ile
values of usually.
The fix to LookupVarOutgoing
( https://go-review.googlesource.com/#/c/22790/ )
deals with some of the same problems addressed by this CL,
but on at least one bug ( #15537 ) this change is still
a significant help.
With this CL:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 4m35.200s
user 13m16.644s
sys 0m36.712s
and pprof reports 3.4GB allocated in one of the larger profiles
With tip:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 10m36.569s
user 25m52.286s
sys 4m3.696s
and pprof reports 8.3GB allocated in the same larger profile
With this CL, most of the compilation time on the benchmarked
input is spent in register/stack allocation (cumulative 53%)
and in the sparse lookup algorithm itself (cumulative 20%).
Fixes #15537.
Change-Id: Ia0299dda6a291534d8b08e5f9883216ded677a00
Reviewed-on: https://go-review.googlesource.com/22342
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-04-21 13:24:58 -04:00
|
|
|
s.resolveFwdRef(v, dm)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-01-14 16:02:23 -08:00
|
|
|
// resolveFwdRef modifies v to be the variable's value at the start of its block.
|
|
|
|
// v must be a FwdRef op.
|
cmd/compile: use sparse algorithm for phis in large program
This adds a sparse method for locating nearest ancestors
in a dominator tree, and checks blocks with more than one
predecessor for differences and inserts phi functions where
there are.
Uses reversed post order to cut number of passes, running
it from first def to last use ("last use" for paramout and
mem is end-of-program; last use for a phi input from a
backedge is the source of the back edge)
Includes a cutover from old algorithm to new to avoid paying
large constant factor for small programs. This keeps normal
builds running at about the same time, while not running
over-long on large machine-generated inputs.
Add "phase" flags for ssa/build -- ssa/build/stats prints
number of blocks, values (before and after linking references
and inserting phis, so expansion can be measured), and their
product; the product governs the cutover, where a good value
seems to be somewhere between 1 and 5 million.
Among the files compiled by make.bash, this is the shape of
the tail of the distribution for #blocks, #vars, and their
product:
#blocks #vars product
max 6171 28180 173,898,780
99.9% 1641 6548 10,401,878
99% 463 1909 873,721
95% 152 639 95,235
90% 84 359 30,021
The old algorithm is indeed usually fastest, for 99%ile
values of usually.
The fix to LookupVarOutgoing
( https://go-review.googlesource.com/#/c/22790/ )
deals with some of the same problems addressed by this CL,
but on at least one bug ( #15537 ) this change is still
a significant help.
With this CL:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 4m35.200s
user 13m16.644s
sys 0m36.712s
and pprof reports 3.4GB allocated in one of the larger profiles
With tip:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 10m36.569s
user 25m52.286s
sys 4m3.696s
and pprof reports 8.3GB allocated in the same larger profile
With this CL, most of the compilation time on the benchmarked
input is spent in register/stack allocation (cumulative 53%)
and in the sparse lookup algorithm itself (cumulative 20%).
Fixes #15537.
Change-Id: Ia0299dda6a291534d8b08e5f9883216ded677a00
Reviewed-on: https://go-review.googlesource.com/22342
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-04-21 13:24:58 -04:00
|
|
|
func (s *state) resolveFwdRef(v *ssa.Value, dm *sparseDefState) {
|
2016-01-14 16:02:23 -08:00
|
|
|
b := v.Block
|
|
|
|
name := v.Aux.(*Node)
|
|
|
|
v.Aux = nil
|
2015-04-15 15:51:25 -07:00
|
|
|
if b == s.f.Entry {
|
2016-01-14 16:02:23 -08:00
|
|
|
// Live variable at start of function.
|
2016-02-27 17:49:31 -08:00
|
|
|
if s.canSSA(name) {
|
2016-04-08 13:33:43 -04:00
|
|
|
if strings.HasPrefix(name.Sym.Name, "autotmp_") {
|
|
|
|
// It's likely that this is an uninitialized variable in the entry block.
|
|
|
|
s.Fatalf("Treating auto as if it were arg, func %s, node %v, value %v", b.Func.Name, name, v)
|
|
|
|
}
|
2016-01-14 16:02:23 -08:00
|
|
|
v.Op = ssa.OpArg
|
|
|
|
v.Aux = name
|
|
|
|
return
|
2015-11-02 08:10:26 -08:00
|
|
|
}
|
2016-03-01 23:21:55 +00:00
|
|
|
// Not SSAable. Load it.
|
2015-06-19 21:02:28 -07:00
|
|
|
addr := s.decladdrs[name]
|
|
|
|
if addr == nil {
|
|
|
|
// TODO: closure args reach here.
|
2015-10-08 17:14:12 -04:00
|
|
|
s.Unimplementedf("unhandled closure arg %s at entry to function %s", name, b.Func.Name)
|
2015-06-19 21:02:28 -07:00
|
|
|
}
|
|
|
|
if _, ok := addr.Aux.(*ssa.ArgSymbol); !ok {
|
|
|
|
s.Fatalf("variable live at start of function %s is not an argument %s", b.Func.Name, name)
|
|
|
|
}
|
2016-01-14 16:02:23 -08:00
|
|
|
v.Op = ssa.OpLoad
|
|
|
|
v.AddArgs(addr, s.startmem)
|
|
|
|
return
|
|
|
|
}
|
|
|
|
if len(b.Preds) == 0 {
|
|
|
|
// This block is dead; we have no predecessors and we're not the entry block.
|
|
|
|
// It doesn't matter what we use here as long as it is well-formed.
|
|
|
|
v.Op = ssa.OpUnknown
|
|
|
|
return
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
2016-01-14 16:02:23 -08:00
|
|
|
// Find variable value on each predecessor.
|
|
|
|
var argstore [4]*ssa.Value
|
|
|
|
args := argstore[:0]
|
2016-04-28 16:52:47 -07:00
|
|
|
for _, e := range b.Preds {
|
|
|
|
p := e.Block()
|
cmd/compile: use sparse algorithm for phis in large program
This adds a sparse method for locating nearest ancestors
in a dominator tree, and checks blocks with more than one
predecessor for differences and inserts phi functions where
there are.
Uses reversed post order to cut number of passes, running
it from first def to last use ("last use" for paramout and
mem is end-of-program; last use for a phi input from a
backedge is the source of the back edge)
Includes a cutover from old algorithm to new to avoid paying
large constant factor for small programs. This keeps normal
builds running at about the same time, while not running
over-long on large machine-generated inputs.
Add "phase" flags for ssa/build -- ssa/build/stats prints
number of blocks, values (before and after linking references
and inserting phis, so expansion can be measured), and their
product; the product governs the cutover, where a good value
seems to be somewhere between 1 and 5 million.
Among the files compiled by make.bash, this is the shape of
the tail of the distribution for #blocks, #vars, and their
product:
#blocks #vars product
max 6171 28180 173,898,780
99.9% 1641 6548 10,401,878
99% 463 1909 873,721
95% 152 639 95,235
90% 84 359 30,021
The old algorithm is indeed usually fastest, for 99%ile
values of usually.
The fix to LookupVarOutgoing
( https://go-review.googlesource.com/#/c/22790/ )
deals with some of the same problems addressed by this CL,
but on at least one bug ( #15537 ) this change is still
a significant help.
With this CL:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 4m35.200s
user 13m16.644s
sys 0m36.712s
and pprof reports 3.4GB allocated in one of the larger profiles
With tip:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
github.com/gogo/protobuf/test/theproto3/combos/...
...
real 10m36.569s
user 25m52.286s
sys 4m3.696s
and pprof reports 8.3GB allocated in the same larger profile
With this CL, most of the compilation time on the benchmarked
input is spent in register/stack allocation (cumulative 53%)
and in the sparse lookup algorithm itself (cumulative 20%).
Fixes #15537.
Change-Id: Ia0299dda6a291534d8b08e5f9883216ded677a00
Reviewed-on: https://go-review.googlesource.com/22342
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-04-21 13:24:58 -04:00
|
|
|
p = dm.FindBetterDefiningBlock(name, p) // try sparse improvement on p
|
2016-01-14 16:02:23 -08:00
|
|
|
args = append(args, s.lookupVarOutgoing(p, v.Type, name, v.Line))
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
2016-01-14 16:02:23 -08:00
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// Decide if we need a phi or not. We need a phi if there
|
2016-01-14 16:02:23 -08:00
|
|
|
// are two different args (which are both not v).
|
|
|
|
var w *ssa.Value
|
|
|
|
for _, a := range args {
|
|
|
|
if a == v {
|
|
|
|
continue // self-reference
|
2015-07-20 15:39:14 -07:00
|
|
|
}
|
2016-01-14 16:02:23 -08:00
|
|
|
if a == w {
|
|
|
|
continue // already have this witness
|
|
|
|
}
|
|
|
|
if w != nil {
|
|
|
|
// two witnesses, need a phi value
|
|
|
|
v.Op = ssa.OpPhi
|
|
|
|
v.AddArgs(args...)
|
|
|
|
return
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
2016-01-14 16:02:23 -08:00
|
|
|
w = a // save witness
|
|
|
|
}
|
|
|
|
if w == nil {
|
|
|
|
s.Fatalf("no witness for reachable phi %s", v)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
2016-03-01 23:21:55 +00:00
|
|
|
// One witness. Make v a copy of w.
|
2016-01-14 16:02:23 -08:00
|
|
|
v.Op = ssa.OpCopy
|
|
|
|
v.AddArg(w)
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
// lookupVarOutgoing finds the variable's value at the end of block b.
|
2016-01-14 16:02:23 -08:00
|
|
|
func (s *state) lookupVarOutgoing(b *ssa.Block, t ssa.Type, name *Node, line int32) *ssa.Value {
|
2016-05-04 14:04:03 -07:00
|
|
|
for {
|
|
|
|
if v, ok := s.defvars[b.ID][name]; ok {
|
|
|
|
return v
|
|
|
|
}
|
|
|
|
// The variable is not defined by b and we haven't looked it up yet.
|
|
|
|
// If b has exactly one predecessor, loop to look it up there.
|
|
|
|
// Otherwise, give up and insert a new FwdRef and resolve it later.
|
|
|
|
if len(b.Preds) != 1 {
|
|
|
|
break
|
|
|
|
}
|
|
|
|
b = b.Preds[0].Block()
|
2015-04-15 15:51:25 -07:00
|
|
|
}
|
2016-05-04 14:04:03 -07:00
|
|
|
// Generate a FwdRef for the variable and return that.
|
2016-01-14 16:02:23 -08:00
|
|
|
v := b.NewValue0A(line, ssa.OpFwdRef, t, name)
|
|
|
|
s.fwdRefs = append(s.fwdRefs, v)
|
2016-05-04 14:04:03 -07:00
|
|
|
s.defvars[b.ID][name] = v
|
2016-01-14 16:02:23 -08:00
|
|
|
s.addNamedValue(name, v)
|
2015-04-15 15:51:25 -07:00
|
|
|
return v
|
|
|
|
}
|
|
|
|
|
2015-10-22 14:22:38 -07:00
|
|
|
func (s *state) addNamedValue(n *Node, v *ssa.Value) {
|
|
|
|
if n.Class == Pxxx {
|
|
|
|
// Don't track our dummy nodes (&memVar etc.).
|
|
|
|
return
|
|
|
|
}
|
|
|
|
if strings.HasPrefix(n.Sym.Name, "autotmp_") {
|
|
|
|
// Don't track autotmp_ variables.
|
|
|
|
return
|
|
|
|
}
|
2016-03-08 20:09:48 -08:00
|
|
|
if n.Class == PPARAMOUT {
|
|
|
|
// Don't track named output values. This prevents return values
|
|
|
|
// from being assigned too early. See #14591 and #14762. TODO: allow this.
|
|
|
|
return
|
|
|
|
}
|
2015-10-22 14:22:38 -07:00
|
|
|
if n.Class == PAUTO && n.Xoffset != 0 {
|
|
|
|
s.Fatalf("AUTO var with offset %s %d", n, n.Xoffset)
|
|
|
|
}
|
2015-11-02 08:10:26 -08:00
|
|
|
loc := ssa.LocalSlot{N: n, Type: n.Type, Off: 0}
|
|
|
|
values, ok := s.f.NamedValues[loc]
|
2015-10-22 14:22:38 -07:00
|
|
|
if !ok {
|
2015-11-02 08:10:26 -08:00
|
|
|
s.f.Names = append(s.f.Names, loc)
|
2015-10-22 14:22:38 -07:00
|
|
|
}
|
2015-11-02 08:10:26 -08:00
|
|
|
s.f.NamedValues[loc] = append(values, v)
|
2015-10-22 14:22:38 -07:00
|
|
|
}
|
|
|
|
|
2016-03-12 14:07:40 -08:00
|
|
|
// Branch is an unresolved branch.
|
|
|
|
type Branch struct {
|
|
|
|
P *obj.Prog // branch instruction
|
|
|
|
B *ssa.Block // target
|
2015-05-12 11:06:44 -07:00
|
|
|
}
|
|
|
|
|
2016-03-12 14:07:40 -08:00
|
|
|
// SSAGenState contains state needed during Prog generation.
|
|
|
|
type SSAGenState struct {
|
|
|
|
// Branches remembers all the branch instructions we've seen
|
2015-08-28 22:51:01 -07:00
|
|
|
// and where they would like to go.
|
2016-03-12 14:07:40 -08:00
|
|
|
Branches []Branch
|
2015-08-28 22:51:01 -07:00
|
|
|
|
|
|
|
// bstart remembers where each block starts (indexed by block ID)
|
|
|
|
bstart []*obj.Prog
|
2016-07-26 11:51:33 -07:00
|
|
|
|
|
|
|
// 387 port: maps from SSE registers (REG_X?) to 387 registers (REG_F?)
|
2016-08-10 11:44:57 -07:00
|
|
|
SSEto387 map[int16]int16
|
|
|
|
// Some architectures require a 64-bit temporary for FP-related register shuffling. Examples include x86-387, PPC, and Sparc V8.
|
|
|
|
ScratchFpMem *Node
|
2015-08-28 22:51:01 -07:00
|
|
|
}
|
|
|
|
|
2016-03-12 14:07:40 -08:00
|
|
|
// Pc returns the current Prog.
|
|
|
|
func (s *SSAGenState) Pc() *obj.Prog {
|
|
|
|
return Pc
|
|
|
|
}
|
|
|
|
|
|
|
|
// SetLineno sets the current source line number.
|
|
|
|
func (s *SSAGenState) SetLineno(l int32) {
|
|
|
|
lineno = l
|
|
|
|
}
|
|
|
|
|
2015-05-12 11:06:44 -07:00
|
|
|
// genssa appends entries to ptxt for each instruction in f.
|
|
|
|
// gcargs and gclocals are filled in with pointer maps for the frame.
|
|
|
|
func genssa(f *ssa.Func, ptxt *obj.Prog, gcargs, gclocals *Sym) {
|
2016-03-12 14:07:40 -08:00
|
|
|
var s SSAGenState
|
2015-08-28 22:51:01 -07:00
|
|
|
|
2015-07-22 13:13:53 -07:00
|
|
|
e := f.Config.Frontend().(*ssaExport)
|
|
|
|
// We're about to emit a bunch of Progs.
|
|
|
|
// Since the only way to get here is to explicitly request it,
|
|
|
|
// just fail on unimplemented instead of trying to unwind our mess.
|
|
|
|
e.mustImplement = true
|
|
|
|
|
2015-05-12 11:06:44 -07:00
|
|
|
// Remember where each block starts.
|
2015-08-28 22:51:01 -07:00
|
|
|
s.bstart = make([]*obj.Prog, f.NumBlocks())
|
2015-05-12 11:06:44 -07:00
|
|
|
|
2015-07-31 14:37:15 -07:00
|
|
|
var valueProgs map[*obj.Prog]*ssa.Value
|
|
|
|
var blockProgs map[*obj.Prog]*ssa.Block
|
2016-03-17 13:46:43 +11:00
|
|
|
var logProgs = e.log
|
2015-07-31 14:37:15 -07:00
|
|
|
if logProgs {
|
|
|
|
valueProgs = make(map[*obj.Prog]*ssa.Value, f.NumValues())
|
|
|
|
blockProgs = make(map[*obj.Prog]*ssa.Block, f.NumBlocks())
|
|
|
|
f.Logf("genssa %s\n", f.Name)
|
|
|
|
blockProgs[Pc] = f.Blocks[0]
|
|
|
|
}
|
|
|
|
|
2016-07-26 11:51:33 -07:00
|
|
|
if Thearch.Use387 {
|
|
|
|
s.SSEto387 = map[int16]int16{}
|
2016-08-10 11:44:57 -07:00
|
|
|
}
|
|
|
|
if f.Config.NeedsFpScratch {
|
|
|
|
s.ScratchFpMem = temp(Types[TUINT64])
|
2016-07-26 11:51:33 -07:00
|
|
|
}
|
|
|
|
|
2015-05-12 11:06:44 -07:00
|
|
|
// Emit basic blocks
|
|
|
|
for i, b := range f.Blocks {
|
2015-08-28 22:51:01 -07:00
|
|
|
s.bstart[b.ID] = Pc
|
2015-05-12 11:06:44 -07:00
|
|
|
// Emit values in block
|
2016-03-12 14:07:40 -08:00
|
|
|
Thearch.SSAMarkMoves(&s, b)
|
2015-05-12 11:06:44 -07:00
|
|
|
for _, v := range b.Values {
|
2015-07-31 14:37:15 -07:00
|
|
|
x := Pc
|
2016-03-12 14:07:40 -08:00
|
|
|
Thearch.SSAGenValue(&s, v)
|
2015-07-31 14:37:15 -07:00
|
|
|
if logProgs {
|
|
|
|
for ; x != Pc; x = x.Link {
|
|
|
|
valueProgs[x] = v
|
|
|
|
}
|
|
|
|
}
|
2015-05-12 11:06:44 -07:00
|
|
|
}
|
|
|
|
// Emit control flow instructions for block
|
|
|
|
var next *ssa.Block
|
2016-02-26 16:32:01 -08:00
|
|
|
if i < len(f.Blocks)-1 && (Debug['N'] == 0 || b.Kind == ssa.BlockCall) {
|
2016-02-22 23:19:00 -08:00
|
|
|
// If -N, leave next==nil so every block with successors
|
2016-02-26 16:32:01 -08:00
|
|
|
// ends in a JMP (except call blocks - plive doesn't like
|
|
|
|
// select{send,recv} followed by a JMP call). Helps keep
|
|
|
|
// line numbers for otherwise empty blocks.
|
2015-05-12 11:06:44 -07:00
|
|
|
next = f.Blocks[i+1]
|
|
|
|
}
|
2015-07-31 14:37:15 -07:00
|
|
|
x := Pc
|
2016-03-12 14:07:40 -08:00
|
|
|
Thearch.SSAGenBlock(&s, b, next)
|
2015-07-31 14:37:15 -07:00
|
|
|
if logProgs {
|
|
|
|
for ; x != Pc; x = x.Link {
|
|
|
|
blockProgs[x] = b
|
|
|
|
}
|
|
|
|
}
|
2015-05-12 11:06:44 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
// Resolve branches
|
2016-03-12 14:07:40 -08:00
|
|
|
for _, br := range s.Branches {
|
|
|
|
br.P.To.Val = s.bstart[br.B.ID]
|
2015-08-28 22:51:01 -07:00
|
|
|
}
|
2015-05-12 11:06:44 -07:00
|
|
|
|
2015-07-31 14:37:15 -07:00
|
|
|
if logProgs {
|
|
|
|
for p := ptxt; p != nil; p = p.Link {
|
|
|
|
var s string
|
|
|
|
if v, ok := valueProgs[p]; ok {
|
|
|
|
s = v.String()
|
|
|
|
} else if b, ok := blockProgs[p]; ok {
|
|
|
|
s = b.String()
|
|
|
|
} else {
|
|
|
|
s = " " // most value and branch strings are 2-3 characters long
|
|
|
|
}
|
|
|
|
f.Logf("%s\t%s\n", s, p)
|
|
|
|
}
|
2015-08-10 12:15:52 -07:00
|
|
|
if f.Config.HTML != nil {
|
|
|
|
saved := ptxt.Ctxt.LineHist.PrintFilenameOnly
|
|
|
|
ptxt.Ctxt.LineHist.PrintFilenameOnly = true
|
|
|
|
var buf bytes.Buffer
|
|
|
|
buf.WriteString("<code>")
|
|
|
|
buf.WriteString("<dl class=\"ssa-gen\">")
|
|
|
|
for p := ptxt; p != nil; p = p.Link {
|
|
|
|
buf.WriteString("<dt class=\"ssa-prog-src\">")
|
|
|
|
if v, ok := valueProgs[p]; ok {
|
|
|
|
buf.WriteString(v.HTML())
|
|
|
|
} else if b, ok := blockProgs[p]; ok {
|
|
|
|
buf.WriteString(b.HTML())
|
|
|
|
}
|
|
|
|
buf.WriteString("</dt>")
|
|
|
|
buf.WriteString("<dd class=\"ssa-prog\">")
|
|
|
|
buf.WriteString(html.EscapeString(p.String()))
|
|
|
|
buf.WriteString("</dd>")
|
|
|
|
buf.WriteString("</li>")
|
|
|
|
}
|
|
|
|
buf.WriteString("</dl>")
|
|
|
|
buf.WriteString("</code>")
|
|
|
|
f.Config.HTML.WriteColumn("genssa", buf.String())
|
|
|
|
ptxt.Ctxt.LineHist.PrintFilenameOnly = saved
|
|
|
|
}
|
2015-07-31 14:37:15 -07:00
|
|
|
}
|
|
|
|
|
2015-07-28 10:56:39 -07:00
|
|
|
// Emit static data
|
|
|
|
if f.StaticData != nil {
|
|
|
|
for _, n := range f.StaticData.([]*Node) {
|
|
|
|
if !gen_as_init(n, false) {
|
2016-04-29 09:02:27 -07:00
|
|
|
Fatalf("non-static data marked as static: %v\n\n", n)
|
2015-07-28 10:56:39 -07:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-08-24 02:16:19 -07:00
|
|
|
// Allocate stack frame
|
|
|
|
allocauto(ptxt)
|
|
|
|
|
|
|
|
// Generate gc bitmaps.
|
|
|
|
liveness(Curfn, ptxt, gcargs, gclocals)
|
2015-05-12 11:06:44 -07:00
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// Add frame prologue. Zero ambiguously live variables.
|
2015-08-24 02:16:19 -07:00
|
|
|
Thearch.Defframe(ptxt)
|
|
|
|
if Debug['f'] != 0 {
|
|
|
|
frame(0)
|
|
|
|
}
|
2015-05-12 11:06:44 -07:00
|
|
|
|
2015-08-24 02:16:19 -07:00
|
|
|
// Remove leftover instrumentation from the instruction stream.
|
|
|
|
removevardef(ptxt)
|
2015-08-10 12:15:52 -07:00
|
|
|
|
|
|
|
f.Config.HTML.Close()
|
2015-05-12 11:06:44 -07:00
|
|
|
}
|
|
|
|
|
2016-03-12 14:07:40 -08:00
|
|
|
type FloatingEQNEJump struct {
|
|
|
|
Jump obj.As
|
|
|
|
Index int
|
2015-08-18 14:39:26 -04:00
|
|
|
}
|
|
|
|
|
2016-03-12 14:07:40 -08:00
|
|
|
func oneFPJump(b *ssa.Block, jumps *FloatingEQNEJump, likely ssa.BranchPrediction, branches []Branch) []Branch {
|
|
|
|
p := Prog(jumps.Jump)
|
2015-08-18 14:39:26 -04:00
|
|
|
p.To.Type = obj.TYPE_BRANCH
|
2016-03-12 14:07:40 -08:00
|
|
|
to := jumps.Index
|
2016-04-28 16:52:47 -07:00
|
|
|
branches = append(branches, Branch{p, b.Succs[to].Block()})
|
2015-08-18 14:39:26 -04:00
|
|
|
if to == 1 {
|
|
|
|
likely = -likely
|
|
|
|
}
|
|
|
|
// liblink reorders the instruction stream as it sees fit.
|
|
|
|
// Pass along what we know so liblink can make use of it.
|
|
|
|
// TODO: Once we've fully switched to SSA,
|
|
|
|
// make liblink leave our output alone.
|
|
|
|
switch likely {
|
|
|
|
case ssa.BranchUnlikely:
|
|
|
|
p.From.Type = obj.TYPE_CONST
|
|
|
|
p.From.Offset = 0
|
|
|
|
case ssa.BranchLikely:
|
|
|
|
p.From.Type = obj.TYPE_CONST
|
|
|
|
p.From.Offset = 1
|
|
|
|
}
|
|
|
|
return branches
|
|
|
|
}
|
|
|
|
|
2016-03-12 14:07:40 -08:00
|
|
|
func SSAGenFPJump(s *SSAGenState, b, next *ssa.Block, jumps *[2][2]FloatingEQNEJump) {
|
2015-08-18 14:39:26 -04:00
|
|
|
likely := b.Likely
|
|
|
|
switch next {
|
2016-04-28 16:52:47 -07:00
|
|
|
case b.Succs[0].Block():
|
2016-03-12 14:07:40 -08:00
|
|
|
s.Branches = oneFPJump(b, &jumps[0][0], likely, s.Branches)
|
|
|
|
s.Branches = oneFPJump(b, &jumps[0][1], likely, s.Branches)
|
2016-04-28 16:52:47 -07:00
|
|
|
case b.Succs[1].Block():
|
2016-03-12 14:07:40 -08:00
|
|
|
s.Branches = oneFPJump(b, &jumps[1][0], likely, s.Branches)
|
|
|
|
s.Branches = oneFPJump(b, &jumps[1][1], likely, s.Branches)
|
2015-08-18 14:39:26 -04:00
|
|
|
default:
|
2016-03-12 14:07:40 -08:00
|
|
|
s.Branches = oneFPJump(b, &jumps[1][0], likely, s.Branches)
|
|
|
|
s.Branches = oneFPJump(b, &jumps[1][1], likely, s.Branches)
|
2015-08-18 14:39:26 -04:00
|
|
|
q := Prog(obj.AJMP)
|
|
|
|
q.To.Type = obj.TYPE_BRANCH
|
2016-04-28 16:52:47 -07:00
|
|
|
s.Branches = append(s.Branches, Branch{q, b.Succs[1].Block()})
|
2015-05-12 11:06:44 -07:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-06-24 14:37:17 -05:00
|
|
|
func AuxOffset(v *ssa.Value) (offset int64) {
|
|
|
|
if v.Aux == nil {
|
|
|
|
return 0
|
|
|
|
}
|
|
|
|
switch sym := v.Aux.(type) {
|
|
|
|
|
|
|
|
case *ssa.AutoSymbol:
|
|
|
|
n := sym.Node.(*Node)
|
|
|
|
return n.Xoffset
|
|
|
|
}
|
|
|
|
return 0
|
|
|
|
}
|
|
|
|
|
2016-03-12 14:07:40 -08:00
|
|
|
// AddAux adds the offset in the aux fields (AuxInt and Aux) of v to a.
|
|
|
|
func AddAux(a *obj.Addr, v *ssa.Value) {
|
|
|
|
AddAux2(a, v, v.AuxInt)
|
2015-10-21 13:13:56 -07:00
|
|
|
}
|
2016-03-12 14:07:40 -08:00
|
|
|
func AddAux2(a *obj.Addr, v *ssa.Value, offset int64) {
|
2016-05-06 10:11:41 -07:00
|
|
|
if a.Type != obj.TYPE_MEM && a.Type != obj.TYPE_ADDR {
|
2016-04-29 09:02:27 -07:00
|
|
|
v.Fatalf("bad AddAux addr %v", a)
|
2015-06-19 21:02:28 -07:00
|
|
|
}
|
|
|
|
// add integer offset
|
2015-10-21 13:13:56 -07:00
|
|
|
a.Offset += offset
|
2015-06-19 21:02:28 -07:00
|
|
|
|
|
|
|
// If no additional symbol offset, we're done.
|
|
|
|
if v.Aux == nil {
|
|
|
|
return
|
|
|
|
}
|
|
|
|
// Add symbol's offset from its base register.
|
|
|
|
switch sym := v.Aux.(type) {
|
|
|
|
case *ssa.ExternSymbol:
|
|
|
|
a.Name = obj.NAME_EXTERN
|
2016-03-16 22:22:58 -07:00
|
|
|
switch s := sym.Sym.(type) {
|
|
|
|
case *Sym:
|
|
|
|
a.Sym = Linksym(s)
|
|
|
|
case *obj.LSym:
|
|
|
|
a.Sym = s
|
|
|
|
default:
|
|
|
|
v.Fatalf("ExternSymbol.Sym is %T", s)
|
|
|
|
}
|
2015-06-19 21:02:28 -07:00
|
|
|
case *ssa.ArgSymbol:
|
2015-08-24 02:16:19 -07:00
|
|
|
n := sym.Node.(*Node)
|
|
|
|
a.Name = obj.NAME_PARAM
|
|
|
|
a.Node = n
|
|
|
|
a.Sym = Linksym(n.Orig.Sym)
|
|
|
|
a.Offset += n.Xoffset // TODO: why do I have to add this here? I don't for auto variables.
|
2015-06-19 21:02:28 -07:00
|
|
|
case *ssa.AutoSymbol:
|
2015-08-24 02:16:19 -07:00
|
|
|
n := sym.Node.(*Node)
|
|
|
|
a.Name = obj.NAME_AUTO
|
|
|
|
a.Node = n
|
|
|
|
a.Sym = Linksym(n.Sym)
|
2015-06-19 21:02:28 -07:00
|
|
|
default:
|
|
|
|
v.Fatalf("aux in %s not implemented %#v", v, v.Aux)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-06-27 16:54:57 -04:00
|
|
|
// SizeAlignAuxInt returns an AuxInt encoding the size and alignment of type t.
|
|
|
|
func SizeAlignAuxInt(t *Type) int64 {
|
|
|
|
return ssa.MakeSizeAndAlign(t.Size(), t.Alignment()).Int64()
|
|
|
|
}
|
|
|
|
|
2015-11-02 21:28:13 -08:00
|
|
|
// extendIndex extends v to a full int width.
|
2016-05-25 09:49:28 -04:00
|
|
|
// panic using the given function if v does not fit in an int (only on 32-bit archs).
|
|
|
|
func (s *state) extendIndex(v *ssa.Value, panicfn *Node) *ssa.Value {
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
size := v.Type.Size()
|
2015-11-02 21:28:13 -08:00
|
|
|
if size == s.config.IntSize {
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
return v
|
|
|
|
}
|
2015-11-02 21:28:13 -08:00
|
|
|
if size > s.config.IntSize {
|
2016-05-25 09:49:28 -04:00
|
|
|
// truncate 64-bit indexes on 32-bit pointer archs. Test the
|
|
|
|
// high word and branch to out-of-bounds failure if it is not 0.
|
|
|
|
if Debug['B'] == 0 {
|
|
|
|
hi := s.newValue1(ssa.OpInt64Hi, Types[TUINT32], v)
|
|
|
|
cmp := s.newValue2(ssa.OpEq32, Types[TBOOL], hi, s.constInt32(Types[TUINT32], 0))
|
|
|
|
s.check(cmp, panicfn)
|
|
|
|
}
|
|
|
|
return s.newValue1(ssa.OpTrunc64to32, Types[TINT], v)
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
// Extend value to the required size
|
|
|
|
var op ssa.Op
|
|
|
|
if v.Type.IsSigned() {
|
2015-11-02 21:28:13 -08:00
|
|
|
switch 10*size + s.config.IntSize {
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
case 14:
|
|
|
|
op = ssa.OpSignExt8to32
|
|
|
|
case 18:
|
|
|
|
op = ssa.OpSignExt8to64
|
|
|
|
case 24:
|
|
|
|
op = ssa.OpSignExt16to32
|
|
|
|
case 28:
|
|
|
|
op = ssa.OpSignExt16to64
|
|
|
|
case 48:
|
|
|
|
op = ssa.OpSignExt32to64
|
|
|
|
default:
|
|
|
|
s.Fatalf("bad signed index extension %s", v.Type)
|
|
|
|
}
|
|
|
|
} else {
|
2015-11-02 21:28:13 -08:00
|
|
|
switch 10*size + s.config.IntSize {
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
case 14:
|
|
|
|
op = ssa.OpZeroExt8to32
|
|
|
|
case 18:
|
|
|
|
op = ssa.OpZeroExt8to64
|
|
|
|
case 24:
|
|
|
|
op = ssa.OpZeroExt16to32
|
|
|
|
case 28:
|
|
|
|
op = ssa.OpZeroExt16to64
|
|
|
|
case 48:
|
|
|
|
op = ssa.OpZeroExt32to64
|
|
|
|
default:
|
|
|
|
s.Fatalf("bad unsigned index extension %s", v.Type)
|
|
|
|
}
|
|
|
|
}
|
2015-11-02 21:28:13 -08:00
|
|
|
return s.newValue1(op, Types[TINT], v)
|
[dev.ssa] cmd/compile/internal/ssa: redo how sign extension is handled
For integer types less than a machine register, we have to decide
what the invariants are for the high bits of the register. We used
to set the high bits to the correct extension (sign or zero, as
determined by the type) of the low bits.
This CL makes the compiler ignore the high bits of the register
altogether (they are junk).
On this plus side, this means ops that generate subword results don't
have to worry about correctly extending them. On the minus side,
ops that consume subword arguments have to deal with the input
registers not being correctly extended.
For x86, this tradeoff is probably worth it. Almost all opcodes
have versions that use only the correct subword piece of their
inputs. (The one big exception is array indexing.) Not many opcodes
can correctly sign extend on output.
For other architectures, the tradeoff is probably not so clear, as
they don't have many subword-safe opcodes (e.g. 16-bit compare,
ignoring the high 16/48 bits). Fortunately we can decide whether
we do this per-architecture.
For the machine-independent opcodes, we pretend that the "register"
size is equal to the type width, so sign extension is immaterial.
Opcodes that care about the signedness of the input (e.g. compare,
right shift) have two different variants.
Change-Id: I465484c5734545ee697afe83bc8bf4b53bd9df8d
Reviewed-on: https://go-review.googlesource.com/12600
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
2015-07-23 14:35:02 -07:00
|
|
|
}
|
|
|
|
|
2016-06-15 15:17:45 -07:00
|
|
|
// SSAReg returns the register to which v has been allocated.
|
|
|
|
func SSAReg(v *ssa.Value) *ssa.Register {
|
2015-08-05 16:06:39 -07:00
|
|
|
reg := v.Block.Func.RegAlloc[v.ID]
|
|
|
|
if reg == nil {
|
2016-06-15 15:17:45 -07:00
|
|
|
v.Fatalf("nil register for value: %s\n%s\n", v.LongString(), v.Block.Func)
|
2015-08-05 16:06:39 -07:00
|
|
|
}
|
2016-06-15 15:17:45 -07:00
|
|
|
return reg.(*ssa.Register)
|
|
|
|
}
|
|
|
|
|
2016-07-13 16:15:54 -07:00
|
|
|
// SSAReg0 returns the register to which the first output of v has been allocated.
|
|
|
|
func SSAReg0(v *ssa.Value) *ssa.Register {
|
|
|
|
reg := v.Block.Func.RegAlloc[v.ID].(ssa.LocPair)[0]
|
|
|
|
if reg == nil {
|
|
|
|
v.Fatalf("nil first register for value: %s\n%s\n", v.LongString(), v.Block.Func)
|
|
|
|
}
|
|
|
|
return reg.(*ssa.Register)
|
|
|
|
}
|
|
|
|
|
|
|
|
// SSAReg1 returns the register to which the second output of v has been allocated.
|
|
|
|
func SSAReg1(v *ssa.Value) *ssa.Register {
|
|
|
|
reg := v.Block.Func.RegAlloc[v.ID].(ssa.LocPair)[1]
|
|
|
|
if reg == nil {
|
|
|
|
v.Fatalf("nil second register for value: %s\n%s\n", v.LongString(), v.Block.Func)
|
|
|
|
}
|
|
|
|
return reg.(*ssa.Register)
|
|
|
|
}
|
|
|
|
|
2016-06-15 15:17:45 -07:00
|
|
|
// SSARegNum returns the register number (in cmd/internal/obj numbering) to which v has been allocated.
|
|
|
|
func SSARegNum(v *ssa.Value) int16 {
|
|
|
|
return Thearch.SSARegToReg[SSAReg(v).Num]
|
2015-05-12 11:06:44 -07:00
|
|
|
}
|
|
|
|
|
2016-07-13 16:15:54 -07:00
|
|
|
// SSARegNum0 returns the register number (in cmd/internal/obj numbering) to which the first output of v has been allocated.
|
|
|
|
func SSARegNum0(v *ssa.Value) int16 {
|
|
|
|
return Thearch.SSARegToReg[SSAReg0(v).Num]
|
|
|
|
}
|
|
|
|
|
|
|
|
// SSARegNum1 returns the register number (in cmd/internal/obj numbering) to which the second output of v has been allocated.
|
|
|
|
func SSARegNum1(v *ssa.Value) int16 {
|
|
|
|
return Thearch.SSARegToReg[SSAReg1(v).Num]
|
|
|
|
}
|
|
|
|
|
2016-06-15 15:26:47 -07:00
|
|
|
// CheckLoweredPhi checks that regalloc and stackalloc correctly handled phi values.
|
|
|
|
// Called during ssaGenValue.
|
|
|
|
func CheckLoweredPhi(v *ssa.Value) {
|
|
|
|
if v.Op != ssa.OpPhi {
|
|
|
|
v.Fatalf("CheckLoweredPhi called with non-phi value: %v", v.LongString())
|
|
|
|
}
|
|
|
|
if v.Type.IsMemory() {
|
|
|
|
return
|
|
|
|
}
|
|
|
|
f := v.Block.Func
|
|
|
|
loc := f.RegAlloc[v.ID]
|
|
|
|
for _, a := range v.Args {
|
|
|
|
if aloc := f.RegAlloc[a.ID]; aloc != loc { // TODO: .Equal() instead?
|
|
|
|
v.Fatalf("phi arg at different location than phi: %v @ %v, but arg %v @ %v\n%s\n", v, loc, a, aloc, v.Block.Func)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-07-03 13:40:03 -07:00
|
|
|
// CheckLoweredGetClosurePtr checks that v is the first instruction in the function's entry block.
|
|
|
|
// The output of LoweredGetClosurePtr is generally hardwired to the correct register.
|
|
|
|
// That register contains the closure pointer on closure entry.
|
|
|
|
func CheckLoweredGetClosurePtr(v *ssa.Value) {
|
|
|
|
entry := v.Block.Func.Entry
|
|
|
|
if entry != v.Block || entry.Values[0] != v {
|
2016-07-21 10:37:59 -07:00
|
|
|
Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
|
2016-07-03 13:40:03 -07:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-03-12 14:07:40 -08:00
|
|
|
// AutoVar returns a *Node and int64 representing the auto variable and offset within it
|
2015-11-02 08:10:26 -08:00
|
|
|
// where v should be spilled.
|
2016-03-12 14:07:40 -08:00
|
|
|
func AutoVar(v *ssa.Value) (*Node, int64) {
|
2015-11-02 08:10:26 -08:00
|
|
|
loc := v.Block.Func.RegAlloc[v.ID].(ssa.LocalSlot)
|
2016-01-04 13:34:54 -08:00
|
|
|
if v.Type.Size() > loc.Type.Size() {
|
|
|
|
v.Fatalf("spill/restore type %s doesn't fit in slot type %s", v.Type, loc.Type)
|
|
|
|
}
|
2015-11-02 08:10:26 -08:00
|
|
|
return loc.N.(*Node), loc.Off
|
2015-05-12 11:06:44 -07:00
|
|
|
}
|
2015-05-27 14:52:22 -07:00
|
|
|
|
2016-01-11 21:05:33 -08:00
|
|
|
// fieldIdx finds the index of the field referred to by the ODOT node n.
|
2016-03-14 12:45:18 -07:00
|
|
|
func fieldIdx(n *Node) int {
|
2016-01-11 21:05:33 -08:00
|
|
|
t := n.Left.Type
|
cmd/compile: change ODOT and friends to use Sym, not Right
The Node type ODOT and its variants all represent a selector, with a
simple name to the right of the dot. Before this change this was
represented by using an ONAME Node in the Right field. This ONAME node
served no useful purpose. This CL changes these Node types to store the
symbol in the Sym field instead, thus not requiring allocating a Node
for each selector.
When compiling x/tools/go/types this CL eliminates nearly 5000 calls to
newname and reduces the total number of Nodes allocated by about 6.6%.
It seems to cut compilation time by 1 to 2 percent.
Getting this right was somewhat subtle, and I added two dubious changes
to produce the exact same output as before. One is to ishairy in
inl.go: the ONAME node increased the cost of ODOT and friends by 1, and
I retained that, although really ODOT is not more expensive than any
other node. The other is to varexpr in walk.go: because the ONAME in
the Right field of an ODOT has no class, varexpr would always return
false for an ODOT, although in fact for some ODOT's it seemingly ought
to return true; I added an && false for now. I will send separate CLs,
that will break toolstash -cmp, to clean these up.
This CL passes toolstash -cmp.
Change-Id: I4af8a10cc59078c436130ce472f25abc3a9b2f80
Reviewed-on: https://go-review.googlesource.com/20890
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
2016-03-18 16:52:30 -07:00
|
|
|
f := n.Sym
|
2016-03-30 14:56:08 -07:00
|
|
|
if !t.IsStruct() {
|
2016-01-11 21:05:33 -08:00
|
|
|
panic("ODOT's LHS is not a struct")
|
|
|
|
}
|
|
|
|
|
2016-03-14 12:45:18 -07:00
|
|
|
var i int
|
2016-03-17 01:32:18 -07:00
|
|
|
for _, t1 := range t.Fields().Slice() {
|
cmd/compile: change ODOT and friends to use Sym, not Right
The Node type ODOT and its variants all represent a selector, with a
simple name to the right of the dot. Before this change this was
represented by using an ONAME Node in the Right field. This ONAME node
served no useful purpose. This CL changes these Node types to store the
symbol in the Sym field instead, thus not requiring allocating a Node
for each selector.
When compiling x/tools/go/types this CL eliminates nearly 5000 calls to
newname and reduces the total number of Nodes allocated by about 6.6%.
It seems to cut compilation time by 1 to 2 percent.
Getting this right was somewhat subtle, and I added two dubious changes
to produce the exact same output as before. One is to ishairy in
inl.go: the ONAME node increased the cost of ODOT and friends by 1, and
I retained that, although really ODOT is not more expensive than any
other node. The other is to varexpr in walk.go: because the ONAME in
the Right field of an ODOT has no class, varexpr would always return
false for an ODOT, although in fact for some ODOT's it seemingly ought
to return true; I added an && false for now. I will send separate CLs,
that will break toolstash -cmp, to clean these up.
This CL passes toolstash -cmp.
Change-Id: I4af8a10cc59078c436130ce472f25abc3a9b2f80
Reviewed-on: https://go-review.googlesource.com/20890
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
2016-03-18 16:52:30 -07:00
|
|
|
if t1.Sym != f {
|
2016-01-11 21:05:33 -08:00
|
|
|
i++
|
|
|
|
continue
|
|
|
|
}
|
2016-03-28 09:40:53 -07:00
|
|
|
if t1.Offset != n.Xoffset {
|
2016-01-11 21:05:33 -08:00
|
|
|
panic("field offset doesn't match")
|
|
|
|
}
|
|
|
|
return i
|
|
|
|
}
|
|
|
|
panic(fmt.Sprintf("can't find field in expr %s\n", n))
|
|
|
|
|
2016-04-03 12:43:27 +01:00
|
|
|
// TODO: keep the result of this function somewhere in the ODOT Node
|
2016-01-11 21:05:33 -08:00
|
|
|
// so we don't have to recompute it each time we need it.
|
|
|
|
}
|
|
|
|
|
2015-05-27 14:52:22 -07:00
|
|
|
// ssaExport exports a bunch of compiler services for the ssa backend.
|
2015-06-12 11:01:13 -07:00
|
|
|
type ssaExport struct {
|
|
|
|
log bool
|
|
|
|
unimplemented bool
|
2015-07-22 13:13:53 -07:00
|
|
|
mustImplement bool
|
2015-06-12 11:01:13 -07:00
|
|
|
}
|
2015-05-27 14:52:22 -07:00
|
|
|
|
2015-07-30 11:03:05 -07:00
|
|
|
func (s *ssaExport) TypeBool() ssa.Type { return Types[TBOOL] }
|
|
|
|
func (s *ssaExport) TypeInt8() ssa.Type { return Types[TINT8] }
|
|
|
|
func (s *ssaExport) TypeInt16() ssa.Type { return Types[TINT16] }
|
|
|
|
func (s *ssaExport) TypeInt32() ssa.Type { return Types[TINT32] }
|
|
|
|
func (s *ssaExport) TypeInt64() ssa.Type { return Types[TINT64] }
|
|
|
|
func (s *ssaExport) TypeUInt8() ssa.Type { return Types[TUINT8] }
|
|
|
|
func (s *ssaExport) TypeUInt16() ssa.Type { return Types[TUINT16] }
|
|
|
|
func (s *ssaExport) TypeUInt32() ssa.Type { return Types[TUINT32] }
|
|
|
|
func (s *ssaExport) TypeUInt64() ssa.Type { return Types[TUINT64] }
|
2015-08-28 14:24:10 -04:00
|
|
|
func (s *ssaExport) TypeFloat32() ssa.Type { return Types[TFLOAT32] }
|
|
|
|
func (s *ssaExport) TypeFloat64() ssa.Type { return Types[TFLOAT64] }
|
2015-07-30 11:03:05 -07:00
|
|
|
func (s *ssaExport) TypeInt() ssa.Type { return Types[TINT] }
|
|
|
|
func (s *ssaExport) TypeUintptr() ssa.Type { return Types[TUINTPTR] }
|
|
|
|
func (s *ssaExport) TypeString() ssa.Type { return Types[TSTRING] }
|
|
|
|
func (s *ssaExport) TypeBytePtr() ssa.Type { return Ptrto(Types[TUINT8]) }
|
|
|
|
|
2015-07-24 11:28:12 -07:00
|
|
|
// StringData returns a symbol (a *Sym wrapped in an interface) which
|
|
|
|
// is the data component of a global string constant containing s.
|
|
|
|
func (*ssaExport) StringData(s string) interface{} {
|
2015-06-19 21:02:28 -07:00
|
|
|
// TODO: is idealstring correct? It might not matter...
|
2015-07-24 11:28:12 -07:00
|
|
|
_, data := stringsym(s)
|
|
|
|
return &ssa.ExternSymbol{Typ: idealstring, Sym: data}
|
2015-05-27 14:52:22 -07:00
|
|
|
}
|
2015-06-12 11:01:13 -07:00
|
|
|
|
2015-10-22 14:22:38 -07:00
|
|
|
func (e *ssaExport) Auto(t ssa.Type) ssa.GCNode {
|
2015-08-24 02:16:19 -07:00
|
|
|
n := temp(t.(*Type)) // Note: adds new auto to Curfn.Func.Dcl list
|
|
|
|
e.mustImplement = true // This modifies the input to SSA, so we want to make sure we succeed from here!
|
|
|
|
return n
|
|
|
|
}
|
|
|
|
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
func (e *ssaExport) SplitString(name ssa.LocalSlot) (ssa.LocalSlot, ssa.LocalSlot) {
|
|
|
|
n := name.N.(*Node)
|
|
|
|
ptrType := Ptrto(Types[TUINT8])
|
|
|
|
lenType := Types[TINT]
|
|
|
|
if n.Class == PAUTO && !n.Addrtaken {
|
|
|
|
// Split this string up into two separate variables.
|
|
|
|
p := e.namedAuto(n.Sym.Name+".ptr", ptrType)
|
|
|
|
l := e.namedAuto(n.Sym.Name+".len", lenType)
|
2016-04-29 09:02:27 -07:00
|
|
|
return ssa.LocalSlot{N: p, Type: ptrType, Off: 0}, ssa.LocalSlot{N: l, Type: lenType, Off: 0}
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
}
|
|
|
|
// Return the two parts of the larger variable.
|
2016-04-29 09:02:27 -07:00
|
|
|
return ssa.LocalSlot{N: n, Type: ptrType, Off: name.Off}, ssa.LocalSlot{N: n, Type: lenType, Off: name.Off + int64(Widthptr)}
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
func (e *ssaExport) SplitInterface(name ssa.LocalSlot) (ssa.LocalSlot, ssa.LocalSlot) {
|
|
|
|
n := name.N.(*Node)
|
|
|
|
t := Ptrto(Types[TUINT8])
|
|
|
|
if n.Class == PAUTO && !n.Addrtaken {
|
|
|
|
// Split this interface up into two separate variables.
|
|
|
|
f := ".itab"
|
2016-04-01 13:36:24 -07:00
|
|
|
if n.Type.IsEmptyInterface() {
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
f = ".type"
|
|
|
|
}
|
|
|
|
c := e.namedAuto(n.Sym.Name+f, t)
|
|
|
|
d := e.namedAuto(n.Sym.Name+".data", t)
|
2016-04-29 09:02:27 -07:00
|
|
|
return ssa.LocalSlot{N: c, Type: t, Off: 0}, ssa.LocalSlot{N: d, Type: t, Off: 0}
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
}
|
|
|
|
// Return the two parts of the larger variable.
|
2016-04-29 09:02:27 -07:00
|
|
|
return ssa.LocalSlot{N: n, Type: t, Off: name.Off}, ssa.LocalSlot{N: n, Type: t, Off: name.Off + int64(Widthptr)}
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
func (e *ssaExport) SplitSlice(name ssa.LocalSlot) (ssa.LocalSlot, ssa.LocalSlot, ssa.LocalSlot) {
|
|
|
|
n := name.N.(*Node)
|
2016-03-31 21:24:10 -07:00
|
|
|
ptrType := Ptrto(name.Type.ElemType().(*Type))
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
lenType := Types[TINT]
|
|
|
|
if n.Class == PAUTO && !n.Addrtaken {
|
|
|
|
// Split this slice up into three separate variables.
|
|
|
|
p := e.namedAuto(n.Sym.Name+".ptr", ptrType)
|
|
|
|
l := e.namedAuto(n.Sym.Name+".len", lenType)
|
|
|
|
c := e.namedAuto(n.Sym.Name+".cap", lenType)
|
2016-04-29 09:02:27 -07:00
|
|
|
return ssa.LocalSlot{N: p, Type: ptrType, Off: 0}, ssa.LocalSlot{N: l, Type: lenType, Off: 0}, ssa.LocalSlot{N: c, Type: lenType, Off: 0}
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
}
|
|
|
|
// Return the three parts of the larger variable.
|
2016-04-29 09:02:27 -07:00
|
|
|
return ssa.LocalSlot{N: n, Type: ptrType, Off: name.Off},
|
|
|
|
ssa.LocalSlot{N: n, Type: lenType, Off: name.Off + int64(Widthptr)},
|
|
|
|
ssa.LocalSlot{N: n, Type: lenType, Off: name.Off + int64(2*Widthptr)}
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
func (e *ssaExport) SplitComplex(name ssa.LocalSlot) (ssa.LocalSlot, ssa.LocalSlot) {
|
|
|
|
n := name.N.(*Node)
|
|
|
|
s := name.Type.Size() / 2
|
|
|
|
var t *Type
|
|
|
|
if s == 8 {
|
|
|
|
t = Types[TFLOAT64]
|
|
|
|
} else {
|
|
|
|
t = Types[TFLOAT32]
|
|
|
|
}
|
|
|
|
if n.Class == PAUTO && !n.Addrtaken {
|
|
|
|
// Split this complex up into two separate variables.
|
|
|
|
c := e.namedAuto(n.Sym.Name+".real", t)
|
|
|
|
d := e.namedAuto(n.Sym.Name+".imag", t)
|
2016-04-29 09:02:27 -07:00
|
|
|
return ssa.LocalSlot{N: c, Type: t, Off: 0}, ssa.LocalSlot{N: d, Type: t, Off: 0}
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
}
|
|
|
|
// Return the two parts of the larger variable.
|
2016-04-29 09:02:27 -07:00
|
|
|
return ssa.LocalSlot{N: n, Type: t, Off: name.Off}, ssa.LocalSlot{N: n, Type: t, Off: name.Off + s}
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
}
|
|
|
|
|
[dev.ssa] cmd/compile: decompose 64-bit integer on ARM
Introduce dec64 rules to (generically) decompose 64-bit integer on
32-bit architectures. 64-bit integer is composed/decomposed with
Int64Make/Hi/Lo ops, as for complex types.
The idea of dealing with Add64 is the following:
(Add64 (Int64Make xh xl) (Int64Make yh yl))
->
(Int64Make
(Add32withcarry xh yh (Select0 (Add32carry xl yl)))
(Select1 (Add32carry xl yl)))
where Add32carry returns a tuple (flags,uint32). Select0 and Select1
read the first and the second component of the tuple, respectively.
The two Add32carry will be CSE'd.
Similarly for multiplication, Mul32uhilo returns a tuple (hi, lo).
Also add support of KeepAlive, to fix build after merge.
Tests addressed_ssa.go, array_ssa.go, break_ssa.go, chan_ssa.go,
cmp_ssa.go, ctl_ssa.go, map_ssa.go, and string_ssa.go in
cmd/compile/internal/gc/testdata passed.
Progress on SSA for ARM. Still not complete.
Updates #15365.
Change-Id: I7867c76785a456312de5d8398a6b3f7ca5a4f7ec
Reviewed-on: https://go-review.googlesource.com/23213
Reviewed-by: Keith Randall <khr@golang.org>
2016-05-18 18:14:36 -04:00
|
|
|
func (e *ssaExport) SplitInt64(name ssa.LocalSlot) (ssa.LocalSlot, ssa.LocalSlot) {
|
|
|
|
n := name.N.(*Node)
|
|
|
|
var t *Type
|
|
|
|
if name.Type.IsSigned() {
|
|
|
|
t = Types[TINT32]
|
|
|
|
} else {
|
|
|
|
t = Types[TUINT32]
|
|
|
|
}
|
|
|
|
if n.Class == PAUTO && !n.Addrtaken {
|
|
|
|
// Split this int64 up into two separate variables.
|
|
|
|
h := e.namedAuto(n.Sym.Name+".hi", t)
|
|
|
|
l := e.namedAuto(n.Sym.Name+".lo", Types[TUINT32])
|
|
|
|
return ssa.LocalSlot{N: h, Type: t, Off: 0}, ssa.LocalSlot{N: l, Type: Types[TUINT32], Off: 0}
|
|
|
|
}
|
|
|
|
// Return the two parts of the larger variable.
|
|
|
|
// Assuming little endian (we don't support big endian 32-bit architecture yet)
|
|
|
|
return ssa.LocalSlot{N: n, Type: t, Off: name.Off + 4}, ssa.LocalSlot{N: n, Type: Types[TUINT32], Off: name.Off}
|
|
|
|
}
|
|
|
|
|
2016-03-31 21:24:10 -07:00
|
|
|
func (e *ssaExport) SplitStruct(name ssa.LocalSlot, i int) ssa.LocalSlot {
|
|
|
|
n := name.N.(*Node)
|
|
|
|
st := name.Type
|
|
|
|
ft := st.FieldType(i)
|
|
|
|
if n.Class == PAUTO && !n.Addrtaken {
|
|
|
|
// Note: the _ field may appear several times. But
|
|
|
|
// have no fear, identically-named but distinct Autos are
|
|
|
|
// ok, albeit maybe confusing for a debugger.
|
|
|
|
x := e.namedAuto(n.Sym.Name+"."+st.FieldName(i), ft)
|
2016-04-29 09:02:27 -07:00
|
|
|
return ssa.LocalSlot{N: x, Type: ft, Off: 0}
|
2016-03-31 21:24:10 -07:00
|
|
|
}
|
2016-04-29 09:02:27 -07:00
|
|
|
return ssa.LocalSlot{N: n, Type: ft, Off: name.Off + st.FieldOff(i)}
|
2016-03-31 21:24:10 -07:00
|
|
|
}
|
|
|
|
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
// namedAuto returns a new AUTO variable with the given name and type.
|
|
|
|
func (e *ssaExport) namedAuto(name string, typ ssa.Type) ssa.GCNode {
|
|
|
|
t := typ.(*Type)
|
2016-08-10 10:36:42 -07:00
|
|
|
s := &Sym{Name: name, Pkg: localpkg}
|
cmd/compile: better job of naming compound types
Compound AUTO types weren't named previously. That was because live
variable analysis (plive.go) doesn't handle spilling to compound types.
It can't handle them because there is no valid place to put VARDEFs when
regalloc is spilling compound types.
compound types = multiword builtin types: complex, string, slice, and
interface.
Instead, we split named AUTOs into individual one-word variables. For
example, a string s gets split into a byte ptr s.ptr and an integer
s.len. Those two variables can be spilled to / restored from
independently. As a result, live variable analysis can handle them
because they are one-word objects.
This CL will change how AUTOs are described in DWARF information.
Consider the code:
func f(s string, i int) int {
x := s[i:i+5]
g()
return lookup(x)
}
The old compiler would spill x to two consecutive slots on the stack,
both named x (at offsets 0 and 8). The new compiler spills the pointer
of x to a slot named x.ptr. It doesn't spill x.len at all, as it is a
constant (5) and can be rematerialized for the call to lookup.
So compound objects may not be spilled in their entirety, and even if
they are they won't necessarily be contiguous. Such is the price of
optimization.
Re-enable live variable analysis tests. One test remains disabled, it
fails because of #14904.
Change-Id: I8ef2b5ab91e43a0d2136bfc231c05d100ec0b801
Reviewed-on: https://go-review.googlesource.com/21233
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2016-03-28 11:25:17 -07:00
|
|
|
n := Nod(ONAME, nil, nil)
|
|
|
|
s.Def = n
|
|
|
|
s.Def.Used = true
|
|
|
|
n.Sym = s
|
|
|
|
n.Type = t
|
|
|
|
n.Class = PAUTO
|
|
|
|
n.Addable = true
|
|
|
|
n.Ullman = 1
|
|
|
|
n.Esc = EscNever
|
|
|
|
n.Xoffset = 0
|
|
|
|
n.Name.Curfn = Curfn
|
|
|
|
Curfn.Func.Dcl = append(Curfn.Func.Dcl, n)
|
|
|
|
|
|
|
|
dowidth(t)
|
|
|
|
e.mustImplement = true
|
|
|
|
|
|
|
|
return n
|
|
|
|
}
|
|
|
|
|
2015-10-22 13:07:38 -07:00
|
|
|
func (e *ssaExport) CanSSA(t ssa.Type) bool {
|
2015-09-18 22:58:10 -07:00
|
|
|
return canSSAType(t.(*Type))
|
|
|
|
}
|
|
|
|
|
2016-01-14 16:02:23 -08:00
|
|
|
func (e *ssaExport) Line(line int32) string {
|
2016-03-02 11:30:29 -08:00
|
|
|
return linestr(line)
|
2016-01-14 16:02:23 -08:00
|
|
|
}
|
|
|
|
|
2015-06-12 11:01:13 -07:00
|
|
|
// Log logs a message from the compiler.
|
2015-06-24 14:03:39 -07:00
|
|
|
func (e *ssaExport) Logf(msg string, args ...interface{}) {
|
2015-06-12 11:01:13 -07:00
|
|
|
// If e was marked as unimplemented, anything could happen. Ignore.
|
|
|
|
if e.log && !e.unimplemented {
|
|
|
|
fmt.Printf(msg, args...)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-01-29 14:44:15 -05:00
|
|
|
func (e *ssaExport) Log() bool {
|
|
|
|
return e.log
|
|
|
|
}
|
|
|
|
|
2015-06-12 11:01:13 -07:00
|
|
|
// Fatal reports a compiler error and exits.
|
2016-01-13 11:14:57 -08:00
|
|
|
func (e *ssaExport) Fatalf(line int32, msg string, args ...interface{}) {
|
2015-06-12 11:01:13 -07:00
|
|
|
// If e was marked as unimplemented, anything could happen. Ignore.
|
|
|
|
if !e.unimplemented {
|
2016-01-13 11:14:57 -08:00
|
|
|
lineno = line
|
2015-09-08 15:42:53 -07:00
|
|
|
Fatalf(msg, args...)
|
2015-06-12 11:01:13 -07:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Unimplemented reports that the function cannot be compiled.
|
|
|
|
// It will be removed once SSA work is complete.
|
2016-01-13 11:14:57 -08:00
|
|
|
func (e *ssaExport) Unimplementedf(line int32, msg string, args ...interface{}) {
|
2015-07-22 13:13:53 -07:00
|
|
|
if e.mustImplement {
|
2016-01-13 11:14:57 -08:00
|
|
|
lineno = line
|
2015-09-08 15:42:53 -07:00
|
|
|
Fatalf(msg, args...)
|
2015-07-22 13:13:53 -07:00
|
|
|
}
|
2015-06-12 11:01:13 -07:00
|
|
|
const alwaysLog = false // enable to calculate top unimplemented features
|
|
|
|
if !e.unimplemented && (e.log || alwaysLog) {
|
|
|
|
// first implementation failure, print explanation
|
|
|
|
fmt.Printf("SSA unimplemented: "+msg+"\n", args...)
|
|
|
|
}
|
|
|
|
e.unimplemented = true
|
|
|
|
}
|
2015-10-22 14:22:38 -07:00
|
|
|
|
2015-10-26 17:34:06 -04:00
|
|
|
// Warnl reports a "warning", which is usually flag-triggered
|
|
|
|
// logging output for the benefit of tests.
|
2016-03-13 23:04:31 -05:00
|
|
|
func (e *ssaExport) Warnl(line int32, fmt_ string, args ...interface{}) {
|
|
|
|
Warnl(line, fmt_, args...)
|
2015-10-26 17:34:06 -04:00
|
|
|
}
|
|
|
|
|
|
|
|
func (e *ssaExport) Debug_checknil() bool {
|
|
|
|
return Debug_checknil != 0
|
|
|
|
}
|
|
|
|
|
2015-10-22 14:22:38 -07:00
|
|
|
func (n *Node) Typ() ssa.Type {
|
|
|
|
return n.Type
|
|
|
|
}
|