2015-02-13 14:40:36 -05:00
|
|
|
// Copyright 2011 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.
|
|
|
|
//
|
|
|
|
// The inlining facility makes 2 passes: first caninl determines which
|
|
|
|
// functions are suitable for inlining, and for those that are it
|
|
|
|
// saves a copy of the body. Then inlcalls walks each function body to
|
|
|
|
// expand calls to inlinable functions.
|
|
|
|
//
|
2016-02-06 20:35:29 +09:00
|
|
|
// The debug['l'] flag controls the aggressiveness. Note that main() swaps level 0 and 1,
|
2015-02-13 14:40:36 -05:00
|
|
|
// making 1 the default and -l disable. -ll and more is useful to flush out bugs.
|
|
|
|
// These additional levels (beyond -l) may be buggy and are not supported.
|
|
|
|
// 0: disabled
|
|
|
|
// 1: 40-nodes leaf functions, oneliners, lazy typechecking (default)
|
|
|
|
// 2: early typechecking of all imported bodies
|
|
|
|
// 3: allow variadic functions
|
|
|
|
// 4: allow non-leaf functions , (breaks runtime.Caller)
|
|
|
|
//
|
|
|
|
// At some point this may get another default and become switch-offable with -N.
|
|
|
|
//
|
|
|
|
// The debug['m'] flag enables diagnostic output. a single -m is useful for verifying
|
|
|
|
// which calls get inlined or not, more is for debugging, and may go away at any point.
|
|
|
|
//
|
|
|
|
// TODO:
|
|
|
|
// - inline functions with ... args
|
|
|
|
// - handle T.meth(f()) with func f() (t T, arg, arg, )
|
|
|
|
|
|
|
|
package gc
|
|
|
|
|
|
|
|
import (
|
|
|
|
"fmt"
|
|
|
|
)
|
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// Get the function's package. For ordinary functions it's on the ->sym, but for imported methods
|
2015-02-13 14:40:36 -05:00
|
|
|
// the ->sym can be re-used in the local package, so peel it off the receiver's type.
|
|
|
|
func fnpkg(fn *Node) *Pkg {
|
2016-03-17 01:47:16 -07:00
|
|
|
if fn.Type.Recv() != nil {
|
2015-02-13 14:40:36 -05:00
|
|
|
// method
|
2016-03-09 20:54:59 -08:00
|
|
|
rcvr := fn.Type.Recv().Type
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2015-03-01 07:54:01 +00:00
|
|
|
if Isptr[rcvr.Etype] {
|
2015-02-13 14:40:36 -05:00
|
|
|
rcvr = rcvr.Type
|
|
|
|
}
|
2015-02-17 22:13:49 -05:00
|
|
|
if rcvr.Sym == nil {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("receiver with no sym: [%v] %v (%v)", fn.Sym, Nconv(fn, FmtLong), rcvr)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
return rcvr.Sym.Pkg
|
|
|
|
}
|
|
|
|
|
|
|
|
// non-method
|
|
|
|
return fn.Sym.Pkg
|
|
|
|
}
|
|
|
|
|
2016-03-01 23:21:55 +00:00
|
|
|
// Lazy typechecking of imported bodies. For local functions, caninl will set ->typecheck
|
2015-02-13 14:40:36 -05:00
|
|
|
// because they're a copy of an already checked body.
|
|
|
|
func typecheckinl(fn *Node) {
|
2016-03-02 12:49:37 -08:00
|
|
|
lno := setlineno(fn)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
// typecheckinl is only for imported functions;
|
|
|
|
// their bodies may refer to unsafe as long as the package
|
|
|
|
// was marked safe during import (which was checked then).
|
|
|
|
// the ->inl of a local function has been typechecked before caninl copied it.
|
2015-02-23 16:07:24 -05:00
|
|
|
pkg := fnpkg(fn)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
if pkg == localpkg || pkg == nil {
|
|
|
|
return // typecheckinl on local function
|
|
|
|
}
|
|
|
|
|
|
|
|
if Debug['m'] > 2 {
|
2016-03-15 13:06:58 -07:00
|
|
|
fmt.Printf("typecheck import [%v] %v { %v }\n", fn.Sym, Nconv(fn, FmtLong), Hconv(fn.Func.Inl, FmtSharp))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2015-02-23 16:07:24 -05:00
|
|
|
save_safemode := safemode
|
2015-02-13 14:40:36 -05:00
|
|
|
safemode = 0
|
|
|
|
|
2015-02-23 16:07:24 -05:00
|
|
|
savefn := Curfn
|
2015-02-13 14:40:36 -05:00
|
|
|
Curfn = fn
|
2016-03-19 17:02:01 -07:00
|
|
|
typecheckslice(fn.Func.Inl.Slice(), Etop)
|
2015-02-13 14:40:36 -05:00
|
|
|
Curfn = savefn
|
|
|
|
|
|
|
|
safemode = save_safemode
|
|
|
|
|
2016-03-02 12:49:37 -08:00
|
|
|
lineno = lno
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// Caninl determines whether fn is inlineable.
|
|
|
|
// If so, caninl saves fn->nbody in fn->inl and substitutes it with a copy.
|
|
|
|
// fn and ->nbody will already have been typechecked.
|
|
|
|
func caninl(fn *Node) {
|
|
|
|
if fn.Op != ODCLFUNC {
|
2015-08-30 23:10:03 +02:00
|
|
|
Fatalf("caninl %v", fn)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2015-05-27 10:42:55 -04:00
|
|
|
if fn.Func.Nname == nil {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("caninl no nname %v", Nconv(fn, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2015-08-24 19:45:59 -05:00
|
|
|
// If marked "go:noinline", don't inline
|
2016-02-26 13:32:28 -08:00
|
|
|
if fn.Func.Pragma&Noinline != 0 {
|
2015-08-24 19:45:59 -05:00
|
|
|
return
|
|
|
|
}
|
|
|
|
|
2015-02-13 14:40:36 -05:00
|
|
|
// If fn has no body (is defined outside of Go), cannot inline it.
|
2016-02-27 14:31:33 -08:00
|
|
|
if len(fn.Nbody.Slice()) == 0 {
|
2015-02-13 14:40:36 -05:00
|
|
|
return
|
|
|
|
}
|
|
|
|
|
|
|
|
if fn.Typecheck == 0 {
|
2015-08-30 23:10:03 +02:00
|
|
|
Fatalf("caninl on non-typechecked function %v", fn)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// can't handle ... args yet
|
|
|
|
if Debug['l'] < 3 {
|
2016-03-17 01:32:18 -07:00
|
|
|
for _, t := range fn.Type.Params().Fields().Slice() {
|
2015-03-09 16:24:07 +11:00
|
|
|
if t.Isddd {
|
2015-02-13 14:40:36 -05:00
|
|
|
return
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-10-20 10:00:07 -07:00
|
|
|
// Runtime package must not be instrumented.
|
|
|
|
// Instrument skips runtime package. However, some runtime code can be
|
2015-04-09 10:08:29 +03:00
|
|
|
// inlined into other packages and instrumented there. To avoid this,
|
2015-10-20 10:00:07 -07:00
|
|
|
// we disable inlining of runtime functions when instrumenting.
|
2015-04-09 10:08:29 +03:00
|
|
|
// The example that we observed is inlining of LockOSThread,
|
|
|
|
// which lead to false race reports on m contents.
|
2015-10-20 10:00:07 -07:00
|
|
|
if instrumenting && myimportpath == "runtime" {
|
2015-04-09 10:08:29 +03:00
|
|
|
return
|
|
|
|
}
|
|
|
|
|
2015-02-24 12:19:01 -05:00
|
|
|
const maxBudget = 80
|
|
|
|
budget := maxBudget // allowed hairyness
|
2016-03-04 09:37:58 -08:00
|
|
|
if ishairylist(fn.Nbody, &budget) || budget < 0 {
|
2015-02-13 14:40:36 -05:00
|
|
|
return
|
|
|
|
}
|
|
|
|
|
2015-02-23 16:07:24 -05:00
|
|
|
savefn := Curfn
|
2015-02-13 14:40:36 -05:00
|
|
|
Curfn = fn
|
|
|
|
|
2016-02-27 14:31:33 -08:00
|
|
|
fn.Func.Nname.Func.Inl.Set(fn.Nbody.Slice())
|
2016-03-04 09:37:58 -08:00
|
|
|
fn.Nbody.Set(inlcopylist(fn.Func.Nname.Func.Inl.Slice()))
|
|
|
|
inldcl := inlcopylist(fn.Func.Nname.Name.Defn.Func.Dcl)
|
2016-03-10 10:42:28 -08:00
|
|
|
fn.Func.Nname.Func.Inldcl.Set(inldcl)
|
2015-05-27 10:42:55 -04:00
|
|
|
fn.Func.Nname.Func.InlCost = int32(maxBudget - budget)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
// hack, TODO, check for better way to link method nodes back to the thing with the ->inl
|
|
|
|
// this is so export can find the body of a method
|
2015-05-27 10:42:55 -04:00
|
|
|
fn.Type.Nname = fn.Func.Nname
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
if Debug['m'] > 1 {
|
2016-03-15 13:06:58 -07:00
|
|
|
fmt.Printf("%v: can inline %v as: %v { %v }\n", fn.Line(), Nconv(fn.Func.Nname, FmtSharp), Tconv(fn.Type, FmtSharp), Hconv(fn.Func.Nname.Func.Inl, FmtSharp))
|
2015-02-13 14:40:36 -05:00
|
|
|
} else if Debug['m'] != 0 {
|
2015-05-27 10:42:55 -04:00
|
|
|
fmt.Printf("%v: can inline %v\n", fn.Line(), fn.Func.Nname)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
Curfn = savefn
|
|
|
|
}
|
|
|
|
|
|
|
|
// Look for anything we want to punt on.
|
2016-03-08 10:26:20 -08:00
|
|
|
func ishairylist(ll Nodes, budget *int) bool {
|
2016-03-08 15:10:26 -08:00
|
|
|
for _, n := range ll.Slice() {
|
|
|
|
if ishairy(n, budget) {
|
2016-02-27 14:31:33 -08:00
|
|
|
return true
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
|
2015-02-17 22:13:49 -05:00
|
|
|
func ishairy(n *Node, budget *int) bool {
|
|
|
|
if n == nil {
|
|
|
|
return false
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
switch n.Op {
|
2015-02-24 12:19:01 -05:00
|
|
|
// Call is okay if inlinable and we have the budget for the body.
|
|
|
|
case OCALLFUNC:
|
2016-02-27 14:31:33 -08:00
|
|
|
if n.Left.Func != nil && len(n.Left.Func.Inl.Slice()) != 0 {
|
2015-03-25 19:33:01 -07:00
|
|
|
*budget -= int(n.Left.Func.InlCost)
|
2015-02-24 12:19:01 -05:00
|
|
|
break
|
|
|
|
}
|
|
|
|
if n.Left.Op == ONAME && n.Left.Left != nil && n.Left.Left.Op == OTYPE && n.Left.Right != nil && n.Left.Right.Op == ONAME { // methods called as functions
|
2016-02-27 14:31:33 -08:00
|
|
|
if n.Left.Sym.Def != nil && len(n.Left.Sym.Def.Func.Inl.Slice()) != 0 {
|
2015-03-25 19:33:01 -07:00
|
|
|
*budget -= int(n.Left.Sym.Def.Func.InlCost)
|
2015-02-24 12:19:01 -05:00
|
|
|
break
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if Debug['l'] < 4 {
|
|
|
|
return true
|
|
|
|
}
|
|
|
|
|
|
|
|
// Call is okay if inlinable and we have the budget for the body.
|
|
|
|
case OCALLMETH:
|
|
|
|
if n.Left.Type == nil {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("no function type for [%p] %v\n", n.Left, Nconv(n.Left, FmtSign))
|
2015-02-24 12:19:01 -05:00
|
|
|
}
|
|
|
|
if n.Left.Type.Nname == nil {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("no function definition for [%p] %v\n", n.Left.Type, Tconv(n.Left.Type, FmtSign))
|
2015-02-24 12:19:01 -05:00
|
|
|
}
|
2016-02-27 14:31:33 -08:00
|
|
|
if len(n.Left.Type.Nname.Func.Inl.Slice()) != 0 {
|
2015-03-25 19:33:01 -07:00
|
|
|
*budget -= int(n.Left.Type.Nname.Func.InlCost)
|
2015-02-24 12:19:01 -05:00
|
|
|
break
|
|
|
|
}
|
|
|
|
if Debug['l'] < 4 {
|
|
|
|
return true
|
|
|
|
}
|
|
|
|
|
|
|
|
// Things that are too hairy, irrespective of the budget
|
2015-04-01 09:38:44 -07:00
|
|
|
case OCALL, OCALLINTER, OPANIC, ORECOVER:
|
2015-02-13 14:40:36 -05:00
|
|
|
if Debug['l'] < 4 {
|
2015-02-17 22:13:49 -05:00
|
|
|
return true
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
case OCLOSURE,
|
|
|
|
OCALLPART,
|
|
|
|
ORANGE,
|
|
|
|
OFOR,
|
|
|
|
OSELECT,
|
2016-03-16 21:29:17 -05:00
|
|
|
OTYPESW,
|
2015-02-13 14:40:36 -05:00
|
|
|
OPROC,
|
|
|
|
ODEFER,
|
2016-03-16 18:44:17 -05:00
|
|
|
ODCLTYPE, // can't print yet
|
2016-03-16 21:29:17 -05:00
|
|
|
OBREAK,
|
2015-02-13 14:40:36 -05:00
|
|
|
ORETJMP:
|
2015-02-17 22:13:49 -05:00
|
|
|
return true
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
(*budget)--
|
|
|
|
|
2016-03-04 09:37:58 -08:00
|
|
|
return *budget < 0 || ishairy(n.Left, budget) || ishairy(n.Right, budget) || ishairylist(n.List, budget) || ishairylist(n.Rlist, budget) || ishairylist(n.Ninit, budget) || ishairylist(n.Nbody, budget)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// Inlcopy and inlcopylist recursively copy the body of a function.
|
|
|
|
// Any name-like node of non-local class is marked for re-export by adding it to
|
|
|
|
// the exportlist.
|
2016-03-08 10:26:20 -08:00
|
|
|
func inlcopylist(ll []*Node) []*Node {
|
2016-03-09 20:29:21 -08:00
|
|
|
s := make([]*Node, 0, len(ll))
|
2016-03-08 15:10:26 -08:00
|
|
|
for _, n := range ll {
|
|
|
|
s = append(s, inlcopy(n))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2016-03-04 09:37:58 -08:00
|
|
|
return s
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
func inlcopy(n *Node) *Node {
|
|
|
|
if n == nil {
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
switch n.Op {
|
2015-04-01 09:38:44 -07:00
|
|
|
case ONAME, OTYPE, OLITERAL:
|
2015-02-13 14:40:36 -05:00
|
|
|
return n
|
|
|
|
}
|
|
|
|
|
2015-02-23 16:07:24 -05:00
|
|
|
m := Nod(OXXX, nil, nil)
|
2015-02-13 14:40:36 -05:00
|
|
|
*m = *n
|
2015-03-10 21:37:13 -07:00
|
|
|
if m.Func != nil {
|
2016-02-27 14:31:33 -08:00
|
|
|
m.Func.Inl.Set(nil)
|
2015-03-10 21:37:13 -07:00
|
|
|
}
|
2015-02-13 14:40:36 -05:00
|
|
|
m.Left = inlcopy(n.Left)
|
|
|
|
m.Right = inlcopy(n.Right)
|
2016-03-08 15:10:26 -08:00
|
|
|
m.List.Set(inlcopylist(n.List.Slice()))
|
|
|
|
m.Rlist.Set(inlcopylist(n.Rlist.Slice()))
|
|
|
|
m.Ninit.Set(inlcopylist(n.Ninit.Slice()))
|
2016-03-04 09:37:58 -08:00
|
|
|
m.Nbody.Set(inlcopylist(n.Nbody.Slice()))
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
return m
|
|
|
|
}
|
|
|
|
|
|
|
|
// Inlcalls/nodelist/node walks fn's statements and expressions and substitutes any
|
2016-03-01 23:21:55 +00:00
|
|
|
// calls made to inlineable functions. This is the external entry point.
|
2015-02-13 14:40:36 -05:00
|
|
|
func inlcalls(fn *Node) {
|
2015-02-23 16:07:24 -05:00
|
|
|
savefn := Curfn
|
2015-02-13 14:40:36 -05:00
|
|
|
Curfn = fn
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
fn = inlnode(fn)
|
2015-02-13 14:40:36 -05:00
|
|
|
if fn != Curfn {
|
2015-08-30 23:10:03 +02:00
|
|
|
Fatalf("inlnode replaced curfn")
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
Curfn = savefn
|
|
|
|
}
|
|
|
|
|
|
|
|
// Turn an OINLCALL into a statement.
|
|
|
|
func inlconv2stmt(n *Node) {
|
|
|
|
n.Op = OBLOCK
|
|
|
|
|
|
|
|
// n->ninit stays
|
2016-03-08 15:10:26 -08:00
|
|
|
n.List.Set(n.Nbody.Slice())
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-02-27 14:31:33 -08:00
|
|
|
n.Nbody.Set(nil)
|
2016-03-08 15:10:26 -08:00
|
|
|
n.Rlist.Set(nil)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// Turn an OINLCALL into a single valued expression.
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
// The result of inlconv2expr MUST be assigned back to n, e.g.
|
|
|
|
// n.Left = inlconv2expr(n.Left)
|
|
|
|
func inlconv2expr(n *Node) *Node {
|
2016-03-08 15:10:26 -08:00
|
|
|
r := n.Rlist.First()
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
r = addinit(r, append(n.Ninit.Slice(), n.Nbody.Slice()...))
|
|
|
|
return r
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// Turn the rlist (with the return values) of the OINLCALL in
|
|
|
|
// n into an expression list lumping the ninit and body
|
|
|
|
// containing the inlined statements on the first list element so
|
|
|
|
// order will be preserved Used in return, oas2func and call
|
|
|
|
// statements.
|
2016-03-04 09:37:58 -08:00
|
|
|
func inlconv2list(n *Node) []*Node {
|
2016-03-08 15:10:26 -08:00
|
|
|
if n.Op != OINLCALL || n.Rlist.Len() == 0 {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("inlconv2list %v\n", Nconv(n, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2016-03-08 15:10:26 -08:00
|
|
|
s := n.Rlist.Slice()
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
s[0] = addinit(s[0], append(n.Ninit.Slice(), n.Nbody.Slice()...))
|
2016-03-04 09:37:58 -08:00
|
|
|
return s
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2016-03-08 10:26:20 -08:00
|
|
|
func inlnodelist(l Nodes) {
|
2016-03-09 12:39:36 -08:00
|
|
|
s := l.Slice()
|
|
|
|
for i := range s {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
s[i] = inlnode(s[i])
|
2016-02-27 14:31:33 -08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-02-13 14:40:36 -05:00
|
|
|
// inlnode recurses over the tree to find inlineable calls, which will
|
2016-03-01 23:21:55 +00:00
|
|
|
// be turned into OINLCALLs by mkinlcall. When the recursion comes
|
2015-02-13 14:40:36 -05:00
|
|
|
// back up will examine left, right, list, rlist, ninit, ntest, nincr,
|
|
|
|
// nbody and nelse and use one of the 4 inlconv/glue functions above
|
|
|
|
// to turn the OINLCALL into an expression, a statement, or patch it
|
|
|
|
// in to this nodes list or rlist as appropriate.
|
|
|
|
// NOTE it makes no sense to pass the glue functions down the
|
|
|
|
// recursion to the level where the OINLCALL gets created because they
|
|
|
|
// have to edit /this/ n, so you'd have to push that one down as well,
|
|
|
|
// but then you may as well do it here. so this is cleaner and
|
|
|
|
// shorter and less complicated.
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
// The result of inlnode MUST be assigned back to n, e.g.
|
|
|
|
// n.Left = inlnode(n.Left)
|
|
|
|
func inlnode(n *Node) *Node {
|
|
|
|
if n == nil {
|
|
|
|
return n
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
switch n.Op {
|
|
|
|
// inhibit inlining of their argument
|
2015-04-01 09:38:44 -07:00
|
|
|
case ODEFER, OPROC:
|
2015-02-13 14:40:36 -05:00
|
|
|
switch n.Left.Op {
|
2015-04-01 09:38:44 -07:00
|
|
|
case OCALLFUNC, OCALLMETH:
|
2015-09-24 23:21:18 +02:00
|
|
|
// TODO(marvin): Fix Node.EType type union.
|
|
|
|
n.Left.Etype = EType(n.Op)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
fallthrough
|
|
|
|
|
|
|
|
// TODO do them here (or earlier),
|
|
|
|
// so escape analysis can avoid more heapmoves.
|
|
|
|
case OCLOSURE:
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
return n
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2016-03-02 12:49:37 -08:00
|
|
|
lno := setlineno(n)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
inlnodelist(n.Ninit)
|
2016-03-08 15:10:26 -08:00
|
|
|
for _, n1 := range n.Ninit.Slice() {
|
|
|
|
if n1.Op == OINLCALL {
|
|
|
|
inlconv2stmt(n1)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
n.Left = inlnode(n.Left)
|
2015-02-13 14:40:36 -05:00
|
|
|
if n.Left != nil && n.Left.Op == OINLCALL {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
n.Left = inlconv2expr(n.Left)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
n.Right = inlnode(n.Right)
|
2015-02-13 14:40:36 -05:00
|
|
|
if n.Right != nil && n.Right.Op == OINLCALL {
|
2015-05-22 01:16:52 -04:00
|
|
|
if n.Op == OFOR {
|
|
|
|
inlconv2stmt(n.Right)
|
|
|
|
} else {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
n.Right = inlconv2expr(n.Right)
|
2015-05-22 01:16:52 -04:00
|
|
|
}
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
inlnodelist(n.List)
|
|
|
|
switch n.Op {
|
|
|
|
case OBLOCK:
|
2016-03-08 15:10:26 -08:00
|
|
|
for _, n2 := range n.List.Slice() {
|
|
|
|
if n2.Op == OINLCALL {
|
|
|
|
inlconv2stmt(n2)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// if we just replaced arg in f(arg()) or return arg with an inlined call
|
|
|
|
// and arg returns multiple values, glue as list
|
|
|
|
case ORETURN,
|
|
|
|
OCALLFUNC,
|
|
|
|
OCALLMETH,
|
|
|
|
OCALLINTER,
|
|
|
|
OAPPEND,
|
|
|
|
OCOMPLEX:
|
2016-03-08 15:10:26 -08:00
|
|
|
if n.List.Len() == 1 && n.List.First().Op == OINLCALL && n.List.First().Rlist.Len() > 1 {
|
|
|
|
n.List.Set(inlconv2list(n.List.First()))
|
2015-02-13 14:40:36 -05:00
|
|
|
break
|
|
|
|
}
|
|
|
|
fallthrough
|
|
|
|
|
|
|
|
default:
|
2016-03-09 12:39:36 -08:00
|
|
|
s := n.List.Slice()
|
|
|
|
for i1, n1 := range s {
|
|
|
|
if n1.Op == OINLCALL {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
s[i1] = inlconv2expr(s[i1])
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
inlnodelist(n.Rlist)
|
|
|
|
switch n.Op {
|
|
|
|
case OAS2FUNC:
|
2016-03-08 15:10:26 -08:00
|
|
|
if n.Rlist.First().Op == OINLCALL {
|
|
|
|
n.Rlist.Set(inlconv2list(n.Rlist.First()))
|
2015-02-13 14:40:36 -05:00
|
|
|
n.Op = OAS2
|
|
|
|
n.Typecheck = 0
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
n = typecheck(n, Etop)
|
2015-02-13 14:40:36 -05:00
|
|
|
break
|
|
|
|
}
|
|
|
|
fallthrough
|
|
|
|
|
|
|
|
default:
|
2016-03-09 12:39:36 -08:00
|
|
|
s := n.Rlist.Slice()
|
|
|
|
for i1, n1 := range s {
|
|
|
|
if n1.Op == OINLCALL {
|
2015-05-22 01:16:52 -04:00
|
|
|
if n.Op == OIF {
|
2016-03-09 12:39:36 -08:00
|
|
|
inlconv2stmt(n1)
|
2015-05-22 01:16:52 -04:00
|
|
|
} else {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
s[i1] = inlconv2expr(s[i1])
|
2015-05-22 01:16:52 -04:00
|
|
|
}
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-03-04 09:37:58 -08:00
|
|
|
inlnodelist(n.Nbody)
|
2016-02-27 14:31:33 -08:00
|
|
|
for _, n := range n.Nbody.Slice() {
|
|
|
|
if n.Op == OINLCALL {
|
|
|
|
inlconv2stmt(n)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// with all the branches out of the way, it is now time to
|
|
|
|
// transmogrify this node itself unless inhibited by the
|
|
|
|
// switch at the top of this function.
|
|
|
|
switch n.Op {
|
2015-04-01 09:38:44 -07:00
|
|
|
case OCALLFUNC, OCALLMETH:
|
2015-09-24 23:21:18 +02:00
|
|
|
// TODO(marvin): Fix Node.EType type union.
|
|
|
|
if n.Etype == EType(OPROC) || n.Etype == EType(ODEFER) {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
return n
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
switch n.Op {
|
|
|
|
case OCALLFUNC:
|
|
|
|
if Debug['m'] > 3 {
|
2016-03-15 13:06:58 -07:00
|
|
|
fmt.Printf("%v:call to func %v\n", n.Line(), Nconv(n.Left, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2016-02-27 14:31:33 -08:00
|
|
|
if n.Left.Func != nil && len(n.Left.Func.Inl.Slice()) != 0 { // normal case
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
n = mkinlcall(n, n.Left, n.Isddd)
|
2015-02-13 14:40:36 -05:00
|
|
|
} else if n.Left.Op == ONAME && n.Left.Left != nil && n.Left.Left.Op == OTYPE && n.Left.Right != nil && n.Left.Right.Op == ONAME { // methods called as functions
|
|
|
|
if n.Left.Sym.Def != nil {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
n = mkinlcall(n, n.Left.Sym.Def, n.Isddd)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
case OCALLMETH:
|
|
|
|
if Debug['m'] > 3 {
|
2016-03-15 13:06:58 -07:00
|
|
|
fmt.Printf("%v:call to meth %v\n", n.Line(), Nconv(n.Left.Right, FmtLong))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// typecheck should have resolved ODOTMETH->type, whose nname points to the actual function.
|
|
|
|
if n.Left.Type == nil {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("no function type for [%p] %v\n", n.Left, Nconv(n.Left, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
if n.Left.Type.Nname == nil {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("no function definition for [%p] %v\n", n.Left.Type, Tconv(n.Left.Type, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
n = mkinlcall(n, n.Left.Type.Nname, n.Isddd)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2016-03-02 12:49:37 -08:00
|
|
|
lineno = lno
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
return n
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
// The result of mkinlcall MUST be assigned back to n, e.g.
|
|
|
|
// n.Left = mkinlcall(n.Left, fn, isddd)
|
|
|
|
func mkinlcall(n *Node, fn *Node, isddd bool) *Node {
|
2015-02-23 16:07:24 -05:00
|
|
|
save_safemode := safemode
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
// imported functions may refer to unsafe as long as the
|
|
|
|
// package was marked safe during import (already checked).
|
2015-02-23 16:07:24 -05:00
|
|
|
pkg := fnpkg(fn)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
if pkg != localpkg && pkg != nil {
|
|
|
|
safemode = 0
|
|
|
|
}
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
n = mkinlcall1(n, fn, isddd)
|
2015-02-13 14:40:36 -05:00
|
|
|
safemode = save_safemode
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
return n
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2016-03-14 01:20:49 -07:00
|
|
|
func tinlvar(t *Field) *Node {
|
2015-02-13 14:40:36 -05:00
|
|
|
if t.Nname != nil && !isblank(t.Nname) {
|
2015-05-15 10:02:19 -07:00
|
|
|
if t.Nname.Name.Inlvar == nil {
|
2015-08-30 23:10:03 +02:00
|
|
|
Fatalf("missing inlvar for %v\n", t.Nname)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2015-05-15 10:02:19 -07:00
|
|
|
return t.Nname.Name.Inlvar
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
nblank = typecheck(nblank, Erv|Easgn)
|
2015-02-13 14:40:36 -05:00
|
|
|
return nblank
|
|
|
|
}
|
|
|
|
|
|
|
|
var inlgen int
|
|
|
|
|
|
|
|
// if *np is a call, and fn is a function with an inlinable body, substitute *np with an OINLCALL.
|
|
|
|
// On return ninit has the parameter assignments, the nbody is the
|
|
|
|
// inlined function body and list, rlist contain the input, output
|
|
|
|
// parameters.
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
// The result of mkinlcall1 MUST be assigned back to n, e.g.
|
|
|
|
// n.Left = mkinlcall1(n.Left, fn, isddd)
|
|
|
|
func mkinlcall1(n *Node, fn *Node, isddd bool) *Node {
|
2015-02-13 14:40:36 -05:00
|
|
|
// For variadic fn.
|
2016-02-27 14:31:33 -08:00
|
|
|
if len(fn.Func.Inl.Slice()) == 0 {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
return n
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2015-05-26 22:19:27 -04:00
|
|
|
if fn == Curfn || fn.Name.Defn == Curfn {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
return n
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
if Debug['l'] < 2 {
|
|
|
|
typecheckinl(fn)
|
|
|
|
}
|
|
|
|
|
|
|
|
// Bingo, we have a function node, and it has an inlineable body
|
|
|
|
if Debug['m'] > 1 {
|
2016-03-15 13:06:58 -07:00
|
|
|
fmt.Printf("%v: inlining call to %v %v { %v }\n", n.Line(), fn.Sym, Tconv(fn.Type, FmtSharp), Hconv(fn.Func.Inl, FmtSharp))
|
2015-02-13 14:40:36 -05:00
|
|
|
} else if Debug['m'] != 0 {
|
2015-04-17 12:03:22 -04:00
|
|
|
fmt.Printf("%v: inlining call to %v\n", n.Line(), fn)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
if Debug['m'] > 2 {
|
2016-03-15 13:06:58 -07:00
|
|
|
fmt.Printf("%v: Before inlining: %v\n", n.Line(), Nconv(n, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2015-02-23 16:07:24 -05:00
|
|
|
ninit := n.Ninit
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
//dumplist("ninit pre", ninit);
|
|
|
|
|
2016-02-25 10:35:19 -08:00
|
|
|
var dcl []*Node
|
2016-02-26 13:48:24 -08:00
|
|
|
if fn.Name.Defn != nil {
|
|
|
|
// local function
|
2016-03-10 10:42:28 -08:00
|
|
|
dcl = fn.Func.Inldcl.Slice()
|
2015-02-13 14:40:36 -05:00
|
|
|
} else {
|
2016-02-26 13:48:24 -08:00
|
|
|
// imported function
|
2015-03-25 19:33:01 -07:00
|
|
|
dcl = fn.Func.Dcl
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2016-03-10 11:49:20 -08:00
|
|
|
var retvars []*Node
|
2015-02-23 16:07:24 -05:00
|
|
|
i := 0
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
// Make temp names to use instead of the originals
|
2016-02-25 10:35:19 -08:00
|
|
|
for _, ln := range dcl {
|
|
|
|
if ln.Class == PPARAMOUT { // return values handled below.
|
2015-02-13 14:40:36 -05:00
|
|
|
continue
|
|
|
|
}
|
2016-02-25 10:35:19 -08:00
|
|
|
if ln.Op == ONAME {
|
|
|
|
ln.Name.Inlvar = inlvar(ln)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
// Typecheck because inlvar is not necessarily a function parameter.
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
ln.Name.Inlvar = typecheck(ln.Name.Inlvar, Erv)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-02-25 10:35:19 -08:00
|
|
|
if ln.Class&^PHEAP != PAUTO {
|
2016-03-08 15:10:26 -08:00
|
|
|
ninit.Append(Nod(ODCL, ln.Name.Inlvar, nil)) // otherwise gen won't emit the allocations for heapallocs
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// temporaries for return values.
|
2015-02-23 16:07:24 -05:00
|
|
|
var m *Node
|
2016-03-17 01:32:18 -07:00
|
|
|
for _, t := range fn.Type.Results().Fields().Slice() {
|
2015-02-13 14:40:36 -05:00
|
|
|
if t != nil && t.Nname != nil && !isblank(t.Nname) {
|
|
|
|
m = inlvar(t.Nname)
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
m = typecheck(m, Erv)
|
2015-05-15 10:02:19 -07:00
|
|
|
t.Nname.Name.Inlvar = m
|
2015-02-13 14:40:36 -05:00
|
|
|
} else {
|
|
|
|
// anonymous return values, synthesize names for use in assignment that replaces return
|
|
|
|
m = retvar(t, i)
|
|
|
|
i++
|
|
|
|
}
|
|
|
|
|
2016-03-08 15:10:26 -08:00
|
|
|
ninit.Append(Nod(ODCL, m, nil))
|
2016-03-10 11:49:20 -08:00
|
|
|
retvars = append(retvars, m)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// assign receiver.
|
2016-03-17 01:47:16 -07:00
|
|
|
if fn.Type.Recv() != nil && n.Left.Op == ODOTMETH {
|
2015-02-13 14:40:36 -05:00
|
|
|
// method call with a receiver.
|
2016-03-09 20:54:59 -08:00
|
|
|
t := fn.Type.Recv()
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2015-05-15 10:02:19 -07:00
|
|
|
if t != nil && t.Nname != nil && !isblank(t.Nname) && t.Nname.Name.Inlvar == nil {
|
2015-08-30 23:10:03 +02:00
|
|
|
Fatalf("missing inlvar for %v\n", t.Nname)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2015-02-17 22:13:49 -05:00
|
|
|
if n.Left.Left == nil {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("method call without receiver: %v", Nconv(n, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
if t == nil {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("method call unknown receiver type: %v", Nconv(n, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2016-03-09 20:29:21 -08:00
|
|
|
as := Nod(OAS, tinlvar(t), n.Left.Left)
|
2015-02-13 14:40:36 -05:00
|
|
|
if as != nil {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
as = typecheck(as, Etop)
|
2016-03-08 15:10:26 -08:00
|
|
|
ninit.Append(as)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// check if inlined function is variadic.
|
2015-02-23 16:07:24 -05:00
|
|
|
variadic := false
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2015-03-02 14:22:05 -05:00
|
|
|
var varargtype *Type
|
2015-02-23 16:07:24 -05:00
|
|
|
varargcount := 0
|
2016-03-17 01:32:18 -07:00
|
|
|
for _, t := range fn.Type.Params().Fields().Slice() {
|
2015-03-09 16:24:07 +11:00
|
|
|
if t.Isddd {
|
2015-02-17 22:13:49 -05:00
|
|
|
variadic = true
|
2015-02-13 14:40:36 -05:00
|
|
|
varargtype = t.Type
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// but if argument is dotted too forget about variadicity.
|
2015-03-09 16:24:07 +11:00
|
|
|
if variadic && isddd {
|
2015-02-17 22:13:49 -05:00
|
|
|
variadic = false
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// check if argument is actually a returned tuple from call.
|
2015-02-23 16:07:24 -05:00
|
|
|
multiret := 0
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-03-08 15:10:26 -08:00
|
|
|
if n.List.Len() == 1 {
|
|
|
|
switch n.List.First().Op {
|
2015-04-01 09:38:44 -07:00
|
|
|
case OCALL, OCALLFUNC, OCALLINTER, OCALLMETH:
|
2016-03-17 01:47:16 -07:00
|
|
|
if n.List.First().Left.Type.Results().NumFields() > 1 {
|
|
|
|
multiret = n.List.First().Left.Type.Results().NumFields() - 1
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-02-17 22:13:49 -05:00
|
|
|
if variadic {
|
2016-03-08 15:10:26 -08:00
|
|
|
varargcount = n.List.Len() + multiret
|
2015-02-13 14:40:36 -05:00
|
|
|
if n.Left.Op != ODOTMETH {
|
2016-03-17 01:47:16 -07:00
|
|
|
varargcount -= fn.Type.Recvs().NumFields()
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2016-03-17 01:47:16 -07:00
|
|
|
varargcount -= fn.Type.Params().NumFields() - 1
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// assign arguments to the parameters' temp names
|
2016-03-09 20:29:21 -08:00
|
|
|
as := Nod(OAS2, nil, nil)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-03-08 15:10:26 -08:00
|
|
|
as.Rlist.Set(n.List.Slice())
|
2016-03-09 12:39:36 -08:00
|
|
|
li := 0
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
// TODO: if len(nlist) == 1 but multiple args, check that n->list->n is a call?
|
2016-03-17 01:47:16 -07:00
|
|
|
if fn.Type.Recv() != nil && n.Left.Op != ODOTMETH {
|
2015-02-13 14:40:36 -05:00
|
|
|
// non-method call to method
|
2016-03-08 15:10:26 -08:00
|
|
|
if n.List.Len() == 0 {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("non-method call to method without first arg: %v", Nconv(n, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// append receiver inlvar to LHS.
|
2016-03-09 20:54:59 -08:00
|
|
|
t := fn.Type.Recv()
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2015-05-15 10:02:19 -07:00
|
|
|
if t != nil && t.Nname != nil && !isblank(t.Nname) && t.Nname.Name.Inlvar == nil {
|
2015-08-30 23:10:03 +02:00
|
|
|
Fatalf("missing inlvar for %v\n", t.Nname)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
if t == nil {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("method call unknown receiver type: %v", Nconv(n, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2016-03-08 15:10:26 -08:00
|
|
|
as.List.Append(tinlvar(t))
|
2016-03-09 12:39:36 -08:00
|
|
|
li++
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// append ordinary arguments to LHS.
|
2016-03-08 15:10:26 -08:00
|
|
|
chkargcount := n.List.Len() > 1
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-03-04 09:37:58 -08:00
|
|
|
var vararg *Node // the slice argument to a variadic call
|
|
|
|
var varargs []*Node // the list of LHS names to put in vararg.
|
2015-02-17 22:13:49 -05:00
|
|
|
if !chkargcount {
|
2015-02-13 14:40:36 -05:00
|
|
|
// 0 or 1 expression on RHS.
|
2015-02-23 16:07:24 -05:00
|
|
|
var i int
|
2016-03-17 01:32:18 -07:00
|
|
|
for _, t := range fn.Type.Params().Fields().Slice() {
|
2015-03-09 16:24:07 +11:00
|
|
|
if variadic && t.Isddd {
|
2015-02-13 14:40:36 -05:00
|
|
|
vararg = tinlvar(t)
|
2016-03-09 12:39:36 -08:00
|
|
|
for i = 0; i < varargcount && li < n.List.Len(); i++ {
|
2015-02-13 14:40:36 -05:00
|
|
|
m = argvar(varargtype, i)
|
2016-03-04 09:37:58 -08:00
|
|
|
varargs = append(varargs, m)
|
2016-03-08 15:10:26 -08:00
|
|
|
as.List.Append(m)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
break
|
|
|
|
}
|
|
|
|
|
2016-03-08 15:10:26 -08:00
|
|
|
as.List.Append(tinlvar(t))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// match arguments except final variadic (unless the call is dotted itself)
|
2016-03-13 23:02:38 -07:00
|
|
|
t, it := IterFields(fn.Type.Params())
|
|
|
|
for t != nil {
|
2016-03-09 12:39:36 -08:00
|
|
|
if li >= n.List.Len() {
|
2015-02-13 14:40:36 -05:00
|
|
|
break
|
|
|
|
}
|
2015-03-09 16:24:07 +11:00
|
|
|
if variadic && t.Isddd {
|
2015-02-13 14:40:36 -05:00
|
|
|
break
|
|
|
|
}
|
2016-03-08 15:10:26 -08:00
|
|
|
as.List.Append(tinlvar(t))
|
2016-03-13 23:02:38 -07:00
|
|
|
t = it.Next()
|
2016-03-09 12:39:36 -08:00
|
|
|
li++
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// match varargcount arguments with variadic parameters.
|
2015-03-09 16:24:07 +11:00
|
|
|
if variadic && t != nil && t.Isddd {
|
2015-02-13 14:40:36 -05:00
|
|
|
vararg = tinlvar(t)
|
2015-02-23 16:07:24 -05:00
|
|
|
var i int
|
2016-03-09 12:39:36 -08:00
|
|
|
for i = 0; i < varargcount && li < n.List.Len(); i++ {
|
2015-02-13 14:40:36 -05:00
|
|
|
m = argvar(varargtype, i)
|
2016-03-04 09:37:58 -08:00
|
|
|
varargs = append(varargs, m)
|
2016-03-08 15:10:26 -08:00
|
|
|
as.List.Append(m)
|
2016-03-09 12:39:36 -08:00
|
|
|
li++
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
if i == varargcount {
|
2016-03-13 23:02:38 -07:00
|
|
|
t = it.Next()
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-03-09 12:39:36 -08:00
|
|
|
if li < n.List.Len() || t != nil {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("arg count mismatch: %v vs %v\n", Tconv(fn.Type.Params(), FmtSharp), Hconv(n.List, FmtComma))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-03-08 15:10:26 -08:00
|
|
|
if as.Rlist.Len() != 0 {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
as = typecheck(as, Etop)
|
2016-03-08 15:10:26 -08:00
|
|
|
ninit.Append(as)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// turn the variadic args into a slice.
|
2015-02-17 22:13:49 -05:00
|
|
|
if variadic {
|
2015-02-13 14:40:36 -05:00
|
|
|
as = Nod(OAS, vararg, nil)
|
2015-02-17 22:13:49 -05:00
|
|
|
if varargcount == 0 {
|
2015-02-13 14:40:36 -05:00
|
|
|
as.Right = nodnil()
|
|
|
|
as.Right.Type = varargtype
|
|
|
|
} else {
|
2015-02-23 16:07:24 -05:00
|
|
|
vararrtype := typ(TARRAY)
|
2015-02-13 14:40:36 -05:00
|
|
|
vararrtype.Type = varargtype.Type
|
|
|
|
vararrtype.Bound = int64(varargcount)
|
|
|
|
|
|
|
|
as.Right = Nod(OCOMPLIT, nil, typenod(varargtype))
|
2016-03-08 15:10:26 -08:00
|
|
|
as.Right.List.Set(varargs)
|
2015-02-13 14:40:36 -05:00
|
|
|
as.Right = Nod(OSLICE, as.Right, Nod(OKEY, nil, nil))
|
|
|
|
}
|
|
|
|
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
as = typecheck(as, Etop)
|
2016-03-08 15:10:26 -08:00
|
|
|
ninit.Append(as)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// zero the outparams
|
2016-03-10 11:49:20 -08:00
|
|
|
for _, n := range retvars {
|
2016-03-09 20:29:21 -08:00
|
|
|
as = Nod(OAS, n, nil)
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
as = typecheck(as, Etop)
|
2016-03-08 15:10:26 -08:00
|
|
|
ninit.Append(as)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2016-03-10 11:49:20 -08:00
|
|
|
retlabel := newlabel_inl()
|
2015-02-13 14:40:36 -05:00
|
|
|
inlgen++
|
|
|
|
|
2016-03-10 11:49:20 -08:00
|
|
|
subst := inlsubst{
|
|
|
|
retlabel: retlabel,
|
|
|
|
retvars: retvars,
|
|
|
|
}
|
|
|
|
|
|
|
|
body := subst.list(fn.Func.Inl)
|
|
|
|
|
|
|
|
body = append(body, Nod(OGOTO, retlabel, nil)) // avoid 'not used' when function doesn't have return
|
|
|
|
body = append(body, Nod(OLABEL, retlabel, nil))
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-02-27 14:31:33 -08:00
|
|
|
typecheckslice(body, Etop)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
//dumplist("ninit post", ninit);
|
|
|
|
|
2015-02-23 16:07:24 -05:00
|
|
|
call := Nod(OINLCALL, nil, nil)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-03-08 15:10:26 -08:00
|
|
|
call.Ninit.Set(ninit.Slice())
|
2016-02-27 14:31:33 -08:00
|
|
|
call.Nbody.Set(body)
|
2016-03-10 11:49:20 -08:00
|
|
|
call.Rlist.Set(retvars)
|
2015-02-13 14:40:36 -05:00
|
|
|
call.Type = n.Type
|
|
|
|
call.Typecheck = 1
|
|
|
|
|
2015-06-29 16:30:19 -04:00
|
|
|
// Hide the args from setlno -- the parameters to the inlined
|
|
|
|
// call already have good line numbers that should be preserved.
|
|
|
|
args := as.Rlist
|
2016-03-08 15:10:26 -08:00
|
|
|
as.Rlist.Set(nil)
|
2015-06-29 16:30:19 -04:00
|
|
|
|
2016-03-02 12:49:37 -08:00
|
|
|
setlno(call, n.Lineno)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-03-08 15:10:26 -08:00
|
|
|
as.Rlist.Set(args.Slice())
|
2015-06-29 16:30:19 -04:00
|
|
|
|
2015-02-13 14:40:36 -05:00
|
|
|
//dumplist("call body", body);
|
|
|
|
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
n = call
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
// transitive inlining
|
2015-02-24 12:19:01 -05:00
|
|
|
// might be nice to do this before exporting the body,
|
|
|
|
// but can't emit the body with inlining expanded.
|
|
|
|
// instead we emit the things that the body needs
|
|
|
|
// and each use must redo the inlining.
|
|
|
|
// luckily these are small.
|
2016-02-27 14:31:33 -08:00
|
|
|
body = fn.Func.Inl.Slice()
|
|
|
|
fn.Func.Inl.Set(nil) // prevent infinite recursion (shouldn't happen anyway)
|
2016-03-04 09:37:58 -08:00
|
|
|
inlnodelist(call.Nbody)
|
2016-02-27 14:31:33 -08:00
|
|
|
for _, n := range call.Nbody.Slice() {
|
|
|
|
if n.Op == OINLCALL {
|
|
|
|
inlconv2stmt(n)
|
2015-02-24 12:19:01 -05:00
|
|
|
}
|
|
|
|
}
|
2016-02-27 14:31:33 -08:00
|
|
|
fn.Func.Inl.Set(body)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
if Debug['m'] > 2 {
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
fmt.Printf("%v: After inlining %v\n\n", n.Line(), Nconv(n, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
|
|
|
|
return n
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// Every time we expand a function we generate a new set of tmpnames,
|
|
|
|
// PAUTO's in the calling functions, and link them off of the
|
|
|
|
// PPARAM's, PAUTOS and PPARAMOUTs of the called function.
|
|
|
|
func inlvar(var_ *Node) *Node {
|
|
|
|
if Debug['m'] > 3 {
|
2016-03-15 13:06:58 -07:00
|
|
|
fmt.Printf("inlvar %v\n", Nconv(var_, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2015-02-23 16:07:24 -05:00
|
|
|
n := newname(var_.Sym)
|
2015-02-13 14:40:36 -05:00
|
|
|
n.Type = var_.Type
|
|
|
|
n.Class = PAUTO
|
2015-03-06 21:18:41 +11:00
|
|
|
n.Used = true
|
2015-05-27 07:31:56 -04:00
|
|
|
n.Name.Curfn = Curfn // the calling function, not the called one
|
2015-02-13 14:40:36 -05:00
|
|
|
n.Addrtaken = var_.Addrtaken
|
|
|
|
|
2015-06-16 18:28:01 -04:00
|
|
|
// This may no longer be necessary now that we run escape analysis
|
|
|
|
// after wrapper generation, but for 1.5 this is conservatively left
|
2016-03-01 23:21:55 +00:00
|
|
|
// unchanged. See bugs 11053 and 9537.
|
2015-06-16 18:28:01 -04:00
|
|
|
if var_.Esc == EscHeap {
|
2015-02-13 14:40:36 -05:00
|
|
|
addrescapes(n)
|
|
|
|
}
|
|
|
|
|
2016-02-25 10:35:19 -08:00
|
|
|
Curfn.Func.Dcl = append(Curfn.Func.Dcl, n)
|
2015-02-13 14:40:36 -05:00
|
|
|
return n
|
|
|
|
}
|
|
|
|
|
|
|
|
// Synthesize a variable to store the inlined function's results in.
|
2016-03-14 01:20:49 -07:00
|
|
|
func retvar(t *Field, i int) *Node {
|
2016-03-19 18:17:58 -07:00
|
|
|
n := newname(LookupN("~r", i))
|
2015-02-13 14:40:36 -05:00
|
|
|
n.Type = t.Type
|
|
|
|
n.Class = PAUTO
|
2015-03-06 21:18:41 +11:00
|
|
|
n.Used = true
|
2015-05-27 07:31:56 -04:00
|
|
|
n.Name.Curfn = Curfn // the calling function, not the called one
|
2016-02-25 10:35:19 -08:00
|
|
|
Curfn.Func.Dcl = append(Curfn.Func.Dcl, n)
|
2015-02-13 14:40:36 -05:00
|
|
|
return n
|
|
|
|
}
|
|
|
|
|
|
|
|
// Synthesize a variable to store the inlined function's arguments
|
|
|
|
// when they come from a multiple return call.
|
|
|
|
func argvar(t *Type, i int) *Node {
|
2016-03-19 18:17:58 -07:00
|
|
|
n := newname(LookupN("~arg", i))
|
2015-02-13 14:40:36 -05:00
|
|
|
n.Type = t.Type
|
|
|
|
n.Class = PAUTO
|
2015-03-06 21:18:41 +11:00
|
|
|
n.Used = true
|
2015-05-27 07:31:56 -04:00
|
|
|
n.Name.Curfn = Curfn // the calling function, not the called one
|
2016-02-25 10:35:19 -08:00
|
|
|
Curfn.Func.Dcl = append(Curfn.Func.Dcl, n)
|
2015-02-13 14:40:36 -05:00
|
|
|
return n
|
|
|
|
}
|
|
|
|
|
|
|
|
var newlabel_inl_label int
|
|
|
|
|
|
|
|
func newlabel_inl() *Node {
|
|
|
|
newlabel_inl_label++
|
2016-03-19 18:17:58 -07:00
|
|
|
n := newname(LookupN(".inlret", newlabel_inl_label))
|
2015-02-13 14:40:36 -05:00
|
|
|
n.Etype = 1 // flag 'safe' for escape analysis (no backjumps)
|
|
|
|
return n
|
|
|
|
}
|
|
|
|
|
2016-03-10 11:49:20 -08:00
|
|
|
// The inlsubst type implements the actual inlining of a single
|
|
|
|
// function call.
|
|
|
|
type inlsubst struct {
|
|
|
|
// Target of the goto substituted in place of a return.
|
|
|
|
retlabel *Node
|
|
|
|
|
|
|
|
// Temporary result variables.
|
|
|
|
retvars []*Node
|
|
|
|
}
|
|
|
|
|
|
|
|
// list inlines a list of nodes.
|
|
|
|
func (subst *inlsubst) list(ll Nodes) []*Node {
|
2016-03-08 15:10:26 -08:00
|
|
|
s := make([]*Node, 0, ll.Len())
|
|
|
|
for _, n := range ll.Slice() {
|
2016-03-10 11:49:20 -08:00
|
|
|
s = append(s, subst.node(n))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2016-03-04 09:37:58 -08:00
|
|
|
return s
|
2016-02-27 14:31:33 -08:00
|
|
|
}
|
|
|
|
|
2016-03-10 11:49:20 -08:00
|
|
|
// node recursively copies a node from the saved pristine body of the
|
|
|
|
// inlined function, substituting references to input/output
|
|
|
|
// parameters with ones to the tmpnames, and substituting returns with
|
|
|
|
// assignments to the output.
|
|
|
|
func (subst *inlsubst) node(n *Node) *Node {
|
2015-02-13 14:40:36 -05:00
|
|
|
if n == nil {
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
switch n.Op {
|
|
|
|
case ONAME:
|
2015-05-15 10:02:19 -07:00
|
|
|
if n.Name.Inlvar != nil { // These will be set during inlnode
|
2015-02-13 14:40:36 -05:00
|
|
|
if Debug['m'] > 2 {
|
2016-03-15 13:06:58 -07:00
|
|
|
fmt.Printf("substituting name %v -> %v\n", Nconv(n, FmtSign), Nconv(n.Name.Inlvar, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2015-05-15 10:02:19 -07:00
|
|
|
return n.Name.Inlvar
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
if Debug['m'] > 2 {
|
2016-03-15 13:06:58 -07:00
|
|
|
fmt.Printf("not substituting name %v\n", Nconv(n, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
return n
|
|
|
|
|
2015-04-01 09:38:44 -07:00
|
|
|
case OLITERAL, OTYPE:
|
2015-02-13 14:40:36 -05:00
|
|
|
return n
|
|
|
|
|
|
|
|
// Since we don't handle bodies with closures, this return is guaranteed to belong to the current inlined function.
|
|
|
|
|
|
|
|
// dump("Return before substitution", n);
|
|
|
|
case ORETURN:
|
2016-03-10 11:49:20 -08:00
|
|
|
m := Nod(OGOTO, subst.retlabel, nil)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-03-10 11:49:20 -08:00
|
|
|
m.Ninit.Set(subst.list(n.Ninit))
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-03-10 11:49:20 -08:00
|
|
|
if len(subst.retvars) != 0 && n.List.Len() != 0 {
|
2015-02-23 16:07:24 -05:00
|
|
|
as := Nod(OAS2, nil, nil)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
2016-03-10 11:49:20 -08:00
|
|
|
// Make a shallow copy of retvars.
|
|
|
|
// Otherwise OINLCALL.Rlist will be the same list,
|
|
|
|
// and later walk and typecheck may clobber it.
|
|
|
|
for _, n := range subst.retvars {
|
2016-03-09 20:29:21 -08:00
|
|
|
as.List.Append(n)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
2016-03-10 11:49:20 -08:00
|
|
|
as.Rlist.Set(subst.list(n.List))
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
as = typecheck(as, Etop)
|
2016-03-08 15:10:26 -08:00
|
|
|
m.Ninit.Append(as)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2016-03-19 17:02:01 -07:00
|
|
|
typecheckslice(m.Ninit.Slice(), Etop)
|
cmd/compile: reduce use of **Node parameters
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
2016-03-20 08:03:31 -07:00
|
|
|
m = typecheck(m, Etop)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
// dump("Return after substitution", m);
|
|
|
|
return m
|
|
|
|
|
2015-04-01 09:38:44 -07:00
|
|
|
case OGOTO, OLABEL:
|
2015-02-23 16:07:24 -05:00
|
|
|
m := Nod(OXXX, nil, nil)
|
2015-02-13 14:40:36 -05:00
|
|
|
*m = *n
|
2016-03-08 15:10:26 -08:00
|
|
|
m.Ninit.Set(nil)
|
2015-02-23 16:07:24 -05:00
|
|
|
p := fmt.Sprintf("%s·%d", n.Left.Sym.Name, inlgen)
|
2015-02-13 14:40:36 -05:00
|
|
|
m.Left = newname(Lookup(p))
|
|
|
|
|
|
|
|
return m
|
|
|
|
}
|
|
|
|
|
2015-02-23 16:07:24 -05:00
|
|
|
m := Nod(OXXX, nil, nil)
|
2015-02-13 14:40:36 -05:00
|
|
|
*m = *n
|
2016-03-08 15:10:26 -08:00
|
|
|
m.Ninit.Set(nil)
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
if n.Op == OCLOSURE {
|
2016-03-15 13:06:58 -07:00
|
|
|
Fatalf("cannot inline function containing closure: %v", Nconv(n, FmtSign))
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
2016-03-10 11:49:20 -08:00
|
|
|
m.Left = subst.node(n.Left)
|
|
|
|
m.Right = subst.node(n.Right)
|
|
|
|
m.List.Set(subst.list(n.List))
|
|
|
|
m.Rlist.Set(subst.list(n.Rlist))
|
|
|
|
m.Ninit.Set(append(m.Ninit.Slice(), subst.list(n.Ninit)...))
|
|
|
|
m.Nbody.Set(subst.list(n.Nbody))
|
2015-02-13 14:40:36 -05:00
|
|
|
|
|
|
|
return m
|
|
|
|
}
|
|
|
|
|
|
|
|
// Plaster over linenumbers
|
2016-03-08 10:26:20 -08:00
|
|
|
func setlnolist(ll Nodes, lno int32) {
|
2016-03-08 15:10:26 -08:00
|
|
|
for _, n := range ll.Slice() {
|
|
|
|
setlno(n, lno)
|
2016-02-27 14:31:33 -08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-03-02 12:49:37 -08:00
|
|
|
func setlno(n *Node, lno int32) {
|
2015-02-17 22:13:49 -05:00
|
|
|
if n == nil {
|
2015-02-13 14:40:36 -05:00
|
|
|
return
|
|
|
|
}
|
|
|
|
|
|
|
|
// don't clobber names, unless they're freshly synthesized
|
|
|
|
if n.Op != ONAME || n.Lineno == 0 {
|
2016-03-02 12:49:37 -08:00
|
|
|
n.Lineno = lno
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
setlno(n.Left, lno)
|
|
|
|
setlno(n.Right, lno)
|
|
|
|
setlnolist(n.List, lno)
|
|
|
|
setlnolist(n.Rlist, lno)
|
|
|
|
setlnolist(n.Ninit, lno)
|
2016-03-04 09:37:58 -08:00
|
|
|
setlnolist(n.Nbody, lno)
|
2015-02-13 14:40:36 -05:00
|
|
|
}
|