Stowage/go
2
0
Fork 0
mirror of https://github.com/golang/go.git synced 2025-12-08 06:10:04 +00:00
Code Issues Projects Releases Packages Wiki Activity

[dev.simd] cmd/compile, simd/_gen: add rewrite for const load ops

Browse source 
This CL adds rewrite rules for ops with const imm8 that takes a load to
its memory form.

Change-Id: I74d0df48715ab48b88b04c8e1bfb3c6b8e528aeb
Reviewed-on: https://go-review.googlesource.com/c/go/+/704635
TryBot-Bypass: Junyang Shao <shaojunyang@google.com>
Reviewed-by: David Chase <drchase@google.com>
This commit is contained in:
Junyang Shao 2025-09-17 14:44:49 +00:00
parent 443b7aeddb
commit 4eb5c6e07b
3 changed files with 4359 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

138
src/cmd/compile/internal/ssa/_gen/simdAMD64.rules
View file

@ -1527,6 +1527,30 @@
(VPANDNQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPANDNQMasked128load {sym} [off] x ptr mask mem)
(VPANDNQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPANDNQMasked256load {sym} [off] x ptr mask mem)
(VPANDNQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPANDNQMasked512load {sym} [off] x ptr mask mem)
(VRNDSCALEPS128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPS128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VRNDSCALEPS256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPS256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VRNDSCALEPS512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPS512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VRNDSCALEPD128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPD128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VRNDSCALEPD256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPD256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VRNDSCALEPD512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPD512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VRNDSCALEPSMasked128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPSMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VRNDSCALEPSMasked256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPSMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VRNDSCALEPSMasked512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPSMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VRNDSCALEPDMasked128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPDMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VRNDSCALEPDMasked256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPDMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VRNDSCALEPDMasked512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VRNDSCALEPDMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VREDUCEPS128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPS128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VREDUCEPS256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPS256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VREDUCEPS512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPS512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VREDUCEPD128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPD128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VREDUCEPD256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPD256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VREDUCEPD512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPD512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VREDUCEPSMasked128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPSMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VREDUCEPSMasked256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPSMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VREDUCEPSMasked512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPSMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VREDUCEPDMasked128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPDMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VREDUCEPDMasked256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPDMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VREDUCEPDMasked512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VREDUCEPDMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPACKSSDW512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPACKSSDW512load {sym} [off] x ptr mem)
(VPACKSSDWMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPACKSSDWMasked128load {sym} [off] x ptr mask mem)
(VPACKSSDWMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPACKSSDWMasked256load {sym} [off] x ptr mask mem)
@ -1555,8 +1579,44 @@
(VDIVPDMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPDMasked512load {sym} [off] x ptr mask mem)
(VPCMPEQD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPEQD512load {sym} [off] x ptr mem)
(VPCMPEQQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPEQQ512load {sym} [off] x ptr mem)
(VCMPPS512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VCMPPS512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VCMPPD512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VCMPPD512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VCMPPSMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCMPPSMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VCMPPSMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCMPPSMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VCMPPSMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCMPPSMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VCMPPDMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCMPPDMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VCMPPDMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCMPPDMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VCMPPDMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCMPPDMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPDMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPDMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPDMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPDMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPDMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPDMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPQMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPQMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPQMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPQMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPQMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPQMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPUDMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPUDMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPUDMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPUDMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPUDMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPUDMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPUQMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPUQMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPUQMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPUQMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPUQMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPCMPUQMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VGF2P8AFFINEQB128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEQB128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VGF2P8AFFINEQB256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEQB256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VGF2P8AFFINEQB512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEQB512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VGF2P8AFFINEINVQB128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEINVQB128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VGF2P8AFFINEINVQB256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEINVQB256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VGF2P8AFFINEINVQB512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEINVQB512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VGF2P8AFFINEINVQBMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEINVQBMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VGF2P8AFFINEINVQBMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEINVQBMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VGF2P8AFFINEINVQBMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEINVQBMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VGF2P8AFFINEQBMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEQBMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VGF2P8AFFINEQBMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEQBMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VGF2P8AFFINEQBMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VGF2P8AFFINEQBMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPCMPGTD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPGTD512load {sym} [off] x ptr mem)
(VPCMPGTQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPGTQ512load {sym} [off] x ptr mem)
(VPCMPUD512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPUD512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPCMPUQ512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPUQ512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPCMPD512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPD512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPCMPQ512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPQ512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPUNPCKHDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKHDQ512load {sym} [off] x ptr mem)
(VPUNPCKHQDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKHQDQ512load {sym} [off] x ptr mem)
(VPUNPCKLDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKLDQ512load {sym} [off] x ptr mem)
@ -1733,6 +1793,10 @@
(VPERMI2QMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPERMI2QMasked256load {sym} [off] x y ptr mask mem)
(VPERMI2PDMasked512 x y l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPERMI2PDMasked512load {sym} [off] x y ptr mask mem)
(VPERMI2QMasked512 x y l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPERMI2QMasked512load {sym} [off] x y ptr mask mem)
(VPSHUFD512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHUFD512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPSHUFDMasked256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHUFDMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSHUFDMasked512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHUFDMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSHUFDMasked128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHUFDMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPERMPSMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPERMPSMasked256load {sym} [off] x ptr mask mem)
(VPERMDMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPERMDMasked256load {sym} [off] x ptr mask mem)
(VPERMPSMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPERMPSMasked512load {sym} [off] x ptr mask mem)
@ -1761,6 +1825,30 @@
(VRSQRT14PDMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VRSQRT14PDMasked128load {sym} [off] ptr mask mem)
(VRSQRT14PDMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VRSQRT14PDMasked256load {sym} [off] ptr mask mem)
(VRSQRT14PDMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VRSQRT14PDMasked512load {sym} [off] ptr mask mem)
(VPROLD128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPROLD128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPROLD256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPROLD256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPROLD512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPROLD512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPROLQ128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPROLQ128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPROLQ256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPROLQ256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPROLQ512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPROLQ512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPROLDMasked128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPROLDMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPROLDMasked256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPROLDMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPROLDMasked512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPROLDMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPROLQMasked128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPROLQMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPROLQMasked256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPROLQMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPROLQMasked512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPROLQMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPRORD128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPRORD128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPRORD256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPRORD256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPRORD512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPRORD512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPRORQ128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPRORQ128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPRORQ256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPRORQ256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPRORQ512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPRORQ512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPRORDMasked128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPRORDMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPRORDMasked256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPRORDMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPRORDMasked512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPRORDMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPRORQMasked128 [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPRORQMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPRORQMasked256 [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPRORQMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPRORQMasked512 [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPRORQMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPROLVD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPROLVD128load {sym} [off] x ptr mem)
(VPROLVD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPROLVD256load {sym} [off] x ptr mem)
(VPROLVD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPROLVD512load {sym} [off] x ptr mem)
@ -1797,6 +1885,30 @@
(VSCALEFPDMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSCALEFPDMasked128load {sym} [off] x ptr mask mem)
(VSCALEFPDMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSCALEFPDMasked256load {sym} [off] x ptr mask mem)
(VSCALEFPDMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSCALEFPDMasked512load {sym} [off] x ptr mask mem)
(VPSHLDD128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHLDD128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHLDD256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHLDD256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHLDD512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHLDD512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHLDQ128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHLDQ128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHLDQ256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHLDQ256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHLDQ512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHLDQ512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHLDDMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHLDDMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHLDDMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHLDDMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHLDDMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHLDDMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHLDQMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHLDQMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHLDQMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHLDQMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHLDQMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHLDQMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHRDD128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHRDD128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHRDD256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHRDD256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHRDD512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHRDD512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHRDQ128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHRDQ128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHRDQ256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHRDQ256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHRDQ512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHRDQ512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mem)
(VPSHRDDMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHRDDMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHRDDMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHRDDMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHRDDMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHRDDMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHRDQMasked128 [c] x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHRDQMasked128load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHRDQMasked256 [c] x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHRDQMasked256load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSHRDQMasked512 [c] x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSHRDQMasked512load {sym} [makeValAndOff(int32(int8(c)),off)] x ptr mask mem)
(VPSLLVD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLVD512load {sym} [off] x ptr mem)
(VPSLLVQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLVQ512load {sym} [off] x ptr mem)
(VPSHLDVD128 x y l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHLDVD128load {sym} [off] x y ptr mem)
@ -1881,3 +1993,29 @@
(VPXORQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPXORQMasked512load {sym} [off] x ptr mask mem)
(VPBLENDMDMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPBLENDMDMasked512load {sym} [off] x ptr mask mem)
(VPBLENDMQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPBLENDMQMasked512load {sym} [off] x ptr mask mem)
(VPSLLD512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLD512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPSLLQ512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLQ512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPSLLDMasked128const [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLDMasked128constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSLLDMasked256const [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLDMasked256constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSLLDMasked512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLDMasked512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSLLQMasked128const [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLQMasked128constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSLLQMasked256const [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLQMasked256constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSLLQMasked512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLQMasked512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRLD512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRLD512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPSRLQ512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRLQ512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPSRAD512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAD512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPSRAQ128const [c] l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAQ128constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPSRAQ256const [c] l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAQ256constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPSRAQ512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAQ512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
(VPSRLDMasked128const [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLDMasked128constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRLDMasked256const [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLDMasked256constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRLDMasked512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLDMasked512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRLQMasked128const [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLQMasked128constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRLQMasked256const [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLQMasked256constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRLQMasked512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLQMasked512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRADMasked128const [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRADMasked128constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRADMasked256const [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRADMasked256constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRADMasked512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRADMasked512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRAQMasked128const [c] l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRAQMasked128constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRAQMasked256const [c] l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRAQMasked256constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)
(VPSRAQMasked512const [c] l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRAQMasked512constload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mask mem)

4206
src/cmd/compile/internal/ssa/rewriteAMD64.go
View file

File diff suppressed because it is too large Load diff

18
src/simd/_gen/simdgen/gen_simdrules.go
View file

@ -236,7 +236,7 @@ func writeSIMDRules(ops []Operation) *bytes.Buffer {
panic("simdgen sees unknwon special lower " + *gOp.SpecialLower + ", maybe implement it?")
}
}
if gOp.MemFeatures != nil && *gOp.MemFeatures == "vbcst" && immType == NoImm {
if gOp.MemFeatures != nil && *gOp.MemFeatures == "vbcst" {
// sanity check
selected := true
for _, a := range gOp.In {
@ -257,9 +257,21 @@ func writeSIMDRules(ops []Operation) *bytes.Buffer {
}
memOpData := data
// Remove the last vreg from the arg and change it to a load.
memOpData.ArgsLoadAddr = data.Args[:len(data.Args)-1] + fmt.Sprintf("l:(VMOVDQUload%d {sym} [off] ptr mem)", *lastVreg.Bits)
origArgs := data.Args[:len(data.Args)-1]
// Prepare imm args.
immArg := ""
immArgCombineOff := " [off] "
if immType != NoImm && immType != InvalidImm {
_, after, found := strings.Cut(origArgs, "]")
if found {
origArgs = after
}
immArg = "[c] "
immArgCombineOff = " [makeValAndOff(int32(int8(c)),off)] "
}
memOpData.ArgsLoadAddr = immArg + origArgs + fmt.Sprintf("l:(VMOVDQUload%d {sym} [off] ptr mem)", *lastVreg.Bits)
// Remove the last vreg from the arg and change it to "ptr".
memOpData.ArgsAddr = "{sym} [off] " + data.Args[:len(data.Args)-1] + "ptr"
memOpData.ArgsAddr = "{sym}" + immArgCombineOff + origArgs + "ptr"
if maskType == OneMask {
memOpData.ArgsAddr += " mask"
memOpData.ArgsLoadAddr += " mask"