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: make rewrite rules consistent on CPU Features

Browse source 
The previous CL left a bug in the xed parser so that the generator can
generate rules rewriting an AVX instruction to AVX512 instruction.

This CL fixes that.

Change-Id: I0df7e7dc6c936ce7add24a757ce7f44a15917fef
Reviewed-on: https://go-review.googlesource.com/c/go/+/703399
Reviewed-by: David Chase <drchase@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
This commit is contained in:
Junyang Shao 2025-09-12 18:45:39 +00:00
parent bdd30e25ca
commit 443b7aeddb
7 changed files with 41 additions and 5036 deletions
Download patch file Download diff file Expand all files Collapse all files

118
src/cmd/compile/internal/amd64/simdssa.go
View file

@ -1397,110 +1397,50 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
ssa.OpAMD64VSQRTPDMasked512load: ssa.OpAMD64VSQRTPDMasked512load:
p = simdVkvload(s, v) p = simdVkvload(s, v)
case ssa.OpAMD64VADDPS128load, case ssa.OpAMD64VADDPS512load,
ssa.OpAMD64VADDPS256load,
ssa.OpAMD64VADDPS512load,
ssa.OpAMD64VADDPD128load,
ssa.OpAMD64VADDPD256load,
ssa.OpAMD64VADDPD512load, ssa.OpAMD64VADDPD512load,
ssa.OpAMD64VPADDD128load,
ssa.OpAMD64VPADDD256load,
ssa.OpAMD64VPADDD512load, ssa.OpAMD64VPADDD512load,
ssa.OpAMD64VPADDQ128load,
ssa.OpAMD64VPADDQ256load,
ssa.OpAMD64VPADDQ512load, ssa.OpAMD64VPADDQ512load,
ssa.OpAMD64VPANDD512load, ssa.OpAMD64VPANDD512load,
ssa.OpAMD64VPANDQ512load, ssa.OpAMD64VPANDQ512load,
ssa.OpAMD64VPANDND512load, ssa.OpAMD64VPANDND512load,
ssa.OpAMD64VPANDNQ512load, ssa.OpAMD64VPANDNQ512load,
ssa.OpAMD64VPACKSSDW128load,
ssa.OpAMD64VPACKSSDW256load,
ssa.OpAMD64VPACKSSDW512load, ssa.OpAMD64VPACKSSDW512load,
ssa.OpAMD64VPACKUSDW128load,
ssa.OpAMD64VPACKUSDW256load,
ssa.OpAMD64VPACKUSDW512load, ssa.OpAMD64VPACKUSDW512load,
ssa.OpAMD64VDIVPS128load,
ssa.OpAMD64VDIVPS256load,
ssa.OpAMD64VDIVPS512load, ssa.OpAMD64VDIVPS512load,
ssa.OpAMD64VDIVPD128load,
ssa.OpAMD64VDIVPD256load,
ssa.OpAMD64VDIVPD512load, ssa.OpAMD64VDIVPD512load,
ssa.OpAMD64VPCMPEQD128load,
ssa.OpAMD64VPCMPEQD256load,
ssa.OpAMD64VPCMPEQQ128load,
ssa.OpAMD64VPCMPEQQ256load,
ssa.OpAMD64VPCMPGTD128load,
ssa.OpAMD64VPCMPGTD256load,
ssa.OpAMD64VPCMPGTQ128load,
ssa.OpAMD64VPCMPGTQ256load,
ssa.OpAMD64VPUNPCKHDQ128load,
ssa.OpAMD64VPUNPCKHQDQ128load,
ssa.OpAMD64VPUNPCKHDQ256load,
ssa.OpAMD64VPUNPCKHDQ512load, ssa.OpAMD64VPUNPCKHDQ512load,
ssa.OpAMD64VPUNPCKHQDQ256load,
ssa.OpAMD64VPUNPCKHQDQ512load, ssa.OpAMD64VPUNPCKHQDQ512load,
ssa.OpAMD64VPUNPCKLDQ128load,
ssa.OpAMD64VPUNPCKLQDQ128load,
ssa.OpAMD64VPUNPCKLDQ256load,
ssa.OpAMD64VPUNPCKLDQ512load, ssa.OpAMD64VPUNPCKLDQ512load,
ssa.OpAMD64VPUNPCKLQDQ256load,
ssa.OpAMD64VPUNPCKLQDQ512load, ssa.OpAMD64VPUNPCKLQDQ512load,
ssa.OpAMD64VMAXPS128load,
ssa.OpAMD64VMAXPS256load,
ssa.OpAMD64VMAXPS512load, ssa.OpAMD64VMAXPS512load,
ssa.OpAMD64VMAXPD128load,
ssa.OpAMD64VMAXPD256load,
ssa.OpAMD64VMAXPD512load, ssa.OpAMD64VMAXPD512load,
ssa.OpAMD64VPMAXSD128load,
ssa.OpAMD64VPMAXSD256load,
ssa.OpAMD64VPMAXSD512load, ssa.OpAMD64VPMAXSD512load,
ssa.OpAMD64VPMAXSQ128load, ssa.OpAMD64VPMAXSQ128load,
ssa.OpAMD64VPMAXSQ256load, ssa.OpAMD64VPMAXSQ256load,
ssa.OpAMD64VPMAXSQ512load, ssa.OpAMD64VPMAXSQ512load,
ssa.OpAMD64VPMAXUD128load,
ssa.OpAMD64VPMAXUD256load,
ssa.OpAMD64VPMAXUD512load, ssa.OpAMD64VPMAXUD512load,
ssa.OpAMD64VPMAXUQ128load, ssa.OpAMD64VPMAXUQ128load,
ssa.OpAMD64VPMAXUQ256load, ssa.OpAMD64VPMAXUQ256load,
ssa.OpAMD64VPMAXUQ512load, ssa.OpAMD64VPMAXUQ512load,
ssa.OpAMD64VMINPS128load,
ssa.OpAMD64VMINPS256load,
ssa.OpAMD64VMINPS512load, ssa.OpAMD64VMINPS512load,
ssa.OpAMD64VMINPD128load,
ssa.OpAMD64VMINPD256load,
ssa.OpAMD64VMINPD512load, ssa.OpAMD64VMINPD512load,
ssa.OpAMD64VPMINSD128load,
ssa.OpAMD64VPMINSD256load,
ssa.OpAMD64VPMINSD512load, ssa.OpAMD64VPMINSD512load,
ssa.OpAMD64VPMINSQ128load, ssa.OpAMD64VPMINSQ128load,
ssa.OpAMD64VPMINSQ256load, ssa.OpAMD64VPMINSQ256load,
ssa.OpAMD64VPMINSQ512load, ssa.OpAMD64VPMINSQ512load,
ssa.OpAMD64VPMINUD128load,
ssa.OpAMD64VPMINUD256load,
ssa.OpAMD64VPMINUD512load, ssa.OpAMD64VPMINUD512load,
ssa.OpAMD64VPMINUQ128load, ssa.OpAMD64VPMINUQ128load,
ssa.OpAMD64VPMINUQ256load, ssa.OpAMD64VPMINUQ256load,
ssa.OpAMD64VPMINUQ512load, ssa.OpAMD64VPMINUQ512load,
ssa.OpAMD64VMULPS128load,
ssa.OpAMD64VMULPS256load,
ssa.OpAMD64VMULPS512load, ssa.OpAMD64VMULPS512load,
ssa.OpAMD64VMULPD128load,
ssa.OpAMD64VMULPD256load,
ssa.OpAMD64VMULPD512load, ssa.OpAMD64VMULPD512load,
ssa.OpAMD64VPMULLD128load,
ssa.OpAMD64VPMULLD256load,
ssa.OpAMD64VPMULLD512load, ssa.OpAMD64VPMULLD512load,
ssa.OpAMD64VPMULLQ128load, ssa.OpAMD64VPMULLQ128load,
ssa.OpAMD64VPMULLQ256load, ssa.OpAMD64VPMULLQ256load,
ssa.OpAMD64VPMULLQ512load, ssa.OpAMD64VPMULLQ512load,
ssa.OpAMD64VPMULDQ128load,
ssa.OpAMD64VPMULDQ256load,
ssa.OpAMD64VPMULUDQ128load,
ssa.OpAMD64VPMULUDQ256load,
ssa.OpAMD64VPORD512load, ssa.OpAMD64VPORD512load,
ssa.OpAMD64VPORQ512load, ssa.OpAMD64VPORQ512load,
ssa.OpAMD64VPERMPS256load,
ssa.OpAMD64VPERMD256load,
ssa.OpAMD64VPERMPS512load, ssa.OpAMD64VPERMPS512load,
ssa.OpAMD64VPERMD512load, ssa.OpAMD64VPERMD512load,
ssa.OpAMD64VPERMPD256load, ssa.OpAMD64VPERMPD256load,
@ -1525,51 +1465,25 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
ssa.OpAMD64VSCALEFPD128load, ssa.OpAMD64VSCALEFPD128load,
ssa.OpAMD64VSCALEFPD256load, ssa.OpAMD64VSCALEFPD256load,
ssa.OpAMD64VSCALEFPD512load, ssa.OpAMD64VSCALEFPD512load,
ssa.OpAMD64VPSLLVD128load,
ssa.OpAMD64VPSLLVD256load,
ssa.OpAMD64VPSLLVD512load, ssa.OpAMD64VPSLLVD512load,
ssa.OpAMD64VPSLLVQ128load,
ssa.OpAMD64VPSLLVQ256load,
ssa.OpAMD64VPSLLVQ512load, ssa.OpAMD64VPSLLVQ512load,
ssa.OpAMD64VPSRAVD128load,
ssa.OpAMD64VPSRAVD256load,
ssa.OpAMD64VPSRAVD512load, ssa.OpAMD64VPSRAVD512load,
ssa.OpAMD64VPSRAVQ128load, ssa.OpAMD64VPSRAVQ128load,
ssa.OpAMD64VPSRAVQ256load, ssa.OpAMD64VPSRAVQ256load,
ssa.OpAMD64VPSRAVQ512load, ssa.OpAMD64VPSRAVQ512load,
ssa.OpAMD64VPSRLVD128load,
ssa.OpAMD64VPSRLVD256load,
ssa.OpAMD64VPSRLVD512load, ssa.OpAMD64VPSRLVD512load,
ssa.OpAMD64VPSRLVQ128load,
ssa.OpAMD64VPSRLVQ256load,
ssa.OpAMD64VPSRLVQ512load, ssa.OpAMD64VPSRLVQ512load,
ssa.OpAMD64VSUBPS128load,
ssa.OpAMD64VSUBPS256load,
ssa.OpAMD64VSUBPS512load, ssa.OpAMD64VSUBPS512load,
ssa.OpAMD64VSUBPD128load,
ssa.OpAMD64VSUBPD256load,
ssa.OpAMD64VSUBPD512load, ssa.OpAMD64VSUBPD512load,
ssa.OpAMD64VPSUBD128load,
ssa.OpAMD64VPSUBD256load,
ssa.OpAMD64VPSUBD512load, ssa.OpAMD64VPSUBD512load,
ssa.OpAMD64VPSUBQ128load,
ssa.OpAMD64VPSUBQ256load,
ssa.OpAMD64VPSUBQ512load, ssa.OpAMD64VPSUBQ512load,
ssa.OpAMD64VPXORD512load, ssa.OpAMD64VPXORD512load,
ssa.OpAMD64VPXORQ512load: ssa.OpAMD64VPXORQ512load:
p = simdV21load(s, v) p = simdV21load(s, v)
case ssa.OpAMD64VPDPWSSD128load, case ssa.OpAMD64VPDPWSSD512load,
ssa.OpAMD64VPDPWSSD256load,
ssa.OpAMD64VPDPWSSD512load,
ssa.OpAMD64VPDPWSSDS128load,
ssa.OpAMD64VPDPWSSDS256load,
ssa.OpAMD64VPDPWSSDS512load, ssa.OpAMD64VPDPWSSDS512load,
ssa.OpAMD64VPDPBUSD128load,
ssa.OpAMD64VPDPBUSD256load,
ssa.OpAMD64VPDPBUSD512load, ssa.OpAMD64VPDPBUSD512load,
ssa.OpAMD64VPDPBUSDS128load,
ssa.OpAMD64VPDPBUSDS256load,
ssa.OpAMD64VPDPBUSDS512load, ssa.OpAMD64VPDPBUSDS512load,
ssa.OpAMD64VFMADD213PS128load, ssa.OpAMD64VFMADD213PS128load,
ssa.OpAMD64VFMADD213PS256load, ssa.OpAMD64VFMADD213PS256load,
@ -1833,14 +1747,10 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
ssa.OpAMD64VPCMPGTQ512load: ssa.OpAMD64VPCMPGTQ512load:
p = simdV2kload(s, v) p = simdV2kload(s, v)
case ssa.OpAMD64VPABSD128load, case ssa.OpAMD64VPABSD512load,
ssa.OpAMD64VPABSD256load,
ssa.OpAMD64VPABSD512load,
ssa.OpAMD64VPABSQ128load, ssa.OpAMD64VPABSQ128load,
ssa.OpAMD64VPABSQ256load, ssa.OpAMD64VPABSQ256load,
ssa.OpAMD64VPABSQ512load, ssa.OpAMD64VPABSQ512load,
ssa.OpAMD64VCVTTPS2DQ128load,
ssa.OpAMD64VCVTTPS2DQ256load,
ssa.OpAMD64VCVTTPS2DQ512load, ssa.OpAMD64VCVTTPS2DQ512load,
ssa.OpAMD64VCVTPS2UDQ128load, ssa.OpAMD64VCVTPS2UDQ128load,
ssa.OpAMD64VCVTPS2UDQ256load, ssa.OpAMD64VCVTPS2UDQ256load,
@ -1865,11 +1775,7 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
ssa.OpAMD64VRSQRT14PD128load, ssa.OpAMD64VRSQRT14PD128load,
ssa.OpAMD64VRSQRT14PD256load, ssa.OpAMD64VRSQRT14PD256load,
ssa.OpAMD64VRSQRT14PD512load, ssa.OpAMD64VRSQRT14PD512load,
ssa.OpAMD64VSQRTPS128load,
ssa.OpAMD64VSQRTPS256load,
ssa.OpAMD64VSQRTPS512load, ssa.OpAMD64VSQRTPS512load,
ssa.OpAMD64VSQRTPD128load,
ssa.OpAMD64VSQRTPD256load,
ssa.OpAMD64VSQRTPD512load: ssa.OpAMD64VSQRTPD512load:
p = simdV11load(s, v) p = simdV11load(s, v)
@ -1885,8 +1791,6 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
ssa.OpAMD64VREDUCEPD128load, ssa.OpAMD64VREDUCEPD128load,
ssa.OpAMD64VREDUCEPD256load, ssa.OpAMD64VREDUCEPD256load,
ssa.OpAMD64VREDUCEPD512load, ssa.OpAMD64VREDUCEPD512load,
ssa.OpAMD64VPSHUFD128load,
ssa.OpAMD64VPSHUFD256load,
ssa.OpAMD64VPSHUFD512load, ssa.OpAMD64VPSHUFD512load,
ssa.OpAMD64VPROLD128load, ssa.OpAMD64VPROLD128load,
ssa.OpAMD64VPROLD256load, ssa.OpAMD64VPROLD256load,
@ -1900,20 +1804,10 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
ssa.OpAMD64VPRORQ128load, ssa.OpAMD64VPRORQ128load,
ssa.OpAMD64VPRORQ256load, ssa.OpAMD64VPRORQ256load,
ssa.OpAMD64VPRORQ512load, ssa.OpAMD64VPRORQ512load,
ssa.OpAMD64VPSLLD128constload,
ssa.OpAMD64VPSLLD256constload,
ssa.OpAMD64VPSLLD512constload, ssa.OpAMD64VPSLLD512constload,
ssa.OpAMD64VPSLLQ128constload,
ssa.OpAMD64VPSLLQ256constload,
ssa.OpAMD64VPSLLQ512constload, ssa.OpAMD64VPSLLQ512constload,
ssa.OpAMD64VPSRLD128constload,
ssa.OpAMD64VPSRLD256constload,
ssa.OpAMD64VPSRLD512constload, ssa.OpAMD64VPSRLD512constload,
ssa.OpAMD64VPSRLQ128constload,
ssa.OpAMD64VPSRLQ256constload,
ssa.OpAMD64VPSRLQ512constload, ssa.OpAMD64VPSRLQ512constload,
ssa.OpAMD64VPSRAD128constload,
ssa.OpAMD64VPSRAD256constload,
ssa.OpAMD64VPSRAD512constload, ssa.OpAMD64VPSRAD512constload,
ssa.OpAMD64VPSRAQ128constload, ssa.OpAMD64VPSRAQ128constload,
ssa.OpAMD64VPSRAQ256constload, ssa.OpAMD64VPSRAQ256constload,
@ -1967,11 +1861,7 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
ssa.OpAMD64VPSRAQMasked512constload: ssa.OpAMD64VPSRAQMasked512constload:
p = simdVkvloadImm8(s, v) p = simdVkvloadImm8(s, v)
case ssa.OpAMD64VCMPPS128load, case ssa.OpAMD64VGF2P8AFFINEQB128load,
ssa.OpAMD64VCMPPS256load,
ssa.OpAMD64VCMPPD128load,
ssa.OpAMD64VCMPPD256load,
ssa.OpAMD64VGF2P8AFFINEQB128load,
ssa.OpAMD64VGF2P8AFFINEQB256load, ssa.OpAMD64VGF2P8AFFINEQB256load,
ssa.OpAMD64VGF2P8AFFINEQB512load, ssa.OpAMD64VGF2P8AFFINEQB512load,
ssa.OpAMD64VGF2P8AFFINEINVQB128load, ssa.OpAMD64VGF2P8AFFINEINVQB128load,

94
src/cmd/compile/internal/ssa/_gen/simdAMD64.rules
View file

@ -1469,8 +1469,6 @@
(VMOVDQU16Masked512 (VPSRAW512const [a] x) mask) => (VPSRAWMasked512const [a] x mask) (VMOVDQU16Masked512 (VPSRAW512const [a] x) mask) => (VPSRAWMasked512const [a] x mask)
(VMOVDQU32Masked512 (VPSRAD512const [a] x) mask) => (VPSRADMasked512const [a] x mask) (VMOVDQU32Masked512 (VPSRAD512const [a] x) mask) => (VPSRADMasked512const [a] x mask)
(VMOVDQU64Masked512 (VPSRAQ512const [a] x) mask) => (VPSRAQMasked512const [a] x mask) (VMOVDQU64Masked512 (VPSRAQ512const [a] x) mask) => (VPSRAQMasked512const [a] x mask)
(VPABSD128 l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSD128load {sym} [off] ptr mem)
(VPABSD256 l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSD256load {sym} [off] ptr mem)
(VPABSD512 l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSD512load {sym} [off] ptr mem) (VPABSD512 l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSD512load {sym} [off] ptr mem)
(VPABSQ128 l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSQ128load {sym} [off] ptr mem) (VPABSQ128 l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSQ128load {sym} [off] ptr mem)
(VPABSQ256 l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSQ256load {sym} [off] ptr mem) (VPABSQ256 l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSQ256load {sym} [off] ptr mem)
@ -1481,38 +1479,22 @@
(VPABSQMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPABSQMasked128load {sym} [off] ptr mask mem) (VPABSQMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPABSQMasked128load {sym} [off] ptr mask mem)
(VPABSQMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPABSQMasked256load {sym} [off] ptr mask mem) (VPABSQMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPABSQMasked256load {sym} [off] ptr mask mem)
(VPABSQMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPABSQMasked512load {sym} [off] ptr mask mem) (VPABSQMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPABSQMasked512load {sym} [off] ptr mask mem)
(VADDPS128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VADDPS128load {sym} [off] x ptr mem)
(VADDPS256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VADDPS256load {sym} [off] x ptr mem)
(VADDPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VADDPS512load {sym} [off] x ptr mem) (VADDPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VADDPS512load {sym} [off] x ptr mem)
(VADDPD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VADDPD128load {sym} [off] x ptr mem)
(VADDPD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VADDPD256load {sym} [off] x ptr mem)
(VADDPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VADDPD512load {sym} [off] x ptr mem) (VADDPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VADDPD512load {sym} [off] x ptr mem)
(VPADDD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPADDD128load {sym} [off] x ptr mem)
(VPADDD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPADDD256load {sym} [off] x ptr mem)
(VPADDD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPADDD512load {sym} [off] x ptr mem) (VPADDD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPADDD512load {sym} [off] x ptr mem)
(VPADDQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPADDQ128load {sym} [off] x ptr mem)
(VPADDQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPADDQ256load {sym} [off] x ptr mem)
(VPADDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPADDQ512load {sym} [off] x ptr mem) (VPADDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPADDQ512load {sym} [off] x ptr mem)
(VPDPWSSD128 x y l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSD128load {sym} [off] x y ptr mem)
(VPDPWSSD256 x y l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSD256load {sym} [off] x y ptr mem)
(VPDPWSSD512 x y l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSD512load {sym} [off] x y ptr mem) (VPDPWSSD512 x y l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSD512load {sym} [off] x y ptr mem)
(VPDPWSSDMasked128 x y l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDMasked128load {sym} [off] x y ptr mask mem) (VPDPWSSDMasked128 x y l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDMasked128load {sym} [off] x y ptr mask mem)
(VPDPWSSDMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDMasked256load {sym} [off] x y ptr mask mem) (VPDPWSSDMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDMasked256load {sym} [off] x y ptr mask mem)
(VPDPWSSDMasked512 x y l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDMasked512load {sym} [off] x y ptr mask mem) (VPDPWSSDMasked512 x y l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDMasked512load {sym} [off] x y ptr mask mem)
(VPDPWSSDS128 x y l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDS128load {sym} [off] x y ptr mem)
(VPDPWSSDS256 x y l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDS256load {sym} [off] x y ptr mem)
(VPDPWSSDS512 x y l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDS512load {sym} [off] x y ptr mem) (VPDPWSSDS512 x y l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDS512load {sym} [off] x y ptr mem)
(VPDPWSSDSMasked128 x y l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDSMasked128load {sym} [off] x y ptr mask mem) (VPDPWSSDSMasked128 x y l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDSMasked128load {sym} [off] x y ptr mask mem)
(VPDPWSSDSMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDSMasked256load {sym} [off] x y ptr mask mem) (VPDPWSSDSMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDSMasked256load {sym} [off] x y ptr mask mem)
(VPDPWSSDSMasked512 x y l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDSMasked512load {sym} [off] x y ptr mask mem) (VPDPWSSDSMasked512 x y l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPWSSDSMasked512load {sym} [off] x y ptr mask mem)
(VPDPBUSD128 x y l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSD128load {sym} [off] x y ptr mem)
(VPDPBUSD256 x y l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSD256load {sym} [off] x y ptr mem)
(VPDPBUSD512 x y l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSD512load {sym} [off] x y ptr mem) (VPDPBUSD512 x y l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSD512load {sym} [off] x y ptr mem)
(VPDPBUSDMasked128 x y l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDMasked128load {sym} [off] x y ptr mask mem) (VPDPBUSDMasked128 x y l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDMasked128load {sym} [off] x y ptr mask mem)
(VPDPBUSDMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDMasked256load {sym} [off] x y ptr mask mem) (VPDPBUSDMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDMasked256load {sym} [off] x y ptr mask mem)
(VPDPBUSDMasked512 x y l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDMasked512load {sym} [off] x y ptr mask mem) (VPDPBUSDMasked512 x y l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDMasked512load {sym} [off] x y ptr mask mem)
(VPDPBUSDS128 x y l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDS128load {sym} [off] x y ptr mem)
(VPDPBUSDS256 x y l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDS256load {sym} [off] x y ptr mem)
(VPDPBUSDS512 x y l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDS512load {sym} [off] x y ptr mem) (VPDPBUSDS512 x y l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDS512load {sym} [off] x y ptr mem)
(VPDPBUSDSMasked128 x y l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDSMasked128load {sym} [off] x y ptr mask mem) (VPDPBUSDSMasked128 x y l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDSMasked128load {sym} [off] x y ptr mask mem)
(VPDPBUSDSMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDSMasked256load {sym} [off] x y ptr mask mem) (VPDPBUSDSMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPDPBUSDSMasked256load {sym} [off] x y ptr mask mem)
@ -1545,20 +1527,14 @@
(VPANDNQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPANDNQMasked128load {sym} [off] x ptr mask mem) (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) (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) (VPANDNQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPANDNQMasked512load {sym} [off] x ptr mask mem)
(VPACKSSDW128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPACKSSDW128load {sym} [off] x ptr mem)
(VPACKSSDW256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPACKSSDW256load {sym} [off] x ptr mem)
(VPACKSSDW512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPACKSSDW512load {sym} [off] x ptr 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) (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) (VPACKSSDWMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPACKSSDWMasked256load {sym} [off] x ptr mask mem)
(VPACKSSDWMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPACKSSDWMasked512load {sym} [off] x ptr mask mem) (VPACKSSDWMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPACKSSDWMasked512load {sym} [off] x ptr mask mem)
(VCVTTPS2DQ128 l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VCVTTPS2DQ128load {sym} [off] ptr mem)
(VCVTTPS2DQ256 l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VCVTTPS2DQ256load {sym} [off] ptr mem)
(VCVTTPS2DQ512 l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VCVTTPS2DQ512load {sym} [off] ptr mem) (VCVTTPS2DQ512 l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VCVTTPS2DQ512load {sym} [off] ptr mem)
(VCVTTPS2DQMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTTPS2DQMasked128load {sym} [off] ptr mask mem) (VCVTTPS2DQMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTTPS2DQMasked128load {sym} [off] ptr mask mem)
(VCVTTPS2DQMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTTPS2DQMasked256load {sym} [off] ptr mask mem) (VCVTTPS2DQMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTTPS2DQMasked256load {sym} [off] ptr mask mem)
(VCVTTPS2DQMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTTPS2DQMasked512load {sym} [off] ptr mask mem) (VCVTTPS2DQMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTTPS2DQMasked512load {sym} [off] ptr mask mem)
(VPACKUSDW128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPACKUSDW128load {sym} [off] x ptr mem)
(VPACKUSDW256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPACKUSDW256load {sym} [off] x ptr mem)
(VPACKUSDW512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPACKUSDW512load {sym} [off] x ptr mem) (VPACKUSDW512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPACKUSDW512load {sym} [off] x ptr mem)
(VPACKUSDWMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPACKUSDWMasked128load {sym} [off] x ptr mask mem) (VPACKUSDWMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPACKUSDWMasked128load {sym} [off] x ptr mask mem)
(VPACKUSDWMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPACKUSDWMasked256load {sym} [off] x ptr mask mem) (VPACKUSDWMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPACKUSDWMasked256load {sym} [off] x ptr mask mem)
@ -1569,11 +1545,7 @@
(VCVTPS2UDQMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTPS2UDQMasked128load {sym} [off] ptr mask mem) (VCVTPS2UDQMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTPS2UDQMasked128load {sym} [off] ptr mask mem)
(VCVTPS2UDQMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTPS2UDQMasked256load {sym} [off] ptr mask mem) (VCVTPS2UDQMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTPS2UDQMasked256load {sym} [off] ptr mask mem)
(VCVTPS2UDQMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTPS2UDQMasked512load {sym} [off] ptr mask mem) (VCVTPS2UDQMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VCVTPS2UDQMasked512load {sym} [off] ptr mask mem)
(VDIVPS128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VDIVPS128load {sym} [off] x ptr mem)
(VDIVPS256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VDIVPS256load {sym} [off] x ptr mem)
(VDIVPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VDIVPS512load {sym} [off] x ptr mem) (VDIVPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VDIVPS512load {sym} [off] x ptr mem)
(VDIVPD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VDIVPD128load {sym} [off] x ptr mem)
(VDIVPD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VDIVPD256load {sym} [off] x ptr mem)
(VDIVPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VDIVPD512load {sym} [off] x ptr mem) (VDIVPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VDIVPD512load {sym} [off] x ptr mem)
(VDIVPSMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPSMasked128load {sym} [off] x ptr mask mem) (VDIVPSMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPSMasked128load {sym} [off] x ptr mask mem)
(VDIVPSMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPSMasked256load {sym} [off] x ptr mask mem) (VDIVPSMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPSMasked256load {sym} [off] x ptr mask mem)
@ -1581,29 +1553,13 @@
(VDIVPDMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPDMasked128load {sym} [off] x ptr mask mem) (VDIVPDMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPDMasked128load {sym} [off] x ptr mask mem)
(VDIVPDMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPDMasked256load {sym} [off] x ptr mask mem) (VDIVPDMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPDMasked256load {sym} [off] x ptr mask mem)
(VDIVPDMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPDMasked512load {sym} [off] x ptr mask mem) (VDIVPDMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VDIVPDMasked512load {sym} [off] x ptr mask mem)
(VPCMPEQD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPEQD128load {sym} [off] x ptr mem)
(VPCMPEQD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPEQD256load {sym} [off] x ptr mem)
(VPCMPEQD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPEQD512load {sym} [off] x ptr mem) (VPCMPEQD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPEQD512load {sym} [off] x ptr mem)
(VPCMPEQQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPEQQ128load {sym} [off] x ptr mem)
(VPCMPEQQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPEQQ256load {sym} [off] x ptr mem)
(VPCMPEQQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPEQQ512load {sym} [off] x ptr mem) (VPCMPEQQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPEQQ512load {sym} [off] x ptr mem)
(VPCMPGTD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPGTD128load {sym} [off] x ptr mem)
(VPCMPGTD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPGTD256load {sym} [off] x ptr mem)
(VPCMPGTD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPGTD512load {sym} [off] x ptr mem) (VPCMPGTD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPGTD512load {sym} [off] x ptr mem)
(VPCMPGTQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPGTQ128load {sym} [off] x ptr mem)
(VPCMPGTQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPGTQ256load {sym} [off] x ptr mem)
(VPCMPGTQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPGTQ512load {sym} [off] x ptr mem) (VPCMPGTQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPCMPGTQ512load {sym} [off] x ptr mem)
(VPUNPCKHDQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKHDQ128load {sym} [off] x ptr mem)
(VPUNPCKHQDQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKHQDQ128load {sym} [off] x ptr mem)
(VPUNPCKHDQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKHDQ256load {sym} [off] x ptr mem)
(VPUNPCKHDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKHDQ512load {sym} [off] x ptr mem) (VPUNPCKHDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKHDQ512load {sym} [off] x ptr mem)
(VPUNPCKHQDQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKHQDQ256load {sym} [off] x ptr mem)
(VPUNPCKHQDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKHQDQ512load {sym} [off] x ptr mem) (VPUNPCKHQDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKHQDQ512load {sym} [off] x ptr mem)
(VPUNPCKLDQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKLDQ128load {sym} [off] x ptr mem)
(VPUNPCKLQDQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKLQDQ128load {sym} [off] x ptr mem)
(VPUNPCKLDQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKLDQ256load {sym} [off] x ptr mem)
(VPUNPCKLDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKLDQ512load {sym} [off] x ptr mem) (VPUNPCKLDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKLDQ512load {sym} [off] x ptr mem)
(VPUNPCKLQDQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKLQDQ256load {sym} [off] x ptr mem)
(VPUNPCKLQDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKLQDQ512load {sym} [off] x ptr mem) (VPUNPCKLQDQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPUNPCKLQDQ512load {sym} [off] x ptr mem)
(VPLZCNTD128 l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPLZCNTD128load {sym} [off] ptr mem) (VPLZCNTD128 l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPLZCNTD128load {sym} [off] ptr mem)
(VPLZCNTD256 l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPLZCNTD256load {sym} [off] ptr mem) (VPLZCNTD256 l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPLZCNTD256load {sym} [off] ptr mem)
@ -1617,20 +1573,12 @@
(VPLZCNTQMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPLZCNTQMasked128load {sym} [off] ptr mask mem) (VPLZCNTQMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPLZCNTQMasked128load {sym} [off] ptr mask mem)
(VPLZCNTQMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPLZCNTQMasked256load {sym} [off] ptr mask mem) (VPLZCNTQMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPLZCNTQMasked256load {sym} [off] ptr mask mem)
(VPLZCNTQMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPLZCNTQMasked512load {sym} [off] ptr mask mem) (VPLZCNTQMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPLZCNTQMasked512load {sym} [off] ptr mask mem)
(VMAXPS128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMAXPS128load {sym} [off] x ptr mem)
(VMAXPS256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMAXPS256load {sym} [off] x ptr mem)
(VMAXPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMAXPS512load {sym} [off] x ptr mem) (VMAXPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMAXPS512load {sym} [off] x ptr mem)
(VMAXPD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMAXPD128load {sym} [off] x ptr mem)
(VMAXPD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMAXPD256load {sym} [off] x ptr mem)
(VMAXPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMAXPD512load {sym} [off] x ptr mem) (VMAXPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMAXPD512load {sym} [off] x ptr mem)
(VPMAXSD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXSD128load {sym} [off] x ptr mem)
(VPMAXSD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXSD256load {sym} [off] x ptr mem)
(VPMAXSD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXSD512load {sym} [off] x ptr mem) (VPMAXSD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXSD512load {sym} [off] x ptr mem)
(VPMAXSQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXSQ128load {sym} [off] x ptr mem) (VPMAXSQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXSQ128load {sym} [off] x ptr mem)
(VPMAXSQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXSQ256load {sym} [off] x ptr mem) (VPMAXSQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXSQ256load {sym} [off] x ptr mem)
(VPMAXSQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXSQ512load {sym} [off] x ptr mem) (VPMAXSQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXSQ512load {sym} [off] x ptr mem)
(VPMAXUD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXUD128load {sym} [off] x ptr mem)
(VPMAXUD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXUD256load {sym} [off] x ptr mem)
(VPMAXUD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXUD512load {sym} [off] x ptr mem) (VPMAXUD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXUD512load {sym} [off] x ptr mem)
(VPMAXUQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXUQ128load {sym} [off] x ptr mem) (VPMAXUQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXUQ128load {sym} [off] x ptr mem)
(VPMAXUQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXUQ256load {sym} [off] x ptr mem) (VPMAXUQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMAXUQ256load {sym} [off] x ptr mem)
@ -1653,20 +1601,12 @@
(VPMAXUQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMAXUQMasked128load {sym} [off] x ptr mask mem) (VPMAXUQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMAXUQMasked128load {sym} [off] x ptr mask mem)
(VPMAXUQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMAXUQMasked256load {sym} [off] x ptr mask mem) (VPMAXUQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMAXUQMasked256load {sym} [off] x ptr mask mem)
(VPMAXUQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMAXUQMasked512load {sym} [off] x ptr mask mem) (VPMAXUQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMAXUQMasked512load {sym} [off] x ptr mask mem)
(VMINPS128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMINPS128load {sym} [off] x ptr mem)
(VMINPS256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMINPS256load {sym} [off] x ptr mem)
(VMINPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMINPS512load {sym} [off] x ptr mem) (VMINPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMINPS512load {sym} [off] x ptr mem)
(VMINPD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMINPD128load {sym} [off] x ptr mem)
(VMINPD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMINPD256load {sym} [off] x ptr mem)
(VMINPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMINPD512load {sym} [off] x ptr mem) (VMINPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMINPD512load {sym} [off] x ptr mem)
(VPMINSD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINSD128load {sym} [off] x ptr mem)
(VPMINSD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINSD256load {sym} [off] x ptr mem)
(VPMINSD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINSD512load {sym} [off] x ptr mem) (VPMINSD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINSD512load {sym} [off] x ptr mem)
(VPMINSQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINSQ128load {sym} [off] x ptr mem) (VPMINSQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINSQ128load {sym} [off] x ptr mem)
(VPMINSQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINSQ256load {sym} [off] x ptr mem) (VPMINSQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINSQ256load {sym} [off] x ptr mem)
(VPMINSQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINSQ512load {sym} [off] x ptr mem) (VPMINSQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINSQ512load {sym} [off] x ptr mem)
(VPMINUD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINUD128load {sym} [off] x ptr mem)
(VPMINUD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINUD256load {sym} [off] x ptr mem)
(VPMINUD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINUD512load {sym} [off] x ptr mem) (VPMINUD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINUD512load {sym} [off] x ptr mem)
(VPMINUQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINUQ128load {sym} [off] x ptr mem) (VPMINUQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINUQ128load {sym} [off] x ptr mem)
(VPMINUQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINUQ256load {sym} [off] x ptr mem) (VPMINUQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMINUQ256load {sym} [off] x ptr mem)
@ -1689,14 +1629,8 @@
(VPMINUQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMINUQMasked128load {sym} [off] x ptr mask mem) (VPMINUQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMINUQMasked128load {sym} [off] x ptr mask mem)
(VPMINUQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMINUQMasked256load {sym} [off] x ptr mask mem) (VPMINUQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMINUQMasked256load {sym} [off] x ptr mask mem)
(VPMINUQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMINUQMasked512load {sym} [off] x ptr mask mem) (VPMINUQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPMINUQMasked512load {sym} [off] x ptr mask mem)
(VMULPS128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMULPS128load {sym} [off] x ptr mem)
(VMULPS256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMULPS256load {sym} [off] x ptr mem)
(VMULPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMULPS512load {sym} [off] x ptr mem) (VMULPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMULPS512load {sym} [off] x ptr mem)
(VMULPD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMULPD128load {sym} [off] x ptr mem)
(VMULPD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMULPD256load {sym} [off] x ptr mem)
(VMULPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMULPD512load {sym} [off] x ptr mem) (VMULPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VMULPD512load {sym} [off] x ptr mem)
(VPMULLD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULLD128load {sym} [off] x ptr mem)
(VPMULLD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULLD256load {sym} [off] x ptr mem)
(VPMULLD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULLD512load {sym} [off] x ptr mem) (VPMULLD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULLD512load {sym} [off] x ptr mem)
(VPMULLQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULLQ128load {sym} [off] x ptr mem) (VPMULLQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULLQ128load {sym} [off] x ptr mem)
(VPMULLQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULLQ256load {sym} [off] x ptr mem) (VPMULLQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULLQ256load {sym} [off] x ptr mem)
@ -1725,10 +1659,6 @@
(VFMADDSUB213PDMasked128 x y l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VFMADDSUB213PDMasked128load {sym} [off] x y ptr mask mem) (VFMADDSUB213PDMasked128 x y l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VFMADDSUB213PDMasked128load {sym} [off] x y ptr mask mem)
(VFMADDSUB213PDMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VFMADDSUB213PDMasked256load {sym} [off] x y ptr mask mem) (VFMADDSUB213PDMasked256 x y l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VFMADDSUB213PDMasked256load {sym} [off] x y ptr mask mem)
(VFMADDSUB213PDMasked512 x y l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VFMADDSUB213PDMasked512load {sym} [off] x y ptr mask mem) (VFMADDSUB213PDMasked512 x y l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VFMADDSUB213PDMasked512load {sym} [off] x y ptr mask mem)
(VPMULDQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULDQ128load {sym} [off] x ptr mem)
(VPMULDQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULDQ256load {sym} [off] x ptr mem)
(VPMULUDQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULUDQ128load {sym} [off] x ptr mem)
(VPMULUDQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPMULUDQ256load {sym} [off] x ptr mem)
(VMULPSMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VMULPSMasked128load {sym} [off] x ptr mask mem) (VMULPSMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VMULPSMasked128load {sym} [off] x ptr mask mem)
(VMULPSMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VMULPSMasked256load {sym} [off] x ptr mask mem) (VMULPSMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VMULPSMasked256load {sym} [off] x ptr mask mem)
(VMULPSMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VMULPSMasked512load {sym} [off] x ptr mask mem) (VMULPSMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VMULPSMasked512load {sym} [off] x ptr mask mem)
@ -1773,8 +1703,6 @@
(VPORQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPORQMasked128load {sym} [off] x ptr mask mem) (VPORQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPORQMasked128load {sym} [off] x ptr mask mem)
(VPORQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPORQMasked256load {sym} [off] x ptr mask mem) (VPORQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPORQMasked256load {sym} [off] x ptr mask mem)
(VPORQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPORQMasked512load {sym} [off] x ptr mask mem) (VPORQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPORQMasked512load {sym} [off] x ptr mask mem)
(VPERMPS256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPERMPS256load {sym} [off] x ptr mem)
(VPERMD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPERMD256load {sym} [off] x ptr mem)
(VPERMPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPERMPS512load {sym} [off] x ptr mem) (VPERMPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPERMPS512load {sym} [off] x ptr mem)
(VPERMD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPERMD512load {sym} [off] x ptr mem) (VPERMD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPERMD512load {sym} [off] x ptr mem)
(VPERMPD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPERMPD256load {sym} [off] x ptr mem) (VPERMPD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPERMPD256load {sym} [off] x ptr mem)
@ -1869,11 +1797,7 @@
(VSCALEFPDMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSCALEFPDMasked128load {sym} [off] x ptr mask mem) (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) (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) (VSCALEFPDMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSCALEFPDMasked512load {sym} [off] x ptr mask mem)
(VPSLLVD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLVD128load {sym} [off] x ptr mem)
(VPSLLVD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLVD256load {sym} [off] x ptr mem)
(VPSLLVD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLVD512load {sym} [off] x ptr mem) (VPSLLVD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLVD512load {sym} [off] x ptr mem)
(VPSLLVQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLVQ128load {sym} [off] x ptr mem)
(VPSLLVQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLVQ256load {sym} [off] x ptr mem)
(VPSLLVQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSLLVQ512load {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) (VPSHLDVD128 x y l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHLDVD128load {sym} [off] x y ptr mem)
(VPSHLDVD256 x y l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHLDVD256load {sym} [off] x y ptr mem) (VPSHLDVD256 x y l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHLDVD256load {sym} [off] x y ptr mem)
@ -1893,17 +1817,11 @@
(VPSLLVQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLVQMasked128load {sym} [off] x ptr mask mem) (VPSLLVQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLVQMasked128load {sym} [off] x ptr mask mem)
(VPSLLVQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLVQMasked256load {sym} [off] x ptr mask mem) (VPSLLVQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLVQMasked256load {sym} [off] x ptr mask mem)
(VPSLLVQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLVQMasked512load {sym} [off] x ptr mask mem) (VPSLLVQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSLLVQMasked512load {sym} [off] x ptr mask mem)
(VPSRAVD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAVD128load {sym} [off] x ptr mem)
(VPSRAVD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAVD256load {sym} [off] x ptr mem)
(VPSRAVD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAVD512load {sym} [off] x ptr mem) (VPSRAVD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAVD512load {sym} [off] x ptr mem)
(VPSRAVQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAVQ128load {sym} [off] x ptr mem) (VPSRAVQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAVQ128load {sym} [off] x ptr mem)
(VPSRAVQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAVQ256load {sym} [off] x ptr mem) (VPSRAVQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAVQ256load {sym} [off] x ptr mem)
(VPSRAVQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAVQ512load {sym} [off] x ptr mem) (VPSRAVQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRAVQ512load {sym} [off] x ptr mem)
(VPSRLVD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRLVD128load {sym} [off] x ptr mem)
(VPSRLVD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRLVD256load {sym} [off] x ptr mem)
(VPSRLVD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRLVD512load {sym} [off] x ptr mem) (VPSRLVD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRLVD512load {sym} [off] x ptr mem)
(VPSRLVQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRLVQ128load {sym} [off] x ptr mem)
(VPSRLVQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRLVQ256load {sym} [off] x ptr mem)
(VPSRLVQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRLVQ512load {sym} [off] x ptr mem) (VPSRLVQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSRLVQ512load {sym} [off] x ptr mem)
(VPSHRDVD128 x y l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHRDVD128load {sym} [off] x y ptr mem) (VPSHRDVD128 x y l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHRDVD128load {sym} [off] x y ptr mem)
(VPSHRDVD256 x y l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHRDVD256load {sym} [off] x y ptr mem) (VPSHRDVD256 x y l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSHRDVD256load {sym} [off] x y ptr mem)
@ -1929,11 +1847,7 @@
(VPSRLVQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLVQMasked128load {sym} [off] x ptr mask mem) (VPSRLVQMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLVQMasked128load {sym} [off] x ptr mask mem)
(VPSRLVQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLVQMasked256load {sym} [off] x ptr mask mem) (VPSRLVQMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLVQMasked256load {sym} [off] x ptr mask mem)
(VPSRLVQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLVQMasked512load {sym} [off] x ptr mask mem) (VPSRLVQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VPSRLVQMasked512load {sym} [off] x ptr mask mem)
(VSQRTPS128 l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSQRTPS128load {sym} [off] ptr mem)
(VSQRTPS256 l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSQRTPS256load {sym} [off] ptr mem)
(VSQRTPS512 l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSQRTPS512load {sym} [off] ptr mem) (VSQRTPS512 l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSQRTPS512load {sym} [off] ptr mem)
(VSQRTPD128 l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSQRTPD128load {sym} [off] ptr mem)
(VSQRTPD256 l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSQRTPD256load {sym} [off] ptr mem)
(VSQRTPD512 l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSQRTPD512load {sym} [off] ptr mem) (VSQRTPD512 l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSQRTPD512load {sym} [off] ptr mem)
(VSQRTPSMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSQRTPSMasked128load {sym} [off] ptr mask mem) (VSQRTPSMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSQRTPSMasked128load {sym} [off] ptr mask mem)
(VSQRTPSMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSQRTPSMasked256load {sym} [off] ptr mask mem) (VSQRTPSMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSQRTPSMasked256load {sym} [off] ptr mask mem)
@ -1941,17 +1855,9 @@
(VSQRTPDMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSQRTPDMasked128load {sym} [off] ptr mask mem) (VSQRTPDMasked128 l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSQRTPDMasked128load {sym} [off] ptr mask mem)
(VSQRTPDMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSQRTPDMasked256load {sym} [off] ptr mask mem) (VSQRTPDMasked256 l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSQRTPDMasked256load {sym} [off] ptr mask mem)
(VSQRTPDMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSQRTPDMasked512load {sym} [off] ptr mask mem) (VSQRTPDMasked512 l:(VMOVDQUload512 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSQRTPDMasked512load {sym} [off] ptr mask mem)
(VSUBPS128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSUBPS128load {sym} [off] x ptr mem)
(VSUBPS256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSUBPS256load {sym} [off] x ptr mem)
(VSUBPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSUBPS512load {sym} [off] x ptr mem) (VSUBPS512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSUBPS512load {sym} [off] x ptr mem)
(VSUBPD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSUBPD128load {sym} [off] x ptr mem)
(VSUBPD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSUBPD256load {sym} [off] x ptr mem)
(VSUBPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSUBPD512load {sym} [off] x ptr mem) (VSUBPD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VSUBPD512load {sym} [off] x ptr mem)
(VPSUBD128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSUBD128load {sym} [off] x ptr mem)
(VPSUBD256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSUBD256load {sym} [off] x ptr mem)
(VPSUBD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSUBD512load {sym} [off] x ptr mem) (VPSUBD512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSUBD512load {sym} [off] x ptr mem)
(VPSUBQ128 x l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSUBQ128load {sym} [off] x ptr mem)
(VPSUBQ256 x l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSUBQ256load {sym} [off] x ptr mem)
(VPSUBQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSUBQ512load {sym} [off] x ptr mem) (VPSUBQ512 x l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPSUBQ512load {sym} [off] x ptr mem)
(VSUBPSMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSUBPSMasked128load {sym} [off] x ptr mask mem) (VSUBPSMasked128 x l:(VMOVDQUload128 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSUBPSMasked128load {sym} [off] x ptr mask mem)
(VSUBPSMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSUBPSMasked256load {sym} [off] x ptr mask mem) (VSUBPSMasked256 x l:(VMOVDQUload256 {sym} [off] ptr mem) mask) && canMergeLoad(v, l) && clobber(l) => (VSUBPSMasked256load {sym} [off] x ptr mask mem)

110
src/cmd/compile/internal/ssa/_gen/simdAMD64ops.go
View file

@ -1310,8 +1310,6 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPSRAQMasked128const", argLength: 2, reg: wkw, asm: "VPSRAQ", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: false}, {name: "VPSRAQMasked128const", argLength: 2, reg: wkw, asm: "VPSRAQ", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: false},
{name: "VPSRAQMasked256const", argLength: 2, reg: wkw, asm: "VPSRAQ", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: false}, {name: "VPSRAQMasked256const", argLength: 2, reg: wkw, asm: "VPSRAQ", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: false},
{name: "VPSRAQMasked512const", argLength: 2, reg: wkw, asm: "VPSRAQ", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: false}, {name: "VPSRAQMasked512const", argLength: 2, reg: wkw, asm: "VPSRAQ", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: false},
{name: "VPABSD128load", argLength: 2, reg: v11load, asm: "VPABSD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPABSD256load", argLength: 2, reg: v11load, asm: "VPABSD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPABSD512load", argLength: 2, reg: w11load, asm: "VPABSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPABSD512load", argLength: 2, reg: w11load, asm: "VPABSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPABSQ128load", argLength: 2, reg: w11load, asm: "VPABSQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPABSQ128load", argLength: 2, reg: w11load, asm: "VPABSQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPABSQ256load", argLength: 2, reg: w11load, asm: "VPABSQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPABSQ256load", argLength: 2, reg: w11load, asm: "VPABSQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1322,38 +1320,22 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPABSQMasked128load", argLength: 3, reg: wkwload, asm: "VPABSQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPABSQMasked128load", argLength: 3, reg: wkwload, asm: "VPABSQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPABSQMasked256load", argLength: 3, reg: wkwload, asm: "VPABSQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPABSQMasked256load", argLength: 3, reg: wkwload, asm: "VPABSQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPABSQMasked512load", argLength: 3, reg: wkwload, asm: "VPABSQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPABSQMasked512load", argLength: 3, reg: wkwload, asm: "VPABSQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VADDPS128load", argLength: 3, reg: v21load, asm: "VADDPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VADDPS256load", argLength: 3, reg: v21load, asm: "VADDPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VADDPS512load", argLength: 3, reg: w21load, asm: "VADDPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VADDPS512load", argLength: 3, reg: w21load, asm: "VADDPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VADDPD128load", argLength: 3, reg: v21load, asm: "VADDPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VADDPD256load", argLength: 3, reg: v21load, asm: "VADDPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VADDPD512load", argLength: 3, reg: w21load, asm: "VADDPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VADDPD512load", argLength: 3, reg: w21load, asm: "VADDPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPADDD128load", argLength: 3, reg: v21load, asm: "VPADDD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPADDD256load", argLength: 3, reg: v21load, asm: "VPADDD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPADDD512load", argLength: 3, reg: w21load, asm: "VPADDD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPADDD512load", argLength: 3, reg: w21load, asm: "VPADDD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPADDQ128load", argLength: 3, reg: v21load, asm: "VPADDQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPADDQ256load", argLength: 3, reg: v21load, asm: "VPADDQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPADDQ512load", argLength: 3, reg: w21load, asm: "VPADDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPADDQ512load", argLength: 3, reg: w21load, asm: "VPADDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPDPWSSD128load", argLength: 4, reg: v31load, asm: "VPDPWSSD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSD256load", argLength: 4, reg: v31load, asm: "VPDPWSSD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSD512load", argLength: 4, reg: w31load, asm: "VPDPWSSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPWSSD512load", argLength: 4, reg: w31load, asm: "VPDPWSSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSDMasked128load", argLength: 5, reg: w3kwload, asm: "VPDPWSSD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPWSSDMasked128load", argLength: 5, reg: w3kwload, asm: "VPDPWSSD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSDMasked256load", argLength: 5, reg: w3kwload, asm: "VPDPWSSD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPWSSDMasked256load", argLength: 5, reg: w3kwload, asm: "VPDPWSSD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSDMasked512load", argLength: 5, reg: w3kwload, asm: "VPDPWSSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPWSSDMasked512load", argLength: 5, reg: w3kwload, asm: "VPDPWSSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSDS128load", argLength: 4, reg: v31load, asm: "VPDPWSSDS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSDS256load", argLength: 4, reg: v31load, asm: "VPDPWSSDS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSDS512load", argLength: 4, reg: w31load, asm: "VPDPWSSDS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPWSSDS512load", argLength: 4, reg: w31load, asm: "VPDPWSSDS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSDSMasked128load", argLength: 5, reg: w3kwload, asm: "VPDPWSSDS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPWSSDSMasked128load", argLength: 5, reg: w3kwload, asm: "VPDPWSSDS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSDSMasked256load", argLength: 5, reg: w3kwload, asm: "VPDPWSSDS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPWSSDSMasked256load", argLength: 5, reg: w3kwload, asm: "VPDPWSSDS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPWSSDSMasked512load", argLength: 5, reg: w3kwload, asm: "VPDPWSSDS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPWSSDSMasked512load", argLength: 5, reg: w3kwload, asm: "VPDPWSSDS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSD128load", argLength: 4, reg: v31load, asm: "VPDPBUSD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSD256load", argLength: 4, reg: v31load, asm: "VPDPBUSD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSD512load", argLength: 4, reg: w31load, asm: "VPDPBUSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPBUSD512load", argLength: 4, reg: w31load, asm: "VPDPBUSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSDMasked128load", argLength: 5, reg: w3kwload, asm: "VPDPBUSD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPBUSDMasked128load", argLength: 5, reg: w3kwload, asm: "VPDPBUSD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSDMasked256load", argLength: 5, reg: w3kwload, asm: "VPDPBUSD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPBUSDMasked256load", argLength: 5, reg: w3kwload, asm: "VPDPBUSD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSDMasked512load", argLength: 5, reg: w3kwload, asm: "VPDPBUSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPBUSDMasked512load", argLength: 5, reg: w3kwload, asm: "VPDPBUSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSDS128load", argLength: 4, reg: v31load, asm: "VPDPBUSDS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSDS256load", argLength: 4, reg: v31load, asm: "VPDPBUSDS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSDS512load", argLength: 4, reg: w31load, asm: "VPDPBUSDS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPBUSDS512load", argLength: 4, reg: w31load, asm: "VPDPBUSDS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSDSMasked128load", argLength: 5, reg: w3kwload, asm: "VPDPBUSDS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPBUSDSMasked128load", argLength: 5, reg: w3kwload, asm: "VPDPBUSDS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPDPBUSDSMasked256load", argLength: 5, reg: w3kwload, asm: "VPDPBUSDS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPDPBUSDSMasked256load", argLength: 5, reg: w3kwload, asm: "VPDPBUSDS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
@ -1386,20 +1368,14 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPANDNQMasked128load", argLength: 4, reg: w2kwload, asm: "VPANDNQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPANDNQMasked128load", argLength: 4, reg: w2kwload, asm: "VPANDNQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPANDNQMasked256load", argLength: 4, reg: w2kwload, asm: "VPANDNQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPANDNQMasked256load", argLength: 4, reg: w2kwload, asm: "VPANDNQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPANDNQMasked512load", argLength: 4, reg: w2kwload, asm: "VPANDNQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPANDNQMasked512load", argLength: 4, reg: w2kwload, asm: "VPANDNQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKSSDW128load", argLength: 3, reg: v21load, asm: "VPACKSSDW", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKSSDW256load", argLength: 3, reg: v21load, asm: "VPACKSSDW", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKSSDW512load", argLength: 3, reg: w21load, asm: "VPACKSSDW", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPACKSSDW512load", argLength: 3, reg: w21load, asm: "VPACKSSDW", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKSSDWMasked128load", argLength: 4, reg: w2kwload, asm: "VPACKSSDW", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPACKSSDWMasked128load", argLength: 4, reg: w2kwload, asm: "VPACKSSDW", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKSSDWMasked256load", argLength: 4, reg: w2kwload, asm: "VPACKSSDW", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPACKSSDWMasked256load", argLength: 4, reg: w2kwload, asm: "VPACKSSDW", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKSSDWMasked512load", argLength: 4, reg: w2kwload, asm: "VPACKSSDW", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPACKSSDWMasked512load", argLength: 4, reg: w2kwload, asm: "VPACKSSDW", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VCVTTPS2DQ128load", argLength: 2, reg: v11load, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VCVTTPS2DQ256load", argLength: 2, reg: v11load, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VCVTTPS2DQ512load", argLength: 2, reg: w11load, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VCVTTPS2DQ512load", argLength: 2, reg: w11load, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VCVTTPS2DQMasked128load", argLength: 3, reg: wkwload, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VCVTTPS2DQMasked128load", argLength: 3, reg: wkwload, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VCVTTPS2DQMasked256load", argLength: 3, reg: wkwload, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VCVTTPS2DQMasked256load", argLength: 3, reg: wkwload, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VCVTTPS2DQMasked512load", argLength: 3, reg: wkwload, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VCVTTPS2DQMasked512load", argLength: 3, reg: wkwload, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKUSDW128load", argLength: 3, reg: v21load, asm: "VPACKUSDW", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKUSDW256load", argLength: 3, reg: v21load, asm: "VPACKUSDW", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKUSDW512load", argLength: 3, reg: w21load, asm: "VPACKUSDW", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPACKUSDW512load", argLength: 3, reg: w21load, asm: "VPACKUSDW", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKUSDWMasked128load", argLength: 4, reg: w2kwload, asm: "VPACKUSDW", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPACKUSDWMasked128load", argLength: 4, reg: w2kwload, asm: "VPACKUSDW", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPACKUSDWMasked256load", argLength: 4, reg: w2kwload, asm: "VPACKUSDW", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPACKUSDWMasked256load", argLength: 4, reg: w2kwload, asm: "VPACKUSDW", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1410,11 +1386,7 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VCVTPS2UDQMasked128load", argLength: 3, reg: wkwload, asm: "VCVTPS2UDQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VCVTPS2UDQMasked128load", argLength: 3, reg: wkwload, asm: "VCVTPS2UDQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VCVTPS2UDQMasked256load", argLength: 3, reg: wkwload, asm: "VCVTPS2UDQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VCVTPS2UDQMasked256load", argLength: 3, reg: wkwload, asm: "VCVTPS2UDQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VCVTPS2UDQMasked512load", argLength: 3, reg: wkwload, asm: "VCVTPS2UDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VCVTPS2UDQMasked512load", argLength: 3, reg: wkwload, asm: "VCVTPS2UDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VDIVPS128load", argLength: 3, reg: v21load, asm: "VDIVPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VDIVPS256load", argLength: 3, reg: v21load, asm: "VDIVPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VDIVPS512load", argLength: 3, reg: w21load, asm: "VDIVPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VDIVPS512load", argLength: 3, reg: w21load, asm: "VDIVPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VDIVPD128load", argLength: 3, reg: v21load, asm: "VDIVPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VDIVPD256load", argLength: 3, reg: v21load, asm: "VDIVPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VDIVPD512load", argLength: 3, reg: w21load, asm: "VDIVPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VDIVPD512load", argLength: 3, reg: w21load, asm: "VDIVPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VDIVPSMasked128load", argLength: 4, reg: w2kwload, asm: "VDIVPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VDIVPSMasked128load", argLength: 4, reg: w2kwload, asm: "VDIVPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VDIVPSMasked256load", argLength: 4, reg: w2kwload, asm: "VDIVPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VDIVPSMasked256load", argLength: 4, reg: w2kwload, asm: "VDIVPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1422,29 +1394,13 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VDIVPDMasked128load", argLength: 4, reg: w2kwload, asm: "VDIVPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VDIVPDMasked128load", argLength: 4, reg: w2kwload, asm: "VDIVPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VDIVPDMasked256load", argLength: 4, reg: w2kwload, asm: "VDIVPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VDIVPDMasked256load", argLength: 4, reg: w2kwload, asm: "VDIVPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VDIVPDMasked512load", argLength: 4, reg: w2kwload, asm: "VDIVPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VDIVPDMasked512load", argLength: 4, reg: w2kwload, asm: "VDIVPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPEQD128load", argLength: 3, reg: v21load, asm: "VPCMPEQD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPEQD256load", argLength: 3, reg: v21load, asm: "VPCMPEQD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPEQD512load", argLength: 3, reg: w2kload, asm: "VPCMPEQD", commutative: false, typ: "Mask", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPCMPEQD512load", argLength: 3, reg: w2kload, asm: "VPCMPEQD", commutative: false, typ: "Mask", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPEQQ128load", argLength: 3, reg: v21load, asm: "VPCMPEQQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPEQQ256load", argLength: 3, reg: v21load, asm: "VPCMPEQQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPEQQ512load", argLength: 3, reg: w2kload, asm: "VPCMPEQQ", commutative: false, typ: "Mask", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPCMPEQQ512load", argLength: 3, reg: w2kload, asm: "VPCMPEQQ", commutative: false, typ: "Mask", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPGTD128load", argLength: 3, reg: v21load, asm: "VPCMPGTD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPGTD256load", argLength: 3, reg: v21load, asm: "VPCMPGTD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPGTD512load", argLength: 3, reg: w2kload, asm: "VPCMPGTD", commutative: false, typ: "Mask", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPCMPGTD512load", argLength: 3, reg: w2kload, asm: "VPCMPGTD", commutative: false, typ: "Mask", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPGTQ128load", argLength: 3, reg: v21load, asm: "VPCMPGTQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPGTQ256load", argLength: 3, reg: v21load, asm: "VPCMPGTQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPGTQ512load", argLength: 3, reg: w2kload, asm: "VPCMPGTQ", commutative: false, typ: "Mask", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPCMPGTQ512load", argLength: 3, reg: w2kload, asm: "VPCMPGTQ", commutative: false, typ: "Mask", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKHDQ128load", argLength: 3, reg: v21load, asm: "VPUNPCKHDQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKHQDQ128load", argLength: 3, reg: v21load, asm: "VPUNPCKHQDQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKHDQ256load", argLength: 3, reg: v21load, asm: "VPUNPCKHDQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKHDQ512load", argLength: 3, reg: w21load, asm: "VPUNPCKHDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPUNPCKHDQ512load", argLength: 3, reg: w21load, asm: "VPUNPCKHDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKHQDQ256load", argLength: 3, reg: v21load, asm: "VPUNPCKHQDQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKHQDQ512load", argLength: 3, reg: w21load, asm: "VPUNPCKHQDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPUNPCKHQDQ512load", argLength: 3, reg: w21load, asm: "VPUNPCKHQDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKLDQ128load", argLength: 3, reg: v21load, asm: "VPUNPCKLDQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKLQDQ128load", argLength: 3, reg: v21load, asm: "VPUNPCKLQDQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKLDQ256load", argLength: 3, reg: v21load, asm: "VPUNPCKLDQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKLDQ512load", argLength: 3, reg: w21load, asm: "VPUNPCKLDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPUNPCKLDQ512load", argLength: 3, reg: w21load, asm: "VPUNPCKLDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKLQDQ256load", argLength: 3, reg: v21load, asm: "VPUNPCKLQDQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPUNPCKLQDQ512load", argLength: 3, reg: w21load, asm: "VPUNPCKLQDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPUNPCKLQDQ512load", argLength: 3, reg: w21load, asm: "VPUNPCKLQDQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPLZCNTD128load", argLength: 2, reg: w11load, asm: "VPLZCNTD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPLZCNTD128load", argLength: 2, reg: w11load, asm: "VPLZCNTD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPLZCNTD256load", argLength: 2, reg: w11load, asm: "VPLZCNTD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPLZCNTD256load", argLength: 2, reg: w11load, asm: "VPLZCNTD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1458,20 +1414,12 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPLZCNTQMasked128load", argLength: 3, reg: wkwload, asm: "VPLZCNTQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPLZCNTQMasked128load", argLength: 3, reg: wkwload, asm: "VPLZCNTQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPLZCNTQMasked256load", argLength: 3, reg: wkwload, asm: "VPLZCNTQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPLZCNTQMasked256load", argLength: 3, reg: wkwload, asm: "VPLZCNTQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPLZCNTQMasked512load", argLength: 3, reg: wkwload, asm: "VPLZCNTQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPLZCNTQMasked512load", argLength: 3, reg: wkwload, asm: "VPLZCNTQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMAXPS128load", argLength: 3, reg: v21load, asm: "VMAXPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMAXPS256load", argLength: 3, reg: v21load, asm: "VMAXPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMAXPS512load", argLength: 3, reg: w21load, asm: "VMAXPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VMAXPS512load", argLength: 3, reg: w21load, asm: "VMAXPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMAXPD128load", argLength: 3, reg: v21load, asm: "VMAXPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMAXPD256load", argLength: 3, reg: v21load, asm: "VMAXPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMAXPD512load", argLength: 3, reg: w21load, asm: "VMAXPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VMAXPD512load", argLength: 3, reg: w21load, asm: "VMAXPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXSD128load", argLength: 3, reg: v21load, asm: "VPMAXSD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXSD256load", argLength: 3, reg: v21load, asm: "VPMAXSD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXSD512load", argLength: 3, reg: w21load, asm: "VPMAXSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMAXSD512load", argLength: 3, reg: w21load, asm: "VPMAXSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXSQ128load", argLength: 3, reg: w21load, asm: "VPMAXSQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMAXSQ128load", argLength: 3, reg: w21load, asm: "VPMAXSQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXSQ256load", argLength: 3, reg: w21load, asm: "VPMAXSQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMAXSQ256load", argLength: 3, reg: w21load, asm: "VPMAXSQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXSQ512load", argLength: 3, reg: w21load, asm: "VPMAXSQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMAXSQ512load", argLength: 3, reg: w21load, asm: "VPMAXSQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXUD128load", argLength: 3, reg: v21load, asm: "VPMAXUD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXUD256load", argLength: 3, reg: v21load, asm: "VPMAXUD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXUD512load", argLength: 3, reg: w21load, asm: "VPMAXUD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMAXUD512load", argLength: 3, reg: w21load, asm: "VPMAXUD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXUQ128load", argLength: 3, reg: w21load, asm: "VPMAXUQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMAXUQ128load", argLength: 3, reg: w21load, asm: "VPMAXUQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXUQ256load", argLength: 3, reg: w21load, asm: "VPMAXUQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMAXUQ256load", argLength: 3, reg: w21load, asm: "VPMAXUQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1494,20 +1442,12 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPMAXUQMasked128load", argLength: 4, reg: w2kwload, asm: "VPMAXUQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMAXUQMasked128load", argLength: 4, reg: w2kwload, asm: "VPMAXUQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXUQMasked256load", argLength: 4, reg: w2kwload, asm: "VPMAXUQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMAXUQMasked256load", argLength: 4, reg: w2kwload, asm: "VPMAXUQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMAXUQMasked512load", argLength: 4, reg: w2kwload, asm: "VPMAXUQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMAXUQMasked512load", argLength: 4, reg: w2kwload, asm: "VPMAXUQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMINPS128load", argLength: 3, reg: v21load, asm: "VMINPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMINPS256load", argLength: 3, reg: v21load, asm: "VMINPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMINPS512load", argLength: 3, reg: w21load, asm: "VMINPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VMINPS512load", argLength: 3, reg: w21load, asm: "VMINPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMINPD128load", argLength: 3, reg: v21load, asm: "VMINPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMINPD256load", argLength: 3, reg: v21load, asm: "VMINPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMINPD512load", argLength: 3, reg: w21load, asm: "VMINPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VMINPD512load", argLength: 3, reg: w21load, asm: "VMINPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINSD128load", argLength: 3, reg: v21load, asm: "VPMINSD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINSD256load", argLength: 3, reg: v21load, asm: "VPMINSD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINSD512load", argLength: 3, reg: w21load, asm: "VPMINSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMINSD512load", argLength: 3, reg: w21load, asm: "VPMINSD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINSQ128load", argLength: 3, reg: w21load, asm: "VPMINSQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMINSQ128load", argLength: 3, reg: w21load, asm: "VPMINSQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINSQ256load", argLength: 3, reg: w21load, asm: "VPMINSQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMINSQ256load", argLength: 3, reg: w21load, asm: "VPMINSQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINSQ512load", argLength: 3, reg: w21load, asm: "VPMINSQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMINSQ512load", argLength: 3, reg: w21load, asm: "VPMINSQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINUD128load", argLength: 3, reg: v21load, asm: "VPMINUD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINUD256load", argLength: 3, reg: v21load, asm: "VPMINUD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINUD512load", argLength: 3, reg: w21load, asm: "VPMINUD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMINUD512load", argLength: 3, reg: w21load, asm: "VPMINUD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINUQ128load", argLength: 3, reg: w21load, asm: "VPMINUQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMINUQ128load", argLength: 3, reg: w21load, asm: "VPMINUQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINUQ256load", argLength: 3, reg: w21load, asm: "VPMINUQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMINUQ256load", argLength: 3, reg: w21load, asm: "VPMINUQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1530,14 +1470,8 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPMINUQMasked128load", argLength: 4, reg: w2kwload, asm: "VPMINUQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMINUQMasked128load", argLength: 4, reg: w2kwload, asm: "VPMINUQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINUQMasked256load", argLength: 4, reg: w2kwload, asm: "VPMINUQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMINUQMasked256load", argLength: 4, reg: w2kwload, asm: "VPMINUQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMINUQMasked512load", argLength: 4, reg: w2kwload, asm: "VPMINUQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMINUQMasked512load", argLength: 4, reg: w2kwload, asm: "VPMINUQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMULPS128load", argLength: 3, reg: v21load, asm: "VMULPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMULPS256load", argLength: 3, reg: v21load, asm: "VMULPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMULPS512load", argLength: 3, reg: w21load, asm: "VMULPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VMULPS512load", argLength: 3, reg: w21load, asm: "VMULPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMULPD128load", argLength: 3, reg: v21load, asm: "VMULPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMULPD256load", argLength: 3, reg: v21load, asm: "VMULPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMULPD512load", argLength: 3, reg: w21load, asm: "VMULPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VMULPD512load", argLength: 3, reg: w21load, asm: "VMULPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMULLD128load", argLength: 3, reg: v21load, asm: "VPMULLD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMULLD256load", argLength: 3, reg: v21load, asm: "VPMULLD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMULLD512load", argLength: 3, reg: w21load, asm: "VPMULLD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMULLD512load", argLength: 3, reg: w21load, asm: "VPMULLD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMULLQ128load", argLength: 3, reg: w21load, asm: "VPMULLQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMULLQ128load", argLength: 3, reg: w21load, asm: "VPMULLQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMULLQ256load", argLength: 3, reg: w21load, asm: "VPMULLQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPMULLQ256load", argLength: 3, reg: w21load, asm: "VPMULLQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1566,10 +1500,6 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VFMADDSUB213PDMasked128load", argLength: 5, reg: w3kwload, asm: "VFMADDSUB213PD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VFMADDSUB213PDMasked128load", argLength: 5, reg: w3kwload, asm: "VFMADDSUB213PD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VFMADDSUB213PDMasked256load", argLength: 5, reg: w3kwload, asm: "VFMADDSUB213PD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VFMADDSUB213PDMasked256load", argLength: 5, reg: w3kwload, asm: "VFMADDSUB213PD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VFMADDSUB213PDMasked512load", argLength: 5, reg: w3kwload, asm: "VFMADDSUB213PD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VFMADDSUB213PDMasked512load", argLength: 5, reg: w3kwload, asm: "VFMADDSUB213PD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPMULDQ128load", argLength: 3, reg: v21load, asm: "VPMULDQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMULDQ256load", argLength: 3, reg: v21load, asm: "VPMULDQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMULUDQ128load", argLength: 3, reg: v21load, asm: "VPMULUDQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPMULUDQ256load", argLength: 3, reg: v21load, asm: "VPMULUDQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMULPSMasked128load", argLength: 4, reg: w2kwload, asm: "VMULPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VMULPSMasked128load", argLength: 4, reg: w2kwload, asm: "VMULPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMULPSMasked256load", argLength: 4, reg: w2kwload, asm: "VMULPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VMULPSMasked256load", argLength: 4, reg: w2kwload, asm: "VMULPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VMULPSMasked512load", argLength: 4, reg: w2kwload, asm: "VMULPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VMULPSMasked512load", argLength: 4, reg: w2kwload, asm: "VMULPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1614,8 +1544,6 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPORQMasked128load", argLength: 4, reg: w2kwload, asm: "VPORQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPORQMasked128load", argLength: 4, reg: w2kwload, asm: "VPORQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPORQMasked256load", argLength: 4, reg: w2kwload, asm: "VPORQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPORQMasked256load", argLength: 4, reg: w2kwload, asm: "VPORQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPORQMasked512load", argLength: 4, reg: w2kwload, asm: "VPORQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPORQMasked512load", argLength: 4, reg: w2kwload, asm: "VPORQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPERMPS256load", argLength: 3, reg: v21load, asm: "VPERMPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPERMD256load", argLength: 3, reg: v21load, asm: "VPERMD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPERMPS512load", argLength: 3, reg: w21load, asm: "VPERMPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPERMPS512load", argLength: 3, reg: w21load, asm: "VPERMPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPERMD512load", argLength: 3, reg: w21load, asm: "VPERMD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPERMD512load", argLength: 3, reg: w21load, asm: "VPERMD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPERMPD256load", argLength: 3, reg: w21load, asm: "VPERMPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPERMPD256load", argLength: 3, reg: w21load, asm: "VPERMPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1710,11 +1638,7 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VSCALEFPDMasked128load", argLength: 4, reg: w2kwload, asm: "VSCALEFPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSCALEFPDMasked128load", argLength: 4, reg: w2kwload, asm: "VSCALEFPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSCALEFPDMasked256load", argLength: 4, reg: w2kwload, asm: "VSCALEFPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSCALEFPDMasked256load", argLength: 4, reg: w2kwload, asm: "VSCALEFPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSCALEFPDMasked512load", argLength: 4, reg: w2kwload, asm: "VSCALEFPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSCALEFPDMasked512load", argLength: 4, reg: w2kwload, asm: "VSCALEFPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLVD128load", argLength: 3, reg: v21load, asm: "VPSLLVD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLVD256load", argLength: 3, reg: v21load, asm: "VPSLLVD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLVD512load", argLength: 3, reg: w21load, asm: "VPSLLVD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLVD512load", argLength: 3, reg: w21load, asm: "VPSLLVD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLVQ128load", argLength: 3, reg: v21load, asm: "VPSLLVQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLVQ256load", argLength: 3, reg: v21load, asm: "VPSLLVQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLVQ512load", argLength: 3, reg: w21load, asm: "VPSLLVQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLVQ512load", argLength: 3, reg: w21load, asm: "VPSLLVQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSHLDVD128load", argLength: 4, reg: w31load, asm: "VPSHLDVD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPSHLDVD128load", argLength: 4, reg: w31load, asm: "VPSHLDVD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPSHLDVD256load", argLength: 4, reg: w31load, asm: "VPSHLDVD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPSHLDVD256load", argLength: 4, reg: w31load, asm: "VPSHLDVD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
@ -1734,17 +1658,11 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPSLLVQMasked128load", argLength: 4, reg: w2kwload, asm: "VPSLLVQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLVQMasked128load", argLength: 4, reg: w2kwload, asm: "VPSLLVQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLVQMasked256load", argLength: 4, reg: w2kwload, asm: "VPSLLVQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLVQMasked256load", argLength: 4, reg: w2kwload, asm: "VPSLLVQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLVQMasked512load", argLength: 4, reg: w2kwload, asm: "VPSLLVQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLVQMasked512load", argLength: 4, reg: w2kwload, asm: "VPSLLVQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAVD128load", argLength: 3, reg: v21load, asm: "VPSRAVD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAVD256load", argLength: 3, reg: v21load, asm: "VPSRAVD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAVD512load", argLength: 3, reg: w21load, asm: "VPSRAVD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRAVD512load", argLength: 3, reg: w21load, asm: "VPSRAVD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAVQ128load", argLength: 3, reg: w21load, asm: "VPSRAVQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRAVQ128load", argLength: 3, reg: w21load, asm: "VPSRAVQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAVQ256load", argLength: 3, reg: w21load, asm: "VPSRAVQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRAVQ256load", argLength: 3, reg: w21load, asm: "VPSRAVQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAVQ512load", argLength: 3, reg: w21load, asm: "VPSRAVQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRAVQ512load", argLength: 3, reg: w21load, asm: "VPSRAVQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLVD128load", argLength: 3, reg: v21load, asm: "VPSRLVD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLVD256load", argLength: 3, reg: v21load, asm: "VPSRLVD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLVD512load", argLength: 3, reg: w21load, asm: "VPSRLVD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRLVD512load", argLength: 3, reg: w21load, asm: "VPSRLVD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLVQ128load", argLength: 3, reg: v21load, asm: "VPSRLVQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLVQ256load", argLength: 3, reg: v21load, asm: "VPSRLVQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLVQ512load", argLength: 3, reg: w21load, asm: "VPSRLVQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRLVQ512load", argLength: 3, reg: w21load, asm: "VPSRLVQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSHRDVD128load", argLength: 4, reg: w31load, asm: "VPSHRDVD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPSHRDVD128load", argLength: 4, reg: w31load, asm: "VPSHRDVD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: true},
{name: "VPSHRDVD256load", argLength: 4, reg: w31load, asm: "VPSHRDVD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true}, {name: "VPSHRDVD256load", argLength: 4, reg: w31load, asm: "VPSHRDVD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: true},
@ -1770,11 +1688,7 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPSRLVQMasked128load", argLength: 4, reg: w2kwload, asm: "VPSRLVQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRLVQMasked128load", argLength: 4, reg: w2kwload, asm: "VPSRLVQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLVQMasked256load", argLength: 4, reg: w2kwload, asm: "VPSRLVQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRLVQMasked256load", argLength: 4, reg: w2kwload, asm: "VPSRLVQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLVQMasked512load", argLength: 4, reg: w2kwload, asm: "VPSRLVQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRLVQMasked512load", argLength: 4, reg: w2kwload, asm: "VPSRLVQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSQRTPS128load", argLength: 2, reg: v11load, asm: "VSQRTPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSQRTPS256load", argLength: 2, reg: v11load, asm: "VSQRTPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSQRTPS512load", argLength: 2, reg: w11load, asm: "VSQRTPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSQRTPS512load", argLength: 2, reg: w11load, asm: "VSQRTPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSQRTPD128load", argLength: 2, reg: v11load, asm: "VSQRTPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSQRTPD256load", argLength: 2, reg: v11load, asm: "VSQRTPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSQRTPD512load", argLength: 2, reg: w11load, asm: "VSQRTPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSQRTPD512load", argLength: 2, reg: w11load, asm: "VSQRTPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSQRTPSMasked128load", argLength: 3, reg: wkwload, asm: "VSQRTPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSQRTPSMasked128load", argLength: 3, reg: wkwload, asm: "VSQRTPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSQRTPSMasked256load", argLength: 3, reg: wkwload, asm: "VSQRTPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSQRTPSMasked256load", argLength: 3, reg: wkwload, asm: "VSQRTPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1782,17 +1696,9 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VSQRTPDMasked128load", argLength: 3, reg: wkwload, asm: "VSQRTPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSQRTPDMasked128load", argLength: 3, reg: wkwload, asm: "VSQRTPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSQRTPDMasked256load", argLength: 3, reg: wkwload, asm: "VSQRTPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSQRTPDMasked256load", argLength: 3, reg: wkwload, asm: "VSQRTPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSQRTPDMasked512load", argLength: 3, reg: wkwload, asm: "VSQRTPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSQRTPDMasked512load", argLength: 3, reg: wkwload, asm: "VSQRTPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSUBPS128load", argLength: 3, reg: v21load, asm: "VSUBPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSUBPS256load", argLength: 3, reg: v21load, asm: "VSUBPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSUBPS512load", argLength: 3, reg: w21load, asm: "VSUBPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSUBPS512load", argLength: 3, reg: w21load, asm: "VSUBPS", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSUBPD128load", argLength: 3, reg: v21load, asm: "VSUBPD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSUBPD256load", argLength: 3, reg: v21load, asm: "VSUBPD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSUBPD512load", argLength: 3, reg: w21load, asm: "VSUBPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSUBPD512load", argLength: 3, reg: w21load, asm: "VSUBPD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSUBD128load", argLength: 3, reg: v21load, asm: "VPSUBD", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSUBD256load", argLength: 3, reg: v21load, asm: "VPSUBD", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSUBD512load", argLength: 3, reg: w21load, asm: "VPSUBD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSUBD512load", argLength: 3, reg: w21load, asm: "VPSUBD", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSUBQ128load", argLength: 3, reg: v21load, asm: "VPSUBQ", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSUBQ256load", argLength: 3, reg: v21load, asm: "VPSUBQ", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VPSUBQ512load", argLength: 3, reg: w21load, asm: "VPSUBQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VPSUBQ512load", argLength: 3, reg: w21load, asm: "VPSUBQ", commutative: false, typ: "Vec512", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSUBPSMasked128load", argLength: 4, reg: w2kwload, asm: "VSUBPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSUBPSMasked128load", argLength: 4, reg: w2kwload, asm: "VSUBPS", commutative: false, typ: "Vec128", aux: "SymOff", symEffect: "Read", resultInArg0: false},
{name: "VSUBPSMasked256load", argLength: 4, reg: w2kwload, asm: "VSUBPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false}, {name: "VSUBPSMasked256load", argLength: 4, reg: w2kwload, asm: "VSUBPS", commutative: false, typ: "Vec256", aux: "SymOff", symEffect: "Read", resultInArg0: false},
@ -1840,11 +1746,7 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VREDUCEPDMasked128load", argLength: 3, reg: wkwload, asm: "VREDUCEPD", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VREDUCEPDMasked128load", argLength: 3, reg: wkwload, asm: "VREDUCEPD", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VREDUCEPDMasked256load", argLength: 3, reg: wkwload, asm: "VREDUCEPD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VREDUCEPDMasked256load", argLength: 3, reg: wkwload, asm: "VREDUCEPD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VREDUCEPDMasked512load", argLength: 3, reg: wkwload, asm: "VREDUCEPD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VREDUCEPDMasked512load", argLength: 3, reg: wkwload, asm: "VREDUCEPD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VCMPPS128load", argLength: 3, reg: v21load, asm: "VCMPPS", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VCMPPS256load", argLength: 3, reg: v21load, asm: "VCMPPS", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VCMPPS512load", argLength: 3, reg: w2kload, asm: "VCMPPS", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VCMPPS512load", argLength: 3, reg: w2kload, asm: "VCMPPS", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VCMPPD128load", argLength: 3, reg: v21load, asm: "VCMPPD", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VCMPPD256load", argLength: 3, reg: v21load, asm: "VCMPPD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VCMPPD512load", argLength: 3, reg: w2kload, asm: "VCMPPD", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VCMPPD512load", argLength: 3, reg: w2kload, asm: "VCMPPD", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VCMPPSMasked128load", argLength: 4, reg: w2kkload, asm: "VCMPPS", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VCMPPSMasked128load", argLength: 4, reg: w2kkload, asm: "VCMPPS", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VCMPPSMasked256load", argLength: 4, reg: w2kkload, asm: "VCMPPS", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VCMPPSMasked256load", argLength: 4, reg: w2kkload, asm: "VCMPPS", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
@ -1880,8 +1782,6 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPCMPUQ512load", argLength: 3, reg: w2kload, asm: "VPCMPUQ", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPCMPUQ512load", argLength: 3, reg: w2kload, asm: "VPCMPUQ", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPD512load", argLength: 3, reg: w2kload, asm: "VPCMPD", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPCMPD512load", argLength: 3, reg: w2kload, asm: "VPCMPD", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPCMPQ512load", argLength: 3, reg: w2kload, asm: "VPCMPQ", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPCMPQ512load", argLength: 3, reg: w2kload, asm: "VPCMPQ", commutative: false, typ: "Mask", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSHUFD128load", argLength: 2, reg: v11load, asm: "VPSHUFD", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSHUFD256load", argLength: 2, reg: v11load, asm: "VPSHUFD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSHUFD512load", argLength: 2, reg: w11load, asm: "VPSHUFD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSHUFD512load", argLength: 2, reg: w11load, asm: "VPSHUFD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSHUFDMasked256load", argLength: 3, reg: wkwload, asm: "VPSHUFD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSHUFDMasked256load", argLength: 3, reg: wkwload, asm: "VPSHUFD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSHUFDMasked512load", argLength: 3, reg: wkwload, asm: "VPSHUFD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSHUFDMasked512load", argLength: 3, reg: wkwload, asm: "VPSHUFD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
@ -1934,11 +1834,7 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPSHRDQMasked128load", argLength: 4, reg: w2kwload, asm: "VPSHRDQ", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSHRDQMasked128load", argLength: 4, reg: w2kwload, asm: "VPSHRDQ", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSHRDQMasked256load", argLength: 4, reg: w2kwload, asm: "VPSHRDQ", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSHRDQMasked256load", argLength: 4, reg: w2kwload, asm: "VPSHRDQ", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSHRDQMasked512load", argLength: 4, reg: w2kwload, asm: "VPSHRDQ", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSHRDQMasked512load", argLength: 4, reg: w2kwload, asm: "VPSHRDQ", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLD128constload", argLength: 2, reg: v11load, asm: "VPSLLD", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLD256constload", argLength: 2, reg: v11load, asm: "VPSLLD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLD512constload", argLength: 2, reg: w11load, asm: "VPSLLD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLD512constload", argLength: 2, reg: w11load, asm: "VPSLLD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLQ128constload", argLength: 2, reg: v11load, asm: "VPSLLQ", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLQ256constload", argLength: 2, reg: v11load, asm: "VPSLLQ", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLQ512constload", argLength: 2, reg: w11load, asm: "VPSLLQ", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLQ512constload", argLength: 2, reg: w11load, asm: "VPSLLQ", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLDMasked128constload", argLength: 3, reg: wkwload, asm: "VPSLLD", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLDMasked128constload", argLength: 3, reg: wkwload, asm: "VPSLLD", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLDMasked256constload", argLength: 3, reg: wkwload, asm: "VPSLLD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLDMasked256constload", argLength: 3, reg: wkwload, asm: "VPSLLD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
@ -1946,14 +1842,8 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
{name: "VPSLLQMasked128constload", argLength: 3, reg: wkwload, asm: "VPSLLQ", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLQMasked128constload", argLength: 3, reg: wkwload, asm: "VPSLLQ", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLQMasked256constload", argLength: 3, reg: wkwload, asm: "VPSLLQ", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLQMasked256constload", argLength: 3, reg: wkwload, asm: "VPSLLQ", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSLLQMasked512constload", argLength: 3, reg: wkwload, asm: "VPSLLQ", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSLLQMasked512constload", argLength: 3, reg: wkwload, asm: "VPSLLQ", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLD128constload", argLength: 2, reg: v11load, asm: "VPSRLD", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLD256constload", argLength: 2, reg: v11load, asm: "VPSRLD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLD512constload", argLength: 2, reg: w11load, asm: "VPSRLD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRLD512constload", argLength: 2, reg: w11load, asm: "VPSRLD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLQ128constload", argLength: 2, reg: v11load, asm: "VPSRLQ", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLQ256constload", argLength: 2, reg: v11load, asm: "VPSRLQ", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRLQ512constload", argLength: 2, reg: w11load, asm: "VPSRLQ", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRLQ512constload", argLength: 2, reg: w11load, asm: "VPSRLQ", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAD128constload", argLength: 2, reg: v11load, asm: "VPSRAD", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAD256constload", argLength: 2, reg: v11load, asm: "VPSRAD", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAD512constload", argLength: 2, reg: w11load, asm: "VPSRAD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRAD512constload", argLength: 2, reg: w11load, asm: "VPSRAD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAQ128constload", argLength: 2, reg: w11load, asm: "VPSRAQ", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRAQ128constload", argLength: 2, reg: w11load, asm: "VPSRAQ", commutative: false, typ: "Vec128", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
{name: "VPSRAQ256constload", argLength: 2, reg: w11load, asm: "VPSRAQ", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false}, {name: "VPSRAQ256constload", argLength: 2, reg: w11load, asm: "VPSRAQ", commutative: false, typ: "Vec256", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},

1866
src/cmd/compile/internal/ssa/opGen.go
View file

File diff suppressed because it is too large Load diff

2840
src/cmd/compile/internal/ssa/rewriteAMD64.go
View file

File diff suppressed because it is too large Load diff

16
src/simd/_gen/simdgen/gen_utility.go
View file

@ -632,7 +632,21 @@ func dedupGodef(ops []Operation) ([]Operation, error) {
if isAVX512(i) && !isAVX512(j) { if isAVX512(i) && !isAVX512(j) {
return 1 return 1
} }
return strings.Compare(i.CPUFeature, j.CPUFeature) if i.CPUFeature != j.CPUFeature {
return strings.Compare(i.CPUFeature, j.CPUFeature)
}
// Weirdly Intel sometimes has duplicated definitions for the same instruction,
// this confuses the XED mem-op merge logic: [MemFeature] will only be attached to an instruction
// for only once, which means that for essentially duplicated instructions only one will have the
// proper [MemFeature] set. We have to make this sort deterministic for [MemFeature].
if i.MemFeatures != nil && j.MemFeatures == nil {
return -1
}
if i.MemFeatures == nil && j.MemFeatures != nil {
return 1
}
// Their order does not matter anymore, at least for now.
return 0
}) })
} }
deduped = append(deduped, dup[0]) deduped = append(deduped, dup[0])

33
src/simd/_gen/simdgen/xed.go
View file

@ -9,6 +9,7 @@ import (
"fmt" "fmt"
"log" "log"
"maps" "maps"
"reflect"
"regexp" "regexp"
"slices" "slices"
"strconv" "strconv"
@ -137,14 +138,24 @@ func loadXED(xedPath string) []*unify.Value {
} }
if len(o.ops) == len(m.ops) { if len(o.ops) == len(m.ops) {
for j := range o.ops { for j := range o.ops {
v1, ok3 := o.ops[j].(operandVReg) if reflect.TypeOf(o.ops[j]) == reflect.TypeOf(m.ops[j]) {
v2, ok4 := m.ops[j].(operandVReg) v1, ok3 := o.ops[j].(operandVReg)
if !ok3 || !ok4 { v2, _ := m.ops[j].(operandVReg)
continue if !ok3 {
} continue
if v1.vecShape != v2.vecShape { }
// A mismatch, skip this memOp if v1.vecShape != v2.vecShape {
continue outer // A mismatch, skip this memOp
continue outer
}
} else {
_, ok3 := o.ops[j].(operandVReg)
_, ok4 := m.ops[j].(operandMem)
// The only difference must be the vreg and mem, no other cases.
if !ok3 || !ok4 {
// A mismatch, skip this memOp
continue outer
}
} }
} }
// Found a match, break early // Found a match, break early
@ -155,10 +166,10 @@ func loadXED(xedPath string) []*unify.Value {
// Remove the match from memOps, it's now merged to this pure vreg operation // Remove the match from memOps, it's now merged to this pure vreg operation
if matchIdx != -1 { if matchIdx != -1 {
memOps[opcode] = append(memOps[opcode][:matchIdx], memOps[opcode][matchIdx+1:]...) memOps[opcode] = append(memOps[opcode][:matchIdx], memOps[opcode][matchIdx+1:]...)
// Merge is done by adding a new field
// Right now we only have vbcst
addFields["memFeatures"] = "vbcst"
} }
// Merge is done by adding a new field
// Right now we only have vbcst
addFields["memFeatures"] = "vbcst"
} }
} }
appendDefs(o.inst, o.ops, addFields) appendDefs(o.inst, o.ops, addFields)