[dev.simd] simd: add emulations for missing AVX2 comparisons

this also removes AVX512 versions of the operations
that would use the same names, but not run on AVX2-only

includes files generated by simdgen CL 692355

Change-Id: Iff29042245b7688133fed49a03e681e85235b8a8
Reviewed-on: https://go-review.googlesource.com/c/go/+/692335
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Junyang Shao <shaojunyang@google.com>

src/cmd/compile/internal/amd64/simdssa.go

@@ -886,29 +886,13 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
 
 	case ssa.OpAMD64VCMPPS512,
 		ssa.OpAMD64VCMPPD512,
-		ssa.OpAMD64VPCMPUB128,
-		ssa.OpAMD64VPCMPUB256,
 		ssa.OpAMD64VPCMPUB512,
-		ssa.OpAMD64VPCMPUW128,
-		ssa.OpAMD64VPCMPUW256,
 		ssa.OpAMD64VPCMPUW512,
-		ssa.OpAMD64VPCMPUD128,
-		ssa.OpAMD64VPCMPUD256,
 		ssa.OpAMD64VPCMPUD512,
-		ssa.OpAMD64VPCMPUQ128,
-		ssa.OpAMD64VPCMPUQ256,
 		ssa.OpAMD64VPCMPUQ512,
-		ssa.OpAMD64VPCMPB128,
-		ssa.OpAMD64VPCMPB256,
 		ssa.OpAMD64VPCMPB512,
-		ssa.OpAMD64VPCMPW128,
-		ssa.OpAMD64VPCMPW256,
 		ssa.OpAMD64VPCMPW512,
-		ssa.OpAMD64VPCMPD128,
-		ssa.OpAMD64VPCMPD256,
 		ssa.OpAMD64VPCMPD512,
-		ssa.OpAMD64VPCMPQ128,
-		ssa.OpAMD64VPCMPQ256,
 		ssa.OpAMD64VPCMPQ512:
 		p = simdV2kImm8(s, v)
 

src/cmd/compile/internal/ssa/_gen/simdAMD64.rules

@@ -590,17 +590,9 @@
 (GreaterInt64x2 ...) => (VPCMPGTQ128 ...)
 (GreaterInt64x4 ...) => (VPCMPGTQ256 ...)
 (GreaterInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPGTQ512 x y))
-(GreaterUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [14] x y))
-(GreaterUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [14] x y))
 (GreaterUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [14] x y))
-(GreaterUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [14] x y))
-(GreaterUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [14] x y))
 (GreaterUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [14] x y))
-(GreaterUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [14] x y))
-(GreaterUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [14] x y))
 (GreaterUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [14] x y))
-(GreaterUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [14] x y))
-(GreaterUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [14] x y))
 (GreaterUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [14] x y))
 (GreaterEqualFloat32x4 x y) => (VCMPPS128 [13] x y)
 (GreaterEqualFloat32x8 x y) => (VCMPPS256 [13] x y)
@@ -608,29 +600,13 @@
 (GreaterEqualFloat64x2 x y) => (VCMPPD128 [13] x y)
 (GreaterEqualFloat64x4 x y) => (VCMPPD256 [13] x y)
 (GreaterEqualFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [13] x y))
-(GreaterEqualInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [13] x y))
-(GreaterEqualInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [13] x y))
 (GreaterEqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [13] x y))
-(GreaterEqualInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [13] x y))
-(GreaterEqualInt16x16 x y) => (VPMOVMToVec16x16 (VPCMPW256 [13] x y))
 (GreaterEqualInt16x32 x y) => (VPMOVMToVec16x32 (VPCMPW512 [13] x y))
-(GreaterEqualInt32x4 x y) => (VPMOVMToVec32x4 (VPCMPD128 [13] x y))
-(GreaterEqualInt32x8 x y) => (VPMOVMToVec32x8 (VPCMPD256 [13] x y))
 (GreaterEqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [13] x y))
-(GreaterEqualInt64x2 x y) => (VPMOVMToVec64x2 (VPCMPQ128 [13] x y))
-(GreaterEqualInt64x4 x y) => (VPMOVMToVec64x4 (VPCMPQ256 [13] x y))
 (GreaterEqualInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPQ512 [13] x y))
-(GreaterEqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [13] x y))
-(GreaterEqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [13] x y))
 (GreaterEqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [13] x y))
-(GreaterEqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [13] x y))
-(GreaterEqualUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [13] x y))
 (GreaterEqualUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [13] x y))
-(GreaterEqualUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [13] x y))
-(GreaterEqualUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [13] x y))
 (GreaterEqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [13] x y))
-(GreaterEqualUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [13] x y))
-(GreaterEqualUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [13] x y))
 (GreaterEqualUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [13] x y))
 (GreaterEqualMaskedFloat32x4 x y mask) => (VPMOVMToVec32x4 (VCMPPSMasked128 [13] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (GreaterEqualMaskedFloat32x8 x y mask) => (VPMOVMToVec32x8 (VCMPPSMasked256 [13] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
@@ -710,29 +686,13 @@
 (LessFloat64x2 x y) => (VCMPPD128 [1] x y)
 (LessFloat64x4 x y) => (VCMPPD256 [1] x y)
 (LessFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [1] x y))
-(LessInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [1] x y))
-(LessInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [1] x y))
 (LessInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [1] x y))
-(LessInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [1] x y))
-(LessInt16x16 x y) => (VPMOVMToVec16x16 (VPCMPW256 [1] x y))
 (LessInt16x32 x y) => (VPMOVMToVec16x32 (VPCMPW512 [1] x y))
-(LessInt32x4 x y) => (VPMOVMToVec32x4 (VPCMPD128 [1] x y))
-(LessInt32x8 x y) => (VPMOVMToVec32x8 (VPCMPD256 [1] x y))
 (LessInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [1] x y))
-(LessInt64x2 x y) => (VPMOVMToVec64x2 (VPCMPQ128 [1] x y))
-(LessInt64x4 x y) => (VPMOVMToVec64x4 (VPCMPQ256 [1] x y))
 (LessInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPQ512 [1] x y))
-(LessUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [1] x y))
-(LessUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [1] x y))
 (LessUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [1] x y))
-(LessUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [1] x y))
-(LessUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [1] x y))
 (LessUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [1] x y))
-(LessUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [1] x y))
-(LessUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [1] x y))
 (LessUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [1] x y))
-(LessUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [1] x y))
-(LessUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [1] x y))
 (LessUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [1] x y))
 (LessEqualFloat32x4 x y) => (VCMPPS128 [2] x y)
 (LessEqualFloat32x8 x y) => (VCMPPS256 [2] x y)
@@ -740,29 +700,13 @@
 (LessEqualFloat64x2 x y) => (VCMPPD128 [2] x y)
 (LessEqualFloat64x4 x y) => (VCMPPD256 [2] x y)
 (LessEqualFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [2] x y))
-(LessEqualInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [2] x y))
-(LessEqualInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [2] x y))
 (LessEqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [2] x y))
-(LessEqualInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [2] x y))
-(LessEqualInt16x16 x y) => (VPMOVMToVec16x16 (VPCMPW256 [2] x y))
 (LessEqualInt16x32 x y) => (VPMOVMToVec16x32 (VPCMPW512 [2] x y))
-(LessEqualInt32x4 x y) => (VPMOVMToVec32x4 (VPCMPD128 [2] x y))
-(LessEqualInt32x8 x y) => (VPMOVMToVec32x8 (VPCMPD256 [2] x y))
 (LessEqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [2] x y))
-(LessEqualInt64x2 x y) => (VPMOVMToVec64x2 (VPCMPQ128 [2] x y))
-(LessEqualInt64x4 x y) => (VPMOVMToVec64x4 (VPCMPQ256 [2] x y))
 (LessEqualInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPQ512 [2] x y))
-(LessEqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [2] x y))
-(LessEqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [2] x y))
 (LessEqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [2] x y))
-(LessEqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [2] x y))
-(LessEqualUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [2] x y))
 (LessEqualUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [2] x y))
-(LessEqualUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [2] x y))
-(LessEqualUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [2] x y))
 (LessEqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [2] x y))
-(LessEqualUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [2] x y))
-(LessEqualUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [2] x y))
 (LessEqualUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [2] x y))
 (LessEqualMaskedFloat32x4 x y mask) => (VPMOVMToVec32x4 (VCMPPSMasked128 [2] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (LessEqualMaskedFloat32x8 x y mask) => (VPMOVMToVec32x8 (VCMPPSMasked256 [2] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
@@ -1050,29 +994,13 @@
 (NotEqualFloat64x2 x y) => (VCMPPD128 [4] x y)
 (NotEqualFloat64x4 x y) => (VCMPPD256 [4] x y)
 (NotEqualFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [4] x y))
-(NotEqualInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [4] x y))
-(NotEqualInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [4] x y))
 (NotEqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [4] x y))
-(NotEqualInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [4] x y))
-(NotEqualInt16x16 x y) => (VPMOVMToVec16x16 (VPCMPW256 [4] x y))
 (NotEqualInt16x32 x y) => (VPMOVMToVec16x32 (VPCMPW512 [4] x y))
-(NotEqualInt32x4 x y) => (VPMOVMToVec32x4 (VPCMPD128 [4] x y))
-(NotEqualInt32x8 x y) => (VPMOVMToVec32x8 (VPCMPD256 [4] x y))
 (NotEqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [4] x y))
-(NotEqualInt64x2 x y) => (VPMOVMToVec64x2 (VPCMPQ128 [4] x y))
-(NotEqualInt64x4 x y) => (VPMOVMToVec64x4 (VPCMPQ256 [4] x y))
 (NotEqualInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPQ512 [4] x y))
-(NotEqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [4] x y))
-(NotEqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [4] x y))
 (NotEqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [4] x y))
-(NotEqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [4] x y))
-(NotEqualUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [4] x y))
 (NotEqualUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [4] x y))
-(NotEqualUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [4] x y))
-(NotEqualUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [4] x y))
 (NotEqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [4] x y))
-(NotEqualUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [4] x y))
-(NotEqualUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [4] x y))
 (NotEqualUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [4] x y))
 (NotEqualMaskedFloat32x4 x y mask) => (VPMOVMToVec32x4 (VCMPPSMasked128 [4] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (NotEqualMaskedFloat32x8 x y mask) => (VPMOVMToVec32x8 (VCMPPSMasked256 [4] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))

src/cmd/compile/internal/ssa/_gen/simdAMD64ops.go

@@ -986,29 +986,13 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
 		{name: "VEXTRACTF64X4256", argLength: 1, reg: w11, asm: "VEXTRACTF64X4", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: false},
 		{name: "VEXTRACTI128128", argLength: 1, reg: v11, asm: "VEXTRACTI128", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: false},
 		{name: "VEXTRACTI64X4256", argLength: 1, reg: w11, asm: "VEXTRACTI64X4", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: false},
-		{name: "VPCMPUB128", argLength: 2, reg: w2k, asm: "VPCMPUB", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPUB256", argLength: 2, reg: w2k, asm: "VPCMPUB", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
 		{name: "VPCMPUB512", argLength: 2, reg: w2k, asm: "VPCMPUB", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPUW128", argLength: 2, reg: w2k, asm: "VPCMPUW", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPUW256", argLength: 2, reg: w2k, asm: "VPCMPUW", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
 		{name: "VPCMPUW512", argLength: 2, reg: w2k, asm: "VPCMPUW", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPUD128", argLength: 2, reg: w2k, asm: "VPCMPUD", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPUD256", argLength: 2, reg: w2k, asm: "VPCMPUD", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
 		{name: "VPCMPUD512", argLength: 2, reg: w2k, asm: "VPCMPUD", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPUQ128", argLength: 2, reg: w2k, asm: "VPCMPUQ", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPUQ256", argLength: 2, reg: w2k, asm: "VPCMPUQ", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
 		{name: "VPCMPUQ512", argLength: 2, reg: w2k, asm: "VPCMPUQ", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPB128", argLength: 2, reg: w2k, asm: "VPCMPB", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPB256", argLength: 2, reg: w2k, asm: "VPCMPB", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
 		{name: "VPCMPB512", argLength: 2, reg: w2k, asm: "VPCMPB", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPW128", argLength: 2, reg: w2k, asm: "VPCMPW", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPW256", argLength: 2, reg: w2k, asm: "VPCMPW", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
 		{name: "VPCMPW512", argLength: 2, reg: w2k, asm: "VPCMPW", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPD128", argLength: 2, reg: w2k, asm: "VPCMPD", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPD256", argLength: 2, reg: w2k, asm: "VPCMPD", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
 		{name: "VPCMPD512", argLength: 2, reg: w2k, asm: "VPCMPD", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPQ128", argLength: 2, reg: w2k, asm: "VPCMPQ", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
-		{name: "VPCMPQ256", argLength: 2, reg: w2k, asm: "VPCMPQ", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
 		{name: "VPCMPQ512", argLength: 2, reg: w2k, asm: "VPCMPQ", aux: "UInt8", commutative: false, typ: "Mask", resultInArg0: false},
 		{name: "VPROLD128", argLength: 1, reg: w11, asm: "VPROLD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: false},
 		{name: "VPROLD256", argLength: 1, reg: w11, asm: "VPROLD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: false},

src/cmd/compile/internal/ssa/_gen/simdgenericOps.go

@@ -514,17 +514,9 @@ func simdGenericOps() []opData {
 		{name: "GreaterEqualFloat64x2", argLength: 2, commutative: false},
 		{name: "GreaterEqualFloat64x4", argLength: 2, commutative: false},
 		{name: "GreaterEqualFloat64x8", argLength: 2, commutative: false},
-		{name: "GreaterEqualInt8x16", argLength: 2, commutative: false},
-		{name: "GreaterEqualInt8x32", argLength: 2, commutative: false},
 		{name: "GreaterEqualInt8x64", argLength: 2, commutative: false},
-		{name: "GreaterEqualInt16x8", argLength: 2, commutative: false},
-		{name: "GreaterEqualInt16x16", argLength: 2, commutative: false},
 		{name: "GreaterEqualInt16x32", argLength: 2, commutative: false},
-		{name: "GreaterEqualInt32x4", argLength: 2, commutative: false},
-		{name: "GreaterEqualInt32x8", argLength: 2, commutative: false},
 		{name: "GreaterEqualInt32x16", argLength: 2, commutative: false},
-		{name: "GreaterEqualInt64x2", argLength: 2, commutative: false},
-		{name: "GreaterEqualInt64x4", argLength: 2, commutative: false},
 		{name: "GreaterEqualInt64x8", argLength: 2, commutative: false},
 		{name: "GreaterEqualMaskedFloat32x4", argLength: 3, commutative: false},
 		{name: "GreaterEqualMaskedFloat32x8", argLength: 3, commutative: false},
@@ -556,17 +548,9 @@ func simdGenericOps() []opData {
 		{name: "GreaterEqualMaskedUint64x2", argLength: 3, commutative: false},
 		{name: "GreaterEqualMaskedUint64x4", argLength: 3, commutative: false},
 		{name: "GreaterEqualMaskedUint64x8", argLength: 3, commutative: false},
-		{name: "GreaterEqualUint8x16", argLength: 2, commutative: false},
-		{name: "GreaterEqualUint8x32", argLength: 2, commutative: false},
 		{name: "GreaterEqualUint8x64", argLength: 2, commutative: false},
-		{name: "GreaterEqualUint16x8", argLength: 2, commutative: false},
-		{name: "GreaterEqualUint16x16", argLength: 2, commutative: false},
 		{name: "GreaterEqualUint16x32", argLength: 2, commutative: false},
-		{name: "GreaterEqualUint32x4", argLength: 2, commutative: false},
-		{name: "GreaterEqualUint32x8", argLength: 2, commutative: false},
 		{name: "GreaterEqualUint32x16", argLength: 2, commutative: false},
-		{name: "GreaterEqualUint64x2", argLength: 2, commutative: false},
-		{name: "GreaterEqualUint64x4", argLength: 2, commutative: false},
 		{name: "GreaterEqualUint64x8", argLength: 2, commutative: false},
 		{name: "GreaterFloat32x4", argLength: 2, commutative: false},
 		{name: "GreaterFloat32x8", argLength: 2, commutative: false},
@@ -616,17 +600,9 @@ func simdGenericOps() []opData {
 		{name: "GreaterMaskedUint64x2", argLength: 3, commutative: false},
 		{name: "GreaterMaskedUint64x4", argLength: 3, commutative: false},
 		{name: "GreaterMaskedUint64x8", argLength: 3, commutative: false},
-		{name: "GreaterUint8x16", argLength: 2, commutative: false},
-		{name: "GreaterUint8x32", argLength: 2, commutative: false},
 		{name: "GreaterUint8x64", argLength: 2, commutative: false},
-		{name: "GreaterUint16x8", argLength: 2, commutative: false},
-		{name: "GreaterUint16x16", argLength: 2, commutative: false},
 		{name: "GreaterUint16x32", argLength: 2, commutative: false},
-		{name: "GreaterUint32x4", argLength: 2, commutative: false},
-		{name: "GreaterUint32x8", argLength: 2, commutative: false},
 		{name: "GreaterUint32x16", argLength: 2, commutative: false},
-		{name: "GreaterUint64x2", argLength: 2, commutative: false},
-		{name: "GreaterUint64x4", argLength: 2, commutative: false},
 		{name: "GreaterUint64x8", argLength: 2, commutative: false},
 		{name: "IsNanFloat32x4", argLength: 2, commutative: true},
 		{name: "IsNanFloat32x8", argLength: 2, commutative: true},
@@ -646,17 +622,9 @@ func simdGenericOps() []opData {
 		{name: "LessEqualFloat64x2", argLength: 2, commutative: false},
 		{name: "LessEqualFloat64x4", argLength: 2, commutative: false},
 		{name: "LessEqualFloat64x8", argLength: 2, commutative: false},
-		{name: "LessEqualInt8x16", argLength: 2, commutative: false},
-		{name: "LessEqualInt8x32", argLength: 2, commutative: false},
 		{name: "LessEqualInt8x64", argLength: 2, commutative: false},
-		{name: "LessEqualInt16x8", argLength: 2, commutative: false},
-		{name: "LessEqualInt16x16", argLength: 2, commutative: false},
 		{name: "LessEqualInt16x32", argLength: 2, commutative: false},
-		{name: "LessEqualInt32x4", argLength: 2, commutative: false},
-		{name: "LessEqualInt32x8", argLength: 2, commutative: false},
 		{name: "LessEqualInt32x16", argLength: 2, commutative: false},
-		{name: "LessEqualInt64x2", argLength: 2, commutative: false},
-		{name: "LessEqualInt64x4", argLength: 2, commutative: false},
 		{name: "LessEqualInt64x8", argLength: 2, commutative: false},
 		{name: "LessEqualMaskedFloat32x4", argLength: 3, commutative: false},
 		{name: "LessEqualMaskedFloat32x8", argLength: 3, commutative: false},
@@ -688,17 +656,9 @@ func simdGenericOps() []opData {
 		{name: "LessEqualMaskedUint64x2", argLength: 3, commutative: false},
 		{name: "LessEqualMaskedUint64x4", argLength: 3, commutative: false},
 		{name: "LessEqualMaskedUint64x8", argLength: 3, commutative: false},
-		{name: "LessEqualUint8x16", argLength: 2, commutative: false},
-		{name: "LessEqualUint8x32", argLength: 2, commutative: false},
 		{name: "LessEqualUint8x64", argLength: 2, commutative: false},
-		{name: "LessEqualUint16x8", argLength: 2, commutative: false},
-		{name: "LessEqualUint16x16", argLength: 2, commutative: false},
 		{name: "LessEqualUint16x32", argLength: 2, commutative: false},
-		{name: "LessEqualUint32x4", argLength: 2, commutative: false},
-		{name: "LessEqualUint32x8", argLength: 2, commutative: false},
 		{name: "LessEqualUint32x16", argLength: 2, commutative: false},
-		{name: "LessEqualUint64x2", argLength: 2, commutative: false},
-		{name: "LessEqualUint64x4", argLength: 2, commutative: false},
 		{name: "LessEqualUint64x8", argLength: 2, commutative: false},
 		{name: "LessFloat32x4", argLength: 2, commutative: false},
 		{name: "LessFloat32x8", argLength: 2, commutative: false},
@@ -706,17 +666,9 @@ func simdGenericOps() []opData {
 		{name: "LessFloat64x2", argLength: 2, commutative: false},
 		{name: "LessFloat64x4", argLength: 2, commutative: false},
 		{name: "LessFloat64x8", argLength: 2, commutative: false},
-		{name: "LessInt8x16", argLength: 2, commutative: false},
-		{name: "LessInt8x32", argLength: 2, commutative: false},
 		{name: "LessInt8x64", argLength: 2, commutative: false},
-		{name: "LessInt16x8", argLength: 2, commutative: false},
-		{name: "LessInt16x16", argLength: 2, commutative: false},
 		{name: "LessInt16x32", argLength: 2, commutative: false},
-		{name: "LessInt32x4", argLength: 2, commutative: false},
-		{name: "LessInt32x8", argLength: 2, commutative: false},
 		{name: "LessInt32x16", argLength: 2, commutative: false},
-		{name: "LessInt64x2", argLength: 2, commutative: false},
-		{name: "LessInt64x4", argLength: 2, commutative: false},
 		{name: "LessInt64x8", argLength: 2, commutative: false},
 		{name: "LessMaskedFloat32x4", argLength: 3, commutative: false},
 		{name: "LessMaskedFloat32x8", argLength: 3, commutative: false},
@@ -748,17 +700,9 @@ func simdGenericOps() []opData {
 		{name: "LessMaskedUint64x2", argLength: 3, commutative: false},
 		{name: "LessMaskedUint64x4", argLength: 3, commutative: false},
 		{name: "LessMaskedUint64x8", argLength: 3, commutative: false},
-		{name: "LessUint8x16", argLength: 2, commutative: false},
-		{name: "LessUint8x32", argLength: 2, commutative: false},
 		{name: "LessUint8x64", argLength: 2, commutative: false},
-		{name: "LessUint16x8", argLength: 2, commutative: false},
-		{name: "LessUint16x16", argLength: 2, commutative: false},
 		{name: "LessUint16x32", argLength: 2, commutative: false},
-		{name: "LessUint32x4", argLength: 2, commutative: false},
-		{name: "LessUint32x8", argLength: 2, commutative: false},
 		{name: "LessUint32x16", argLength: 2, commutative: false},
-		{name: "LessUint64x2", argLength: 2, commutative: false},
-		{name: "LessUint64x4", argLength: 2, commutative: false},
 		{name: "LessUint64x8", argLength: 2, commutative: false},
 		{name: "MaxFloat32x4", argLength: 2, commutative: true},
 		{name: "MaxFloat32x8", argLength: 2, commutative: true},
@@ -986,17 +930,9 @@ func simdGenericOps() []opData {
 		{name: "NotEqualFloat64x2", argLength: 2, commutative: true},
 		{name: "NotEqualFloat64x4", argLength: 2, commutative: true},
 		{name: "NotEqualFloat64x8", argLength: 2, commutative: true},
-		{name: "NotEqualInt8x16", argLength: 2, commutative: true},
-		{name: "NotEqualInt8x32", argLength: 2, commutative: true},
 		{name: "NotEqualInt8x64", argLength: 2, commutative: true},
-		{name: "NotEqualInt16x8", argLength: 2, commutative: true},
-		{name: "NotEqualInt16x16", argLength: 2, commutative: true},
 		{name: "NotEqualInt16x32", argLength: 2, commutative: true},
-		{name: "NotEqualInt32x4", argLength: 2, commutative: true},
-		{name: "NotEqualInt32x8", argLength: 2, commutative: true},
 		{name: "NotEqualInt32x16", argLength: 2, commutative: true},
-		{name: "NotEqualInt64x2", argLength: 2, commutative: true},
-		{name: "NotEqualInt64x4", argLength: 2, commutative: true},
 		{name: "NotEqualInt64x8", argLength: 2, commutative: true},
 		{name: "NotEqualMaskedFloat32x4", argLength: 3, commutative: true},
 		{name: "NotEqualMaskedFloat32x8", argLength: 3, commutative: true},
@@ -1028,17 +964,9 @@ func simdGenericOps() []opData {
 		{name: "NotEqualMaskedUint64x2", argLength: 3, commutative: true},
 		{name: "NotEqualMaskedUint64x4", argLength: 3, commutative: true},
 		{name: "NotEqualMaskedUint64x8", argLength: 3, commutative: true},
-		{name: "NotEqualUint8x16", argLength: 2, commutative: true},
-		{name: "NotEqualUint8x32", argLength: 2, commutative: true},
 		{name: "NotEqualUint8x64", argLength: 2, commutative: true},
-		{name: "NotEqualUint16x8", argLength: 2, commutative: true},
-		{name: "NotEqualUint16x16", argLength: 2, commutative: true},
 		{name: "NotEqualUint16x32", argLength: 2, commutative: true},
-		{name: "NotEqualUint32x4", argLength: 2, commutative: true},
-		{name: "NotEqualUint32x8", argLength: 2, commutative: true},
 		{name: "NotEqualUint32x16", argLength: 2, commutative: true},
-		{name: "NotEqualUint64x2", argLength: 2, commutative: true},
-		{name: "NotEqualUint64x4", argLength: 2, commutative: true},
 		{name: "NotEqualUint64x8", argLength: 2, commutative: true},
 		{name: "OnesCountInt8x16", argLength: 1, commutative: false},
 		{name: "OnesCountInt8x32", argLength: 1, commutative: false},

src/cmd/compile/internal/ssagen/simdintrinsics.go

@@ -602,17 +602,9 @@ func simdIntrinsics(addF func(pkg, fn string, b intrinsicBuilder, archFamilies .
 	addF(simdPackage, "Float64x2.Greater", opLen2(ssa.OpGreaterFloat64x2, types.TypeVec128), sys.AMD64)
 	addF(simdPackage, "Float64x4.Greater", opLen2(ssa.OpGreaterFloat64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Float64x8.Greater", opLen2(ssa.OpGreaterFloat64x8, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint8x16.Greater", opLen2(ssa.OpGreaterUint8x16, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint8x32.Greater", opLen2(ssa.OpGreaterUint8x32, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint8x64.Greater", opLen2(ssa.OpGreaterUint8x64, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint16x8.Greater", opLen2(ssa.OpGreaterUint16x8, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint16x16.Greater", opLen2(ssa.OpGreaterUint16x16, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint16x32.Greater", opLen2(ssa.OpGreaterUint16x32, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint32x4.Greater", opLen2(ssa.OpGreaterUint32x4, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint32x8.Greater", opLen2(ssa.OpGreaterUint32x8, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint32x16.Greater", opLen2(ssa.OpGreaterUint32x16, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint64x2.Greater", opLen2(ssa.OpGreaterUint64x2, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint64x4.Greater", opLen2(ssa.OpGreaterUint64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint64x8.Greater", opLen2(ssa.OpGreaterUint64x8, types.TypeVec512), sys.AMD64)
 	addF(simdPackage, "Float32x4.GreaterEqual", opLen2(ssa.OpGreaterEqualFloat32x4, types.TypeVec128), sys.AMD64)
 	addF(simdPackage, "Float32x8.GreaterEqual", opLen2(ssa.OpGreaterEqualFloat32x8, types.TypeVec256), sys.AMD64)
@@ -620,29 +612,13 @@ func simdIntrinsics(addF func(pkg, fn string, b intrinsicBuilder, archFamilies .
 	addF(simdPackage, "Float64x2.GreaterEqual", opLen2(ssa.OpGreaterEqualFloat64x2, types.TypeVec128), sys.AMD64)
 	addF(simdPackage, "Float64x4.GreaterEqual", opLen2(ssa.OpGreaterEqualFloat64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Float64x8.GreaterEqual", opLen2(ssa.OpGreaterEqualFloat64x8, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int8x16.GreaterEqual", opLen2(ssa.OpGreaterEqualInt8x16, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int8x32.GreaterEqual", opLen2(ssa.OpGreaterEqualInt8x32, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int8x64.GreaterEqual", opLen2(ssa.OpGreaterEqualInt8x64, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int16x8.GreaterEqual", opLen2(ssa.OpGreaterEqualInt16x8, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int16x16.GreaterEqual", opLen2(ssa.OpGreaterEqualInt16x16, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int16x32.GreaterEqual", opLen2(ssa.OpGreaterEqualInt16x32, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int32x4.GreaterEqual", opLen2(ssa.OpGreaterEqualInt32x4, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int32x8.GreaterEqual", opLen2(ssa.OpGreaterEqualInt32x8, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int32x16.GreaterEqual", opLen2(ssa.OpGreaterEqualInt32x16, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int64x2.GreaterEqual", opLen2(ssa.OpGreaterEqualInt64x2, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int64x4.GreaterEqual", opLen2(ssa.OpGreaterEqualInt64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int64x8.GreaterEqual", opLen2(ssa.OpGreaterEqualInt64x8, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint8x16.GreaterEqual", opLen2(ssa.OpGreaterEqualUint8x16, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint8x32.GreaterEqual", opLen2(ssa.OpGreaterEqualUint8x32, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint8x64.GreaterEqual", opLen2(ssa.OpGreaterEqualUint8x64, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint16x8.GreaterEqual", opLen2(ssa.OpGreaterEqualUint16x8, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint16x16.GreaterEqual", opLen2(ssa.OpGreaterEqualUint16x16, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint16x32.GreaterEqual", opLen2(ssa.OpGreaterEqualUint16x32, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint32x4.GreaterEqual", opLen2(ssa.OpGreaterEqualUint32x4, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint32x8.GreaterEqual", opLen2(ssa.OpGreaterEqualUint32x8, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint32x16.GreaterEqual", opLen2(ssa.OpGreaterEqualUint32x16, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint64x2.GreaterEqual", opLen2(ssa.OpGreaterEqualUint64x2, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint64x4.GreaterEqual", opLen2(ssa.OpGreaterEqualUint64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint64x8.GreaterEqual", opLen2(ssa.OpGreaterEqualUint64x8, types.TypeVec512), sys.AMD64)
 	addF(simdPackage, "Float32x4.GreaterEqualMasked", opLen3(ssa.OpGreaterEqualMaskedFloat32x4, types.TypeVec128), sys.AMD64)
 	addF(simdPackage, "Float32x8.GreaterEqualMasked", opLen3(ssa.OpGreaterEqualMaskedFloat32x8, types.TypeVec256), sys.AMD64)
@@ -722,29 +698,13 @@ func simdIntrinsics(addF func(pkg, fn string, b intrinsicBuilder, archFamilies .
 	addF(simdPackage, "Float64x2.Less", opLen2(ssa.OpLessFloat64x2, types.TypeVec128), sys.AMD64)
 	addF(simdPackage, "Float64x4.Less", opLen2(ssa.OpLessFloat64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Float64x8.Less", opLen2(ssa.OpLessFloat64x8, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int8x16.Less", opLen2(ssa.OpLessInt8x16, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int8x32.Less", opLen2(ssa.OpLessInt8x32, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int8x64.Less", opLen2(ssa.OpLessInt8x64, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int16x8.Less", opLen2(ssa.OpLessInt16x8, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int16x16.Less", opLen2(ssa.OpLessInt16x16, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int16x32.Less", opLen2(ssa.OpLessInt16x32, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int32x4.Less", opLen2(ssa.OpLessInt32x4, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int32x8.Less", opLen2(ssa.OpLessInt32x8, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int32x16.Less", opLen2(ssa.OpLessInt32x16, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int64x2.Less", opLen2(ssa.OpLessInt64x2, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int64x4.Less", opLen2(ssa.OpLessInt64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int64x8.Less", opLen2(ssa.OpLessInt64x8, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint8x16.Less", opLen2(ssa.OpLessUint8x16, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint8x32.Less", opLen2(ssa.OpLessUint8x32, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint8x64.Less", opLen2(ssa.OpLessUint8x64, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint16x8.Less", opLen2(ssa.OpLessUint16x8, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint16x16.Less", opLen2(ssa.OpLessUint16x16, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint16x32.Less", opLen2(ssa.OpLessUint16x32, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint32x4.Less", opLen2(ssa.OpLessUint32x4, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint32x8.Less", opLen2(ssa.OpLessUint32x8, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint32x16.Less", opLen2(ssa.OpLessUint32x16, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint64x2.Less", opLen2(ssa.OpLessUint64x2, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint64x4.Less", opLen2(ssa.OpLessUint64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint64x8.Less", opLen2(ssa.OpLessUint64x8, types.TypeVec512), sys.AMD64)
 	addF(simdPackage, "Float32x4.LessEqual", opLen2(ssa.OpLessEqualFloat32x4, types.TypeVec128), sys.AMD64)
 	addF(simdPackage, "Float32x8.LessEqual", opLen2(ssa.OpLessEqualFloat32x8, types.TypeVec256), sys.AMD64)
@@ -752,29 +712,13 @@ func simdIntrinsics(addF func(pkg, fn string, b intrinsicBuilder, archFamilies .
 	addF(simdPackage, "Float64x2.LessEqual", opLen2(ssa.OpLessEqualFloat64x2, types.TypeVec128), sys.AMD64)
 	addF(simdPackage, "Float64x4.LessEqual", opLen2(ssa.OpLessEqualFloat64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Float64x8.LessEqual", opLen2(ssa.OpLessEqualFloat64x8, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int8x16.LessEqual", opLen2(ssa.OpLessEqualInt8x16, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int8x32.LessEqual", opLen2(ssa.OpLessEqualInt8x32, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int8x64.LessEqual", opLen2(ssa.OpLessEqualInt8x64, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int16x8.LessEqual", opLen2(ssa.OpLessEqualInt16x8, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int16x16.LessEqual", opLen2(ssa.OpLessEqualInt16x16, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int16x32.LessEqual", opLen2(ssa.OpLessEqualInt16x32, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int32x4.LessEqual", opLen2(ssa.OpLessEqualInt32x4, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int32x8.LessEqual", opLen2(ssa.OpLessEqualInt32x8, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int32x16.LessEqual", opLen2(ssa.OpLessEqualInt32x16, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int64x2.LessEqual", opLen2(ssa.OpLessEqualInt64x2, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int64x4.LessEqual", opLen2(ssa.OpLessEqualInt64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int64x8.LessEqual", opLen2(ssa.OpLessEqualInt64x8, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint8x16.LessEqual", opLen2(ssa.OpLessEqualUint8x16, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint8x32.LessEqual", opLen2(ssa.OpLessEqualUint8x32, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint8x64.LessEqual", opLen2(ssa.OpLessEqualUint8x64, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint16x8.LessEqual", opLen2(ssa.OpLessEqualUint16x8, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint16x16.LessEqual", opLen2(ssa.OpLessEqualUint16x16, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint16x32.LessEqual", opLen2(ssa.OpLessEqualUint16x32, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint32x4.LessEqual", opLen2(ssa.OpLessEqualUint32x4, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint32x8.LessEqual", opLen2(ssa.OpLessEqualUint32x8, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint32x16.LessEqual", opLen2(ssa.OpLessEqualUint32x16, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint64x2.LessEqual", opLen2(ssa.OpLessEqualUint64x2, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint64x4.LessEqual", opLen2(ssa.OpLessEqualUint64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint64x8.LessEqual", opLen2(ssa.OpLessEqualUint64x8, types.TypeVec512), sys.AMD64)
 	addF(simdPackage, "Float32x4.LessEqualMasked", opLen3(ssa.OpLessEqualMaskedFloat32x4, types.TypeVec128), sys.AMD64)
 	addF(simdPackage, "Float32x8.LessEqualMasked", opLen3(ssa.OpLessEqualMaskedFloat32x8, types.TypeVec256), sys.AMD64)
@@ -1062,29 +1006,13 @@ func simdIntrinsics(addF func(pkg, fn string, b intrinsicBuilder, archFamilies .
 	addF(simdPackage, "Float64x2.NotEqual", opLen2(ssa.OpNotEqualFloat64x2, types.TypeVec128), sys.AMD64)
 	addF(simdPackage, "Float64x4.NotEqual", opLen2(ssa.OpNotEqualFloat64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Float64x8.NotEqual", opLen2(ssa.OpNotEqualFloat64x8, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int8x16.NotEqual", opLen2(ssa.OpNotEqualInt8x16, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int8x32.NotEqual", opLen2(ssa.OpNotEqualInt8x32, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int8x64.NotEqual", opLen2(ssa.OpNotEqualInt8x64, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int16x8.NotEqual", opLen2(ssa.OpNotEqualInt16x8, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int16x16.NotEqual", opLen2(ssa.OpNotEqualInt16x16, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int16x32.NotEqual", opLen2(ssa.OpNotEqualInt16x32, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int32x4.NotEqual", opLen2(ssa.OpNotEqualInt32x4, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int32x8.NotEqual", opLen2(ssa.OpNotEqualInt32x8, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int32x16.NotEqual", opLen2(ssa.OpNotEqualInt32x16, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Int64x2.NotEqual", opLen2(ssa.OpNotEqualInt64x2, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Int64x4.NotEqual", opLen2(ssa.OpNotEqualInt64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Int64x8.NotEqual", opLen2(ssa.OpNotEqualInt64x8, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint8x16.NotEqual", opLen2(ssa.OpNotEqualUint8x16, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint8x32.NotEqual", opLen2(ssa.OpNotEqualUint8x32, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint8x64.NotEqual", opLen2(ssa.OpNotEqualUint8x64, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint16x8.NotEqual", opLen2(ssa.OpNotEqualUint16x8, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint16x16.NotEqual", opLen2(ssa.OpNotEqualUint16x16, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint16x32.NotEqual", opLen2(ssa.OpNotEqualUint16x32, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint32x4.NotEqual", opLen2(ssa.OpNotEqualUint32x4, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint32x8.NotEqual", opLen2(ssa.OpNotEqualUint32x8, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint32x16.NotEqual", opLen2(ssa.OpNotEqualUint32x16, types.TypeVec512), sys.AMD64)
-	addF(simdPackage, "Uint64x2.NotEqual", opLen2(ssa.OpNotEqualUint64x2, types.TypeVec128), sys.AMD64)
-	addF(simdPackage, "Uint64x4.NotEqual", opLen2(ssa.OpNotEqualUint64x4, types.TypeVec256), sys.AMD64)
 	addF(simdPackage, "Uint64x8.NotEqual", opLen2(ssa.OpNotEqualUint64x8, types.TypeVec512), sys.AMD64)
 	addF(simdPackage, "Float32x4.NotEqualMasked", opLen3(ssa.OpNotEqualMaskedFloat32x4, types.TypeVec128), sys.AMD64)
 	addF(simdPackage, "Float32x8.NotEqualMasked", opLen3(ssa.OpNotEqualMaskedFloat32x8, types.TypeVec256), sys.AMD64)

src/simd/compare_test.go

@@ -59,17 +59,32 @@ func TestLess(t *testing.T) {
 	testFloat64x2Compare(t, simd.Float64x2.Less, lessSlice[float64])
 	testFloat64x4Compare(t, simd.Float64x4.Less, lessSlice[float64])
 
-	if comparisonFixed {
+	testInt16x16Compare(t, simd.Int16x16.Less, lessSlice[int16])
-		testInt16x16Compare(t, simd.Int16x16.Less, lessSlice[int16])
+	testInt16x8Compare(t, simd.Int16x8.Less, lessSlice[int16])
-		testInt16x8Compare(t, simd.Int16x8.Less, lessSlice[int16])
+	testInt32x4Compare(t, simd.Int32x4.Less, lessSlice[int32])
-		testInt32x4Compare(t, simd.Int32x4.Less, lessSlice[int32])
+	testInt32x8Compare(t, simd.Int32x8.Less, lessSlice[int32])
-		testInt32x8Compare(t, simd.Int32x8.Less, lessSlice[int32])
+	testInt64x2Compare(t, simd.Int64x2.Less, lessSlice[int64])
-		testInt64x2Compare(t, simd.Int64x2.Less, lessSlice[int64])
+	testInt64x4Compare(t, simd.Int64x4.Less, lessSlice[int64])
-		testInt64x4Compare(t, simd.Int64x4.Less, lessSlice[int64])
+	testInt8x16Compare(t, simd.Int8x16.Less, lessSlice[int8])
-		testInt8x16Compare(t, simd.Int8x16.Less, lessSlice[int8])
+	testInt8x32Compare(t, simd.Int8x32.Less, lessSlice[int8])
-		testInt8x32Compare(t, simd.Int8x32.Less, lessSlice[int8])
 
-	}
+	testInt16x16Compare(t, simd.Int16x16.Less, lessSlice[int16])
+	testInt16x8Compare(t, simd.Int16x8.Less, lessSlice[int16])
+	testInt32x4Compare(t, simd.Int32x4.Less, lessSlice[int32])
+	testInt32x8Compare(t, simd.Int32x8.Less, lessSlice[int32])
+	testInt64x2Compare(t, simd.Int64x2.Less, lessSlice[int64])
+	testInt64x4Compare(t, simd.Int64x4.Less, lessSlice[int64])
+	testInt8x16Compare(t, simd.Int8x16.Less, lessSlice[int8])
+	testInt8x32Compare(t, simd.Int8x32.Less, lessSlice[int8])
+
+	testUint16x16Compare(t, simd.Uint16x16.Less, lessSlice[uint16])
+	testUint16x8Compare(t, simd.Uint16x8.Less, lessSlice[uint16])
+	testUint32x4Compare(t, simd.Uint32x4.Less, lessSlice[uint32])
+	testUint32x8Compare(t, simd.Uint32x8.Less, lessSlice[uint32])
+	testUint64x2Compare(t, simd.Uint64x2.Less, lessSlice[uint64])
+	testUint64x4Compare(t, simd.Uint64x4.Less, lessSlice[uint64])
+	testUint8x16Compare(t, simd.Uint8x16.Less, lessSlice[uint8])
+	testUint8x32Compare(t, simd.Uint8x32.Less, lessSlice[uint8])
 
 	if simd.HasAVX512() {
 		testUint16x16Compare(t, simd.Uint16x16.Less, lessSlice[uint16])
@@ -100,28 +115,25 @@ func TestLessEqual(t *testing.T) {
 	testFloat64x2Compare(t, simd.Float64x2.LessEqual, lessEqualSlice[float64])
 	testFloat64x4Compare(t, simd.Float64x4.LessEqual, lessEqualSlice[float64])
 
-	if comparisonFixed {
+	testInt16x16Compare(t, simd.Int16x16.LessEqual, lessEqualSlice[int16])
-		testInt16x16Compare(t, simd.Int16x16.LessEqual, lessEqualSlice[int16])
+	testInt16x8Compare(t, simd.Int16x8.LessEqual, lessEqualSlice[int16])
-		testInt16x8Compare(t, simd.Int16x8.LessEqual, lessEqualSlice[int16])
+	testInt32x4Compare(t, simd.Int32x4.LessEqual, lessEqualSlice[int32])
-		testInt32x4Compare(t, simd.Int32x4.LessEqual, lessEqualSlice[int32])
+	testInt32x8Compare(t, simd.Int32x8.LessEqual, lessEqualSlice[int32])
-		testInt32x8Compare(t, simd.Int32x8.LessEqual, lessEqualSlice[int32])
+	testInt64x2Compare(t, simd.Int64x2.LessEqual, lessEqualSlice[int64])
-		testInt64x2Compare(t, simd.Int64x2.LessEqual, lessEqualSlice[int64])
+	testInt64x4Compare(t, simd.Int64x4.LessEqual, lessEqualSlice[int64])
-		testInt64x4Compare(t, simd.Int64x4.LessEqual, lessEqualSlice[int64])
+	testInt8x16Compare(t, simd.Int8x16.LessEqual, lessEqualSlice[int8])
-		testInt8x16Compare(t, simd.Int8x16.LessEqual, lessEqualSlice[int8])
+	testInt8x32Compare(t, simd.Int8x32.LessEqual, lessEqualSlice[int8])
-		testInt8x32Compare(t, simd.Int8x32.LessEqual, lessEqualSlice[int8])
 
-	}
+	testUint16x16Compare(t, simd.Uint16x16.LessEqual, lessEqualSlice[uint16])
+	testUint16x8Compare(t, simd.Uint16x8.LessEqual, lessEqualSlice[uint16])
+	testUint32x4Compare(t, simd.Uint32x4.LessEqual, lessEqualSlice[uint32])
+	testUint32x8Compare(t, simd.Uint32x8.LessEqual, lessEqualSlice[uint32])
+	testUint64x2Compare(t, simd.Uint64x2.LessEqual, lessEqualSlice[uint64])
+	testUint64x4Compare(t, simd.Uint64x4.LessEqual, lessEqualSlice[uint64])
+	testUint8x16Compare(t, simd.Uint8x16.LessEqual, lessEqualSlice[uint8])
+	testUint8x32Compare(t, simd.Uint8x32.LessEqual, lessEqualSlice[uint8])
 
 	if simd.HasAVX512() {
-		testUint16x16Compare(t, simd.Uint16x16.LessEqual, lessEqualSlice[uint16])
-		testUint16x8Compare(t, simd.Uint16x8.LessEqual, lessEqualSlice[uint16])
-		testUint32x4Compare(t, simd.Uint32x4.LessEqual, lessEqualSlice[uint32])
-		testUint32x8Compare(t, simd.Uint32x8.LessEqual, lessEqualSlice[uint32])
-		testUint64x2Compare(t, simd.Uint64x2.LessEqual, lessEqualSlice[uint64])
-		testUint64x4Compare(t, simd.Uint64x4.LessEqual, lessEqualSlice[uint64])
-		testUint8x16Compare(t, simd.Uint8x16.LessEqual, lessEqualSlice[uint8])
-		testUint8x32Compare(t, simd.Uint8x32.LessEqual, lessEqualSlice[uint8])
-
 		testFloat32x16Compare(t, simd.Float32x16.LessEqual, lessEqualSlice[float32])
 		testFloat64x8Compare(t, simd.Float64x8.LessEqual, lessEqualSlice[float64])
 		testInt8x64Compare(t, simd.Int8x64.LessEqual, lessEqualSlice[int8])
@@ -151,16 +163,17 @@ func TestGreater(t *testing.T) {
 	testInt8x16Compare(t, simd.Int8x16.Greater, greaterSlice[int8])
 	testInt8x32Compare(t, simd.Int8x32.Greater, greaterSlice[int8])
 
-	if simd.HasAVX512() {
+	testUint16x16Compare(t, simd.Uint16x16.Greater, greaterSlice[uint16])
-		testUint16x16Compare(t, simd.Uint16x16.Greater, greaterSlice[uint16])
+	testUint16x8Compare(t, simd.Uint16x8.Greater, greaterSlice[uint16])
-		testUint16x8Compare(t, simd.Uint16x8.Greater, greaterSlice[uint16])
+	testUint32x4Compare(t, simd.Uint32x4.Greater, greaterSlice[uint32])
-		testUint32x4Compare(t, simd.Uint32x4.Greater, greaterSlice[uint32])
+	testUint32x8Compare(t, simd.Uint32x8.Greater, greaterSlice[uint32])
-		testUint32x8Compare(t, simd.Uint32x8.Greater, greaterSlice[uint32])
 
-		testUint64x2Compare(t, simd.Uint64x2.Greater, greaterSlice[uint64])
+	testUint64x2Compare(t, simd.Uint64x2.Greater, greaterSlice[uint64])
-		testUint64x4Compare(t, simd.Uint64x4.Greater, greaterSlice[uint64])
+	testUint64x4Compare(t, simd.Uint64x4.Greater, greaterSlice[uint64])
-		testUint8x16Compare(t, simd.Uint8x16.Greater, greaterSlice[uint8])
+	testUint8x16Compare(t, simd.Uint8x16.Greater, greaterSlice[uint8])
-		testUint8x32Compare(t, simd.Uint8x32.Greater, greaterSlice[uint8])
+	testUint8x32Compare(t, simd.Uint8x32.Greater, greaterSlice[uint8])
+
+	if simd.HasAVX512() {
 
 		testFloat32x16Compare(t, simd.Float32x16.Greater, greaterSlice[float32])
 		testFloat64x8Compare(t, simd.Float64x8.Greater, greaterSlice[float64])
@@ -181,28 +194,25 @@ func TestGreaterEqual(t *testing.T) {
 	testFloat64x2Compare(t, simd.Float64x2.GreaterEqual, greaterEqualSlice[float64])
 	testFloat64x4Compare(t, simd.Float64x4.GreaterEqual, greaterEqualSlice[float64])
 
-	if comparisonFixed {
+	testInt16x16Compare(t, simd.Int16x16.GreaterEqual, greaterEqualSlice[int16])
-		testInt16x16Compare(t, simd.Int16x16.GreaterEqual, greaterEqualSlice[int16])
+	testInt16x8Compare(t, simd.Int16x8.GreaterEqual, greaterEqualSlice[int16])
-		testInt16x8Compare(t, simd.Int16x8.GreaterEqual, greaterEqualSlice[int16])
+	testInt32x4Compare(t, simd.Int32x4.GreaterEqual, greaterEqualSlice[int32])
-		testInt32x4Compare(t, simd.Int32x4.GreaterEqual, greaterEqualSlice[int32])
+	testInt32x8Compare(t, simd.Int32x8.GreaterEqual, greaterEqualSlice[int32])
-		testInt32x8Compare(t, simd.Int32x8.GreaterEqual, greaterEqualSlice[int32])
+	testInt64x2Compare(t, simd.Int64x2.GreaterEqual, greaterEqualSlice[int64])
-		testInt64x2Compare(t, simd.Int64x2.GreaterEqual, greaterEqualSlice[int64])
+	testInt64x4Compare(t, simd.Int64x4.GreaterEqual, greaterEqualSlice[int64])
-		testInt64x4Compare(t, simd.Int64x4.GreaterEqual, greaterEqualSlice[int64])
+	testInt8x16Compare(t, simd.Int8x16.GreaterEqual, greaterEqualSlice[int8])
-		testInt8x16Compare(t, simd.Int8x16.GreaterEqual, greaterEqualSlice[int8])
+	testInt8x32Compare(t, simd.Int8x32.GreaterEqual, greaterEqualSlice[int8])
-		testInt8x32Compare(t, simd.Int8x32.GreaterEqual, greaterEqualSlice[int8])
 
-	}
+	testUint16x16Compare(t, simd.Uint16x16.GreaterEqual, greaterEqualSlice[uint16])
+	testUint16x8Compare(t, simd.Uint16x8.GreaterEqual, greaterEqualSlice[uint16])
+	testUint32x4Compare(t, simd.Uint32x4.GreaterEqual, greaterEqualSlice[uint32])
+	testUint32x8Compare(t, simd.Uint32x8.GreaterEqual, greaterEqualSlice[uint32])
+	testUint64x2Compare(t, simd.Uint64x2.GreaterEqual, greaterEqualSlice[uint64])
+	testUint64x4Compare(t, simd.Uint64x4.GreaterEqual, greaterEqualSlice[uint64])
+	testUint8x16Compare(t, simd.Uint8x16.GreaterEqual, greaterEqualSlice[uint8])
+	testUint8x32Compare(t, simd.Uint8x32.GreaterEqual, greaterEqualSlice[uint8])
 
 	if simd.HasAVX512() {
-		testUint16x16Compare(t, simd.Uint16x16.GreaterEqual, greaterEqualSlice[uint16])
-		testUint16x8Compare(t, simd.Uint16x8.GreaterEqual, greaterEqualSlice[uint16])
-		testUint32x4Compare(t, simd.Uint32x4.GreaterEqual, greaterEqualSlice[uint32])
-		testUint32x8Compare(t, simd.Uint32x8.GreaterEqual, greaterEqualSlice[uint32])
-		testUint64x2Compare(t, simd.Uint64x2.GreaterEqual, greaterEqualSlice[uint64])
-		testUint64x4Compare(t, simd.Uint64x4.GreaterEqual, greaterEqualSlice[uint64])
-		testUint8x16Compare(t, simd.Uint8x16.GreaterEqual, greaterEqualSlice[uint8])
-		testUint8x32Compare(t, simd.Uint8x32.GreaterEqual, greaterEqualSlice[uint8])
-
 		testFloat32x16Compare(t, simd.Float32x16.GreaterEqual, greaterEqualSlice[float32])
 		testFloat64x8Compare(t, simd.Float64x8.GreaterEqual, greaterEqualSlice[float64])
 		testInt8x64Compare(t, simd.Int8x64.GreaterEqual, greaterEqualSlice[int8])
@@ -260,25 +270,23 @@ func TestNotEqual(t *testing.T) {
 	testFloat64x2Compare(t, simd.Float64x2.NotEqual, notEqualSlice[float64])
 	testFloat64x4Compare(t, simd.Float64x4.NotEqual, notEqualSlice[float64])
 
-	if comparisonFixed {
+	testInt16x16Compare(t, simd.Int16x16.NotEqual, notEqualSlice[int16])
-		testInt16x16Compare(t, simd.Int16x16.NotEqual, notEqualSlice[int16])
+	testInt16x8Compare(t, simd.Int16x8.NotEqual, notEqualSlice[int16])
-		testInt16x8Compare(t, simd.Int16x8.NotEqual, notEqualSlice[int16])
+	testInt32x4Compare(t, simd.Int32x4.NotEqual, notEqualSlice[int32])
-		testInt32x4Compare(t, simd.Int32x4.NotEqual, notEqualSlice[int32])
+	testInt32x8Compare(t, simd.Int32x8.NotEqual, notEqualSlice[int32])
-		testInt32x8Compare(t, simd.Int32x8.NotEqual, notEqualSlice[int32])
+	testInt64x2Compare(t, simd.Int64x2.NotEqual, notEqualSlice[int64])
-		testInt64x2Compare(t, simd.Int64x2.NotEqual, notEqualSlice[int64])
+	testInt64x4Compare(t, simd.Int64x4.NotEqual, notEqualSlice[int64])
-		testInt64x4Compare(t, simd.Int64x4.NotEqual, notEqualSlice[int64])
+	testInt8x16Compare(t, simd.Int8x16.NotEqual, notEqualSlice[int8])
-		testInt8x16Compare(t, simd.Int8x16.NotEqual, notEqualSlice[int8])
+	testInt8x32Compare(t, simd.Int8x32.NotEqual, notEqualSlice[int8])
-		testInt8x32Compare(t, simd.Int8x32.NotEqual, notEqualSlice[int8])
 
-		testUint16x16Compare(t, simd.Uint16x16.NotEqual, notEqualSlice[uint16])
+	testUint16x16Compare(t, simd.Uint16x16.NotEqual, notEqualSlice[uint16])
-		testUint16x8Compare(t, simd.Uint16x8.NotEqual, notEqualSlice[uint16])
+	testUint16x8Compare(t, simd.Uint16x8.NotEqual, notEqualSlice[uint16])
-		testUint32x4Compare(t, simd.Uint32x4.NotEqual, notEqualSlice[uint32])
+	testUint32x4Compare(t, simd.Uint32x4.NotEqual, notEqualSlice[uint32])
-		testUint32x8Compare(t, simd.Uint32x8.NotEqual, notEqualSlice[uint32])
+	testUint32x8Compare(t, simd.Uint32x8.NotEqual, notEqualSlice[uint32])
-		testUint64x2Compare(t, simd.Uint64x2.NotEqual, notEqualSlice[uint64])
+	testUint64x2Compare(t, simd.Uint64x2.NotEqual, notEqualSlice[uint64])
-		testUint64x4Compare(t, simd.Uint64x4.NotEqual, notEqualSlice[uint64])
+	testUint64x4Compare(t, simd.Uint64x4.NotEqual, notEqualSlice[uint64])
-		testUint8x16Compare(t, simd.Uint8x16.NotEqual, notEqualSlice[uint8])
+	testUint8x16Compare(t, simd.Uint8x16.NotEqual, notEqualSlice[uint8])
-		testUint8x32Compare(t, simd.Uint8x32.NotEqual, notEqualSlice[uint8])
+	testUint8x32Compare(t, simd.Uint8x32.NotEqual, notEqualSlice[uint8])
-	}
 
 	if simd.HasAVX512() {
 		testFloat32x16Compare(t, simd.Float32x16.NotEqual, notEqualSlice[float32])

src/simd/genfiles.go

@@ -87,6 +87,16 @@ var ternaryFlaky = &shapes{ // for tests that support flaky equality
 	floats: []int{32},
 }
 
+var avx2SignedComparisons = &shapes{
+	vecs: []int{128, 256},
+	ints: []int{8, 16, 32, 64},
+}
+
+var avx2UnsignedComparisons = &shapes{
+	vecs:  []int{128, 256},
+	uints: []int{8, 16, 32, 64},
+}
+
 type templateData struct {
 	Vec    string // the type of the vector, e.g. Float32x4
 	AOrAn  string // for documentation, the article "a" or "an"
@@ -486,6 +496,130 @@ func (x {{.Vec}}) StoreSlicePart(s []{{.Type}}) {
 }
 `)
 
+func (t templateData) CPUfeature() string {
+	switch t.Vwidth {
+	case 128:
+		return "AVX"
+	case 256:
+		return "AVX2"
+	case 512:
+		return "AVX512"
+	}
+	panic(fmt.Errorf("unexpected vector width %d", t.Vwidth))
+}
+
+var avx2SignedComparisonsTemplate = shapedTemplateOf(avx2SignedComparisons, "avx2 signed comparisons", `
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature {{.CPUfeature}}
+func (x {{.Vec}}) Less(y {{.Vec}}) Mask{{.WxC}} {
+	return y.Greater(x)
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature {{.CPUfeature}}
+func (x {{.Vec}}) GreaterEqual(y {{.Vec}}) Mask{{.WxC}} {
+	ones := x.Equal(x).AsInt{{.WxC}}()
+	return y.Greater(x).AsInt{{.WxC}}().Xor(ones).AsMask{{.WxC}}()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature {{.CPUfeature}}
+func (x {{.Vec}}) LessEqual(y {{.Vec}}) Mask{{.WxC}} {
+	ones := x.Equal(x).AsInt{{.WxC}}()
+	return x.Greater(y).AsInt{{.WxC}}().Xor(ones).AsMask{{.WxC}}()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature {{.CPUfeature}}
+func (x {{.Vec}}) NotEqual(y {{.Vec}}) Mask{{.WxC}} {
+	ones := x.Equal(x).AsInt{{.WxC}}()
+	return x.Equal(y).AsInt{{.WxC}}().Xor(ones).AsMask{{.WxC}}()	
+}
+`)
+
+// CPUfeatureAVX2if8 return AVX2 if the element width is 8,
+// otherwise, it returns CPUfeature.  This is for the cpufeature
+// of unsigned comparison emulation, which uses shifts for all
+// the sizes > 8 (shifts are AVX) but must use broadcast (AVX2)
+// for bytes.
+func (t templateData) CPUfeatureAVX2if8() string {
+	if t.Width == 8 {
+		return "AVX2"
+	}
+	return t.CPUfeature()
+}
+
+var avx2UnsignedComparisonsTemplate = shapedTemplateOf(avx2UnsignedComparisons, "avx2 unsigned comparisons", `
+// Greater returns a mask whose elements indicate whether x > y
+//
+// Emulated, CPU Feature {{.CPUfeatureAVX2if8}}
+func (x {{.Vec}}) Greater(y {{.Vec}}) Mask{{.WxC}} {
+	a, b := x.AsInt{{.WxC}}(), y.AsInt{{.WxC}}()
+{{- if eq .Width 8}}
+	signs := BroadcastInt{{.WxC}}(-1 << ({{.Width}}-1))
+{{- else}}
+	ones := x.Equal(x).AsInt{{.WxC}}()
+	signs := ones.ShiftAllLeft({{.Width}}-1)
+{{- end }}
+	return a.Xor(signs).Greater(b.Xor(signs))
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature {{.CPUfeatureAVX2if8}}
+func (x {{.Vec}}) Less(y {{.Vec}}) Mask{{.WxC}} {
+	a, b := x.AsInt{{.WxC}}(), y.AsInt{{.WxC}}()
+{{- if eq .Width 8}}
+	signs := BroadcastInt{{.WxC}}(-1 << ({{.Width}}-1))
+{{- else}}
+	ones := x.Equal(x).AsInt{{.WxC}}()
+	signs := ones.ShiftAllLeft({{.Width}}-1)
+{{- end }}
+	return b.Xor(signs).Greater(a.Xor(signs))
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature {{.CPUfeatureAVX2if8}}
+func (x {{.Vec}}) GreaterEqual(y {{.Vec}}) Mask{{.WxC}} {
+	a, b := x.AsInt{{.WxC}}(), y.AsInt{{.WxC}}()
+	ones := x.Equal(x).AsInt{{.WxC}}()
+{{- if eq .Width 8}}
+	signs := BroadcastInt{{.WxC}}(-1 << ({{.Width}}-1))
+{{- else}}
+	signs := ones.ShiftAllLeft({{.Width}}-1)
+{{- end }}
+	return b.Xor(signs).Greater(a.Xor(signs)).AsInt{{.WxC}}().Xor(ones).AsMask{{.WxC}}()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature {{.CPUfeatureAVX2if8}}
+func (x {{.Vec}}) LessEqual(y {{.Vec}}) Mask{{.WxC}} {
+	a, b := x.AsInt{{.WxC}}(), y.AsInt{{.WxC}}()
+	ones := x.Equal(x).AsInt{{.WxC}}()
+{{- if eq .Width 8}}
+	signs := BroadcastInt{{.WxC}}(-1 << ({{.Width}}-1))
+{{- else}}
+	signs := ones.ShiftAllLeft({{.Width}}-1)
+{{- end }}
+	return a.Xor(signs).Greater(b.Xor(signs)).AsInt{{.WxC}}().Xor(ones).AsMask{{.WxC}}()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature {{.CPUfeature}}
+func (x {{.Vec}}) NotEqual(y {{.Vec}}) Mask{{.WxC}} {
+	a, b := x.AsInt{{.WxC}}(), y.AsInt{{.WxC}}()
+	ones := x.Equal(x).AsInt{{.WxC}}()
+	return a.Equal(b).AsInt{{.WxC}}().Xor(ones).AsMask{{.WxC}}()
+}
+`)
+
 var unsafePATemplate = templateOf("unsafe PA helper", `
 // pa{{.Vec}} returns a type-unsafe pointer to array that can
 // only be used with partial load/store operations that only
@@ -591,6 +725,8 @@ func main() {
 			avx2SmallLoadSlicePartTemplate,
 			avx2MaskedTemplate,
 			avx512MaskedTemplate,
+			avx2SignedComparisonsTemplate,
+			avx2UnsignedComparisonsTemplate,
 			broadcastTemplate,
 		)
 	}

src/simd/ops_amd64.go

@@ -3822,61 +3822,21 @@ func (x Float64x4) Greater(y Float64x4) Mask64x4
 // Asm: VCMPPD, CPU Feature: AVX512
 func (x Float64x8) Greater(y Float64x8) Mask64x8
 
-// Greater compares for greater than.
-//
-// Asm: VPCMPUB, CPU Feature: AVX512
-func (x Uint8x16) Greater(y Uint8x16) Mask8x16
-
-// Greater compares for greater than.
-//
-// Asm: VPCMPUB, CPU Feature: AVX512
-func (x Uint8x32) Greater(y Uint8x32) Mask8x32
-
 // Greater compares for greater than.
 //
 // Asm: VPCMPUB, CPU Feature: AVX512
 func (x Uint8x64) Greater(y Uint8x64) Mask8x64
 
-// Greater compares for greater than.
-//
-// Asm: VPCMPUW, CPU Feature: AVX512
-func (x Uint16x8) Greater(y Uint16x8) Mask16x8
-
-// Greater compares for greater than.
-//
-// Asm: VPCMPUW, CPU Feature: AVX512
-func (x Uint16x16) Greater(y Uint16x16) Mask16x16
-
 // Greater compares for greater than.
 //
 // Asm: VPCMPUW, CPU Feature: AVX512
 func (x Uint16x32) Greater(y Uint16x32) Mask16x32
 
-// Greater compares for greater than.
-//
-// Asm: VPCMPUD, CPU Feature: AVX512
-func (x Uint32x4) Greater(y Uint32x4) Mask32x4
-
-// Greater compares for greater than.
-//
-// Asm: VPCMPUD, CPU Feature: AVX512
-func (x Uint32x8) Greater(y Uint32x8) Mask32x8
-
 // Greater compares for greater than.
 //
 // Asm: VPCMPUD, CPU Feature: AVX512
 func (x Uint32x16) Greater(y Uint32x16) Mask32x16
 
-// Greater compares for greater than.
-//
-// Asm: VPCMPUQ, CPU Feature: AVX512
-func (x Uint64x2) Greater(y Uint64x2) Mask64x2
-
-// Greater compares for greater than.
-//
-// Asm: VPCMPUQ, CPU Feature: AVX512
-func (x Uint64x4) Greater(y Uint64x4) Mask64x4
-
 // Greater compares for greater than.
 //
 // Asm: VPCMPUQ, CPU Feature: AVX512
@@ -3914,121 +3874,41 @@ func (x Float64x4) GreaterEqual(y Float64x4) Mask64x4
 // Asm: VCMPPD, CPU Feature: AVX512
 func (x Float64x8) GreaterEqual(y Float64x8) Mask64x8
 
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPB, CPU Feature: AVX512
-func (x Int8x16) GreaterEqual(y Int8x16) Mask8x16
-
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPB, CPU Feature: AVX512
-func (x Int8x32) GreaterEqual(y Int8x32) Mask8x32
-
 // GreaterEqual compares for greater than or equal.
 //
 // Asm: VPCMPB, CPU Feature: AVX512
 func (x Int8x64) GreaterEqual(y Int8x64) Mask8x64
 
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPW, CPU Feature: AVX512
-func (x Int16x8) GreaterEqual(y Int16x8) Mask16x8
-
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPW, CPU Feature: AVX512
-func (x Int16x16) GreaterEqual(y Int16x16) Mask16x16
-
 // GreaterEqual compares for greater than or equal.
 //
 // Asm: VPCMPW, CPU Feature: AVX512
 func (x Int16x32) GreaterEqual(y Int16x32) Mask16x32
 
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPD, CPU Feature: AVX512
-func (x Int32x4) GreaterEqual(y Int32x4) Mask32x4
-
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPD, CPU Feature: AVX512
-func (x Int32x8) GreaterEqual(y Int32x8) Mask32x8
-
 // GreaterEqual compares for greater than or equal.
 //
 // Asm: VPCMPD, CPU Feature: AVX512
 func (x Int32x16) GreaterEqual(y Int32x16) Mask32x16
 
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPQ, CPU Feature: AVX512
-func (x Int64x2) GreaterEqual(y Int64x2) Mask64x2
-
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPQ, CPU Feature: AVX512
-func (x Int64x4) GreaterEqual(y Int64x4) Mask64x4
-
 // GreaterEqual compares for greater than or equal.
 //
 // Asm: VPCMPQ, CPU Feature: AVX512
 func (x Int64x8) GreaterEqual(y Int64x8) Mask64x8
 
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPUB, CPU Feature: AVX512
-func (x Uint8x16) GreaterEqual(y Uint8x16) Mask8x16
-
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPUB, CPU Feature: AVX512
-func (x Uint8x32) GreaterEqual(y Uint8x32) Mask8x32
-
 // GreaterEqual compares for greater than or equal.
 //
 // Asm: VPCMPUB, CPU Feature: AVX512
 func (x Uint8x64) GreaterEqual(y Uint8x64) Mask8x64
 
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPUW, CPU Feature: AVX512
-func (x Uint16x8) GreaterEqual(y Uint16x8) Mask16x8
-
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPUW, CPU Feature: AVX512
-func (x Uint16x16) GreaterEqual(y Uint16x16) Mask16x16
-
 // GreaterEqual compares for greater than or equal.
 //
 // Asm: VPCMPUW, CPU Feature: AVX512
 func (x Uint16x32) GreaterEqual(y Uint16x32) Mask16x32
 
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPUD, CPU Feature: AVX512
-func (x Uint32x4) GreaterEqual(y Uint32x4) Mask32x4
-
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPUD, CPU Feature: AVX512
-func (x Uint32x8) GreaterEqual(y Uint32x8) Mask32x8
-
 // GreaterEqual compares for greater than or equal.
 //
 // Asm: VPCMPUD, CPU Feature: AVX512
 func (x Uint32x16) GreaterEqual(y Uint32x16) Mask32x16
 
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPUQ, CPU Feature: AVX512
-func (x Uint64x2) GreaterEqual(y Uint64x2) Mask64x2
-
-// GreaterEqual compares for greater than or equal.
-//
-// Asm: VPCMPUQ, CPU Feature: AVX512
-func (x Uint64x4) GreaterEqual(y Uint64x4) Mask64x4
-
 // GreaterEqual compares for greater than or equal.
 //
 // Asm: VPCMPUQ, CPU Feature: AVX512
@@ -4566,121 +4446,41 @@ func (x Float64x4) Less(y Float64x4) Mask64x4
 // Asm: VCMPPD, CPU Feature: AVX512
 func (x Float64x8) Less(y Float64x8) Mask64x8
 
-// Less compares for less than.
-//
-// Asm: VPCMPB, CPU Feature: AVX512
-func (x Int8x16) Less(y Int8x16) Mask8x16
-
-// Less compares for less than.
-//
-// Asm: VPCMPB, CPU Feature: AVX512
-func (x Int8x32) Less(y Int8x32) Mask8x32
-
 // Less compares for less than.
 //
 // Asm: VPCMPB, CPU Feature: AVX512
 func (x Int8x64) Less(y Int8x64) Mask8x64
 
-// Less compares for less than.
-//
-// Asm: VPCMPW, CPU Feature: AVX512
-func (x Int16x8) Less(y Int16x8) Mask16x8
-
-// Less compares for less than.
-//
-// Asm: VPCMPW, CPU Feature: AVX512
-func (x Int16x16) Less(y Int16x16) Mask16x16
-
 // Less compares for less than.
 //
 // Asm: VPCMPW, CPU Feature: AVX512
 func (x Int16x32) Less(y Int16x32) Mask16x32
 
-// Less compares for less than.
-//
-// Asm: VPCMPD, CPU Feature: AVX512
-func (x Int32x4) Less(y Int32x4) Mask32x4
-
-// Less compares for less than.
-//
-// Asm: VPCMPD, CPU Feature: AVX512
-func (x Int32x8) Less(y Int32x8) Mask32x8
-
 // Less compares for less than.
 //
 // Asm: VPCMPD, CPU Feature: AVX512
 func (x Int32x16) Less(y Int32x16) Mask32x16
 
-// Less compares for less than.
-//
-// Asm: VPCMPQ, CPU Feature: AVX512
-func (x Int64x2) Less(y Int64x2) Mask64x2
-
-// Less compares for less than.
-//
-// Asm: VPCMPQ, CPU Feature: AVX512
-func (x Int64x4) Less(y Int64x4) Mask64x4
-
 // Less compares for less than.
 //
 // Asm: VPCMPQ, CPU Feature: AVX512
 func (x Int64x8) Less(y Int64x8) Mask64x8
 
-// Less compares for less than.
-//
-// Asm: VPCMPUB, CPU Feature: AVX512
-func (x Uint8x16) Less(y Uint8x16) Mask8x16
-
-// Less compares for less than.
-//
-// Asm: VPCMPUB, CPU Feature: AVX512
-func (x Uint8x32) Less(y Uint8x32) Mask8x32
-
 // Less compares for less than.
 //
 // Asm: VPCMPUB, CPU Feature: AVX512
 func (x Uint8x64) Less(y Uint8x64) Mask8x64
 
-// Less compares for less than.
-//
-// Asm: VPCMPUW, CPU Feature: AVX512
-func (x Uint16x8) Less(y Uint16x8) Mask16x8
-
-// Less compares for less than.
-//
-// Asm: VPCMPUW, CPU Feature: AVX512
-func (x Uint16x16) Less(y Uint16x16) Mask16x16
-
 // Less compares for less than.
 //
 // Asm: VPCMPUW, CPU Feature: AVX512
 func (x Uint16x32) Less(y Uint16x32) Mask16x32
 
-// Less compares for less than.
-//
-// Asm: VPCMPUD, CPU Feature: AVX512
-func (x Uint32x4) Less(y Uint32x4) Mask32x4
-
-// Less compares for less than.
-//
-// Asm: VPCMPUD, CPU Feature: AVX512
-func (x Uint32x8) Less(y Uint32x8) Mask32x8
-
 // Less compares for less than.
 //
 // Asm: VPCMPUD, CPU Feature: AVX512
 func (x Uint32x16) Less(y Uint32x16) Mask32x16
 
-// Less compares for less than.
-//
-// Asm: VPCMPUQ, CPU Feature: AVX512
-func (x Uint64x2) Less(y Uint64x2) Mask64x2
-
-// Less compares for less than.
-//
-// Asm: VPCMPUQ, CPU Feature: AVX512
-func (x Uint64x4) Less(y Uint64x4) Mask64x4
-
 // Less compares for less than.
 //
 // Asm: VPCMPUQ, CPU Feature: AVX512
@@ -4718,121 +4518,41 @@ func (x Float64x4) LessEqual(y Float64x4) Mask64x4
 // Asm: VCMPPD, CPU Feature: AVX512
 func (x Float64x8) LessEqual(y Float64x8) Mask64x8
 
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPB, CPU Feature: AVX512
-func (x Int8x16) LessEqual(y Int8x16) Mask8x16
-
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPB, CPU Feature: AVX512
-func (x Int8x32) LessEqual(y Int8x32) Mask8x32
-
 // LessEqual compares for less than or equal.
 //
 // Asm: VPCMPB, CPU Feature: AVX512
 func (x Int8x64) LessEqual(y Int8x64) Mask8x64
 
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPW, CPU Feature: AVX512
-func (x Int16x8) LessEqual(y Int16x8) Mask16x8
-
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPW, CPU Feature: AVX512
-func (x Int16x16) LessEqual(y Int16x16) Mask16x16
-
 // LessEqual compares for less than or equal.
 //
 // Asm: VPCMPW, CPU Feature: AVX512
 func (x Int16x32) LessEqual(y Int16x32) Mask16x32
 
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPD, CPU Feature: AVX512
-func (x Int32x4) LessEqual(y Int32x4) Mask32x4
-
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPD, CPU Feature: AVX512
-func (x Int32x8) LessEqual(y Int32x8) Mask32x8
-
 // LessEqual compares for less than or equal.
 //
 // Asm: VPCMPD, CPU Feature: AVX512
 func (x Int32x16) LessEqual(y Int32x16) Mask32x16
 
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPQ, CPU Feature: AVX512
-func (x Int64x2) LessEqual(y Int64x2) Mask64x2
-
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPQ, CPU Feature: AVX512
-func (x Int64x4) LessEqual(y Int64x4) Mask64x4
-
 // LessEqual compares for less than or equal.
 //
 // Asm: VPCMPQ, CPU Feature: AVX512
 func (x Int64x8) LessEqual(y Int64x8) Mask64x8
 
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPUB, CPU Feature: AVX512
-func (x Uint8x16) LessEqual(y Uint8x16) Mask8x16
-
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPUB, CPU Feature: AVX512
-func (x Uint8x32) LessEqual(y Uint8x32) Mask8x32
-
 // LessEqual compares for less than or equal.
 //
 // Asm: VPCMPUB, CPU Feature: AVX512
 func (x Uint8x64) LessEqual(y Uint8x64) Mask8x64
 
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPUW, CPU Feature: AVX512
-func (x Uint16x8) LessEqual(y Uint16x8) Mask16x8
-
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPUW, CPU Feature: AVX512
-func (x Uint16x16) LessEqual(y Uint16x16) Mask16x16
-
 // LessEqual compares for less than or equal.
 //
 // Asm: VPCMPUW, CPU Feature: AVX512
 func (x Uint16x32) LessEqual(y Uint16x32) Mask16x32
 
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPUD, CPU Feature: AVX512
-func (x Uint32x4) LessEqual(y Uint32x4) Mask32x4
-
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPUD, CPU Feature: AVX512
-func (x Uint32x8) LessEqual(y Uint32x8) Mask32x8
-
 // LessEqual compares for less than or equal.
 //
 // Asm: VPCMPUD, CPU Feature: AVX512
 func (x Uint32x16) LessEqual(y Uint32x16) Mask32x16
 
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPUQ, CPU Feature: AVX512
-func (x Uint64x2) LessEqual(y Uint64x2) Mask64x2
-
-// LessEqual compares for less than or equal.
-//
-// Asm: VPCMPUQ, CPU Feature: AVX512
-func (x Uint64x4) LessEqual(y Uint64x4) Mask64x4
-
 // LessEqual compares for less than or equal.
 //
 // Asm: VPCMPUQ, CPU Feature: AVX512
@@ -6644,121 +6364,41 @@ func (x Float64x4) NotEqual(y Float64x4) Mask64x4
 // Asm: VCMPPD, CPU Feature: AVX512
 func (x Float64x8) NotEqual(y Float64x8) Mask64x8
 
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPB, CPU Feature: AVX512
-func (x Int8x16) NotEqual(y Int8x16) Mask8x16
-
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPB, CPU Feature: AVX512
-func (x Int8x32) NotEqual(y Int8x32) Mask8x32
-
 // NotEqual compares for inequality.
 //
 // Asm: VPCMPB, CPU Feature: AVX512
 func (x Int8x64) NotEqual(y Int8x64) Mask8x64
 
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPW, CPU Feature: AVX512
-func (x Int16x8) NotEqual(y Int16x8) Mask16x8
-
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPW, CPU Feature: AVX512
-func (x Int16x16) NotEqual(y Int16x16) Mask16x16
-
 // NotEqual compares for inequality.
 //
 // Asm: VPCMPW, CPU Feature: AVX512
 func (x Int16x32) NotEqual(y Int16x32) Mask16x32
 
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPD, CPU Feature: AVX512
-func (x Int32x4) NotEqual(y Int32x4) Mask32x4
-
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPD, CPU Feature: AVX512
-func (x Int32x8) NotEqual(y Int32x8) Mask32x8
-
 // NotEqual compares for inequality.
 //
 // Asm: VPCMPD, CPU Feature: AVX512
 func (x Int32x16) NotEqual(y Int32x16) Mask32x16
 
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPQ, CPU Feature: AVX512
-func (x Int64x2) NotEqual(y Int64x2) Mask64x2
-
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPQ, CPU Feature: AVX512
-func (x Int64x4) NotEqual(y Int64x4) Mask64x4
-
 // NotEqual compares for inequality.
 //
 // Asm: VPCMPQ, CPU Feature: AVX512
 func (x Int64x8) NotEqual(y Int64x8) Mask64x8
 
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPUB, CPU Feature: AVX512
-func (x Uint8x16) NotEqual(y Uint8x16) Mask8x16
-
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPUB, CPU Feature: AVX512
-func (x Uint8x32) NotEqual(y Uint8x32) Mask8x32
-
 // NotEqual compares for inequality.
 //
 // Asm: VPCMPUB, CPU Feature: AVX512
 func (x Uint8x64) NotEqual(y Uint8x64) Mask8x64
 
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPUW, CPU Feature: AVX512
-func (x Uint16x8) NotEqual(y Uint16x8) Mask16x8
-
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPUW, CPU Feature: AVX512
-func (x Uint16x16) NotEqual(y Uint16x16) Mask16x16
-
 // NotEqual compares for inequality.
 //
 // Asm: VPCMPUW, CPU Feature: AVX512
 func (x Uint16x32) NotEqual(y Uint16x32) Mask16x32
 
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPUD, CPU Feature: AVX512
-func (x Uint32x4) NotEqual(y Uint32x4) Mask32x4
-
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPUD, CPU Feature: AVX512
-func (x Uint32x8) NotEqual(y Uint32x8) Mask32x8
-
 // NotEqual compares for inequality.
 //
 // Asm: VPCMPUD, CPU Feature: AVX512
 func (x Uint32x16) NotEqual(y Uint32x16) Mask32x16
 
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPUQ, CPU Feature: AVX512
-func (x Uint64x2) NotEqual(y Uint64x2) Mask64x2
-
-// NotEqual compares for inequality.
-//
-// Asm: VPCMPUQ, CPU Feature: AVX512
-func (x Uint64x4) NotEqual(y Uint64x4) Mask64x4
-
 // NotEqual compares for inequality.
 //
 // Asm: VPCMPUQ, CPU Feature: AVX512

src/simd/slice_amd64.go

@@ -1500,6 +1500,642 @@ func (x Float64x8) Merge(y Float64x8, mask Mask64x8) Float64x8 {
 	return iy.blendMasked(ix, mask).AsFloat64x8()
 }
 
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX
+func (x Int8x16) Less(y Int8x16) Mask8x16 {
+	return y.Greater(x)
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX
+func (x Int8x16) GreaterEqual(y Int8x16) Mask8x16 {
+	ones := x.Equal(x).AsInt8x16()
+	return y.Greater(x).AsInt8x16().Xor(ones).AsMask8x16()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX
+func (x Int8x16) LessEqual(y Int8x16) Mask8x16 {
+	ones := x.Equal(x).AsInt8x16()
+	return x.Greater(y).AsInt8x16().Xor(ones).AsMask8x16()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX
+func (x Int8x16) NotEqual(y Int8x16) Mask8x16 {
+	ones := x.Equal(x).AsInt8x16()
+	return x.Equal(y).AsInt8x16().Xor(ones).AsMask8x16()
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX
+func (x Int16x8) Less(y Int16x8) Mask16x8 {
+	return y.Greater(x)
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX
+func (x Int16x8) GreaterEqual(y Int16x8) Mask16x8 {
+	ones := x.Equal(x).AsInt16x8()
+	return y.Greater(x).AsInt16x8().Xor(ones).AsMask16x8()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX
+func (x Int16x8) LessEqual(y Int16x8) Mask16x8 {
+	ones := x.Equal(x).AsInt16x8()
+	return x.Greater(y).AsInt16x8().Xor(ones).AsMask16x8()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX
+func (x Int16x8) NotEqual(y Int16x8) Mask16x8 {
+	ones := x.Equal(x).AsInt16x8()
+	return x.Equal(y).AsInt16x8().Xor(ones).AsMask16x8()
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX
+func (x Int32x4) Less(y Int32x4) Mask32x4 {
+	return y.Greater(x)
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX
+func (x Int32x4) GreaterEqual(y Int32x4) Mask32x4 {
+	ones := x.Equal(x).AsInt32x4()
+	return y.Greater(x).AsInt32x4().Xor(ones).AsMask32x4()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX
+func (x Int32x4) LessEqual(y Int32x4) Mask32x4 {
+	ones := x.Equal(x).AsInt32x4()
+	return x.Greater(y).AsInt32x4().Xor(ones).AsMask32x4()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX
+func (x Int32x4) NotEqual(y Int32x4) Mask32x4 {
+	ones := x.Equal(x).AsInt32x4()
+	return x.Equal(y).AsInt32x4().Xor(ones).AsMask32x4()
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX
+func (x Int64x2) Less(y Int64x2) Mask64x2 {
+	return y.Greater(x)
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX
+func (x Int64x2) GreaterEqual(y Int64x2) Mask64x2 {
+	ones := x.Equal(x).AsInt64x2()
+	return y.Greater(x).AsInt64x2().Xor(ones).AsMask64x2()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX
+func (x Int64x2) LessEqual(y Int64x2) Mask64x2 {
+	ones := x.Equal(x).AsInt64x2()
+	return x.Greater(y).AsInt64x2().Xor(ones).AsMask64x2()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX
+func (x Int64x2) NotEqual(y Int64x2) Mask64x2 {
+	ones := x.Equal(x).AsInt64x2()
+	return x.Equal(y).AsInt64x2().Xor(ones).AsMask64x2()
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX2
+func (x Int8x32) Less(y Int8x32) Mask8x32 {
+	return y.Greater(x)
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX2
+func (x Int8x32) GreaterEqual(y Int8x32) Mask8x32 {
+	ones := x.Equal(x).AsInt8x32()
+	return y.Greater(x).AsInt8x32().Xor(ones).AsMask8x32()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX2
+func (x Int8x32) LessEqual(y Int8x32) Mask8x32 {
+	ones := x.Equal(x).AsInt8x32()
+	return x.Greater(y).AsInt8x32().Xor(ones).AsMask8x32()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX2
+func (x Int8x32) NotEqual(y Int8x32) Mask8x32 {
+	ones := x.Equal(x).AsInt8x32()
+	return x.Equal(y).AsInt8x32().Xor(ones).AsMask8x32()
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX2
+func (x Int16x16) Less(y Int16x16) Mask16x16 {
+	return y.Greater(x)
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX2
+func (x Int16x16) GreaterEqual(y Int16x16) Mask16x16 {
+	ones := x.Equal(x).AsInt16x16()
+	return y.Greater(x).AsInt16x16().Xor(ones).AsMask16x16()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX2
+func (x Int16x16) LessEqual(y Int16x16) Mask16x16 {
+	ones := x.Equal(x).AsInt16x16()
+	return x.Greater(y).AsInt16x16().Xor(ones).AsMask16x16()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX2
+func (x Int16x16) NotEqual(y Int16x16) Mask16x16 {
+	ones := x.Equal(x).AsInt16x16()
+	return x.Equal(y).AsInt16x16().Xor(ones).AsMask16x16()
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX2
+func (x Int32x8) Less(y Int32x8) Mask32x8 {
+	return y.Greater(x)
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX2
+func (x Int32x8) GreaterEqual(y Int32x8) Mask32x8 {
+	ones := x.Equal(x).AsInt32x8()
+	return y.Greater(x).AsInt32x8().Xor(ones).AsMask32x8()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX2
+func (x Int32x8) LessEqual(y Int32x8) Mask32x8 {
+	ones := x.Equal(x).AsInt32x8()
+	return x.Greater(y).AsInt32x8().Xor(ones).AsMask32x8()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX2
+func (x Int32x8) NotEqual(y Int32x8) Mask32x8 {
+	ones := x.Equal(x).AsInt32x8()
+	return x.Equal(y).AsInt32x8().Xor(ones).AsMask32x8()
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX2
+func (x Int64x4) Less(y Int64x4) Mask64x4 {
+	return y.Greater(x)
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX2
+func (x Int64x4) GreaterEqual(y Int64x4) Mask64x4 {
+	ones := x.Equal(x).AsInt64x4()
+	return y.Greater(x).AsInt64x4().Xor(ones).AsMask64x4()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX2
+func (x Int64x4) LessEqual(y Int64x4) Mask64x4 {
+	ones := x.Equal(x).AsInt64x4()
+	return x.Greater(y).AsInt64x4().Xor(ones).AsMask64x4()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX2
+func (x Int64x4) NotEqual(y Int64x4) Mask64x4 {
+	ones := x.Equal(x).AsInt64x4()
+	return x.Equal(y).AsInt64x4().Xor(ones).AsMask64x4()
+}
+
+// Greater returns a mask whose elements indicate whether x > y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint8x16) Greater(y Uint8x16) Mask8x16 {
+	a, b := x.AsInt8x16(), y.AsInt8x16()
+	signs := BroadcastInt8x16(-1 << (8 - 1))
+	return a.Xor(signs).Greater(b.Xor(signs))
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint8x16) Less(y Uint8x16) Mask8x16 {
+	a, b := x.AsInt8x16(), y.AsInt8x16()
+	signs := BroadcastInt8x16(-1 << (8 - 1))
+	return b.Xor(signs).Greater(a.Xor(signs))
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint8x16) GreaterEqual(y Uint8x16) Mask8x16 {
+	a, b := x.AsInt8x16(), y.AsInt8x16()
+	ones := x.Equal(x).AsInt8x16()
+	signs := BroadcastInt8x16(-1 << (8 - 1))
+	return b.Xor(signs).Greater(a.Xor(signs)).AsInt8x16().Xor(ones).AsMask8x16()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint8x16) LessEqual(y Uint8x16) Mask8x16 {
+	a, b := x.AsInt8x16(), y.AsInt8x16()
+	ones := x.Equal(x).AsInt8x16()
+	signs := BroadcastInt8x16(-1 << (8 - 1))
+	return a.Xor(signs).Greater(b.Xor(signs)).AsInt8x16().Xor(ones).AsMask8x16()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX
+func (x Uint8x16) NotEqual(y Uint8x16) Mask8x16 {
+	a, b := x.AsInt8x16(), y.AsInt8x16()
+	ones := x.Equal(x).AsInt8x16()
+	return a.Equal(b).AsInt8x16().Xor(ones).AsMask8x16()
+}
+
+// Greater returns a mask whose elements indicate whether x > y
+//
+// Emulated, CPU Feature AVX
+func (x Uint16x8) Greater(y Uint16x8) Mask16x8 {
+	a, b := x.AsInt16x8(), y.AsInt16x8()
+	ones := x.Equal(x).AsInt16x8()
+	signs := ones.ShiftAllLeft(16 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs))
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX
+func (x Uint16x8) Less(y Uint16x8) Mask16x8 {
+	a, b := x.AsInt16x8(), y.AsInt16x8()
+	ones := x.Equal(x).AsInt16x8()
+	signs := ones.ShiftAllLeft(16 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs))
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX
+func (x Uint16x8) GreaterEqual(y Uint16x8) Mask16x8 {
+	a, b := x.AsInt16x8(), y.AsInt16x8()
+	ones := x.Equal(x).AsInt16x8()
+	signs := ones.ShiftAllLeft(16 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs)).AsInt16x8().Xor(ones).AsMask16x8()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX
+func (x Uint16x8) LessEqual(y Uint16x8) Mask16x8 {
+	a, b := x.AsInt16x8(), y.AsInt16x8()
+	ones := x.Equal(x).AsInt16x8()
+	signs := ones.ShiftAllLeft(16 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs)).AsInt16x8().Xor(ones).AsMask16x8()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX
+func (x Uint16x8) NotEqual(y Uint16x8) Mask16x8 {
+	a, b := x.AsInt16x8(), y.AsInt16x8()
+	ones := x.Equal(x).AsInt16x8()
+	return a.Equal(b).AsInt16x8().Xor(ones).AsMask16x8()
+}
+
+// Greater returns a mask whose elements indicate whether x > y
+//
+// Emulated, CPU Feature AVX
+func (x Uint32x4) Greater(y Uint32x4) Mask32x4 {
+	a, b := x.AsInt32x4(), y.AsInt32x4()
+	ones := x.Equal(x).AsInt32x4()
+	signs := ones.ShiftAllLeft(32 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs))
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX
+func (x Uint32x4) Less(y Uint32x4) Mask32x4 {
+	a, b := x.AsInt32x4(), y.AsInt32x4()
+	ones := x.Equal(x).AsInt32x4()
+	signs := ones.ShiftAllLeft(32 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs))
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX
+func (x Uint32x4) GreaterEqual(y Uint32x4) Mask32x4 {
+	a, b := x.AsInt32x4(), y.AsInt32x4()
+	ones := x.Equal(x).AsInt32x4()
+	signs := ones.ShiftAllLeft(32 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs)).AsInt32x4().Xor(ones).AsMask32x4()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX
+func (x Uint32x4) LessEqual(y Uint32x4) Mask32x4 {
+	a, b := x.AsInt32x4(), y.AsInt32x4()
+	ones := x.Equal(x).AsInt32x4()
+	signs := ones.ShiftAllLeft(32 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs)).AsInt32x4().Xor(ones).AsMask32x4()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX
+func (x Uint32x4) NotEqual(y Uint32x4) Mask32x4 {
+	a, b := x.AsInt32x4(), y.AsInt32x4()
+	ones := x.Equal(x).AsInt32x4()
+	return a.Equal(b).AsInt32x4().Xor(ones).AsMask32x4()
+}
+
+// Greater returns a mask whose elements indicate whether x > y
+//
+// Emulated, CPU Feature AVX
+func (x Uint64x2) Greater(y Uint64x2) Mask64x2 {
+	a, b := x.AsInt64x2(), y.AsInt64x2()
+	ones := x.Equal(x).AsInt64x2()
+	signs := ones.ShiftAllLeft(64 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs))
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX
+func (x Uint64x2) Less(y Uint64x2) Mask64x2 {
+	a, b := x.AsInt64x2(), y.AsInt64x2()
+	ones := x.Equal(x).AsInt64x2()
+	signs := ones.ShiftAllLeft(64 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs))
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX
+func (x Uint64x2) GreaterEqual(y Uint64x2) Mask64x2 {
+	a, b := x.AsInt64x2(), y.AsInt64x2()
+	ones := x.Equal(x).AsInt64x2()
+	signs := ones.ShiftAllLeft(64 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs)).AsInt64x2().Xor(ones).AsMask64x2()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX
+func (x Uint64x2) LessEqual(y Uint64x2) Mask64x2 {
+	a, b := x.AsInt64x2(), y.AsInt64x2()
+	ones := x.Equal(x).AsInt64x2()
+	signs := ones.ShiftAllLeft(64 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs)).AsInt64x2().Xor(ones).AsMask64x2()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX
+func (x Uint64x2) NotEqual(y Uint64x2) Mask64x2 {
+	a, b := x.AsInt64x2(), y.AsInt64x2()
+	ones := x.Equal(x).AsInt64x2()
+	return a.Equal(b).AsInt64x2().Xor(ones).AsMask64x2()
+}
+
+// Greater returns a mask whose elements indicate whether x > y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint8x32) Greater(y Uint8x32) Mask8x32 {
+	a, b := x.AsInt8x32(), y.AsInt8x32()
+	signs := BroadcastInt8x32(-1 << (8 - 1))
+	return a.Xor(signs).Greater(b.Xor(signs))
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint8x32) Less(y Uint8x32) Mask8x32 {
+	a, b := x.AsInt8x32(), y.AsInt8x32()
+	signs := BroadcastInt8x32(-1 << (8 - 1))
+	return b.Xor(signs).Greater(a.Xor(signs))
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint8x32) GreaterEqual(y Uint8x32) Mask8x32 {
+	a, b := x.AsInt8x32(), y.AsInt8x32()
+	ones := x.Equal(x).AsInt8x32()
+	signs := BroadcastInt8x32(-1 << (8 - 1))
+	return b.Xor(signs).Greater(a.Xor(signs)).AsInt8x32().Xor(ones).AsMask8x32()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint8x32) LessEqual(y Uint8x32) Mask8x32 {
+	a, b := x.AsInt8x32(), y.AsInt8x32()
+	ones := x.Equal(x).AsInt8x32()
+	signs := BroadcastInt8x32(-1 << (8 - 1))
+	return a.Xor(signs).Greater(b.Xor(signs)).AsInt8x32().Xor(ones).AsMask8x32()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint8x32) NotEqual(y Uint8x32) Mask8x32 {
+	a, b := x.AsInt8x32(), y.AsInt8x32()
+	ones := x.Equal(x).AsInt8x32()
+	return a.Equal(b).AsInt8x32().Xor(ones).AsMask8x32()
+}
+
+// Greater returns a mask whose elements indicate whether x > y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint16x16) Greater(y Uint16x16) Mask16x16 {
+	a, b := x.AsInt16x16(), y.AsInt16x16()
+	ones := x.Equal(x).AsInt16x16()
+	signs := ones.ShiftAllLeft(16 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs))
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint16x16) Less(y Uint16x16) Mask16x16 {
+	a, b := x.AsInt16x16(), y.AsInt16x16()
+	ones := x.Equal(x).AsInt16x16()
+	signs := ones.ShiftAllLeft(16 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs))
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint16x16) GreaterEqual(y Uint16x16) Mask16x16 {
+	a, b := x.AsInt16x16(), y.AsInt16x16()
+	ones := x.Equal(x).AsInt16x16()
+	signs := ones.ShiftAllLeft(16 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs)).AsInt16x16().Xor(ones).AsMask16x16()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint16x16) LessEqual(y Uint16x16) Mask16x16 {
+	a, b := x.AsInt16x16(), y.AsInt16x16()
+	ones := x.Equal(x).AsInt16x16()
+	signs := ones.ShiftAllLeft(16 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs)).AsInt16x16().Xor(ones).AsMask16x16()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint16x16) NotEqual(y Uint16x16) Mask16x16 {
+	a, b := x.AsInt16x16(), y.AsInt16x16()
+	ones := x.Equal(x).AsInt16x16()
+	return a.Equal(b).AsInt16x16().Xor(ones).AsMask16x16()
+}
+
+// Greater returns a mask whose elements indicate whether x > y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint32x8) Greater(y Uint32x8) Mask32x8 {
+	a, b := x.AsInt32x8(), y.AsInt32x8()
+	ones := x.Equal(x).AsInt32x8()
+	signs := ones.ShiftAllLeft(32 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs))
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint32x8) Less(y Uint32x8) Mask32x8 {
+	a, b := x.AsInt32x8(), y.AsInt32x8()
+	ones := x.Equal(x).AsInt32x8()
+	signs := ones.ShiftAllLeft(32 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs))
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint32x8) GreaterEqual(y Uint32x8) Mask32x8 {
+	a, b := x.AsInt32x8(), y.AsInt32x8()
+	ones := x.Equal(x).AsInt32x8()
+	signs := ones.ShiftAllLeft(32 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs)).AsInt32x8().Xor(ones).AsMask32x8()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint32x8) LessEqual(y Uint32x8) Mask32x8 {
+	a, b := x.AsInt32x8(), y.AsInt32x8()
+	ones := x.Equal(x).AsInt32x8()
+	signs := ones.ShiftAllLeft(32 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs)).AsInt32x8().Xor(ones).AsMask32x8()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint32x8) NotEqual(y Uint32x8) Mask32x8 {
+	a, b := x.AsInt32x8(), y.AsInt32x8()
+	ones := x.Equal(x).AsInt32x8()
+	return a.Equal(b).AsInt32x8().Xor(ones).AsMask32x8()
+}
+
+// Greater returns a mask whose elements indicate whether x > y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint64x4) Greater(y Uint64x4) Mask64x4 {
+	a, b := x.AsInt64x4(), y.AsInt64x4()
+	ones := x.Equal(x).AsInt64x4()
+	signs := ones.ShiftAllLeft(64 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs))
+}
+
+// Less returns a mask whose elements indicate whether x < y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint64x4) Less(y Uint64x4) Mask64x4 {
+	a, b := x.AsInt64x4(), y.AsInt64x4()
+	ones := x.Equal(x).AsInt64x4()
+	signs := ones.ShiftAllLeft(64 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs))
+}
+
+// GreaterEqual returns a mask whose elements indicate whether x >= y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint64x4) GreaterEqual(y Uint64x4) Mask64x4 {
+	a, b := x.AsInt64x4(), y.AsInt64x4()
+	ones := x.Equal(x).AsInt64x4()
+	signs := ones.ShiftAllLeft(64 - 1)
+	return b.Xor(signs).Greater(a.Xor(signs)).AsInt64x4().Xor(ones).AsMask64x4()
+}
+
+// LessEqual returns a mask whose elements indicate whether x <= y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint64x4) LessEqual(y Uint64x4) Mask64x4 {
+	a, b := x.AsInt64x4(), y.AsInt64x4()
+	ones := x.Equal(x).AsInt64x4()
+	signs := ones.ShiftAllLeft(64 - 1)
+	return a.Xor(signs).Greater(b.Xor(signs)).AsInt64x4().Xor(ones).AsMask64x4()
+}
+
+// NotEqual returns a mask whose elements indicate whether x != y
+//
+// Emulated, CPU Feature AVX2
+func (x Uint64x4) NotEqual(y Uint64x4) Mask64x4 {
+	a, b := x.AsInt64x4(), y.AsInt64x4()
+	ones := x.Equal(x).AsInt64x4()
+	return a.Equal(b).AsInt64x4().Xor(ones).AsMask64x4()
+}
+
 // BroadcastInt8x16 returns a vector with the input
 // x assigned to all elements of the output.
 //