[TargetLowering] check for sign-bit comparisons in SimplifyDemandedBits

[spatel]

Summary

I don't know if anything other than x86 vectors is affected by this change, but this may allow us to remove target-specific intrinsics for blendv* (vector selects). The simplification arises from the fact that blendv* instructions only use the sign-bit when deciding which vector element to choose for the destination vector. The mechanism to fold VSELECT into SHRUNKBLEND nodes already exists in x86 lowering; this demanded bits change just enables the transform to fire more often.

The original motivation starts with a bug for DSE of masked stores that seems completely unrelated, but I've explained the likely steps in this series here:
https://llvm.org/bugs/show_bug.cgi?id=11210

[spatel]

Patch updated:
I'm not sure how to expose this, but I think I was missing a check for BooleanContent::ZeroOrNegativeOneBooleanContent. If the target/type uses a different format for setcc, it probably wouldn't make sense to demand the top bit, but let's check that to be safe?

[RKSimon]

lib/CodeGen/SelectionDAG/TargetLowering.cpp

770 ↗

(On Diff #87530)

Add test cases for these?

test/CodeGen/X86/vselect-pcmp.ll

4 ↗

(On Diff #87530)

Please can you add a avx512vl target to check more mask predicate cases?

[spatel]

lib/CodeGen/SelectionDAG/TargetLowering.cpp

770 ↗

(On Diff #87530)

I'm not sure we want to bloat it up with those cases yet; that's why I made it a TODO? rather than a FIXME.
Instcombine already canonicalizes the X <= -1 variant to X < 0. It doesn't know to change X > -1 to X < 0 and swap the select operands, but I think that's just an IR canonicalization oversight, so I was planning to fix that.

So if we add codegen tests for those variants, it's only because a non-canonical select pattern has been created in the backend. I'd like to find evidence of that happening before adding code or tests for it.

test/CodeGen/X86/vselect-pcmp.ll

4 ↗

(On Diff #87530)

Sure - I'm still not clear on all of the avx512 variants, but I added a 'vl' RUN line here:
rL294462

I'll rebase with those diffs next.

[spatel]

lib/CodeGen/SelectionDAG/TargetLowering.cpp

770 ↗

(On Diff #87530)

Also, there's a hint in the test variable names for this shift/trunc canonicalization to X < 0:

define <4 x i32> @signbit_sel_v4i32(<4 x i32> %x, <4 x i32> %y, <4 x i32> %mask) {
  %sh = lshr <4 x i32> %mask, <i32 31, i32 31, i32 31, i32 31>
  %tr = trunc <4 x i32> %sh to <4 x i1>
  %z = select <4 x i1> %tr, <4 x i32> %x, <4 x i32> %y
  ret <4 x i32> %z
}

$ opt -instcombine -S foo.ll

define <4 x i32> @signbit_sel_v4i32(<4 x i32> %x, <4 x i32> %y, <4 x i32> %mask) {
  %tr = icmp slt <4 x i32> %mask, zeroinitializer
  %z = select <4 x i1> %tr, <4 x i32> %x, <4 x i32> %y
  ret <4 x i32> %z
}

[spatel]

Patch updated:
No code changes, but added a RUN for avx512vl in rL294462 to show that variant.
I think the first test shows a missed opportunity for avx512vl and avx512f; not sure about the rest.

[RKSimon]

LGTM thanks Sanjay

test/CodeGen/X86/vselect-pcmp.ll

12 ↗

(On Diff #87663)

Add a FIXME for the AVX512 test case.

llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp

case ISD::SELECT_CC:
     // If the operands are constants, see if we can simplify them.
     if (TLO.ShrinkDemandedConstant(Op, NewMask))
       return true;
 
     // Only known if known in both the LHS and RHS.
     KnownOne &= KnownOne2;
     KnownZero &= KnownZero2;
     break;
+  case ISD::SETCC:
+    // If (1) we only need the sign-bit, (2) the setcc operands are the same
+    // width as the setcc result, and (3) the result of a setcc conforms to 0 or
+    // -1, we may be able to bypass the setcc.
+    if (NewMask.isSignBit() &&
+        Op.getOperand(0).getScalarValueSizeInBits() == BitWidth &&
+        getBooleanContents(Op.getValueType()) ==
+            BooleanContent::ZeroOrNegativeOneBooleanContent) {
+      ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
+      // If we're testing if X < 0, then this compare isn't needed - just use X!
+      if (CC == ISD::SETLT &&
+          (isNullConstant(Op.getOperand(1)) ||
+           ISD::isBuildVectorAllZeros(Op.getOperand(1).getNode())))
+        return TLO.CombineTo(Op, Op.getOperand(0));
+
+      // TODO: Should we check for other forms of sign-bit comparisons?
+      // Examples: X <= -1, X >= 0
+    }
+    break;
   case ISD::SHL:
     if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
       unsigned ShAmt = SA->getZExtValue();
       SDValue InOp = Op.getOperand(0);
 
       // If the shift count is an invalid immediate, don't do anything.
       if (ShAmt >= BitWidth)
         break;

llvm/trunk/test/CodeGen/X86/vselect-pcmp.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx       | FileCheck %s --check-prefix=AVX --check-prefix=AVX12F --check-prefix=AVX12 --check-prefix=AVX1
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx2      | FileCheck %s --check-prefix=AVX --check-prefix=AVX12F --check-prefix=AVX12 --check-prefix=AVX2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512f   | FileCheck %s --check-prefix=AVX --check-prefix=AVX12F --check-prefix=AVX512 --check-prefix=AVX512F
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512vl  | FileCheck %s --check-prefix=AVX                       --check-prefix=AVX512 --check-prefix=AVX512VL
 
 ; The condition vector for BLENDV* only cares about the sign bit of each element.
 ; So in these tests, if we generate BLENDV*, we should be able to remove the redundant cmp op.
 
 ; Test 128-bit vectors for all legal element types.
 
+; FIXME: Why didn't AVX-512 optimize too?
+
 define <16 x i8> @signbit_sel_v16i8(<16 x i8> %x, <16 x i8> %y, <16 x i8> %mask) {
-; AVX-LABEL: signbit_sel_v16i8:
-; AVX:       # BB#0:
-; AVX-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX-NEXT:    vpcmpgtb %xmm2, %xmm3, %xmm2
-; AVX-NEXT:    vpblendvb %xmm2, %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    retq
+; AVX12-LABEL: signbit_sel_v16i8:
+; AVX12:       # BB#0:
+; AVX12-NEXT:    vpblendvb %xmm2, %xmm0, %xmm1, %xmm0
+; AVX12-NEXT:    retq
+;
+; AVX512-LABEL: signbit_sel_v16i8:
+; AVX512:       # BB#0:
+; AVX512-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512-NEXT:    vpcmpgtb %xmm2, %xmm3, %xmm2
+; AVX512-NEXT:    vpblendvb %xmm2, %xmm0, %xmm1, %xmm0
+; AVX512-NEXT:    retq
   %tr = icmp slt <16 x i8> %mask, zeroinitializer
   %z = select <16 x i1> %tr, <16 x i8> %x, <16 x i8> %y
   ret <16 x i8> %z
 }
 
 ; Sorry 16-bit, you're not important enough to support?
 
 define <8 x i16> @signbit_sel_v8i16(<8 x i16> %x, <8 x i16> %y, <8 x i16> %mask) {
; AVX512VL-NEXT:    retq
   %tr = icmp slt <8 x i16> %mask, zeroinitializer
   %z = select <8 x i1> %tr, <8 x i16> %x, <8 x i16> %y
   ret <8 x i16> %z
 }
 
 define <4 x i32> @signbit_sel_v4i32(<4 x i32> %x, <4 x i32> %y, <4 x i32> %mask) {
 ; AVX12F-LABEL: signbit_sel_v4i32:
 ; AVX12F:       # BB#0:
-; AVX12F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX12F-NEXT:    vpcmpgtd %xmm2, %xmm3, %xmm2
 ; AVX12F-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX12F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: signbit_sel_v4i32:
 ; AVX512VL:       # BB#0:
 ; AVX512VL-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX512VL-NEXT:    vpcmpgtd %xmm2, %xmm3, %k1
 ; AVX512VL-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
 ; AVX512VL-NEXT:    retq
   %tr = icmp slt <4 x i32> %mask, zeroinitializer
   %z = select <4 x i1> %tr, <4 x i32> %x, <4 x i32> %y
   ret <4 x i32> %z
 }
 
 define <2 x i64> @signbit_sel_v2i64(<2 x i64> %x, <2 x i64> %y, <2 x i64> %mask) {
 ; AVX12F-LABEL: signbit_sel_v2i64:
 ; AVX12F:       # BB#0:
-; AVX12F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX12F-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
 ; AVX12F-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX12F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: signbit_sel_v2i64:
 ; AVX512VL:       # BB#0:
 ; AVX512VL-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX512VL-NEXT:    vpcmpgtq %xmm2, %xmm3, %k1
 ; AVX512VL-NEXT:    vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
 ; AVX512VL-NEXT:    retq
   %tr = icmp slt <2 x i64> %mask, zeroinitializer
   %z = select <2 x i1> %tr, <2 x i64> %x, <2 x i64> %y
   ret <2 x i64> %z
 }
 
 define <4 x float> @signbit_sel_v4f32(<4 x float> %x, <4 x float> %y, <4 x i32> %mask) {
 ; AVX12F-LABEL: signbit_sel_v4f32:
 ; AVX12F:       # BB#0:
-; AVX12F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX12F-NEXT:    vpcmpgtd %xmm2, %xmm3, %xmm2
 ; AVX12F-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX12F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: signbit_sel_v4f32:
 ; AVX512VL:       # BB#0:
 ; AVX512VL-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX512VL-NEXT:    vpcmpgtd %xmm2, %xmm3, %k1
 ; AVX512VL-NEXT:    vblendmps %xmm0, %xmm1, %xmm0 {%k1}
 ; AVX512VL-NEXT:    retq
   %tr = icmp slt <4 x i32> %mask, zeroinitializer
   %z = select <4 x i1> %tr, <4 x float> %x, <4 x float> %y
   ret <4 x float> %z
 }
 
 define <2 x double> @signbit_sel_v2f64(<2 x double> %x, <2 x double> %y, <2 x i64> %mask) {
 ; AVX12F-LABEL: signbit_sel_v2f64:
 ; AVX12F:       # BB#0:
-; AVX12F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX12F-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
 ; AVX12F-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX12F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: signbit_sel_v2f64:
 ; AVX512VL:       # BB#0:
 ; AVX512VL-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX512VL-NEXT:    vpcmpgtq %xmm2, %xmm3, %k1
 ; AVX512VL-NEXT:    vblendmpd %xmm0, %xmm1, %xmm0 {%k1}

 ; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
 ; AVX1-NEXT:    vandnps %ymm1, %ymm2, %ymm1
 ; AVX1-NEXT:    vandps %ymm2, %ymm0, %ymm0
 ; AVX1-NEXT:    vorps %ymm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: signbit_sel_v32i8:
 ; AVX2:       # BB#0:
-; AVX2-NEXT:    vpxor %ymm3, %ymm3, %ymm3
-; AVX2-NEXT:    vpcmpgtb %ymm2, %ymm3, %ymm2
 ; AVX2-NEXT:    vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: signbit_sel_v32i8:
 ; AVX512:       # BB#0:
 ; AVX512-NEXT:    vpxor %ymm3, %ymm3, %ymm3
 ; AVX512-NEXT:    vpcmpgtb %ymm2, %ymm3, %ymm2
 ; AVX512-NEXT:    vpblendvb %ymm2, %ymm0, %ymm1, %ymm0

 ; AVX512VL-NEXT:    vpor %ymm1, %ymm0, %ymm0
 ; AVX512VL-NEXT:    retq
   %tr = icmp slt <16 x i16> %mask, zeroinitializer
   %z = select <16 x i1> %tr, <16 x i16> %x, <16 x i16> %y
   ret <16 x i16> %z
 }
 
 define <8 x i32> @signbit_sel_v8i32(<8 x i32> %x, <8 x i32> %y, <8 x i32> %mask) {
-; AVX1-LABEL: signbit_sel_v8i32:
-; AVX1:       # BB#0:
-; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpcmpgtd %xmm3, %xmm4, %xmm3
-; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; AVX1-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: signbit_sel_v8i32:
-; AVX2:       # BB#0:
-; AVX2-NEXT:    vpxor %ymm3, %ymm3, %ymm3
-; AVX2-NEXT:    vpcmpgtd %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    retq
+; AVX12-LABEL: signbit_sel_v8i32:
+; AVX12:       # BB#0:
+; AVX12-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
+; AVX12-NEXT:    retq
 ;
 ; AVX512F-LABEL: signbit_sel_v8i32:
 ; AVX512F:       # BB#0:
 ; AVX512F-NEXT:    # kill: %YMM2<def> %YMM2<kill> %ZMM2<def>
 ; AVX512F-NEXT:    # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
 ; AVX512F-NEXT:    # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
 ; AVX512F-NEXT:    vpxor %ymm3, %ymm3, %ymm3
 ; AVX512F-NEXT:    vpcmpgtd %zmm2, %zmm3, %k1
 ; AVX512F-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
 ; AVX512F-NEXT:    # kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: signbit_sel_v8i32:
 ; AVX512VL:       # BB#0:
 ; AVX512VL-NEXT:    vpxor %ymm3, %ymm3, %ymm3
 ; AVX512VL-NEXT:    vpcmpgtd %ymm2, %ymm3, %k1
 ; AVX512VL-NEXT:    vpblendmd %ymm0, %ymm1, %ymm0 {%k1}
 ; AVX512VL-NEXT:    retq
   %tr = icmp slt <8 x i32> %mask, zeroinitializer
   %z = select <8 x i1> %tr, <8 x i32> %x, <8 x i32> %y
   ret <8 x i32> %z
 }
 
 define <4 x i64> @signbit_sel_v4i64(<4 x i64> %x, <4 x i64> %y, <4 x i64> %mask) {
-; AVX1-LABEL: signbit_sel_v4i64:
-; AVX1:       # BB#0:
-; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: signbit_sel_v4i64:
-; AVX2:       # BB#0:
-; AVX2-NEXT:    vpxor %ymm3, %ymm3, %ymm3
-; AVX2-NEXT:    vpcmpgtq %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512F-LABEL: signbit_sel_v4i64:
-; AVX512F:       # BB#0:
-; AVX512F-NEXT:    vpxor %ymm3, %ymm3, %ymm3
-; AVX512F-NEXT:    vpcmpgtq %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    retq
+; AVX12F-LABEL: signbit_sel_v4i64:
+; AVX12F:       # BB#0:
+; AVX12F-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
+; AVX12F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: signbit_sel_v4i64:
 ; AVX512VL:       # BB#0:
 ; AVX512VL-NEXT:    vpxor %ymm3, %ymm3, %ymm3
 ; AVX512VL-NEXT:    vpcmpgtq %ymm2, %ymm3, %k1
 ; AVX512VL-NEXT:    vpblendmq %ymm0, %ymm1, %ymm0 {%k1}
 ; AVX512VL-NEXT:    retq
   %tr = icmp slt <4 x i64> %mask, zeroinitializer
   %z = select <4 x i1> %tr, <4 x i64> %x, <4 x i64> %y
   ret <4 x i64> %z
 }
 
 define <4 x double> @signbit_sel_v4f64(<4 x double> %x, <4 x double> %y, <4 x i64> %mask) {
-; AVX1-LABEL: signbit_sel_v4f64:
-; AVX1:       # BB#0:
-; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm4, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: signbit_sel_v4f64:
-; AVX2:       # BB#0:
-; AVX2-NEXT:    vpxor %ymm3, %ymm3, %ymm3
-; AVX2-NEXT:    vpcmpgtq %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512F-LABEL: signbit_sel_v4f64:
-; AVX512F:       # BB#0:
-; AVX512F-NEXT:    vpxor %ymm3, %ymm3, %ymm3
-; AVX512F-NEXT:    vpcmpgtq %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    retq
+; AVX12F-LABEL: signbit_sel_v4f64:
+; AVX12F:       # BB#0:
+; AVX12F-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
+; AVX12F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: signbit_sel_v4f64:
 ; AVX512VL:       # BB#0:
 ; AVX512VL-NEXT:    vpxor %ymm3, %ymm3, %ymm3
 ; AVX512VL-NEXT:    vpcmpgtq %ymm2, %ymm3, %k1
 ; AVX512VL-NEXT:    vblendmpd %ymm0, %ymm1, %ymm0 {%k1}
 ; AVX512VL-NEXT:    retq
   %tr = icmp slt <4 x i64> %mask, zeroinitializer

 ; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,0,1]
 ; AVX1-NEXT:    vpmovsxdq %xmm2, %xmm2
 ; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
 ; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: signbit_sel_v4f64_small_mask:
 ; AVX2:       # BB#0:
-; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX2-NEXT:    vpcmpgtd %xmm2, %xmm3, %xmm2
 ; AVX2-NEXT:    vpmovsxdq %xmm2, %ymm2
 ; AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: signbit_sel_v4f64_small_mask:
 ; AVX512F:       # BB#0:
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpcmpgtd %xmm2, %xmm3, %xmm2
 ; AVX512F-NEXT:    vpmovsxdq %xmm2, %ymm2
 ; AVX512F-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: signbit_sel_v4f64_small_mask:
 ; AVX512VL:       # BB#0:
 ; AVX512VL-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX512VL-NEXT:    vpcmpgtd %xmm2, %xmm3, %k1
 ; AVX512VL-NEXT:    vblendmpd %ymm0, %ymm1, %ymm0 {%k1}
 ; AVX512VL-NEXT:    retq
   %tr = icmp slt <4 x i32> %mask, zeroinitializer
   %z = select <4 x i1> %tr, <4 x double> %x, <4 x double> %y
   ret <4 x double> %z
 }
 
 ; Try a 512-bit vector to make sure AVX-512 is handled as expected.
 
 define <8 x double> @signbit_sel_v8f64(<8 x double> %x, <8 x double> %y, <8 x i64> %mask) {
-; AVX1-LABEL: signbit_sel_v8f64:
-; AVX1:       # BB#0:
-; AVX1-NEXT:    vextractf128 $1, %ymm5, %xmm6
-; AVX1-NEXT:    vpxor %xmm7, %xmm7, %xmm7
-; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm7, %xmm6
-; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm7, %xmm5
-; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm5, %ymm5
-; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm6
-; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm7, %xmm6
-; AVX1-NEXT:    vpcmpgtq %xmm4, %xmm7, %xmm4
-; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm4, %ymm4
-; AVX1-NEXT:    vblendvpd %ymm4, %ymm0, %ymm2, %ymm0
-; AVX1-NEXT:    vblendvpd %ymm5, %ymm1, %ymm3, %ymm1
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: signbit_sel_v8f64:
-; AVX2:       # BB#0:
-; AVX2-NEXT:    vpxor %ymm6, %ymm6, %ymm6
-; AVX2-NEXT:    vpcmpgtq %ymm5, %ymm6, %ymm5
-; AVX2-NEXT:    vpcmpgtq %ymm4, %ymm6, %ymm4
-; AVX2-NEXT:    vblendvpd %ymm4, %ymm0, %ymm2, %ymm0
-; AVX2-NEXT:    vblendvpd %ymm5, %ymm1, %ymm3, %ymm1
-; AVX2-NEXT:    retq
+; AVX12-LABEL: signbit_sel_v8f64:
+; AVX12:       # BB#0:
+; AVX12-NEXT:    vblendvpd %ymm4, %ymm0, %ymm2, %ymm0
+; AVX12-NEXT:    vblendvpd %ymm5, %ymm1, %ymm3, %ymm1
+; AVX12-NEXT:    retq
 ;
 ; AVX512-LABEL: signbit_sel_v8f64:
 ; AVX512:       # BB#0:
 ; AVX512-NEXT:    vpxord %zmm3, %zmm3, %zmm3
 ; AVX512-NEXT:    vpcmpgtq %zmm2, %zmm3, %k1
 ; AVX512-NEXT:    vblendmpd %zmm0, %zmm1, %zmm0 {%k1}
 ; AVX512-NEXT:    retq
   %tr = icmp slt <8 x i64> %mask, zeroinitializer
   %z = select <8 x i1> %tr, <8 x double> %x, <8 x double> %y
   ret <8 x double> %z
 }