[x86] split 256-bit vector selects if operands are vector concats

This is similar logic/motivation to the select splitting in D62969.

In D63233, the pattern changes so that we no longer have an extract_subvector of vselect,
but the operands of the select are still being concatenated.

The closest case is represented in either the first or last test diffs here - we have an
extra instruction, but we converted 3-4 ymm instructions into 4-5 xmm instructions.
I think that's the right trade-off for most AVX1 targets.

In the example based on PR37428:
https://bugs.llvm.org/show_bug.cgi?id=37428
...this makes the loop about 30% faster (tested on Haswell by compiling with -mavx).

Differential Revision: https://reviews.llvm.org/D63364

llvm-svn: 363508

4 files changed, 93 insertions, 53 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 78726818927..a0bb3220ba3 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -35542,6 +35542,39 @@ combineVSelectWithAllOnesOrZeros(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+/// If both arms of a vector select are concatenated vectors, split the select,
+/// and concatenate the result to eliminate a wide (256-bit) vector instruction:
+///   vselect Cond, (concat T0, T1), (concat F0, F1) -->
+///   concat (vselect (split Cond), T0, F0), (vselect (split Cond), T1, F1)
+static SDValue narrowVectorSelect(SDNode *N, SelectionDAG &DAG,
+                                  const X86Subtarget &Subtarget) {
+  unsigned Opcode = N->getOpcode();
+  if (Opcode != X86ISD::BLENDV && Opcode != ISD::VSELECT)
+    return SDValue();
+
+  // TODO: Split 512-bit vectors too?
+  EVT VT = N->getValueType(0);
+  if (!VT.is256BitVector())
+    return SDValue();
+
+  // TODO: Split as long as any 2 of the 3 operands are concatenated?
+  SDValue Cond = N->getOperand(0);
+  SDValue TVal = N->getOperand(1);
+  SDValue FVal = N->getOperand(2);
+  SmallVector<SDValue, 4> CatOpsT, CatOpsF;
+  if (!TVal.hasOneUse() || !FVal.hasOneUse() ||
+      !collectConcatOps(TVal.getNode(), CatOpsT) ||
+      !collectConcatOps(FVal.getNode(), CatOpsF))
+    return SDValue();
+
+  auto makeBlend = [Opcode](SelectionDAG &DAG, const SDLoc &DL,
+                            ArrayRef<SDValue> Ops) {
+    return DAG.getNode(Opcode, DL, Ops[1].getValueType(), Ops);
+  };
+  return SplitOpsAndApply(DAG, Subtarget, SDLoc(N), VT, { Cond, TVal, FVal },
+                          makeBlend, /*CheckBWI*/ false);
+}
+
 static SDValue combineSelectOfTwoConstants(SDNode *N, SelectionDAG &DAG) {
   SDValue Cond = N->getOperand(0);
   SDValue LHS = N->getOperand(1);
@@ -36105,6 +36138,9 @@ static SDValue combineSelect(SDNode *N, SelectionDAG &DAG,
   if (SDValue V = combineVSelectToBLENDV(N, DAG, DCI, Subtarget))
     return V;
 
+  if (SDValue V = narrowVectorSelect(N, DAG, Subtarget))
+    return V;
+
   // Custom action for SELECT MMX
   if (VT == MVT::x86mmx) {
     LHS = DAG.getBitcast(MVT::i64, LHS);
diff --git a/llvm/test/CodeGen/X86/cast-vsel.ll b/llvm/test/CodeGen/X86/cast-vsel.ll
index 03efb540216..b02987996a4 100644
--- a/llvm/test/CodeGen/X86/cast-vsel.ll
+++ b/llvm/test/CodeGen/X86/cast-vsel.ll
@@ -38,15 +38,16 @@ define <8 x i32> @sext(<8 x float> %a, <8 x float> %b, <8 x i16> %c, <8 x i16> %
 ; AVX1-LABEL: sext:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vcmpltps %ymm1, %ymm0, %ymm0
-; AVX1-NEXT:    vpmovsxwd %xmm2, %xmm1
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,0,1]
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm4 = xmm2[2,3,0,1]
+; AVX1-NEXT:    vpmovsxwd %xmm4, %xmm4
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm5 = xmm3[2,3,0,1]
+; AVX1-NEXT:    vpmovsxwd %xmm5, %xmm5
+; AVX1-NEXT:    vblendvps %xmm1, %xmm4, %xmm5, %xmm1
 ; AVX1-NEXT:    vpmovsxwd %xmm2, %xmm2
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
-; AVX1-NEXT:    vpmovsxwd %xmm3, %xmm2
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,3,0,1]
 ; AVX1-NEXT:    vpmovsxwd %xmm3, %xmm3
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; AVX1-NEXT:    vblendvps %ymm0, %ymm1, %ymm2, %ymm0
+; AVX1-NEXT:    vblendvps %xmm0, %xmm2, %xmm3, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: sext:
@@ -95,12 +96,13 @@ define <8 x i32> @zext(<8 x float> %a, <8 x float> %b, <8 x i16> %c, <8 x i16> %
 ; AVX1-NEXT:    vcmpltps %ymm1, %ymm0, %ymm0
 ; AVX1-NEXT:    vxorps %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpunpckhwd {{.*#+}} xmm4 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX1-NEXT:    vpmovzxwd {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
-; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm2, %ymm2
 ; AVX1-NEXT:    vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5
+; AVX1-NEXT:    vblendvps %xmm5, %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vpmovzxwd {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
 ; AVX1-NEXT:    vpmovzxwd {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm3, %ymm1
-; AVX1-NEXT:    vblendvps %ymm0, %ymm2, %ymm1, %ymm0
+; AVX1-NEXT:    vblendvps %xmm0, %xmm2, %xmm3, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: zext:
@@ -403,6 +405,8 @@ for.end:
   ret void
 }
 
+; TODO: AVX1 could have used 256-bit ops for a likely improvement.
+
 define void @example24(i16 signext %x, i16 signext %y) nounwind {
 ; SSE2-LABEL: example24:
 ; SSE2:       # %bb.0: # %vector.ph
@@ -469,26 +473,27 @@ define void @example24(i16 signext %x, i16 signext %y) nounwind {
 ; AVX1:       # %bb.0: # %vector.ph
 ; AVX1-NEXT:    vmovd %edi, %xmm0
 ; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,0,2,3,4,5,6,7]
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
-; AVX1-NEXT:    vmovd %esi, %xmm1
-; AVX1-NEXT:    vpshuflw {{.*#+}} xmm1 = xmm1[0,0,2,3,4,5,6,7]
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,0,0,0]
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[0,0,0,0]
+; AVX1-NEXT:    vmovd %esi, %xmm0
+; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[0,0,2,3,4,5,6,7]
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[0,0,0,0]
 ; AVX1-NEXT:    movq $-4096, %rax # imm = 0xF000
-; AVX1-NEXT:    vpmovsxwd %xmm0, %xmm2
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
-; AVX1-NEXT:    vpmovsxwd %xmm0, %xmm0
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
-; AVX1-NEXT:    vpmovsxwd %xmm1, %xmm2
-; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,3,0,1]
-; AVX1-NEXT:    vpmovsxwd %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm2, %ymm1
+; AVX1-NEXT:    vpmovsxwd %xmm2, %xmm0
+; AVX1-NEXT:    vpmovsxwd %xmm3, %xmm1
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,0,1]
+; AVX1-NEXT:    vpmovsxwd %xmm2, %xmm2
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,3,0,1]
+; AVX1-NEXT:    vpmovsxwd %xmm3, %xmm3
 ; AVX1-NEXT:    .p2align 4, 0x90
 ; AVX1-NEXT:  .LBB6_1: # %vector.body
 ; AVX1-NEXT:    # =>This Inner Loop Header: Depth=1
-; AVX1-NEXT:    vmovups da+4096(%rax), %ymm2
-; AVX1-NEXT:    vcmpltps db+4096(%rax), %ymm2, %ymm2
-; AVX1-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm2
-; AVX1-NEXT:    vmovups %ymm2, dj+4096(%rax)
+; AVX1-NEXT:    vmovups da+4096(%rax), %ymm4
+; AVX1-NEXT:    vcmpltps db+4096(%rax), %ymm4, %ymm4
+; AVX1-NEXT:    vblendvps %xmm4, %xmm0, %xmm1, %xmm5
+; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm4
+; AVX1-NEXT:    vblendvps %xmm4, %xmm2, %xmm3, %xmm4
+; AVX1-NEXT:    vmovaps %xmm4, dj+4112(%rax)
+; AVX1-NEXT:    vmovaps %xmm5, dj+4096(%rax)
 ; AVX1-NEXT:    addq $32, %rax
 ; AVX1-NEXT:    jne .LBB6_1
 ; AVX1-NEXT:  # %bb.2: # %for.end
diff --git a/llvm/test/CodeGen/X86/known-signbits-vector.ll b/llvm/test/CodeGen/X86/known-signbits-vector.ll
index 57572caf0a2..eacdd24404f 100644
--- a/llvm/test/CodeGen/X86/known-signbits-vector.ll
+++ b/llvm/test/CodeGen/X86/known-signbits-vector.ll
@@ -310,24 +310,23 @@ define <4 x float> @signbits_ashr_sext_select_shuffle_sitofp(<4 x i64> %a0, <4 x
 ; X32-NEXT:    movl %esp, %ebp
 ; X32-NEXT:    andl $-16, %esp
 ; X32-NEXT:    subl $16, %esp
-; X32-NEXT:    vpmovsxdq 16(%ebp), %xmm3
-; X32-NEXT:    vpmovsxdq 8(%ebp), %xmm4
-; X32-NEXT:    vextractf128 $1, %ymm2, %xmm5
-; X32-NEXT:    vpsrlq $33, %xmm5, %xmm5
+; X32-NEXT:    vpmovsxdq 8(%ebp), %xmm3
+; X32-NEXT:    vpmovsxdq 16(%ebp), %xmm4
+; X32-NEXT:    vpsrlq $33, %xmm2, %xmm5
 ; X32-NEXT:    vmovdqa {{.*#+}} xmm6 = [1073741824,0,1,0]
 ; X32-NEXT:    vpxor %xmm6, %xmm5, %xmm5
 ; X32-NEXT:    vpsubq %xmm6, %xmm5, %xmm5
+; X32-NEXT:    vextractf128 $1, %ymm2, %xmm2
 ; X32-NEXT:    vpsrlq $33, %xmm2, %xmm2
 ; X32-NEXT:    vpxor %xmm6, %xmm2, %xmm2
 ; X32-NEXT:    vpsubq %xmm6, %xmm2, %xmm2
-; X32-NEXT:    vinsertf128 $1, %xmm5, %ymm2, %ymm2
-; X32-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm3
-; X32-NEXT:    vextractf128 $1, %ymm1, %xmm4
-; X32-NEXT:    vextractf128 $1, %ymm0, %xmm5
-; X32-NEXT:    vpcmpeqq %xmm4, %xmm5, %xmm4
+; X32-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm6
+; X32-NEXT:    vblendvpd %xmm6, %xmm5, %xmm3, %xmm3
+; X32-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; X32-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; X32-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
-; X32-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm0
-; X32-NEXT:    vblendvpd %ymm0, %ymm2, %ymm3, %ymm0
+; X32-NEXT:    vblendvpd %xmm0, %xmm2, %xmm4, %xmm0
+; X32-NEXT:    vinsertf128 $1, %xmm0, %ymm3, %ymm0
 ; X32-NEXT:    vmovddup {{.*#+}} ymm0 = ymm0[0,0,2,2]
 ; X32-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X32-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]
@@ -339,25 +338,24 @@ define <4 x float> @signbits_ashr_sext_select_shuffle_sitofp(<4 x i64> %a0, <4 x
 ;
 ; X64-LABEL: signbits_ashr_sext_select_shuffle_sitofp:
 ; X64:       # %bb.0:
-; X64-NEXT:    vextractf128 $1, %ymm2, %xmm4
-; X64-NEXT:    vpsrlq $33, %xmm4, %xmm4
+; X64-NEXT:    vpsrlq $33, %xmm2, %xmm4
 ; X64-NEXT:    vmovdqa {{.*#+}} xmm5 = [1073741824,1]
 ; X64-NEXT:    vpxor %xmm5, %xmm4, %xmm4
 ; X64-NEXT:    vpsubq %xmm5, %xmm4, %xmm4
+; X64-NEXT:    vextractf128 $1, %ymm2, %xmm2
 ; X64-NEXT:    vpsrlq $33, %xmm2, %xmm2
 ; X64-NEXT:    vpxor %xmm5, %xmm2, %xmm2
 ; X64-NEXT:    vpsubq %xmm5, %xmm2, %xmm2
-; X64-NEXT:    vinsertf128 $1, %xmm4, %ymm2, %ymm2
-; X64-NEXT:    vpmovsxdq %xmm3, %xmm4
+; X64-NEXT:    vpmovsxdq %xmm3, %xmm5
 ; X64-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[2,3,0,1]
 ; X64-NEXT:    vpmovsxdq %xmm3, %xmm3
-; X64-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm3
-; X64-NEXT:    vextractf128 $1, %ymm1, %xmm4
-; X64-NEXT:    vextractf128 $1, %ymm0, %xmm5
-; X64-NEXT:    vpcmpeqq %xmm4, %xmm5, %xmm4
+; X64-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm6
+; X64-NEXT:    vblendvpd %xmm6, %xmm4, %xmm5, %xmm4
+; X64-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; X64-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; X64-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
-; X64-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm0
-; X64-NEXT:    vblendvpd %ymm0, %ymm2, %ymm3, %ymm0
+; X64-NEXT:    vblendvpd %xmm0, %xmm2, %xmm3, %xmm0
+; X64-NEXT:    vinsertf128 $1, %xmm0, %ymm4, %ymm0
 ; X64-NEXT:    vmovddup {{.*#+}} ymm0 = ymm0[0,0,2,2]
 ; X64-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X64-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]
diff --git a/llvm/test/CodeGen/X86/vselect-avx.ll b/llvm/test/CodeGen/X86/vselect-avx.ll
index 8252084a1cd..cf84c4b7ec8 100644
--- a/llvm/test/CodeGen/X86/vselect-avx.ll
+++ b/llvm/test/CodeGen/X86/vselect-avx.ll
@@ -167,7 +167,7 @@ define <32 x i8> @PR22706(<32 x i1> %x) {
   ret <32 x i8> %tmp
 }
 
-; TODO: Split a 256-bit select into two 128-bit selects when the operands are concatenated.
+; Split a 256-bit select into two 128-bit selects when the operands are concatenated.
 
 define void @blendv_split(<8 x i32>* %p, <8 x i32> %cond, <8 x i32> %a, <8 x i32> %x, <8 x i32> %y, <8 x i32> %z, <8 x i32> %w) {
 ; AVX1-LABEL: blendv_split:
@@ -177,12 +177,13 @@ define void @blendv_split(<8 x i32>* %p, <8 x i32> %cond, <8 x i32> %a, <8 x i32
 ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm4
 ; AVX1-NEXT:    vpslld %xmm2, %xmm4, %xmm5
 ; AVX1-NEXT:    vpslld %xmm2, %xmm1, %xmm2
-; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm2, %ymm2
 ; AVX1-NEXT:    vpslld %xmm3, %xmm4, %xmm4
 ; AVX1-NEXT:    vpslld %xmm3, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm1, %ymm1
-; AVX1-NEXT:    vblendvps %ymm0, %ymm2, %ymm1, %ymm0
-; AVX1-NEXT:    vmovups %ymm0, (%rdi)
+; AVX1-NEXT:    vblendvps %xmm0, %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vblendvps %xmm0, %xmm5, %xmm4, %xmm0
+; AVX1-NEXT:    vmovups %xmm0, 16(%rdi)
+; AVX1-NEXT:    vmovups %xmm1, (%rdi)
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;