[X86] Invert a vector select IR canonicalization with a binop identity constant

[LuoYuanke]

Summary

This patch follows https://reviews.llvm.org/D118644 to invert fmul and
fdiv in X86 backend when AVX512 is available.

[pengfei]

llvm/lib/Target/X86/X86ISelLowering.cpp

48950

/

49012–49015

Equals to return combineBinopWithSelect(N, DAG, Subtarget)?

53883

Or call combineBinopWithSelect directly?

llvm/test/CodeGen/X86/vector-bo-select.ll

345–347

Curiously: it should be equal to

vmulps %zmm2, %zmm1, %zmm1 {%k1}
vmovaps %zmm1, %zmm0

Why sometimes we use this way, sometime another?

[LuoYuanke]

llvm/lib/Target/X86/X86ISelLowering.cpp

48950

Sorry, I don't understand this comment.

49012–49015

OK, I'll update it.

53883

I prefer to following the current coding convention, so that when there is more to combine we can extent the code in the sub-function.

llvm/test/CodeGen/X86/vector-bo-select.ll

345–347

I guess it is because sometime it is commuted or swapped, sometimes it is not.

[pengfei]

llvm/lib/Target/X86/X86ISelLowering.cpp

48950

X / 1.0 --> X

[LuoYuanke]

llvm/lib/Target/X86/X86ISelLowering.cpp

48950

Got it. :)

[LuoYuanke]

Address Phoebe's comments.

[spatel]

Thanks for working on this.
I have a patch to move the existing code over to DAGCombiner with a TLI hook that is only enabled for x86 currently, so it would be effectively NFC.
That would allow us to avoid adding code for custom combining fmul/fdiv to x86 (just add cases for the new opcodes in DAGCombiner).
I'm not sure if one way is more efficient than the other, but let me post that for review.

[spatel]

See D119150.

I think this patch is good either way, but there would be less diffs if we move the existing code over as a first step.

[LuoYuanke]

In D119111#3301516, @spatel wrote:

See D119150.

I think this patch is good either way, but there would be less diffs if we move the existing code over as a first step.

Ok, we can improve more operator based on D119150. I'll abandon this patch.

[spatel]

Added FMUL/FDIV to the updated code here:
905abc5b7db2 (test diffs are identical)

[LuoYuanke]

In D119111#3304769, @spatel wrote:

Added FMUL/FDIV to the updated code here:
905abc5b7db2 (test diffs are identical)

Thanks, Sanjay.

llvm/lib/Target/X86/X86ISelLowering.cpp

   setTargetDAGCombine(ISD::MUL);
   setTargetDAGCombine(ISD::XOR);
   setTargetDAGCombine(ISD::MSCATTER);
   setTargetDAGCombine(ISD::MGATHER);
   setTargetDAGCombine(ISD::FP16_TO_FP);
   setTargetDAGCombine(ISD::FP_EXTEND);
   setTargetDAGCombine(ISD::STRICT_FP_EXTEND);
   setTargetDAGCombine(ISD::FP_ROUND);
+  setTargetDAGCombine(ISD::FMUL);
+  setTargetDAGCombine(ISD::FDIV);
 
   computeRegisterProperties(Subtarget.getRegisterInfo());
 
   MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
   MaxStoresPerMemsetOptSize = 8;
   MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
   MaxStoresPerMemcpyOptSize = 4;
   MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
static SDValue foldSelectWithIdentityConstant(SDNode *N, SelectionDAG &DAG,
   // TODO: With fast-math (NSZ), allow the opposite-sign form of zero?
   auto isIdentityConstantForOpcode = [](unsigned Opcode, SDValue V) {
     if (ConstantFPSDNode *C = isConstOrConstSplatFP(V)) {
       switch (Opcode) {
       case ISD::FADD: // X + -0.0 --> X
         return C->isZero() && C->isNegative();
       case ISD::FSUB: // X - 0.0 --> X
         return C->isZero() && !C->isNegative();
+      case ISD::FMUL: // X * 1.0 --> X
+      case ISD::FDIV: // X / 1.0 --> X
+        return C->isExactlyValue(1.0);
       }
     }
     return false;
   };
 
   // This transform increases uses of N0, so freeze it to be safe.
   // binop N0, (vselect Cond, IDC, FVal) --> vselect Cond, N0, (binop N0, FVal)
   if (isIdentityConstantForOpcode(Opcode, TVal)) {
if (SDValue COp = combineFaddCFmul(N, DAG, Subtarget))
     return COp;
 
   if (SDValue Sel = combineBinopWithSelect(N, DAG, Subtarget))
     return Sel;
 
   return SDValue();
 }
 
+static SDValue combineFmul(SDNode *N, SelectionDAG &DAG,
+                           const X86Subtarget &Subtarget) {
+  return combineBinopWithSelect(N, DAG, Subtarget);
+}
+
+static SDValue combineFdiv(SDNode *N, SelectionDAG &DAG,
+                           const X86Subtarget &Subtarget) {
+  return combineBinopWithSelect(N, DAG, Subtarget);
+}
+
 /// Attempt to pre-truncate inputs to arithmetic ops if it will simplify
 /// the codegen.
 /// e.g. TRUNC( BINOP( X, Y ) ) --> BINOP( TRUNC( X ), TRUNC( Y ) )
 /// TODO: This overlaps with the generic combiner's visitTRUNCATE. Remove
 ///       anything that is guaranteed to be transformed by DAGCombiner.
 static SDValue combineTruncatedArithmetic(SDNode *N, SelectionDAG &DAG,
                                           const X86Subtarget &Subtarget,
                                           const SDLoc &DL) {
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::SINT_TO_FP:
   case ISD::STRICT_SINT_TO_FP:
     return combineSIntToFP(N, DAG, DCI, Subtarget);
   case ISD::UINT_TO_FP:
   case ISD::STRICT_UINT_TO_FP:
     return combineUIntToFP(N, DAG, Subtarget);
   case ISD::FADD:
   case ISD::FSUB:           return combineFaddFsub(N, DAG, Subtarget);
+  case ISD::FMUL:           return combineFmul(N, DAG, Subtarget);
+  case ISD::FDIV:           return combineFdiv(N, DAG, Subtarget);
   case X86ISD::VFCMULC:
   case X86ISD::VFMULC:      return combineFMulcFCMulc(N, DAG, Subtarget);
   case ISD::FNEG:           return combineFneg(N, DAG, DCI, Subtarget);
   case ISD::TRUNCATE:       return combineTruncate(N, DAG, Subtarget);
   case X86ISD::VTRUNC:      return combineVTRUNC(N, DAG, DCI);
   case X86ISD::ANDNP:       return combineAndnp(N, DAG, DCI, Subtarget);
   case X86ISD::FAND:        return combineFAnd(N, DAG, Subtarget);
   case X86ISD::FANDN:       return combineFAndn(N, DAG, Subtarget);

llvm/test/CodeGen/X86/vector-bo-select.ll

 ; AVX512F-NEXT:    vmulps %xmm0, %xmm1, %xmm0
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: fmul_v4f32:
 ; AVX512VL:       # %bb.0:
 ; AVX512VL-NEXT:    vpslld $31, %xmm0, %xmm0
 ; AVX512VL-NEXT:    vptestmd %xmm0, %xmm0, %k1
-; AVX512VL-NEXT:    vbroadcastss {{.*#+}} xmm0 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
-; AVX512VL-NEXT:    vmovaps %xmm2, %xmm0 {%k1}
-; AVX512VL-NEXT:    vmulps %xmm0, %xmm1, %xmm0
+; AVX512VL-NEXT:    vmulps %xmm2, %xmm1, %xmm1 {%k1}
+; AVX512VL-NEXT:    vmovaps %xmm1, %xmm0
 ; AVX512VL-NEXT:    retq
   %s = select <4 x i1> %b, <4 x float> %y, <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>
   %r = fmul <4 x float> %x, %s
   ret <4 x float> %r
 }
 
 define <8 x float> @fmul_v8f32_commute(<8 x i1> %b, <8 x float> noundef %x, <8 x float> noundef %y) {
 ; AVX2-LABEL: fmul_v8f32_commute:

 ; AVX512F-NEXT:    vmulps %ymm1, %ymm0, %ymm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: fmul_v8f32_commute:
 ; AVX512VL:       # %bb.0:
 ; AVX512VL-NEXT:    vpmovsxwd %xmm0, %ymm0
 ; AVX512VL-NEXT:    vpslld $31, %ymm0, %ymm0
 ; AVX512VL-NEXT:    vptestmd %ymm0, %ymm0, %k1
-; AVX512VL-NEXT:    vbroadcastss {{.*#+}} ymm0 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
-; AVX512VL-NEXT:    vmovaps %ymm2, %ymm0 {%k1}
-; AVX512VL-NEXT:    vmulps %ymm1, %ymm0, %ymm0
+; AVX512VL-NEXT:    vmulps %ymm2, %ymm1, %ymm1 {%k1}
+; AVX512VL-NEXT:    vmovaps %ymm1, %ymm0
 ; AVX512VL-NEXT:    retq
   %s = select <8 x i1> %b, <8 x float> %y, <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>
   %r = fmul <8 x float> %s, %x
   ret <8 x float> %r
 }
 
 define <16 x float> @fmul_v16f32_swap(<16 x i1> %b, <16 x float> noundef %x, <16 x float> noundef %y) {
 ; AVX2-LABEL: fmul_v16f32_swap:

 ; AVX2-NEXT:    vmulps %ymm4, %ymm2, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: fmul_v16f32_swap:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; AVX512-NEXT:    vpslld $31, %zmm0, %zmm0
 ; AVX512-NEXT:    vptestmd %zmm0, %zmm0, %k1
-; AVX512-NEXT:    vbroadcastss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2 {%k1}
 ; AVX512-NEXT:    vmulps %zmm2, %zmm1, %zmm0
+; AVX512-NEXT:    vmovaps %zmm1, %zmm0 {%k1}
 ; AVX512-NEXT:    retq
   %s = select <16 x i1> %b, <16 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, <16 x float> %y
   %r = fmul <16 x float> %x, %s
   ret <16 x float> %r
 }
 
 define <16 x float> @fmul_v16f32_commute_swap(<16 x i1> %b, <16 x float> noundef %x, <16 x float> noundef %y) {
 ; AVX2-LABEL: fmul_v16f32_commute_swap:
 ; AVX2:       # %bb.0:

 ; AVX2-NEXT:    vmulps %ymm2, %ymm4, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: fmul_v16f32_commute_swap:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; AVX512-NEXT:    vpslld $31, %zmm0, %zmm0
 ; AVX512-NEXT:    vptestmd %zmm0, %zmm0, %k1
-; AVX512-NEXT:    vbroadcastss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2 {%k1}
-; AVX512-NEXT:    vmulps %zmm1, %zmm2, %zmm0
+; AVX512-NEXT:    vmulps %zmm2, %zmm1, %zmm0
+; AVX512-NEXT:    vmovaps %zmm1, %zmm0 {%k1}
 ; AVX512-NEXT:    retq
   %s = select <16 x i1> %b, <16 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, <16 x float> %y
   %r = fmul <16 x float> %s, %x
   ret <16 x float> %r
 }
 
 define <4 x float> @fdiv_v4f32(<4 x i1> %b, <4 x float> noundef %x, <4 x float> noundef %y) {
 ; AVX2-LABEL: fdiv_v4f32:

 ; AVX512F-NEXT:    vdivps %xmm0, %xmm1, %xmm0
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: fdiv_v4f32:
 ; AVX512VL:       # %bb.0:
 ; AVX512VL-NEXT:    vpslld $31, %xmm0, %xmm0
 ; AVX512VL-NEXT:    vptestmd %xmm0, %xmm0, %k1
-; AVX512VL-NEXT:    vbroadcastss {{.*#+}} xmm0 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
-; AVX512VL-NEXT:    vmovaps %xmm2, %xmm0 {%k1}
-; AVX512VL-NEXT:    vdivps %xmm0, %xmm1, %xmm0
+; AVX512VL-NEXT:    vdivps %xmm2, %xmm1, %xmm1 {%k1}
+; AVX512VL-NEXT:    vmovaps %xmm1, %xmm0
 ; AVX512VL-NEXT:    retq
   %s = select <4 x i1> %b, <4 x float> %y, <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>
   %r = fdiv <4 x float> %x, %s
   ret <4 x float> %r
 }
 
 define <8 x float> @fdiv_v8f32_commute(<8 x i1> %b, <8 x float> noundef %x, <8 x float> noundef %y) {
 ; AVX2-LABEL: fdiv_v8f32_commute:

 ; AVX2-NEXT:    vdivps %ymm4, %ymm2, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: fdiv_v16f32_swap:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; AVX512-NEXT:    vpslld $31, %zmm0, %zmm0
 ; AVX512-NEXT:    vptestmd %zmm0, %zmm0, %k1
-; AVX512-NEXT:    vbroadcastss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2 {%k1}
 ; AVX512-NEXT:    vdivps %zmm2, %zmm1, %zmm0
+; AVX512-NEXT:    vmovaps %zmm1, %zmm0 {%k1}
 ; AVX512-NEXT:    retq
   %s = select <16 x i1> %b, <16 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, <16 x float> %y
   %r = fdiv <16 x float> %x, %s
   ret <16 x float> %r
 }
 
 define <16 x float> @fdiv_v16f32_commute_swap(<16 x i1> %b, <16 x float> noundef %x, <16 x float> noundef %y) {
 ; AVX2-LABEL: fdiv_v16f32_commute_swap:

 ; AVX512F-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
 ; AVX512F-NEXT:    vmovaps %zmm1, %zmm2 {%k1}
 ; AVX512F-NEXT:    vmulps %ymm2, %ymm0, %ymm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: fmul_v8f32_cast_cond:
 ; AVX512VL:       # %bb.0:
 ; AVX512VL-NEXT:    kmovw %edi, %k1
-; AVX512VL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
-; AVX512VL-NEXT:    vmovaps %ymm1, %ymm2 {%k1}
-; AVX512VL-NEXT:    vmulps %ymm2, %ymm0, %ymm0
+; AVX512VL-NEXT:    vmulps %ymm1, %ymm0, %ymm0 {%k1}
 ; AVX512VL-NEXT:    retq
   %b = bitcast i8 %pb to <8 x i1>
   %s = select <8 x i1> %b, <8 x float> %y, <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>
   %r = fmul <8 x float> %x, %s
   ret <8 x float> %r
 }
 
 define <8 x double> @fmul_v8f64_cast_cond(i8 noundef zeroext %pb, <8 x double> noundef %x, <8 x double> noundef %y) {

 ; AVX2-NEXT:    vblendvpd %ymm4, %ymm2, %ymm6, %ymm2
 ; AVX2-NEXT:    vmulpd %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vmulpd %ymm3, %ymm1, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: fmul_v8f64_cast_cond:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovw %edi, %k1
-; AVX512-NEXT:    vbroadcastsd {{.*#+}} zmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
-; AVX512-NEXT:    vmovapd %zmm1, %zmm2 {%k1}
-; AVX512-NEXT:    vmulpd %zmm2, %zmm0, %zmm0
+; AVX512-NEXT:    vmulpd %zmm1, %zmm0, %zmm0 {%k1}
 ; AVX512-NEXT:    retq
   %b = bitcast i8 %pb to <8 x i1>
   %s = select <8 x i1> %b, <8 x double> %y, <8 x double> <double 1.0, double 1.0, double 1.0, double 1.0, double 1.0, double 1.0, double 1.0, double 1.0>
   %r = fmul <8 x double> %x, %s
   ret <8 x double> %r
 }
 
 define <8 x float> @fdiv_v8f32_cast_cond(i8 noundef zeroext %pb, <8 x float> noundef %x, <8 x float> noundef %y) {

 ; AVX512F-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
 ; AVX512F-NEXT:    vmovaps %zmm1, %zmm2 {%k1}
 ; AVX512F-NEXT:    vdivps %ymm2, %ymm0, %ymm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: fdiv_v8f32_cast_cond:
 ; AVX512VL:       # %bb.0:
 ; AVX512VL-NEXT:    kmovw %edi, %k1
-; AVX512VL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
-; AVX512VL-NEXT:    vmovaps %ymm1, %ymm2 {%k1}
-; AVX512VL-NEXT:    vdivps %ymm2, %ymm0, %ymm0
+; AVX512VL-NEXT:    vdivps %ymm1, %ymm0, %ymm0 {%k1}
 ; AVX512VL-NEXT:    retq
   %b = bitcast i8 %pb to <8 x i1>
   %s = select <8 x i1> %b, <8 x float> %y, <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>
   %r = fdiv <8 x float> %x, %s
   ret <8 x float> %r
 }
 
 define <8 x double> @fdiv_v8f64_cast_cond(i8 noundef zeroext %pb, <8 x double> noundef %x, <8 x double> noundef %y) {

 ; AVX2-NEXT:    vblendvpd %ymm4, %ymm2, %ymm6, %ymm2
 ; AVX2-NEXT:    vdivpd %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vdivpd %ymm3, %ymm1, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: fdiv_v8f64_cast_cond:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    kmovw %edi, %k1
-; AVX512-NEXT:    vbroadcastsd {{.*#+}} zmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
-; AVX512-NEXT:    vmovapd %zmm1, %zmm2 {%k1}
-; AVX512-NEXT:    vdivpd %zmm2, %zmm0, %zmm0
+; AVX512-NEXT:    vdivpd %zmm1, %zmm0, %zmm0 {%k1}
 ; AVX512-NEXT:    retq
   %b = bitcast i8 %pb to <8 x i1>
   %s = select <8 x i1> %b, <8 x double> %y, <8 x double> <double 1.0, double 1.0, double 1.0, double 1.0, double 1.0, double 1.0, double 1.0, double 1.0>
   %r = fdiv <8 x double> %x, %s
   ret <8 x double> %r
 }