[X86] Add a DAG combines to turn vXi64 muls into VPMULDQ/VPMULUDQ if the upper bits...

Normally we catch this during lowering, but vXi64 mul is considered legal when we have AVX512DQ.

This DAG combine allows us to avoid PMULLQ with AVX512DQ if we can prove its unnecessary. PMULLQ is 3 uops that take 4 cycles each. While pmuldq/pmuludq is only one 4 cycle uop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@321437 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 363cdee..10dcf21 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -32423,6 +32423,37 @@ static SDValue combineMulSpecial(uint64_t MulAmt, SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+static SDValue combineVMUL(SDNode *N, SelectionDAG &DAG,
+                           const X86Subtarget &Subtarget) {
+  EVT VT = N->getValueType(0);
+  SDLoc dl(N);
+
+  if (VT.getScalarType() != MVT::i64)
+    return SDValue();
+
+  MVT MulVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
+
+  SDValue LHS = N->getOperand(0);
+  SDValue RHS = N->getOperand(1);
+
+  // MULDQ returns the 64-bit result of the signed multiplication of the lower
+  // 32-bits. We can lower with this if the sign bits stretch that far.
+  if (Subtarget.hasSSE41() && DAG.ComputeNumSignBits(LHS) > 32 &&
+      DAG.ComputeNumSignBits(RHS) > 32) {
+    return DAG.getNode(X86ISD::PMULDQ, dl, VT, DAG.getBitcast(MulVT, LHS),
+                       DAG.getBitcast(MulVT, RHS));
+  }
+
+  // If the upper bits are zero we can use a single pmuludq.
+  APInt Mask = APInt::getHighBitsSet(64, 32);
+  if (DAG.MaskedValueIsZero(LHS, Mask) && DAG.MaskedValueIsZero(RHS, Mask)) {
+    return DAG.getNode(X86ISD::PMULUDQ, dl, VT, DAG.getBitcast(MulVT, LHS),
+                       DAG.getBitcast(MulVT, RHS));
+  }
+
+  return SDValue();
+}
+
 /// Optimize a single multiply with constant into two operations in order to
 /// implement it with two cheaper instructions, e.g. LEA + SHL, LEA + LEA.
 static SDValue combineMul(SDNode *N, SelectionDAG &DAG,
@@ -32432,6 +32463,9 @@ static SDValue combineMul(SDNode *N, SelectionDAG &DAG,
   if (DCI.isBeforeLegalize() && VT.isVector())
     return reduceVMULWidth(N, DAG, Subtarget);
 
+  if (!DCI.isBeforeLegalize() && VT.isVector())
+    return combineVMUL(N, DAG, Subtarget);
+
   if (!MulConstantOptimization)
     return SDValue();
   // An imul is usually smaller than the alternative sequence.

diff --git a/test/CodeGen/X86/combine-pmuldq.ll b/test/CodeGen/X86/combine-pmuldq.ll
index aa318c6..421c948 100644 (file)
--- a/test/CodeGen/X86/combine-pmuldq.ll
+++ b/test/CodeGen/X86/combine-pmuldq.ll
@@ -15,32 +15,14 @@ define <2 x i64> @combine_shuffle_sext_pmuldq(<4 x i32> %a0, <4 x i32> %a1) {
 ; SSE-NEXT:    pmuldq %xmm2, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX2-LABEL: combine_shuffle_sext_pmuldq:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
-; AVX2-NEXT:    vpmovsxdq %xmm0, %xmm0
-; AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
-; AVX2-NEXT:    vpmovsxdq %xmm1, %xmm1
-; AVX2-NEXT:    vpmuldq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    retq
-;
-; AVX512VL-LABEL: combine_shuffle_sext_pmuldq:
-; AVX512VL:       # %bb.0:
-; AVX512VL-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
-; AVX512VL-NEXT:    vpmovsxdq %xmm0, %xmm0
-; AVX512VL-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
-; AVX512VL-NEXT:    vpmovsxdq %xmm1, %xmm1
-; AVX512VL-NEXT:    vpmuldq %xmm1, %xmm0, %xmm0
-; AVX512VL-NEXT:    retq
-;
-; AVX512DQVL-LABEL: combine_shuffle_sext_pmuldq:
-; AVX512DQVL:       # %bb.0:
-; AVX512DQVL-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
-; AVX512DQVL-NEXT:    vpmovsxdq %xmm0, %xmm0
-; AVX512DQVL-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
-; AVX512DQVL-NEXT:    vpmovsxdq %xmm1, %xmm1
-; AVX512DQVL-NEXT:    vpmullq %xmm1, %xmm0, %xmm0
-; AVX512DQVL-NEXT:    retq
+; AVX-LABEL: combine_shuffle_sext_pmuldq:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
+; AVX-NEXT:    vpmovsxdq %xmm0, %xmm0
+; AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
+; AVX-NEXT:    vpmovsxdq %xmm1, %xmm1
+; AVX-NEXT:    vpmuldq %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
   %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <2 x i32> <i32 0, i32 2>
   %2 = shufflevector <4 x i32> %a1, <4 x i32> undef, <2 x i32> <i32 0, i32 2>
   %3 = sext <2 x i32> %1 to <2 x i64>
@@ -60,32 +42,14 @@ define <2 x i64> @combine_shuffle_zext_pmuludq(<4 x i32> %a0, <4 x i32> %a1) {
 ; SSE-NEXT:    pmuludq %xmm2, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX2-LABEL: combine_shuffle_zext_pmuludq:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
-; AVX2-NEXT:    vpmuludq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    retq
-;
-; AVX512VL-LABEL: combine_shuffle_zext_pmuludq:
-; AVX512VL:       # %bb.0:
-; AVX512VL-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
-; AVX512VL-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX512VL-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
-; AVX512VL-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
-; AVX512VL-NEXT:    vpmuludq %xmm1, %xmm0, %xmm0
-; AVX512VL-NEXT:    retq
-;
-; AVX512DQVL-LABEL: combine_shuffle_zext_pmuludq:
-; AVX512DQVL:       # %bb.0:
-; AVX512DQVL-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
-; AVX512DQVL-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX512DQVL-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
-; AVX512DQVL-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
-; AVX512DQVL-NEXT:    vpmullq %xmm1, %xmm0, %xmm0
-; AVX512DQVL-NEXT:    retq
+; AVX-LABEL: combine_shuffle_zext_pmuludq:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
+; AVX-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
+; AVX-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
+; AVX-NEXT:    vpmuludq %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
   %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <2 x i32> <i32 0, i32 2>
   %2 = shufflevector <4 x i32> %a1, <4 x i32> undef, <2 x i32> <i32 0, i32 2>
   %3 = zext <2 x i32> %1 to <2 x i64>

diff --git a/test/CodeGen/X86/masked_gather_scatter.ll b/test/CodeGen/X86/masked_gather_scatter.ll
index d318dde..d3521ca 100644 (file)
--- a/test/CodeGen/X86/masked_gather_scatter.ll
+++ b/test/CodeGen/X86/masked_gather_scatter.ll
@@ -497,7 +497,7 @@ define <8 x i32> @test9(%struct.ST* %base, <8 x i64> %ind1, <8 x i32>%ind5) {
 ; SKX_SMALL-NEXT:    vpbroadcastq %rdi, %zmm2
 ; SKX_SMALL-NEXT:    vpmullq {{.*}}(%rip){1to8}, %zmm0, %zmm0
 ; SKX_SMALL-NEXT:    vpmovsxdq %ymm1, %zmm1
-; SKX_SMALL-NEXT:    vpmullq {{.*}}(%rip){1to8}, %zmm1, %zmm1
+; SKX_SMALL-NEXT:    vpmuldq {{.*}}(%rip){1to8}, %zmm1, %zmm1
 ; SKX_SMALL-NEXT:    vpaddq %zmm1, %zmm0, %zmm0
 ; SKX_SMALL-NEXT:    vpaddq %zmm0, %zmm2, %zmm0
 ; SKX_SMALL-NEXT:    vpaddq {{.*}}(%rip){1to8}, %zmm0, %zmm1
@@ -510,7 +510,7 @@ define <8 x i32> @test9(%struct.ST* %base, <8 x i64> %ind1, <8 x i32>%ind5) {
 ; SKX_LARGE-NEXT:    vpbroadcastq %rdi, %zmm2
 ; SKX_LARGE-NEXT:    vpmovsxdq %ymm1, %zmm1
 ; SKX_LARGE-NEXT:    movabsq ${{\.LCPI.*}}, %rax
-; SKX_LARGE-NEXT:    vpmullq (%rax){1to8}, %zmm1, %zmm1
+; SKX_LARGE-NEXT:    vpmuldq (%rax){1to8}, %zmm1, %zmm1
 ; SKX_LARGE-NEXT:    movabsq ${{\.LCPI.*}}, %rax
 ; SKX_LARGE-NEXT:    vpmullq (%rax){1to8}, %zmm0, %zmm0
 ; SKX_LARGE-NEXT:    vpaddq %zmm1, %zmm0, %zmm0
@@ -582,7 +582,7 @@ define <8 x i32> @test10(%struct.ST* %base, <8 x i64> %i1, <8 x i32>%ind5) {
 ; SKX_SMALL-NEXT:    vpbroadcastq %rdi, %zmm2
 ; SKX_SMALL-NEXT:    vpmullq {{.*}}(%rip){1to8}, %zmm0, %zmm0
 ; SKX_SMALL-NEXT:    vpmovsxdq %ymm1, %zmm1
-; SKX_SMALL-NEXT:    vpmullq {{.*}}(%rip){1to8}, %zmm1, %zmm1
+; SKX_SMALL-NEXT:    vpmuldq {{.*}}(%rip){1to8}, %zmm1, %zmm1
 ; SKX_SMALL-NEXT:    vpaddq %zmm1, %zmm0, %zmm0
 ; SKX_SMALL-NEXT:    vpaddq %zmm0, %zmm2, %zmm0
 ; SKX_SMALL-NEXT:    vpaddq {{.*}}(%rip){1to8}, %zmm0, %zmm1
@@ -595,7 +595,7 @@ define <8 x i32> @test10(%struct.ST* %base, <8 x i64> %i1, <8 x i32>%ind5) {
 ; SKX_LARGE-NEXT:    vpbroadcastq %rdi, %zmm2
 ; SKX_LARGE-NEXT:    vpmovsxdq %ymm1, %zmm1
 ; SKX_LARGE-NEXT:    movabsq ${{\.LCPI.*}}, %rax
-; SKX_LARGE-NEXT:    vpmullq (%rax){1to8}, %zmm1, %zmm1
+; SKX_LARGE-NEXT:    vpmuldq (%rax){1to8}, %zmm1, %zmm1
 ; SKX_LARGE-NEXT:    movabsq ${{\.LCPI.*}}, %rax
 ; SKX_LARGE-NEXT:    vpmullq (%rax){1to8}, %zmm0, %zmm0
 ; SKX_LARGE-NEXT:    vpaddq %zmm1, %zmm0, %zmm0