Automerge: [X86] getScalarMaskingNode - if the mask is zero just return the blended passthrough and preserved source value (#153575)

We already handle the case if the mask is one, so I added the other case where the op is replaced with a MOVSH/S/D blend.

This assumes the scalar passthrough is op0.

I had to adjust the test case for #98306 as AFAICT it'd been over
reduced

Fixes #153570

(cherry picked from commit ba707db)

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -26236,10 +26236,9 @@ static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
                                     SDValue PreservedSrc,
                                     const X86Subtarget &Subtarget,
                                     SelectionDAG &DAG) {
-
-  if (auto *MaskConst = dyn_cast<ConstantSDNode>(Mask))
-    if (MaskConst->getZExtValue() & 0x1)
-      return Op;
+  auto *MaskConst = dyn_cast<ConstantSDNode>(Mask);
+  if (MaskConst && (MaskConst->getZExtValue() & 0x1))
+    return Op;
 
   MVT VT = Op.getSimpleValueType();
   SDLoc dl(Op);
@@ -26255,6 +26254,17 @@ static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
 
   if (PreservedSrc.isUndef())
     PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
+
+  if (MaskConst) {
+    assert((MaskConst->getZExtValue() & 0x1) == 0 && "Expected false mask");
+    // Discard op and blend passthrough with scalar op src/dst.
+    SmallVector<int, 16> ShuffleMask(VT.getVectorNumElements());
+    std::iota(ShuffleMask.begin(), ShuffleMask.end(), 0);
+    ShuffleMask[0] = VT.getVectorNumElements();
+    return DAG.getVectorShuffle(VT, dl, Op.getOperand(0), PreservedSrc,
+                                ShuffleMask);
+  }
+
   return DAG.getNode(X86ISD::SELECTS, dl, VT, IMask, Op, PreservedSrc);
 }
 

llvm/test/CodeGen/X86/avx512cfmulsh-instrinsics.ll

@@ -278,14 +278,14 @@ define <4 x float> @test_int_x86_avx512fp16_maskz_cfcmadd_sh(<4 x float> %x0, <4
   ret <4 x float> %res
 }
 
-define <4 x float> @PR98306() {
+define <4 x float> @PR98306(i8 %m) {
 ; CHECK-LABEL: PR98306:
 ; CHECK:       ## %bb.0:
-; CHECK-NEXT:    kxorw %k0, %k0, %k1
+; CHECK-NEXT:    kmovd %edi, %k1
 ; CHECK-NEXT:    vmovaps {{.*#+}} xmm1 = [7.8125E-3,1.050912E+6,4.203776E+6,1.6815616E+7]
 ; CHECK-NEXT:    vmovaps {{.*#+}} xmm0 = [3.2E+1,4.03288064E+8,8.0658432E+8,1.61318502E+9]
 ; CHECK-NEXT:    vfmaddcsh {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    retq
-  %res = call <4 x float> @llvm.x86.avx512fp16.maskz.vfmadd.csh(<4 x float> <float 7.812500e-03, float 0x4130092000000000, float 0x4150094000000000, float 0x4170096000000000>, <4 x float> <float 2.000000e+00, float 0x4188098000000000, float 0x4198099000000000, float 0x41A809A000000000>, <4 x float> <float 3.200000e+01, float 0x41B809B000000000, float 0x41C809C000000000, float 0x41D809D000000000>, i8 0, i32 4)
+  %res = call <4 x float> @llvm.x86.avx512fp16.maskz.vfmadd.csh(<4 x float> <float 7.812500e-03, float 0x4130092000000000, float 0x4150094000000000, float 0x4170096000000000>, <4 x float> <float 2.000000e+00, float 0x4188098000000000, float 0x4198099000000000, float 0x41A809A000000000>, <4 x float> <float 3.200000e+01, float 0x41B809B000000000, float 0x41C809C000000000, float 0x41D809D000000000>, i8 %m, i32 4)
   ret <4 x float> %res
 }

llvm/test/CodeGen/X86/avx512fp16-intrinsics.ll

@@ -1361,3 +1361,19 @@ define <32 x half> @test_mm512_castph256_ph512_freeze(<16 x half> %a0) nounwind
   %res = shufflevector <16 x half> %a0, <16 x half> %a1, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
   ret <32 x half> %res
 }
+
+define <8 x half> @PR153570(ptr %p) {
+; CHECK-LABEL: PR153570:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0]
+; CHECK-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
+; CHECK-NEXT:    vmulsh {rn-sae}, %xmm0, %xmm1, %xmm0
+; CHECK-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; CHECK-NEXT:    vmovsh %xmm2, %xmm1, %xmm1
+; CHECK-NEXT:    vmovaps %xmm1, (%rdi)
+; CHECK-NEXT:    retq
+  %r = tail call <8 x half> @llvm.x86.avx512fp16.mask.mul.sh.round(<8 x half> <half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00>, <8 x half> <half 0xH4000, half 0xH4000, half 0xH4000, half 0xH4000, half 0xH4000, half 0xH4000, half 0xH4000, half 0xH4000>, <8 x half> <half 0xH8000, half 0xH8000, half 0xH8000, half 0xH8000, half 0xH8000, half 0xH8000, half 0xH8000, half 0xH8000>, i8 0, i32 8)
+  store <8 x half> %r, ptr %p, align 16
+  %r1 = tail call <8 x half> @llvm.x86.avx512fp16.mask.mul.sh.round(<8 x half> <half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00, half 0xH3C00>, <8 x half> <half 0xH4000, half 0xH4000, half 0xH4000, half 0xH4000, half 0xH4000, half 0xH4000, half 0xH4000, half 0xH4000>, <8 x half> <half 0xH8000, half 0xH8000, half 0xH8000, half 0xH8000, half 0xH8000, half 0xH8000, half 0xH8000, half 0xH8000>, i8 1, i32 8)
+  ret <8 x half> %r1
+}