[llvm-branch-commits] [llvm] release/22.x: [X86] getScalarMaskingNode - FIXUPIMM scalar ops take upper elements from second operand (#179101) (PR #179117)

[via llvm-branch-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-x86

Author: None (llvmbot)

<details>
<summary>Changes</summary>

Backport 618d71dc98df760d0c724cff6fa69b780e8c0372 49d2323447aec77c3d1ae8c941f3f8a126ff1480

Requested by: @<!-- -->RKSimon

---
Full diff: https://github.com/llvm/llvm-project/pull/179117.diff


2 Files Affected:

- (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+6-4) 
- (modified) llvm/test/CodeGen/X86/avx512-intrinsics.ll (+38) 


``````````diff
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index a354704c5958b..5935f2eb344e1 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -26574,7 +26574,8 @@ static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
                                     SDValue PreservedSrc,
                                     const X86Subtarget &Subtarget,
-                                    SelectionDAG &DAG) {
+                                    SelectionDAG &DAG,
+                                    unsigned UpperEltOpSrc = 0) {
   auto *MaskConst = dyn_cast<ConstantSDNode>(Mask);
   if (MaskConst && (MaskConst->getZExtValue() & 0x1))
     return Op;
@@ -26600,8 +26601,8 @@ static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
     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.getVectorShuffle(VT, dl, Op.getOperand(UpperEltOpSrc),
+                                PreservedSrc, ShuffleMask);
   }
 
   return DAG.getNode(X86ISD::SELECTS, dl, VT, IMask, Op, PreservedSrc);
@@ -27262,7 +27263,8 @@ SDValue X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
       if (Opc == X86ISD::VFIXUPIMM || Opc == X86ISD::VFIXUPIMM_SAE)
         return getVectorMaskingNode(FixupImm, Mask, Passthru, Subtarget, DAG);
 
-      return getScalarMaskingNode(FixupImm, Mask, Passthru, Subtarget, DAG);
+      return getScalarMaskingNode(FixupImm, Mask, Passthru, Subtarget, DAG,
+                                  /*UpperEltOpSrc=*/1);
     }
     case ROUNDP: {
       assert(IntrData->Opc0 == X86ISD::VRNDSCALE && "Unexpected opcode");
diff --git a/llvm/test/CodeGen/X86/avx512-intrinsics.ll b/llvm/test/CodeGen/X86/avx512-intrinsics.ll
index f9b5994a18d36..b979f7531cd36 100644
--- a/llvm/test/CodeGen/X86/avx512-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512-intrinsics.ll
@@ -5518,6 +5518,15 @@ define <4 x float>@test_int_x86_avx512_mask_fixupimm_ss(<4 x float> %x0, <4 x fl
   ret <4 x float> %res4
 }
 
+define <4 x float>@test_int_x86_avx512_mask_fixupimm_ss_passthrough_zero_mask(<4 x float> %x0, <4 x float> %x1, <4 x i32> %x2) {
+; CHECK-LABEL: test_int_x86_avx512_mask_fixupimm_ss_passthrough_zero_mask:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
+; CHECK-NEXT:    ret{{[l|q]}}
+  %res = call <4 x float> @llvm.x86.avx512.mask.fixupimm.ss(<4 x float> %x0, <4 x float> %x1, <4 x i32> %x2, i32 5, i8 -2, i32 4)
+  ret <4 x float> %res
+}
+
 declare <4 x float> @llvm.x86.avx512.maskz.fixupimm.ss(<4 x float>, <4 x float>, <4 x i32>, i32, i8, i32)
 
 define <4 x float>@test_int_x86_avx512_maskz_fixupimm_ss(<4 x float> %x0, <4 x float> %x1, <4 x i32> %x2, i8 %x4) {
@@ -5555,6 +5564,16 @@ define <4 x float>@test_int_x86_avx512_maskz_fixupimm_ss(<4 x float> %x0, <4 x f
   ret <4 x float> %res4
 }
 
+define <4 x float>@test_int_x86_avx512_maskz_fixupimm_ss_passthrough_zero_mask(<4 x float> %x0, <4 x float> %x1, <4 x i32> %x2) {
+; CHECK-LABEL: test_int_x86_avx512_maskz_fixupimm_ss_passthrough_zero_mask:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vxorps %xmm0, %xmm0, %xmm0
+; CHECK-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
+; CHECK-NEXT:    ret{{[l|q]}}
+  %res = call <4 x float> @llvm.x86.avx512.maskz.fixupimm.ss(<4 x float> %x0, <4 x float> %x1, <4 x i32> %x2, i32 5, i8 -2, i32 4)
+  ret <4 x float> %res
+}
+
 declare <16 x float> @llvm.x86.avx512.mask.fixupimm.ps.512(<16 x float>, <16 x float>, <16 x i32>, i32, i16, i32)
 
 define <16 x float>@test_int_x86_avx512_mask_fixupimm_ps_512(<16 x float> %x0, <16 x float> %x1, <16 x i32> %x2, i16 %x4) {
@@ -5680,6 +5699,15 @@ define <2 x double>@test_int_x86_avx512_mask_fixupimm_sd(<2 x double> %x0, <2 x
   ret <2 x double> %res4
 }
 
+define <2 x double>@test_int_x86_avx512_mask_fixupimm_sd_passthrough_zero_mask(<2 x double> %x0, <2 x double> %x1, <2 x i64> %x2) {
+; CHECK-LABEL: test_int_x86_avx512_mask_fixupimm_sd_passthrough_zero_mask:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vmovsd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
+; CHECK-NEXT:    ret{{[l|q]}}
+  %res = call <2 x double> @llvm.x86.avx512.mask.fixupimm.sd(<2 x double> %x0, <2 x double> %x1, <2 x i64> %x2, i32 5, i8 -2, i32 4)
+  ret <2 x double> %res
+}
+
 declare <2 x double> @llvm.x86.avx512.maskz.fixupimm.sd(<2 x double>, <2 x double>, <2 x i64>, i32, i8, i32)
 
 define <2 x double>@test_int_x86_avx512_maskz_fixupimm_sd(<2 x double> %x0, <2 x double> %x1, <2 x i64> %x2, i8 %x4) {
@@ -5717,6 +5745,16 @@ define <2 x double>@test_int_x86_avx512_maskz_fixupimm_sd(<2 x double> %x0, <2 x
   ret <2 x double> %res4
 }
 
+define <2 x double>@test_int_x86_avx512_maskz_fixupimm_sd_passthrough_zero_mask(<2 x double> %x0, <2 x double> %x1, <2 x i64> %x2) {
+; CHECK-LABEL: test_int_x86_avx512_maskz_fixupimm_sd_passthrough_zero_mask:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vxorps %xmm0, %xmm0, %xmm0
+; CHECK-NEXT:    vmovsd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
+; CHECK-NEXT:    ret{{[l|q]}}
+  %res = call <2 x double> @llvm.x86.avx512.maskz.fixupimm.sd(<2 x double> %x0, <2 x double> %x1, <2 x i64> %x2, i32 5, i8 -2, i32 4)
+  ret <2 x double> %res
+}
+
 declare double @llvm.fma.f64(double, double, double) #1
 declare double @llvm.x86.avx512.vfmadd.f64(double, double, double, i32) #0
 

``````````

</details>


https://github.com/llvm/llvm-project/pull/179117