[llvm] [X86] Widden binary shuffle to unary for i8/i16 (PR #96414)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-x86

Author: Phoebe Wang (phoebewang)

<details>
<summary>Changes</summary>

VPERM[I,T]2[B,W] are 3 uops on Skylake and Icelake so we try to use VPERMV.

---
Full diff: https://github.com/llvm/llvm-project/pull/96414.diff


2 Files Affected:

- (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+19-1) 
- (modified) llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll (+1-2) 


``````````diff
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 82d2b301d854e..c4d08206593f0 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -14123,7 +14123,25 @@ static SDValue lowerShuffleWithPERMV(const SDLoc &DL, MVT VT,
   MVT MaskVT = VT.changeTypeToInteger();
   SDValue MaskNode;
   MVT ShuffleVT = VT;
-  if (!VT.is512BitVector() && !Subtarget.hasVLX()) {
+  MVT ElementVT = VT.getVectorElementType();
+  // VPERM[I,T]2[B,W] are 3 uops on Skylake and Icelake so we try to use VPERMV.
+  if (Subtarget.hasVBMI() && !V2.isUndef() &&
+      (VT.is128BitVector() ||
+       (VT.is256BitVector() && Subtarget.hasEVEX512())) &&
+      (ElementVT == MVT::i8 || ElementVT == MVT::i16) &&
+      V1.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
+      V1.getConstantOperandVal(1) == 0 &&
+      V2.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
+      V2.getConstantOperandVal(1) == VT.getVectorNumElements() &&
+      V1.getOperand(0) == V2.getOperand(0)) {
+    SmallVector<int, 32> AdjustedMask(Mask);
+    AdjustedMask.resize(VT.getVectorNumElements() * 2, SM_SentinelUndef);
+    ShuffleVT = VT.getDoubleNumVectorElementsVT();
+    V1 = V1.getOperand(0);
+    V2 = DAG.getUNDEF(ShuffleVT);
+    MaskNode = getConstVector(
+        AdjustedMask, MaskVT.getDoubleNumVectorElementsVT(), DAG, DL, true);
+  } else if (!VT.is512BitVector() && !Subtarget.hasVLX()) {
     V1 = widenSubVector(V1, false, Subtarget, DAG, DL, 512);
     V2 = widenSubVector(V2, false, Subtarget, DAG, DL, 512);
     ShuffleVT = V1.getSimpleValueType();
diff --git a/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll b/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
index a481aaef4257d..95e249984e184 100644
--- a/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
+++ b/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
@@ -348,8 +348,7 @@ define <16 x i8> @trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_
 ; AVX512VBMIVL-LABEL: trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_57_62:
 ; AVX512VBMIVL:       # %bb.0:
 ; AVX512VBMIVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [1,5,9,13,17,21,25,29,33,37,41,45,49,53,57,62]
-; AVX512VBMIVL-NEXT:    vextracti64x4 $1, %zmm0, %ymm2
-; AVX512VBMIVL-NEXT:    vpermt2b %ymm2, %ymm1, %ymm0
+; AVX512VBMIVL-NEXT:    vpermb %zmm0, %zmm1, %zmm0
 ; AVX512VBMIVL-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
 ; AVX512VBMIVL-NEXT:    vzeroupper
 ; AVX512VBMIVL-NEXT:    retq

``````````

</details>


https://github.com/llvm/llvm-project/pull/96414