[X86] combineINSERT_SUBVECTOR - attempt to constant fold from constant pool loads (if we're not widening).

[RKSimon]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/129682.diff

2 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+15)
• (modified) llvm/test/CodeGen/X86/vector-shuffle-combining-xop.ll (+5-10)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 0f82b541744d6..bbab43d4e92af 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -58684,6 +58684,21 @@ static SDValue combineINSERT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
     }
   }
 
+  // Attempt to constant fold (if we're not widening).
+  if (!Vec.isUndef() && !ISD::isBuildVectorAllZeros(Vec.getNode())) {
+    unsigned EltSizeInBits = OpVT.getScalarSizeInBits();
+    APInt VecUndefElts, SubUndefElts;
+    SmallVector<APInt, 16> VecEltBits, SubEltBits;
+    if (getTargetConstantBitsFromNode(Vec, EltSizeInBits, VecUndefElts,
+                                      VecEltBits) &&
+        getTargetConstantBitsFromNode(SubVec, EltSizeInBits, SubUndefElts,
+                                      SubEltBits)) {
+      VecUndefElts.insertBits(SubUndefElts, IdxVal);
+      llvm::copy(SubEltBits, VecEltBits.begin() + IdxVal);
+      return getConstVector(VecEltBits, VecUndefElts, OpVT, DAG, dl);
+    }
+  }
+
   return SDValue();
 }
 
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-xop.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-xop.ll
index 483191fb32bdf..f53c7a3370174 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-xop.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-xop.ll
@@ -140,16 +140,11 @@ define <4 x double> @demandedelts_vpermil2pd256_as_shufpd(<4 x double> %a0, <4 x
 ; X86-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[1,1,2,3]
 ; X86-NEXT:    retl
 ;
-; X64-AVX-LABEL: demandedelts_vpermil2pd256_as_shufpd:
-; X64-AVX:       # %bb.0:
-; X64-AVX-NEXT:    vpermil2pd {{.*#+}} ymm0 = ymm1[0,0],ymm0[3],ymm1[3]
-; X64-AVX-NEXT:    retq
-;
-; X64-AVX2-LABEL: demandedelts_vpermil2pd256_as_shufpd:
-; X64-AVX2:       # %bb.0:
-; X64-AVX2-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[3],ymm1[3]
-; X64-AVX2-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[1,1,2,3]
-; X64-AVX2-NEXT:    retq
+; X64-LABEL: demandedelts_vpermil2pd256_as_shufpd:
+; X64:       # %bb.0:
+; X64-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[3],ymm1[3]
+; X64-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[1,1,2,3]
+; X64-NEXT:    retq
   %res0 = insertelement <4 x i64> <i64 0, i64 4, i64 2, i64 7>, i64 %a2, i32 0
   %res1 = call <4 x double> @llvm.x86.xop.vpermil2pd.256(<4 x double> %a0, <4 x double> %a1, <4 x i64> %res0, i8 0)
   %res2 = shufflevector <4 x double> %res1, <4 x double> undef, <4 x i32> <i32 1, i32 1, i32 2, i32 3>