[X86][AVX] Narrow VPBROADCASTQ->VPBROADCASTD if we don't need the upper bits.

Helps fix cases where we've splatted smaller types to a wider vector element type without needing the upper bits.

Avoid this on AVX512 targets as that can affect broadcast folding.

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 76b4aaa..0e22301 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -38959,6 +38959,19 @@ bool X86TargetLowering::SimplifyDemandedBitsForTargetNode(
     if (SimplifyDemandedBits(Src, OriginalDemandedBits, DemandedElts, Known,
                              TLO, Depth + 1))
       return true;
+    // If we don't need the upper bits, attempt to narrow the broadcast source.
+    // Don't attempt this on AVX512 as it might affect broadcast folding.
+    // TODO: Should we attempt this for i32/i16 splats? They tend to be slower.
+    if ((BitWidth == 64) && SrcVT.isScalarInteger() && !Subtarget.hasAVX512() &&
+        OriginalDemandedBits.countLeadingZeros() >= (BitWidth / 2)) {
+      MVT NewSrcVT = MVT::getIntegerVT(BitWidth / 2);
+      SDValue NewSrc =
+          TLO.DAG.getNode(ISD::TRUNCATE, SDLoc(Src), NewSrcVT, Src);
+      MVT NewVT = MVT::getVectorVT(NewSrcVT, VT.getVectorNumElements() * 2);
+      SDValue NewBcst =
+          TLO.DAG.getNode(X86ISD::VBROADCAST, SDLoc(Op), NewVT, NewSrc);
+      return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, NewBcst));
+    }
     break;
   }
   case X86ISD::PCMPGT:

diff --git a/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-sext.ll b/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-sext.ll
index 1ce9f01..4712984 100644 (file)
--- a/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-sext.ll
+++ b/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-sext.ll
@@ -220,9 +220,8 @@ define <4 x i64> @ext_i4_4i64(i4 %a0) {
 ;
 ; AVX2-LABEL: ext_i4_4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    # kill: def $edi killed $edi def $rdi
-; AVX2-NEXT:    vmovq %rdi, %xmm0
-; AVX2-NEXT:    vpbroadcastq %xmm0, %ymm0
+; AVX2-NEXT:    vmovd %edi, %xmm0
+; AVX2-NEXT:    vpbroadcastd %xmm0, %ymm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [1,2,4,8]
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0

diff --git a/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-zext.ll b/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-zext.ll
index bbc1c14..d014798 100644 (file)
--- a/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-zext.ll
+++ b/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-zext.ll
@@ -278,9 +278,8 @@ define <4 x i64> @ext_i4_4i64(i4 %a0) {
 ;
 ; AVX2-LABEL: ext_i4_4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    # kill: def $edi killed $edi def $rdi
-; AVX2-NEXT:    vmovq %rdi, %xmm0
-; AVX2-NEXT:    vpbroadcastq %xmm0, %ymm0
+; AVX2-NEXT:    vmovd %edi, %xmm0
+; AVX2-NEXT:    vpbroadcastd %xmm0, %ymm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [1,2,4,8]
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0