[X86][SSE] canonicalizeShuffleWithBinOps - add X86ISD::PSHUFB handling.

Recommit rGcd938ab162b0ac560dd0e9fee290980c7e0e47e5 with an early-out if the pshub would introduce zeros across the binop.

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index a2eb28f..c518c6a 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -36838,6 +36838,15 @@ static SDValue canonicalizeShuffleWithBinOps(SDValue N, SelectionDAG &DAG,
   unsigned Opc = N.getOpcode();
   switch (Opc) {
   // Unary and Unary+Permute Shuffles.
+  case X86ISD::PSHUFB: {
+    // Don't merge PSHUFB if it contains zero'd elements.
+    SmallVector<int> Mask;
+    SmallVector<SDValue> Ops;
+    if (!getTargetShuffleMask(N.getNode(), ShuffleVT.getSimpleVT(), false, Ops,
+                              Mask))
+      break;
+    LLVM_FALLTHROUGH;
+  }
   case X86ISD::VBROADCAST:
   case X86ISD::MOVDDUP:
   case X86ISD::PSHUFD: {

diff --git a/llvm/test/CodeGen/X86/vector-fshr-rot-128.ll b/llvm/test/CodeGen/X86/vector-fshr-rot-128.ll
index ea795d9..1175362 100644 (file)
--- a/llvm/test/CodeGen/X86/vector-fshr-rot-128.ll
+++ b/llvm/test/CodeGen/X86/vector-fshr-rot-128.ll
@@ -1403,8 +1403,8 @@ define <16 x i8> @splatvar_funnnel_v16i8(<16 x i8> %x, <16 x i8> %amt) nounwind
 ; XOPAVX1-LABEL: splatvar_funnnel_v16i8:
 ; XOPAVX1:       # %bb.0:
 ; XOPAVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; XOPAVX1-NEXT:    vpsubb %xmm1, %xmm2, %xmm1
 ; XOPAVX1-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
+; XOPAVX1-NEXT:    vpsubb %xmm1, %xmm2, %xmm1
 ; XOPAVX1-NEXT:    vprotb %xmm1, %xmm0, %xmm0
 ; XOPAVX1-NEXT:    retq
 ;

diff --git a/llvm/test/CodeGen/X86/vector-shift-ashr-128.ll b/llvm/test/CodeGen/X86/vector-shift-ashr-128.ll
index ab84a58..cb3382c 100644 (file)
--- a/llvm/test/CodeGen/X86/vector-shift-ashr-128.ll
+++ b/llvm/test/CodeGen/X86/vector-shift-ashr-128.ll
@@ -848,8 +848,8 @@ define <16 x i8> @splatvar_shift_v16i8(<16 x i8> %a, <16 x i8> %b) nounwind {
 ; XOPAVX1-LABEL: splatvar_shift_v16i8:
 ; XOPAVX1:       # %bb.0:
 ; XOPAVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; XOPAVX1-NEXT:    vpsubb %xmm1, %xmm2, %xmm1
 ; XOPAVX1-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
+; XOPAVX1-NEXT:    vpsubb %xmm1, %xmm2, %xmm1
 ; XOPAVX1-NEXT:    vpshab %xmm1, %xmm0, %xmm0
 ; XOPAVX1-NEXT:    retq
 ;

diff --git a/llvm/test/CodeGen/X86/vector-shift-lshr-128.ll b/llvm/test/CodeGen/X86/vector-shift-lshr-128.ll
index 7cea8eb..530ebbe 100644 (file)
--- a/llvm/test/CodeGen/X86/vector-shift-lshr-128.ll
+++ b/llvm/test/CodeGen/X86/vector-shift-lshr-128.ll
@@ -698,8 +698,8 @@ define <16 x i8> @splatvar_shift_v16i8(<16 x i8> %a, <16 x i8> %b) nounwind {
 ; XOPAVX1-LABEL: splatvar_shift_v16i8:
 ; XOPAVX1:       # %bb.0:
 ; XOPAVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; XOPAVX1-NEXT:    vpsubb %xmm1, %xmm2, %xmm1
 ; XOPAVX1-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
+; XOPAVX1-NEXT:    vpsubb %xmm1, %xmm2, %xmm1
 ; XOPAVX1-NEXT:    vpshlb %xmm1, %xmm0, %xmm0
 ; XOPAVX1-NEXT:    retq
 ;