[X86][SSE] combineVectorSignBitsTruncation - fold trunc(srl(x,c)) -> packss(sra(x,c))

If a srl doesn't introduce any sign bits into the truncated result, then replace with a sra to let us use a PACKSS truncation - fixes a regression noticed in D56387 on pre-SSE41 targets that don't have PACKUSDW.

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 97fcef0..0ee6717 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -46071,9 +46071,23 @@ static SDValue combineVectorSignBitsTruncation(SDNode *N, const SDLoc &DL,
   if (SVT == MVT::i32 && NumSignBits != InSVT.getSizeInBits())
     return SDValue();
 
-  if (NumSignBits > (InSVT.getSizeInBits() - NumPackedSignBits))
+  unsigned MinSignBits = InSVT.getSizeInBits() - NumPackedSignBits;
+  if (NumSignBits > MinSignBits)
     return truncateVectorWithPACK(X86ISD::PACKSS, VT, In, DL, DAG, Subtarget);
 
+  // If we have a srl that only generates signbits that we will discard in
+  // the truncation then we can use PACKSS by converting the srl to a sra.
+  // SimplifyDemandedBits often relaxes sra to srl so we need to reverse it.
+  if (In.getOpcode() == ISD::SRL && N->isOnlyUserOf(In.getNode()))
+    if (const APInt *ShAmt = DAG.getValidShiftAmountConstant(
+            In, APInt::getAllOnesValue(VT.getVectorNumElements()))) {
+      if (*ShAmt == MinSignBits) {
+        SDValue NewIn = DAG.getNode(ISD::SRA, DL, InVT, In->ops());
+        return truncateVectorWithPACK(X86ISD::PACKSS, VT, NewIn, DL, DAG,
+                                      Subtarget);
+      }
+    }
+
   return SDValue();
 }
 

diff --git a/llvm/test/CodeGen/X86/vector-trunc.ll b/llvm/test/CodeGen/X86/vector-trunc.ll
index f35e315..1d8d6f6 100644 (file)
--- a/llvm/test/CodeGen/X86/vector-trunc.ll
+++ b/llvm/test/CodeGen/X86/vector-trunc.ll
@@ -452,10 +452,9 @@ define <8 x i16> @trunc8i32_8i16_lshr(<8 x i32> %a) {
 ;
 ; SSSE3-LABEL: trunc8i32_8i16_lshr:
 ; SSSE3:       # %bb.0: # %entry
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [2,3,6,7,10,11,14,15,10,11,14,15,14,15,128,128]
-; SSSE3-NEXT:    pshufb %xmm2, %xmm1
-; SSSE3-NEXT:    pshufb %xmm2, %xmm0
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSSE3-NEXT:    psrad $16, %xmm1
+; SSSE3-NEXT:    psrad $16, %xmm0
+; SSSE3-NEXT:    packssdw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc8i32_8i16_lshr: