[X86] SimplifyDemandedBitsForTargetNode - generalize X86ISD::VSRAI handling when only demanding 'known signbits'

If we only want bits that already match the signbit then we don't need to shift.

Generalizes an existing pattern that just handled signbit-only demanded bits to match what we do for ISD::SRA.

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -44615,8 +44615,11 @@ bool X86TargetLowering::SimplifyDemandedBitsForTargetNode(
 
     APInt DemandedMask = OriginalDemandedBits << ShAmt;
 
-    // If we just want the sign bit then we don't need to shift it.
-    if (OriginalDemandedBits.isSignMask())
+    // If we only want bits that already match the signbit then we don't need
+    // to shift.
+    unsigned NumHiDemandedBits = BitWidth - OriginalDemandedBits.countr_zero();
+    if (TLO.DAG.ComputeNumSignBits(Op0, OriginalDemandedElts, Depth + 1) >=
+        NumHiDemandedBits)
       return TLO.CombineTo(Op, Op0);
 
     // fold (VSRAI (VSHLI X, C1), C1) --> X iff NumSignBits(X) > C1

llvm/test/CodeGen/X86/combine-pack.ll

@@ -5,22 +5,19 @@
 
 declare <8 x i16> @llvm.x86.sse2.packssdw.128(<4 x i32>, <4 x i32>)
 
-; TODO: Failure to remove unnecessary signsplat
 define <8 x i16> @combine_packss_v4i32_signsplat(<4 x i32> %a0, <4 x i32> %a1) {
 ; SSE-LABEL: combine_packss_v4i32_signsplat:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE-NEXT:    packssdw %xmm1, %xmm0
-; SSE-NEXT:    psraw $15, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: combine_packss_v4i32_signsplat:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
 ; AVX-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vpsraw $15, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %cmp = icmp sgt <4 x i32> %a0, %a1
   %ext = sext <4 x i1> %cmp to <4 x i32>