[X86] Fold PACKSS(NOT(X),NOT(Y)) -> NOT(PACKSS(X,Y))

llvm · Jul 14, 2023 · 720debc · 720debc
1 parent f8ffd67
commit 720debc
Show file tree

Hide file tree

Showing 13 changed files with 378 additions and 533 deletions.
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -49538,6 +49538,23 @@ static SDValue combineVectorPack(SDNode *N, SelectionDAG &DAG,
   if (SDValue V = combineHorizOpWithShuffle(N, DAG, Subtarget))
     return V;
 
+  // Try to fold PACKSS(NOT(X),NOT(Y)) -> NOT(PACKSS(X,Y)).
+  // Currently limit this to allsignbits cases only.
+  if (IsSigned &&
+      (N0.isUndef() || DAG.ComputeNumSignBits(N0) == SrcBitsPerElt) &&
+      (N1.isUndef() || DAG.ComputeNumSignBits(N1) == SrcBitsPerElt)) {
+    SDValue Not0 = N0.isUndef() ? N0 : IsNOT(N0, DAG);
+    SDValue Not1 = N1.isUndef() ? N1 : IsNOT(N1, DAG);
+    if (Not0 && Not1) {
+      SDLoc DL(N);
+      MVT SrcVT = N0.getSimpleValueType();
+      SDValue Pack =
+          DAG.getNode(X86ISD::PACKSS, DL, VT, DAG.getBitcast(SrcVT, Not0),
+                      DAG.getBitcast(SrcVT, Not1));
+      return DAG.getNOT(DL, Pack, VT);
+    }
+  }
+
   // Try to combine a PACKUSWB/PACKSSWB implemented truncate with a regular
   // truncate to create a larger truncate.
   if (Subtarget.hasAVX512() &&

diff --git a/llvm/test/CodeGen/X86/icmp-abs-C-vec.ll b/llvm/test/CodeGen/X86/icmp-abs-C-vec.ll
@@ -245,12 +245,10 @@ define <4 x i1> @illegal_abs_to_ne_and(<4 x i64> %x) {
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm1
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [129,129]
 ; SSE41-NEXT:    pcmpeqq %xmm0, %xmm1
-; SSE41-NEXT:    pcmpeqd %xmm3, %xmm3
-; SSE41-NEXT:    pxor %xmm3, %xmm1
 ; SSE41-NEXT:    pcmpeqq %xmm0, %xmm2
-; SSE41-NEXT:    pxor %xmm3, %xmm2
 ; SSE41-NEXT:    packssdw %xmm1, %xmm2
-; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm0
+; SSE41-NEXT:    pxor %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; SSE2-LABEL: illegal_abs_to_ne_and:
@@ -267,13 +265,12 @@ define <4 x i1> @illegal_abs_to_ne_and(<4 x i64> %x) {
 ; SSE2-NEXT:    pcmpeqd %xmm2, %xmm1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[1,0,3,2]
 ; SSE2-NEXT:    pand %xmm1, %xmm3
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
-; SSE2-NEXT:    pxor %xmm1, %xmm3
 ; SSE2-NEXT:    pcmpeqd %xmm2, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,0,3,2]
-; SSE2-NEXT:    pand %xmm2, %xmm0
-; SSE2-NEXT:    pxor %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
+; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    packssdw %xmm3, %xmm0
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE2-NEXT:    pxor %xmm1, %xmm0
 ; SSE2-NEXT:    retq
   %abs = call <4 x i64> @llvm.abs.v4i64(<4 x i64> %x, i1 true)
   %cmp = icmp ne <4 x i64> %abs, <i64 129, i64 129, i64 129, i64 129>

diff --git a/llvm/test/CodeGen/X86/ispow2.ll b/llvm/test/CodeGen/X86/ispow2.ll
@@ -139,15 +139,14 @@ define <4 x i1> @neither_pow2_non_zero_4xv64(<4 x i64> %xin) {
 ; CHECK-NOBMI-NEXT:    pcmpeqd %xmm1, %xmm3
 ; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,0,3,2]
 ; CHECK-NOBMI-NEXT:    pand %xmm3, %xmm4
-; CHECK-NOBMI-NEXT:    pxor %xmm2, %xmm4
 ; CHECK-NOBMI-NEXT:    movdqa %xmm0, %xmm3
 ; CHECK-NOBMI-NEXT:    paddq %xmm2, %xmm3
 ; CHECK-NOBMI-NEXT:    pand %xmm3, %xmm0
 ; CHECK-NOBMI-NEXT:    pcmpeqd %xmm1, %xmm0
 ; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
 ; CHECK-NOBMI-NEXT:    pand %xmm1, %xmm0
-; CHECK-NOBMI-NEXT:    pxor %xmm2, %xmm0
 ; CHECK-NOBMI-NEXT:    packssdw %xmm4, %xmm0
+; CHECK-NOBMI-NEXT:    pxor %xmm2, %xmm0
 ; CHECK-NOBMI-NEXT:    retq
 ;
 ; CHECK-AVX2-LABEL: neither_pow2_non_zero_4xv64: