[DAG] SimplifyDemandedBits - don't fold sext(x) -> aext(x) if we lose an 0/-1 allsignbits mask

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

@@ -2444,6 +2444,13 @@ bool TargetLowering::SimplifyDemandedBits(
     unsigned InElts = SrcVT.isFixedLengthVector() ? SrcVT.getVectorNumElements() : 1;
     bool IsVecInReg = Op.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG;
 
+    APInt InDemandedElts = DemandedElts.zext(InElts);
+    APInt InDemandedBits = DemandedBits.trunc(InBits);
+
+    // Since some of the sign extended bits are demanded, we know that the sign
+    // bit is demanded.
+    InDemandedBits.setBit(InBits - 1);
+
     // If none of the top bits are demanded, convert this into an any_extend.
     if (DemandedBits.getActiveBits() <= InBits) {
       // If we only need the non-extended bits of the bottom element
@@ -2452,19 +2459,17 @@ bool TargetLowering::SimplifyDemandedBits(
           VT.getSizeInBits() == SrcVT.getSizeInBits())
         return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src));
 
-      unsigned Opc =
-          IsVecInReg ? ISD::ANY_EXTEND_VECTOR_INREG : ISD::ANY_EXTEND;
-      if (!TLO.LegalOperations() || isOperationLegal(Opc, VT))
-        return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src));
+      // Don't lose an all signbits 0/-1 splat on targets with 0/-1 booleans.
+      if (getBooleanContents(VT) != ZeroOrNegativeOneBooleanContent ||
+          TLO.DAG.ComputeNumSignBits(Src, InDemandedElts, Depth + 1) !=
+              InBits) {
+        unsigned Opc =
+            IsVecInReg ? ISD::ANY_EXTEND_VECTOR_INREG : ISD::ANY_EXTEND;
+        if (!TLO.LegalOperations() || isOperationLegal(Opc, VT))
+          return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src));
+      }
     }
 
-    APInt InDemandedBits = DemandedBits.trunc(InBits);
-    APInt InDemandedElts = DemandedElts.zext(InElts);
-
-    // Since some of the sign extended bits are demanded, we know that the sign
-    // bit is demanded.
-    InDemandedBits.setBit(InBits - 1);
-
     if (SimplifyDemandedBits(Src, InDemandedBits, InDemandedElts, Known, TLO,
                              Depth + 1))
       return true;

llvm/test/CodeGen/AArch64/arm64-zip.ll

@@ -328,7 +328,7 @@ define <4 x i32> @shuffle_zip3(<4 x i32> %arg) {
 ; CHECK-NEXT:    zip2.4h v0, v0, v1
 ; CHECK-NEXT:    movi.4s v1, #1
 ; CHECK-NEXT:    zip1.4h v0, v0, v0
-; CHECK-NEXT:    ushll.4s v0, v0, #0
+; CHECK-NEXT:    sshll.4s v0, v0, #0
 ; CHECK-NEXT:    and.16b v0, v0, v1
 ; CHECK-NEXT:    ret
 bb:

llvm/test/CodeGen/AArch64/vselect-ext.ll

@@ -219,17 +219,17 @@ define <3 x i32> @same_zext_used_in_cmp_unsigned_pred_and_select_v3i16(<3 x i8>
 ; CHECK-NEXT:    fmov s0, w0
 ; CHECK-NEXT:  Lloh0:
 ; CHECK-NEXT:    adrp x8, lCPI9_0@PAGE
+; CHECK-NEXT:    movi.2d v3, #0x0000ff000000ff
 ; CHECK-NEXT:  Lloh1:
 ; CHECK-NEXT:    ldr d2, [x8, lCPI9_0@PAGEOFF]
 ; CHECK-NEXT:    mov.h v0[1], w1
 ; CHECK-NEXT:    mov.h v0[2], w2
-; CHECK-NEXT:    fmov d1, d0
-; CHECK-NEXT:    bic.4h v1, #255, lsl #8
-; CHECK-NEXT:    cmhi.4h v1, v1, v2
-; CHECK-NEXT:    and.8b v0, v0, v1
-; CHECK-NEXT:    movi.2d v1, #0x0000ff000000ff
-; CHECK-NEXT:    ushll.4s v0, v0, #0
-; CHECK-NEXT:    and.16b v0, v0, v1
+; CHECK-NEXT:    ushll.4s v1, v0, #0
+; CHECK-NEXT:    bic.4h v0, #255, lsl #8
+; CHECK-NEXT:    cmhi.4h v0, v0, v2
+; CHECK-NEXT:    and.16b v1, v1, v3
+; CHECK-NEXT:    sshll.4s v0, v0, #0
+; CHECK-NEXT:    and.16b v0, v1, v0
 ; CHECK-NEXT:    ret
 ; CHECK-NEXT:    .loh AdrpLdr Lloh0, Lloh1
   %ext = zext <3 x i8> %a to <3 x i32>

llvm/test/CodeGen/SystemZ/vec-perm-14.ll

@@ -1,16 +1,14 @@
 ; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z14 | FileCheck %s
-;
-; Test that only one vperm of the vector compare is needed for both extracts.
 
+; Test that no vperm of the vector compare is needed for the extracts.
 define void @fun() {
 ; CHECK-LABEL: fun:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    vlrepf %v0, 0(%r1)
 ; CHECK-NEXT:    vgbm %v1, 0
-; CHECK-NEXT:    larl %r1, .LCPI0_0
 ; CHECK-NEXT:    vceqb %v0, %v0, %v1
-; CHECK-NEXT:    vl %v1, 0(%r1), 3
-; CHECK-NEXT:    vperm %v0, %v0, %v0, %v1
+; CHECK-NEXT:    vuphb %v0, %v0
+; CHECK-NEXT:    vuphh %v0, %v0
 ; CHECK-NEXT:    vlgvf %r0, %v0, 0
 ; CHECK-NEXT:    tmll %r0, 1
 ; CHECK-NEXT:    je .LBB0_2

llvm/test/CodeGen/X86/test-shrink-bug.ll

@@ -68,8 +68,8 @@ define dso_local void @fail(i16 %a, <2 x i8> %b) {
 ; CHECK-X64-NEXT:    je .LBB1_3
 ; CHECK-X64-NEXT:  # %bb.1:
 ; CHECK-X64-NEXT:    pcmpeqb {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; CHECK-X64-NEXT:    pslldq {{.*#+}} xmm0 = zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7,8]
-; CHECK-X64-NEXT:    pextrw $4, %xmm0, %eax
+; CHECK-X64-NEXT:    pslld $8, %xmm0
+; CHECK-X64-NEXT:    pextrw $1, %xmm0, %eax
 ; CHECK-X64-NEXT:    testb $1, %al
 ; CHECK-X64-NEXT:    jne .LBB1_3
 ; CHECK-X64-NEXT:  # %bb.2: # %no

llvm/test/CodeGen/X86/vec_setcc.ll

@@ -308,9 +308,9 @@ define <3 x i1> @test_setcc_v3i1_v3i16(ptr %a) nounwind {
 ; SSE2-LABEL: test_setcc_v3i1_v3i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
+; SSE2-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
 ; SSE2-NEXT:    pxor %xmm1, %xmm1
 ; SSE2-NEXT:    pcmpeqw %xmm0, %xmm1
-; SSE2-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
 ; SSE2-NEXT:    movdqa %xmm1, -{{[0-9]+}}(%rsp)
 ; SSE2-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
 ; SSE2-NEXT:    movzbl -{{[0-9]+}}(%rsp), %edx