[DAG] SelectionDAG.computeKnownBits - add NSW/NUW flags support to ISD::SHL handling (#89877)

zxc12523 · web-flow · commit 171aeb20ad46 · 2024-05-02T10:31:56.000+01:00
fix #89414
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -3527,16 +3527,23 @@ KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts,
       Known.Zero.setBitsFrom(1);
     break;
   }
-  case ISD::SHL:
+  case ISD::SHL: {
     Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
     Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
-    Known = KnownBits::shl(Known, Known2);
+
+    bool NUW = Op->getFlags().hasNoUnsignedWrap();
+    bool NSW = Op->getFlags().hasNoSignedWrap();
+
+    bool ShAmtNonZero = Known2.isNonZero();
+
+    Known = KnownBits::shl(Known, Known2, NUW, NSW, ShAmtNonZero);
 
     // Minimum shift low bits are known zero.
     if (const APInt *ShMinAmt =
             getValidMinimumShiftAmountConstant(Op, DemandedElts))
       Known.Zero.setLowBits(ShMinAmt->getZExtValue());
     break;
+  }
   case ISD::SRL:
     Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
     Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
diff --git a/llvm/test/CodeGen/X86/fold-int-pow2-with-fmul-or-fdiv.ll b/llvm/test/CodeGen/X86/fold-int-pow2-with-fmul-or-fdiv.ll
@@ -840,44 +840,18 @@ define double @fmul_pow_shl_cnt_fail_maybe_non_pow2(i64 %v, i64 %cnt) nounwind {
 define <2 x float> @fmul_pow_shl_cnt_vec_fail_expensive_cast(<2 x i64> %cnt) nounwind {
 ; CHECK-SSE-LABEL: fmul_pow_shl_cnt_vec_fail_expensive_cast:
 ; CHECK-SSE:       # %bb.0:
-; CHECK-SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
-; CHECK-SSE-NEXT:    movdqa {{.*#+}} xmm3 = [2,2]
-; CHECK-SSE-NEXT:    movdqa %xmm3, %xmm1
-; CHECK-SSE-NEXT:    psllq %xmm2, %xmm1
-; CHECK-SSE-NEXT:    psllq %xmm0, %xmm3
-; CHECK-SSE-NEXT:    movq %xmm3, %rax
-; CHECK-SSE-NEXT:    testq %rax, %rax
-; CHECK-SSE-NEXT:    js .LBB12_1
-; CHECK-SSE-NEXT:  # %bb.2:
-; CHECK-SSE-NEXT:    xorps %xmm0, %xmm0
-; CHECK-SSE-NEXT:    cvtsi2ss %rax, %xmm0
-; CHECK-SSE-NEXT:    jmp .LBB12_3
-; CHECK-SSE-NEXT:  .LBB12_1:
-; CHECK-SSE-NEXT:    movq %rax, %rcx
-; CHECK-SSE-NEXT:    shrq %rcx
-; CHECK-SSE-NEXT:    andl $1, %eax
-; CHECK-SSE-NEXT:    orq %rcx, %rax
+; CHECK-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; CHECK-SSE-NEXT:    movdqa {{.*#+}} xmm2 = [2,2]
+; CHECK-SSE-NEXT:    movdqa %xmm2, %xmm3
+; CHECK-SSE-NEXT:    psllq %xmm1, %xmm3
+; CHECK-SSE-NEXT:    psllq %xmm0, %xmm2
+; CHECK-SSE-NEXT:    movq %xmm2, %rax
 ; CHECK-SSE-NEXT:    xorps %xmm0, %xmm0
 ; CHECK-SSE-NEXT:    cvtsi2ss %rax, %xmm0
-; CHECK-SSE-NEXT:    addss %xmm0, %xmm0
-; CHECK-SSE-NEXT:  .LBB12_3:
-; CHECK-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
+; CHECK-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]
 ; CHECK-SSE-NEXT:    movq %xmm1, %rax
-; CHECK-SSE-NEXT:    testq %rax, %rax
-; CHECK-SSE-NEXT:    js .LBB12_4
-; CHECK-SSE-NEXT:  # %bb.5:
-; CHECK-SSE-NEXT:    xorps %xmm1, %xmm1
-; CHECK-SSE-NEXT:    cvtsi2ss %rax, %xmm1
-; CHECK-SSE-NEXT:    jmp .LBB12_6
-; CHECK-SSE-NEXT:  .LBB12_4:
-; CHECK-SSE-NEXT:    movq %rax, %rcx
-; CHECK-SSE-NEXT:    shrq %rcx
-; CHECK-SSE-NEXT:    andl $1, %eax
-; CHECK-SSE-NEXT:    orq %rcx, %rax
 ; CHECK-SSE-NEXT:    xorps %xmm1, %xmm1
 ; CHECK-SSE-NEXT:    cvtsi2ss %rax, %xmm1
-; CHECK-SSE-NEXT:    addss %xmm1, %xmm1
-; CHECK-SSE-NEXT:  .LBB12_6:
 ; CHECK-SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; CHECK-SSE-NEXT:    mulps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; CHECK-SSE-NEXT:    retq
@@ -886,18 +860,11 @@ define <2 x float> @fmul_pow_shl_cnt_vec_fail_expensive_cast(<2 x i64> %cnt) nou
 ; CHECK-AVX2:       # %bb.0:
 ; CHECK-AVX2-NEXT:    vpmovsxbq {{.*#+}} xmm1 = [2,2]
 ; CHECK-AVX2-NEXT:    vpsllvq %xmm0, %xmm1, %xmm0
-; CHECK-AVX2-NEXT:    vpsrlq $1, %xmm0, %xmm1
-; CHECK-AVX2-NEXT:    vblendvpd %xmm0, %xmm1, %xmm0, %xmm1
-; CHECK-AVX2-NEXT:    vpextrq $1, %xmm1, %rax
-; CHECK-AVX2-NEXT:    vcvtsi2ss %rax, %xmm2, %xmm2
-; CHECK-AVX2-NEXT:    vmovq %xmm1, %rax
-; CHECK-AVX2-NEXT:    vcvtsi2ss %rax, %xmm3, %xmm1
-; CHECK-AVX2-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],zero,zero
-; CHECK-AVX2-NEXT:    vaddps %xmm1, %xmm1, %xmm2
-; CHECK-AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; CHECK-AVX2-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm0
-; CHECK-AVX2-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
-; CHECK-AVX2-NEXT:    vblendvps %xmm0, %xmm2, %xmm1, %xmm0
+; CHECK-AVX2-NEXT:    vpextrq $1, %xmm0, %rax
+; CHECK-AVX2-NEXT:    vcvtsi2ss %rax, %xmm2, %xmm1
+; CHECK-AVX2-NEXT:    vmovq %xmm0, %rax
+; CHECK-AVX2-NEXT:    vcvtsi2ss %rax, %xmm2, %xmm0
+; CHECK-AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],zero,zero
 ; CHECK-AVX2-NEXT:    vbroadcastss {{.*#+}} xmm1 = [1.5E+1,1.5E+1,1.5E+1,1.5E+1]
 ; CHECK-AVX2-NEXT:    vmulps %xmm1, %xmm0, %xmm0
 ; CHECK-AVX2-NEXT:    retq
@@ -907,9 +874,9 @@ define <2 x float> @fmul_pow_shl_cnt_vec_fail_expensive_cast(<2 x i64> %cnt) nou
 ; CHECK-NO-FASTFMA-NEXT:    vpmovsxbq {{.*#+}} xmm1 = [2,2]
 ; CHECK-NO-FASTFMA-NEXT:    vpsllvq %xmm0, %xmm1, %xmm0
 ; CHECK-NO-FASTFMA-NEXT:    vpextrq $1, %xmm0, %rax
-; CHECK-NO-FASTFMA-NEXT:    vcvtusi2ss %rax, %xmm2, %xmm1
+; CHECK-NO-FASTFMA-NEXT:    vcvtsi2ss %rax, %xmm2, %xmm1
 ; CHECK-NO-FASTFMA-NEXT:    vmovq %xmm0, %rax
-; CHECK-NO-FASTFMA-NEXT:    vcvtusi2ss %rax, %xmm2, %xmm0
+; CHECK-NO-FASTFMA-NEXT:    vcvtsi2ss %rax, %xmm2, %xmm0
 ; CHECK-NO-FASTFMA-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],zero,zero
 ; CHECK-NO-FASTFMA-NEXT:    vbroadcastss {{.*#+}} xmm1 = [1.5E+1,1.5E+1,1.5E+1,1.5E+1]
 ; CHECK-NO-FASTFMA-NEXT:    vmulps %xmm1, %xmm0, %xmm0
@@ -919,7 +886,7 @@ define <2 x float> @fmul_pow_shl_cnt_vec_fail_expensive_cast(<2 x i64> %cnt) nou
 ; CHECK-FMA:       # %bb.0:
 ; CHECK-FMA-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [2,2]
 ; CHECK-FMA-NEXT:    vpsllvq %xmm0, %xmm1, %xmm0
-; CHECK-FMA-NEXT:    vcvtuqq2ps %xmm0, %xmm0
+; CHECK-FMA-NEXT:    vcvtqq2ps %xmm0, %xmm0
 ; CHECK-FMA-NEXT:    vmulps {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
 ; CHECK-FMA-NEXT:    retq
   %shl = shl nsw nuw <2 x i64> <i64 2, i64 2>, %cnt
@@ -986,7 +953,7 @@ define <4 x float> @fmul_pow_shl_cnt_vec_preserve_fma(<4 x i32> %cnt, <4 x float
 ; CHECK-FMA:       # %bb.0:
 ; CHECK-FMA-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2,2,2,2]
 ; CHECK-FMA-NEXT:    vpsllvd %xmm0, %xmm2, %xmm0
-; CHECK-FMA-NEXT:    vcvtudq2ps %xmm0, %xmm0
+; CHECK-FMA-NEXT:    vcvtdq2ps %xmm0, %xmm0
 ; CHECK-FMA-NEXT:    vfmadd132ps {{.*#+}} xmm0 = (xmm0 * mem) + xmm1
 ; CHECK-FMA-NEXT:    retq
   %shl = shl nsw nuw <4 x i32> <i32 2, i32 2, i32 2, i32 2>, %cnt
diff --git a/llvm/test/CodeGen/X86/known-never-zero.ll b/llvm/test/CodeGen/X86/known-never-zero.ll
@@ -1612,7 +1612,7 @@ define i32 @sext_known_nonzero(i16 %xx) {
 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
 ; X86-NEXT:    movl $256, %eax # imm = 0x100
 ; X86-NEXT:    shll %cl, %eax
-; X86-NEXT:    cwtl
+; X86-NEXT:    movzwl %ax, %eax
 ; X86-NEXT:    rep bsfl %eax, %eax
 ; X86-NEXT:    retl
 ;
@@ -1622,7 +1622,7 @@ define i32 @sext_known_nonzero(i16 %xx) {
 ; X64-NEXT:    movl $256, %eax # imm = 0x100
 ; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
 ; X64-NEXT:    shll %cl, %eax
-; X64-NEXT:    cwtl
+; X64-NEXT:    movzwl %ax, %eax
 ; X64-NEXT:    rep bsfl %eax, %eax
 ; X64-NEXT:    retq
   %x = shl nuw nsw i16 256, %xx
diff --git a/llvm/test/CodeGen/X86/pr89877.ll b/llvm/test/CodeGen/X86/pr89877.ll
@@ -0,0 +1,115 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=X86
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=X64
+
+define i32 @sext_known_nonzero(i16 %xx) {
+; X86-LABEL: sext_known_nonzero:
+; X86:       # %bb.0:
+; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movl $256, %eax # imm = 0x100
+; X86-NEXT:    shll %cl, %eax
+; X86-NEXT:    cwtl
+; X86-NEXT:    testl %eax, %eax
+; X86-NEXT:    je .LBB0_1
+; X86-NEXT:  # %bb.2: # %cond.false
+; X86-NEXT:    rep bsfl %eax, %eax
+; X86-NEXT:    retl
+; X86-NEXT:  .LBB0_1:
+; X86-NEXT:    movl $32, %eax
+; X86-NEXT:    retl
+;
+; X64-LABEL: sext_known_nonzero:
+; X64:       # %bb.0:
+; X64-NEXT:    movl %edi, %ecx
+; X64-NEXT:    movl $256, %eax # imm = 0x100
+; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X64-NEXT:    shll %cl, %eax
+; X64-NEXT:    cwtl
+; X64-NEXT:    testl %eax, %eax
+; X64-NEXT:    je .LBB0_1
+; X64-NEXT:  # %bb.2: # %cond.false
+; X64-NEXT:    rep bsfl %eax, %eax
+; X64-NEXT:    retq
+; X64-NEXT:  .LBB0_1:
+; X64-NEXT:    movl $32, %eax
+; X64-NEXT:    retq
+  %x = shl i16 256, %xx
+  %z = sext i16 %x to i32
+  %r = call i32 @llvm.cttz.i32(i32 %z, i1 false)
+  ret i32 %r
+}
+
+define i32 @sext_known_nonzero_nuw(i16 %xx) {
+; X86-LABEL: sext_known_nonzero_nuw:
+; X86:       # %bb.0:
+; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movl $256, %eax # imm = 0x100
+; X86-NEXT:    shll %cl, %eax
+; X86-NEXT:    cwtl
+; X86-NEXT:    rep bsfl %eax, %eax
+; X86-NEXT:    retl
+;
+; X64-LABEL: sext_known_nonzero_nuw:
+; X64:       # %bb.0:
+; X64-NEXT:    movl %edi, %ecx
+; X64-NEXT:    movl $256, %eax # imm = 0x100
+; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X64-NEXT:    shll %cl, %eax
+; X64-NEXT:    cwtl
+; X64-NEXT:    rep bsfl %eax, %eax
+; X64-NEXT:    retq
+  %x = shl nuw i16 256, %xx
+  %z = sext i16 %x to i32
+  %r = call i32 @llvm.cttz.i32(i32 %z, i1 false)
+  ret i32 %r
+}
+
+define i32 @sext_known_nonzero_nsw(i16 %xx) {
+; X86-LABEL: sext_known_nonzero_nsw:
+; X86:       # %bb.0:
+; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movl $256, %eax # imm = 0x100
+; X86-NEXT:    shll %cl, %eax
+; X86-NEXT:    movzwl %ax, %eax
+; X86-NEXT:    rep bsfl %eax, %eax
+; X86-NEXT:    retl
+;
+; X64-LABEL: sext_known_nonzero_nsw:
+; X64:       # %bb.0:
+; X64-NEXT:    movl %edi, %ecx
+; X64-NEXT:    movl $256, %eax # imm = 0x100
+; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X64-NEXT:    shll %cl, %eax
+; X64-NEXT:    movzwl %ax, %eax
+; X64-NEXT:    rep bsfl %eax, %eax
+; X64-NEXT:    retq
+  %x = shl nsw i16 256, %xx
+  %z = sext i16 %x to i32
+  %r = call i32 @llvm.cttz.i32(i32 %z, i1 false)
+  ret i32 %r
+}
+
+define i32 @sext_known_nonzero_nuw_nsw(i16 %xx) {
+; X86-LABEL: sext_known_nonzero_nuw_nsw:
+; X86:       # %bb.0:
+; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movl $256, %eax # imm = 0x100
+; X86-NEXT:    shll %cl, %eax
+; X86-NEXT:    movzwl %ax, %eax
+; X86-NEXT:    rep bsfl %eax, %eax
+; X86-NEXT:    retl
+;
+; X64-LABEL: sext_known_nonzero_nuw_nsw:
+; X64:       # %bb.0:
+; X64-NEXT:    movl %edi, %ecx
+; X64-NEXT:    movl $256, %eax # imm = 0x100
+; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
+; X64-NEXT:    shll %cl, %eax
+; X64-NEXT:    movzwl %ax, %eax
+; X64-NEXT:    rep bsfl %eax, %eax
+; X64-NEXT:    retq
+  %x = shl nuw nsw i16 256, %xx
+  %z = sext i16 %x to i32
+  %r = call i32 @llvm.cttz.i32(i32 %z, i1 false)
+  ret i32 %r
+}
