Reland "[ValueTracking] Compute knownbits from known fp classes" #92084

dtcxzyw · 2024-05-14T08:36:19Z

This patch relands #86409.

I mistakenly thought that Known.makeNegative() clears the sign bit of Known.Zero. This patch fixes the assertion failure by explicitly clearing the sign bit.

…es (llvm#86409)"" This reverts commit 2e165a2.

llvmbot · 2024-05-14T08:36:51Z

@llvm/pr-subscribers-llvm-ir

@llvm/pr-subscribers-llvm-transforms

Author: Yingwei Zheng (dtcxzyw)

Changes

This patch relands #86409.

I mistakenly thought that Known.makeNegative() clears the sign bit of Known.Zero. This patch fixes the assertion failure by explicitly clearing the sign bit.

Patch is 21.00 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/92084.diff

6 Files Affected:

(modified) llvm/include/llvm/IR/PatternMatch.h (+1-1)
(modified) llvm/lib/Analysis/ValueTracking.cpp (+36)
(modified) llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll (+7-7)
(modified) llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll (+6-6)
(modified) llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll (+5-5)
(modified) llvm/test/Transforms/InstCombine/known-bits.ll (+284)

diff --git a/llvm/include/llvm/IR/PatternMatch.h b/llvm/include/llvm/IR/PatternMatch.h
index 171ddab977dea..0d6d86cb47e67 100644
--- a/llvm/include/llvm/IR/PatternMatch.h
+++ b/llvm/include/llvm/IR/PatternMatch.h
@@ -1904,7 +1904,7 @@ template <typename Op_t> struct ElementWiseBitCast_match {
   ElementWiseBitCast_match(const Op_t &OpMatch) : Op(OpMatch) {}
 
   template <typename OpTy> bool match(OpTy *V) {
-    BitCastInst *I = dyn_cast<BitCastInst>(V);
+    auto *I = dyn_cast<BitCastInst>(V);
     if (!I)
       return false;
     Type *SrcType = I->getSrcTy();
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 375385aca7a39..99b621461fcdc 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -1118,6 +1118,42 @@ static void computeKnownBitsFromOperator(const Operator *I,
       break;
     }
 
+    const Value *V;
+    // Handle bitcast from floating point to integer.
+    if (match(I, m_ElementWiseBitCast(m_Value(V))) &&
+        V->getType()->isFPOrFPVectorTy()) {
+      Type *FPType = V->getType()->getScalarType();
+      KnownFPClass Result = computeKnownFPClass(V, fcAllFlags, Depth + 1, Q);
+      FPClassTest FPClasses = Result.KnownFPClasses;
+
+      if (Result.isKnownNever(fcNormal | fcSubnormal | fcNan)) {
+        Known.Zero.setAllBits();
+        Known.One.setAllBits();
+
+        if (FPClasses & fcInf)
+          Known = Known.intersectWith(KnownBits::makeConstant(
+              APFloat::getInf(FPType->getFltSemantics()).bitcastToAPInt()));
+
+        if (FPClasses & fcZero)
+          Known = Known.intersectWith(KnownBits::makeConstant(
+              APInt::getZero(FPType->getScalarSizeInBits())));
+
+        Known.Zero.clearSignBit();
+        Known.One.clearSignBit();
+      }
+
+      if (Result.SignBit) {
+        if (*Result.SignBit)
+          Known.makeNegative();
+        else
+          Known.makeNonNegative();
+      }
+
+      assert(!Known.hasConflict() && "Bits known to be one AND zero?");
+
+      break;
+    }
+
     // Handle cast from vector integer type to scalar or vector integer.
     auto *SrcVecTy = dyn_cast<FixedVectorType>(SrcTy);
     if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() ||
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
index c4bd4bc126f73..5db25a59d33fc 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
@@ -2216,7 +2216,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_sitofp(float %x, i32 %y)
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2304,7 +2304,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_uitofp(float %x, i32 %y)
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2353,7 +2353,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_uitofp_i256(float %x, i2
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2376,7 +2376,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_sitofp_i256(float %x, i2
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2399,7 +2399,7 @@ define <2 x float> @test_pow_afn_nnan_ninf_v2f32_known_integral_sitofp(<2 x floa
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <2 x float> [[X]] to <2 x i32>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i32> [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x float> [[__EXP2]] to <2 x i32>
-; CHECK-NEXT:    [[TMP4:%.*]] = or <2 x i32> [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint <2 x i32> [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <2 x float>
 ; CHECK-NEXT:    ret <2 x float> [[TMP5]]
 ;
@@ -2448,7 +2448,7 @@ define <2 x float> @test_pow_afn_nnan_ninf_v2f32_known_integral_uitofp(<2 x floa
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <2 x float> [[X]] to <2 x i32>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i32> [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x float> [[__EXP2]] to <2 x i32>
-; CHECK-NEXT:    [[TMP4:%.*]] = or <2 x i32> [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint <2 x i32> [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <2 x float>
 ; CHECK-NEXT:    ret <2 x float> [[TMP5]]
 ;
@@ -2560,7 +2560,7 @@ define float @test_pow_afn_f32_nnan_ninf__y_known_integral_trunc(float %x, float
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll
index 8ddaf243db92c..e298226ee7ccd 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll
@@ -680,7 +680,7 @@ define float @test_pown_afn_nnan_ninf_f32(float %x, i32 %y) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP2:%.*]] = or i32 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32 [[TMP2]] to float
 ; CHECK-NEXT:    ret float [[TMP3]]
 ;
@@ -703,7 +703,7 @@ define <2 x float> @test_pown_afn_nnan_ninf_v2f32(<2 x float> %x, <2 x i32> %y)
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <2 x float> [[X]] to <2 x i32>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i32> [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <2 x float> [[__EXP2]] to <2 x i32>
-; CHECK-NEXT:    [[TMP2:%.*]] = or <2 x i32> [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint <2 x i32> [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x i32> [[TMP2]] to <2 x float>
 ; CHECK-NEXT:    ret <2 x float> [[TMP3]]
 ;
@@ -772,7 +772,7 @@ define half @test_pown_afn_nnan_ninf_f16(half %x, i32 %y) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast half [[X]] to i16
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i16 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast half [[__EXP2]] to i16
-; CHECK-NEXT:    [[TMP2:%.*]] = or i16 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i16 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i16 [[TMP2]] to half
 ; CHECK-NEXT:    ret half [[TMP3]]
 ;
@@ -795,7 +795,7 @@ define <2 x half> @test_pown_afn_nnan_ninf_v2f16(<2 x half> %x, <2 x i32> %y) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <2 x half> [[X]] to <2 x i16>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i16> [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <2 x half> [[__EXP2]] to <2 x i16>
-; CHECK-NEXT:    [[TMP2:%.*]] = or <2 x i16> [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint <2 x i16> [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x i16> [[TMP2]] to <2 x half>
 ; CHECK-NEXT:    ret <2 x half> [[TMP3]]
 ;
@@ -829,7 +829,7 @@ define float @test_pown_fast_f32_strictfp(float %x, i32 %y) #1 {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP2:%.*]] = or i32 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32 [[TMP2]] to float
 ; CHECK-NEXT:    ret float [[TMP3]]
 ;
@@ -1075,7 +1075,7 @@ define float @test_pown_afn_ninf_nnan_f32__x_known_positive(float nofpclass(ninf
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP2:%.*]] = or i32 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32 [[TMP2]] to float
 ; CHECK-NEXT:    ret float [[TMP3]]
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll b/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
index 204c8140d3f17..54ca33401ccf4 100644
--- a/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
+++ b/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
@@ -360,7 +360,7 @@ declare half @_Z4pownDhi(half, i32)
 ; GCN-NATIVE: %0 = bitcast half %x to i16
 ; GCN-NATIVE: %__pow_sign = and i16 %__yeven, %0
 ; GCN-NATIVE: %1 = bitcast half %__exp2 to i16
-; GCN-NATIVE: %2 = or i16 %__pow_sign, %1
+; GCN-NATIVE: %2 = or disjoint i16 %__pow_sign, %1
 ; GCN-NATIVE: %3 = bitcast i16 %2 to half
 define half @test_pown_f16(half %x, i32 %y) {
 entry:
@@ -378,7 +378,7 @@ declare float @_Z4pownfi(float, i32)
 ; GCN: %[[r0:.*]] = bitcast float %tmp to i32
 ; GCN: %__pow_sign = and i32 %[[r0]], -2147483648
 ; GCN: %[[r1:.*]] = bitcast float %__exp2 to i32
-; GCN: %[[r2:.*]] = or i32 %__pow_sign, %[[r1]]
+; GCN: %[[r2:.*]] = or disjoint i32 %__pow_sign, %[[r1]]
 ; GCN: store i32 %[[r2]], ptr addrspace(1) %a, align 4
 define amdgpu_kernel void @test_pow(ptr addrspace(1) nocapture %a) {
 entry:
@@ -414,7 +414,7 @@ entry:
 ; GCN: %[[r0:.*]] = bitcast float %tmp to i32
 ; GCN: %__pow_sign = and i32 %__yeven, %[[r0]]
 ; GCN: %[[r1:.*]] = bitcast float %__exp2 to i32
-; GCN: %[[r2:.*]] = or i32 %__pow_sign, %[[r1]]
+; GCN: %[[r2:.*]] = or disjoint i32 %__pow_sign, %[[r1]]
 ; GCN: store i32 %[[r2]], ptr addrspace(1) %a, align 4
 define amdgpu_kernel void @test_pown(ptr addrspace(1) nocapture %a) {
 entry:
@@ -438,7 +438,7 @@ declare <2 x half> @_Z3powDv2_DhS_(<2 x half>, <2 x half>)
 ; GCN: %1 = bitcast half %x to i16
 ; GCN: %__pow_sign = and i16 %1, -32768
 ; GCN: %2 = bitcast half %__exp2 to i16
-; GCN: %3 = or i16 %__pow_sign, %2
+; GCN: %3 = or disjoint i16 %__pow_sign, %2
 ; GCN: %4 = bitcast i16 %3 to half
 define half @test_pow_fast_f16__y_13(half %x) {
   %powr = tail call fast half @_Z3powDhDh(half %x, half 13.0)
@@ -453,7 +453,7 @@ define half @test_pow_fast_f16__y_13(half %x) {
 ; GCN: %1 = bitcast <2 x half> %x to <2 x i16>
 ; GCN: %__pow_sign = and <2 x i16> %1, <i16 -32768, i16 -32768>
 ; GCN: %2 = bitcast <2 x half> %__exp2 to <2 x i16>
-; GCN: %3 = or <2 x i16> %__pow_sign, %2
+; GCN: %3 = or disjoint <2 x i16> %__pow_sign, %2
 ; GCN: %4 = bitcast <2 x i16> %3 to <2 x half>
 define <2 x half> @test_pow_fast_v2f16__y_13(<2 x half> %x) {
   %powr = tail call fast <2 x half> @_Z3powDv2_DhS_(<2 x half> %x, <2 x half> <half 13.0, half 13.0>)
diff --git a/llvm/test/Transforms/InstCombine/known-bits.ll b/llvm/test/Transforms/InstCombine/known-bits.ll
index 8b4249b2c25a9..7a38020500517 100644
--- a/llvm/test/Transforms/InstCombine/known-bits.ll
+++ b/llvm/test/Transforms/InstCombine/known-bits.ll
@@ -1374,5 +1374,289 @@ define i8 @nonzero_reduce_xor_vscale_odd(<vscale x 3 x i8> %xx) {
   ret i8 %r
 }
 
+define i1 @test_sign_pos(float %x) {
+; CHECK-LABEL: @test_sign_pos(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call float @llvm.fabs.f32(float %x)
+  %y = bitcast float %fabs to i32
+  %sign = icmp sgt i32 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_pos_half(half %x) {
+; CHECK-LABEL: @test_sign_pos_half(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call half @llvm.fabs.f16(half %x)
+  %y = bitcast half %fabs to i16
+  %sign = icmp sgt i16 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_pos_half_non_elementwise(<2 x half> %x) {
+; CHECK-LABEL: @test_sign_pos_half_non_elementwise(
+; CHECK-NEXT:    [[FABS:%.*]] = call <2 x half> @llvm.fabs.v2f16(<2 x half> [[X:%.*]])
+; CHECK-NEXT:    [[Y:%.*]] = bitcast <2 x half> [[FABS]] to i32
+; CHECK-NEXT:    [[SIGN:%.*]] = icmp sgt i32 [[Y]], -1
+; CHECK-NEXT:    ret i1 [[SIGN]]
+;
+  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %x)
+  %y = bitcast <2 x half> %fabs to i32
+  %sign = icmp sgt i32 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_neg(float %x) {
+; CHECK-LABEL: @test_sign_neg(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call float @llvm.fabs.f32(float %x)
+  %fnabs = fneg float %fabs
+  %y = bitcast float %fnabs to i32
+  %sign = icmp slt i32 %y, 0
+  ret i1 %sign
+}
+
+define <2 x i1> @test_sign_pos_vec(<2 x float> %x) {
+; CHECK-LABEL: @test_sign_pos_vec(
+; CHECK-NEXT:    ret <2 x i1> zeroinitializer
+;
+  %fabs = call <2 x float> @llvm.fabs.v2f32(<2 x float> %x)
+  %y = bitcast <2 x float> %fabs to <2 x i32>
+  %sign = icmp slt <2 x i32> %y, zeroinitializer
+  ret <2 x i1> %sign
+}
+
+define i32 @test_inf_only(float nofpclass(nan sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_only(
+; CHECK-NEXT:    ret i32 2139095040
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2147483647
+  ret i32 %and
+}
+
+define i16 @test_inf_only_bfloat(bfloat nofpclass(nan sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_only_bfloat(
+; CHECK-NEXT:    ret i16 32640
+;
+  %y = bitcast bfloat %x to i16
+  %and = and i16 %y, 32767
+  ret i16 %and
+}
+
+define i128 @test_inf_only_ppc_fp128(ppc_fp128 nofpclass(nan sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_only_ppc_fp128(
+; CHECK-NEXT:    ret i128 9218868437227405312
+;
+  %y = bitcast ppc_fp128 %x to i128
+  %and = and i128 %y, 170141183460469231731687303715884105727
+  ret i128 %and
+}
+
+define i32 @test_zero_only(float nofpclass(nan sub norm inf) %x) {
+; CHECK-LABEL: @test_zero_only(
+; CHECK-NEXT:    ret i32 0
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2147483647
+  ret i32 %and
+}
+
+define i80 @test_zero_only_non_ieee(x86_fp80 nofpclass(nan sub norm inf) %x) {
+; CHECK-LABEL: @test_zero_only_non_ieee(
+; CHECK-NEXT:    ret i80 0
+;
+  %y = bitcast x86_fp80 %x to i80
+  %and = and i80 %y, 604462909807314587353087
+  ret i80 %and
+}
+
+define i32 @test_inf_nan_only(float nofpclass(sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_nan_only(
+; CHECK-NEXT:    [[Y:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[Y]], 2130706432
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2130706432
+  ret i32 %and
+}
+
+define i32 @test_sub_zero_only(float nofpclass(nan norm inf) %x) {
+; CHECK-LABEL: @test_sub_zero_only(
+; CHECK-NEXT:    [[Y:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[Y]], 2130706432
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2130706432
+  ret i32 %and
+}
+
+define i32 @test_inf_zero_only(float nofpclass(nan norm sub) %x) {
+; CHECK-LABEL: @test_inf_zero_only(
+; CHECK-NEXT:    [[Y:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[Y]], 8388608
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 16777215
+  ret i32 %and
+}
+
+; Make sure that the signbit is cleared.
+define i32 @test_ninf_only(double %x) {
+; CHECK-LABEL: @test_ninf_only(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp oeq double [[X:%.*]], 0xFFF0000000000000
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    ret i32 0
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i32 0
+;
+  %cmp = fcmp oeq double %x, 0xFFF0000000000000
+  br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+  %cast = bitcast double %x to i64
+  %trunc = trunc i64 %cast to i32
+  ret i32 %trunc
+
+if.else:
+  ret i32 0
+}
+
+define i1 @test_simplify_icmp(i32 %x) {
+; CHECK-LABEL: @test_simplify_icmp(
+; CHECK-NEXT:    ret i1 false
+;
+  %cast1 = uitofp i32 %x to double
+  %cast2 = bitcast double %cast1 to i64
+  %mask = and i64 %cast2, -140737488355328
+  %cmp = icmp eq i64 %mask, -1970324836974592
+  ret i1 %cmp
+}
+
+define i32 @test_snan_quiet_bit1(float nofpclass(sub norm inf qnan) %x) {
+; CHECK-LABEL: @test_snan_quiet_bit1(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 4194304
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 4194304
+  ret i32 %masked
+}
+
+define i32 @test_snan_quiet_bit2(float nofpclass(sub norm inf qnan) %x) {
+; CHECK-LABEL: @test_snan_quiet_bit2(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 2097152
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 2097152
+  ret i32 %masked
+}
+
+define i32 @test_qnan_quiet_bit1(float nofpclass(sub norm inf snan) %x) {
+; CHECK-LABEL: @test_qnan_quiet_bit1(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 4194304
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 4194304
+  ret i32 %masked
+}
+
+define i32 @test_qnan_quiet_bit2(float nofpclass(sub norm inf snan) %x) {
+; CHECK-LABEL: @test_qnan_quiet_bit2(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 2097152
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 2097152
+  ret i32 %masked
+}
+
+define i16 @test_simplify_mask(i32 %ui, float %x) {
+; CHECK-LABEL: @test_simplify_mask(
+; CHECK-NEXT:    [[CONV:%.*]] = uitofp i32 [[UI:%.*]] to float
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ogt float [[CONV]], [[X:%.*]]
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_ELSE:%.*]], label [[IF_END:%.*]]
+; CHECK:       if.end:
+; CHECK-NEXT:    ret i16 31744
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i16 0
+;
+  %conv = uitofp i32 %ui to float
+  %cmp = fcmp olt float %x, %conv
+  br i1 %cmp, label %if.else, label %if.end
+
+if.end:
+  %cast = bitcast float %conv to i32
+  %shr = lshr i32 %cast, 16
+  %trunc = trunc i32 %shr to i16
+  %and = and i16 %trunc, -32768
+  %or = or disjoint i16 %and, 31744
+  ret i16 %or
+
+if.else:
+  ret i16 0
+}
+
+; TODO: %cmp always evaluates to false
+
+define i1 @test_simplify_icmp2(double %x) {
+; CHECK-LABEL: @test_simplify_icmp2(
+; CHECK-NEXT:    [[ABS:%.*]] = tail call double @llvm.fabs.f64(double [[X:%.*]])
+; CHECK-NEXT:    [[COND:%.*]] = fcmp oeq double [[ABS]], 0x7FF0000000000000
+; CHECK-NEXT:    br i1 [[COND]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    [[CAST:%.*]] = bitcast double [[X]] to i64
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i64 [[CAST]], 3458764513820540928
+; CHECK-NEXT:    ret i1 [[CMP]]
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i1 false
+;
+  %ab...
[truncated]

llvmbot · 2024-05-14T08:36:51Z

@llvm/pr-subscribers-backend-amdgpu

Author: Yingwei Zheng (dtcxzyw)

Changes

This patch relands #86409.

I mistakenly thought that Known.makeNegative() clears the sign bit of Known.Zero. This patch fixes the assertion failure by explicitly clearing the sign bit.

Patch is 21.00 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/92084.diff

6 Files Affected:

(modified) llvm/include/llvm/IR/PatternMatch.h (+1-1)
(modified) llvm/lib/Analysis/ValueTracking.cpp (+36)
(modified) llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll (+7-7)
(modified) llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll (+6-6)
(modified) llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll (+5-5)
(modified) llvm/test/Transforms/InstCombine/known-bits.ll (+284)

diff --git a/llvm/include/llvm/IR/PatternMatch.h b/llvm/include/llvm/IR/PatternMatch.h
index 171ddab977dea..0d6d86cb47e67 100644
--- a/llvm/include/llvm/IR/PatternMatch.h
+++ b/llvm/include/llvm/IR/PatternMatch.h
@@ -1904,7 +1904,7 @@ template <typename Op_t> struct ElementWiseBitCast_match {
   ElementWiseBitCast_match(const Op_t &OpMatch) : Op(OpMatch) {}
 
   template <typename OpTy> bool match(OpTy *V) {
-    BitCastInst *I = dyn_cast<BitCastInst>(V);
+    auto *I = dyn_cast<BitCastInst>(V);
     if (!I)
       return false;
     Type *SrcType = I->getSrcTy();
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 375385aca7a39..99b621461fcdc 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -1118,6 +1118,42 @@ static void computeKnownBitsFromOperator(const Operator *I,
       break;
     }
 
+    const Value *V;
+    // Handle bitcast from floating point to integer.
+    if (match(I, m_ElementWiseBitCast(m_Value(V))) &&
+        V->getType()->isFPOrFPVectorTy()) {
+      Type *FPType = V->getType()->getScalarType();
+      KnownFPClass Result = computeKnownFPClass(V, fcAllFlags, Depth + 1, Q);
+      FPClassTest FPClasses = Result.KnownFPClasses;
+
+      if (Result.isKnownNever(fcNormal | fcSubnormal | fcNan)) {
+        Known.Zero.setAllBits();
+        Known.One.setAllBits();
+
+        if (FPClasses & fcInf)
+          Known = Known.intersectWith(KnownBits::makeConstant(
+              APFloat::getInf(FPType->getFltSemantics()).bitcastToAPInt()));
+
+        if (FPClasses & fcZero)
+          Known = Known.intersectWith(KnownBits::makeConstant(
+              APInt::getZero(FPType->getScalarSizeInBits())));
+
+        Known.Zero.clearSignBit();
+        Known.One.clearSignBit();
+      }
+
+      if (Result.SignBit) {
+        if (*Result.SignBit)
+          Known.makeNegative();
+        else
+          Known.makeNonNegative();
+      }
+
+      assert(!Known.hasConflict() && "Bits known to be one AND zero?");
+
+      break;
+    }
+
     // Handle cast from vector integer type to scalar or vector integer.
     auto *SrcVecTy = dyn_cast<FixedVectorType>(SrcTy);
     if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() ||
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
index c4bd4bc126f73..5db25a59d33fc 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
@@ -2216,7 +2216,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_sitofp(float %x, i32 %y)
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2304,7 +2304,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_uitofp(float %x, i32 %y)
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2353,7 +2353,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_uitofp_i256(float %x, i2
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2376,7 +2376,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_sitofp_i256(float %x, i2
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2399,7 +2399,7 @@ define <2 x float> @test_pow_afn_nnan_ninf_v2f32_known_integral_sitofp(<2 x floa
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <2 x float> [[X]] to <2 x i32>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i32> [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x float> [[__EXP2]] to <2 x i32>
-; CHECK-NEXT:    [[TMP4:%.*]] = or <2 x i32> [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint <2 x i32> [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <2 x float>
 ; CHECK-NEXT:    ret <2 x float> [[TMP5]]
 ;
@@ -2448,7 +2448,7 @@ define <2 x float> @test_pow_afn_nnan_ninf_v2f32_known_integral_uitofp(<2 x floa
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <2 x float> [[X]] to <2 x i32>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i32> [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x float> [[__EXP2]] to <2 x i32>
-; CHECK-NEXT:    [[TMP4:%.*]] = or <2 x i32> [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint <2 x i32> [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <2 x float>
 ; CHECK-NEXT:    ret <2 x float> [[TMP5]]
 ;
@@ -2560,7 +2560,7 @@ define float @test_pow_afn_f32_nnan_ninf__y_known_integral_trunc(float %x, float
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll
index 8ddaf243db92c..e298226ee7ccd 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll
@@ -680,7 +680,7 @@ define float @test_pown_afn_nnan_ninf_f32(float %x, i32 %y) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP2:%.*]] = or i32 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32 [[TMP2]] to float
 ; CHECK-NEXT:    ret float [[TMP3]]
 ;
@@ -703,7 +703,7 @@ define <2 x float> @test_pown_afn_nnan_ninf_v2f32(<2 x float> %x, <2 x i32> %y)
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <2 x float> [[X]] to <2 x i32>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i32> [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <2 x float> [[__EXP2]] to <2 x i32>
-; CHECK-NEXT:    [[TMP2:%.*]] = or <2 x i32> [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint <2 x i32> [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x i32> [[TMP2]] to <2 x float>
 ; CHECK-NEXT:    ret <2 x float> [[TMP3]]
 ;
@@ -772,7 +772,7 @@ define half @test_pown_afn_nnan_ninf_f16(half %x, i32 %y) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast half [[X]] to i16
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i16 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast half [[__EXP2]] to i16
-; CHECK-NEXT:    [[TMP2:%.*]] = or i16 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i16 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i16 [[TMP2]] to half
 ; CHECK-NEXT:    ret half [[TMP3]]
 ;
@@ -795,7 +795,7 @@ define <2 x half> @test_pown_afn_nnan_ninf_v2f16(<2 x half> %x, <2 x i32> %y) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <2 x half> [[X]] to <2 x i16>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i16> [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <2 x half> [[__EXP2]] to <2 x i16>
-; CHECK-NEXT:    [[TMP2:%.*]] = or <2 x i16> [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint <2 x i16> [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x i16> [[TMP2]] to <2 x half>
 ; CHECK-NEXT:    ret <2 x half> [[TMP3]]
 ;
@@ -829,7 +829,7 @@ define float @test_pown_fast_f32_strictfp(float %x, i32 %y) #1 {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP2:%.*]] = or i32 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32 [[TMP2]] to float
 ; CHECK-NEXT:    ret float [[TMP3]]
 ;
@@ -1075,7 +1075,7 @@ define float @test_pown_afn_ninf_nnan_f32__x_known_positive(float nofpclass(ninf
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP2:%.*]] = or i32 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32 [[TMP2]] to float
 ; CHECK-NEXT:    ret float [[TMP3]]
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll b/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
index 204c8140d3f17..54ca33401ccf4 100644
--- a/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
+++ b/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
@@ -360,7 +360,7 @@ declare half @_Z4pownDhi(half, i32)
 ; GCN-NATIVE: %0 = bitcast half %x to i16
 ; GCN-NATIVE: %__pow_sign = and i16 %__yeven, %0
 ; GCN-NATIVE: %1 = bitcast half %__exp2 to i16
-; GCN-NATIVE: %2 = or i16 %__pow_sign, %1
+; GCN-NATIVE: %2 = or disjoint i16 %__pow_sign, %1
 ; GCN-NATIVE: %3 = bitcast i16 %2 to half
 define half @test_pown_f16(half %x, i32 %y) {
 entry:
@@ -378,7 +378,7 @@ declare float @_Z4pownfi(float, i32)
 ; GCN: %[[r0:.*]] = bitcast float %tmp to i32
 ; GCN: %__pow_sign = and i32 %[[r0]], -2147483648
 ; GCN: %[[r1:.*]] = bitcast float %__exp2 to i32
-; GCN: %[[r2:.*]] = or i32 %__pow_sign, %[[r1]]
+; GCN: %[[r2:.*]] = or disjoint i32 %__pow_sign, %[[r1]]
 ; GCN: store i32 %[[r2]], ptr addrspace(1) %a, align 4
 define amdgpu_kernel void @test_pow(ptr addrspace(1) nocapture %a) {
 entry:
@@ -414,7 +414,7 @@ entry:
 ; GCN: %[[r0:.*]] = bitcast float %tmp to i32
 ; GCN: %__pow_sign = and i32 %__yeven, %[[r0]]
 ; GCN: %[[r1:.*]] = bitcast float %__exp2 to i32
-; GCN: %[[r2:.*]] = or i32 %__pow_sign, %[[r1]]
+; GCN: %[[r2:.*]] = or disjoint i32 %__pow_sign, %[[r1]]
 ; GCN: store i32 %[[r2]], ptr addrspace(1) %a, align 4
 define amdgpu_kernel void @test_pown(ptr addrspace(1) nocapture %a) {
 entry:
@@ -438,7 +438,7 @@ declare <2 x half> @_Z3powDv2_DhS_(<2 x half>, <2 x half>)
 ; GCN: %1 = bitcast half %x to i16
 ; GCN: %__pow_sign = and i16 %1, -32768
 ; GCN: %2 = bitcast half %__exp2 to i16
-; GCN: %3 = or i16 %__pow_sign, %2
+; GCN: %3 = or disjoint i16 %__pow_sign, %2
 ; GCN: %4 = bitcast i16 %3 to half
 define half @test_pow_fast_f16__y_13(half %x) {
   %powr = tail call fast half @_Z3powDhDh(half %x, half 13.0)
@@ -453,7 +453,7 @@ define half @test_pow_fast_f16__y_13(half %x) {
 ; GCN: %1 = bitcast <2 x half> %x to <2 x i16>
 ; GCN: %__pow_sign = and <2 x i16> %1, <i16 -32768, i16 -32768>
 ; GCN: %2 = bitcast <2 x half> %__exp2 to <2 x i16>
-; GCN: %3 = or <2 x i16> %__pow_sign, %2
+; GCN: %3 = or disjoint <2 x i16> %__pow_sign, %2
 ; GCN: %4 = bitcast <2 x i16> %3 to <2 x half>
 define <2 x half> @test_pow_fast_v2f16__y_13(<2 x half> %x) {
   %powr = tail call fast <2 x half> @_Z3powDv2_DhS_(<2 x half> %x, <2 x half> <half 13.0, half 13.0>)
diff --git a/llvm/test/Transforms/InstCombine/known-bits.ll b/llvm/test/Transforms/InstCombine/known-bits.ll
index 8b4249b2c25a9..7a38020500517 100644
--- a/llvm/test/Transforms/InstCombine/known-bits.ll
+++ b/llvm/test/Transforms/InstCombine/known-bits.ll
@@ -1374,5 +1374,289 @@ define i8 @nonzero_reduce_xor_vscale_odd(<vscale x 3 x i8> %xx) {
   ret i8 %r
 }
 
+define i1 @test_sign_pos(float %x) {
+; CHECK-LABEL: @test_sign_pos(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call float @llvm.fabs.f32(float %x)
+  %y = bitcast float %fabs to i32
+  %sign = icmp sgt i32 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_pos_half(half %x) {
+; CHECK-LABEL: @test_sign_pos_half(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call half @llvm.fabs.f16(half %x)
+  %y = bitcast half %fabs to i16
+  %sign = icmp sgt i16 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_pos_half_non_elementwise(<2 x half> %x) {
+; CHECK-LABEL: @test_sign_pos_half_non_elementwise(
+; CHECK-NEXT:    [[FABS:%.*]] = call <2 x half> @llvm.fabs.v2f16(<2 x half> [[X:%.*]])
+; CHECK-NEXT:    [[Y:%.*]] = bitcast <2 x half> [[FABS]] to i32
+; CHECK-NEXT:    [[SIGN:%.*]] = icmp sgt i32 [[Y]], -1
+; CHECK-NEXT:    ret i1 [[SIGN]]
+;
+  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %x)
+  %y = bitcast <2 x half> %fabs to i32
+  %sign = icmp sgt i32 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_neg(float %x) {
+; CHECK-LABEL: @test_sign_neg(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call float @llvm.fabs.f32(float %x)
+  %fnabs = fneg float %fabs
+  %y = bitcast float %fnabs to i32
+  %sign = icmp slt i32 %y, 0
+  ret i1 %sign
+}
+
+define <2 x i1> @test_sign_pos_vec(<2 x float> %x) {
+; CHECK-LABEL: @test_sign_pos_vec(
+; CHECK-NEXT:    ret <2 x i1> zeroinitializer
+;
+  %fabs = call <2 x float> @llvm.fabs.v2f32(<2 x float> %x)
+  %y = bitcast <2 x float> %fabs to <2 x i32>
+  %sign = icmp slt <2 x i32> %y, zeroinitializer
+  ret <2 x i1> %sign
+}
+
+define i32 @test_inf_only(float nofpclass(nan sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_only(
+; CHECK-NEXT:    ret i32 2139095040
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2147483647
+  ret i32 %and
+}
+
+define i16 @test_inf_only_bfloat(bfloat nofpclass(nan sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_only_bfloat(
+; CHECK-NEXT:    ret i16 32640
+;
+  %y = bitcast bfloat %x to i16
+  %and = and i16 %y, 32767
+  ret i16 %and
+}
+
+define i128 @test_inf_only_ppc_fp128(ppc_fp128 nofpclass(nan sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_only_ppc_fp128(
+; CHECK-NEXT:    ret i128 9218868437227405312
+;
+  %y = bitcast ppc_fp128 %x to i128
+  %and = and i128 %y, 170141183460469231731687303715884105727
+  ret i128 %and
+}
+
+define i32 @test_zero_only(float nofpclass(nan sub norm inf) %x) {
+; CHECK-LABEL: @test_zero_only(
+; CHECK-NEXT:    ret i32 0
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2147483647
+  ret i32 %and
+}
+
+define i80 @test_zero_only_non_ieee(x86_fp80 nofpclass(nan sub norm inf) %x) {
+; CHECK-LABEL: @test_zero_only_non_ieee(
+; CHECK-NEXT:    ret i80 0
+;
+  %y = bitcast x86_fp80 %x to i80
+  %and = and i80 %y, 604462909807314587353087
+  ret i80 %and
+}
+
+define i32 @test_inf_nan_only(float nofpclass(sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_nan_only(
+; CHECK-NEXT:    [[Y:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[Y]], 2130706432
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2130706432
+  ret i32 %and
+}
+
+define i32 @test_sub_zero_only(float nofpclass(nan norm inf) %x) {
+; CHECK-LABEL: @test_sub_zero_only(
+; CHECK-NEXT:    [[Y:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[Y]], 2130706432
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2130706432
+  ret i32 %and
+}
+
+define i32 @test_inf_zero_only(float nofpclass(nan norm sub) %x) {
+; CHECK-LABEL: @test_inf_zero_only(
+; CHECK-NEXT:    [[Y:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[Y]], 8388608
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 16777215
+  ret i32 %and
+}
+
+; Make sure that the signbit is cleared.
+define i32 @test_ninf_only(double %x) {
+; CHECK-LABEL: @test_ninf_only(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp oeq double [[X:%.*]], 0xFFF0000000000000
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    ret i32 0
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i32 0
+;
+  %cmp = fcmp oeq double %x, 0xFFF0000000000000
+  br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+  %cast = bitcast double %x to i64
+  %trunc = trunc i64 %cast to i32
+  ret i32 %trunc
+
+if.else:
+  ret i32 0
+}
+
+define i1 @test_simplify_icmp(i32 %x) {
+; CHECK-LABEL: @test_simplify_icmp(
+; CHECK-NEXT:    ret i1 false
+;
+  %cast1 = uitofp i32 %x to double
+  %cast2 = bitcast double %cast1 to i64
+  %mask = and i64 %cast2, -140737488355328
+  %cmp = icmp eq i64 %mask, -1970324836974592
+  ret i1 %cmp
+}
+
+define i32 @test_snan_quiet_bit1(float nofpclass(sub norm inf qnan) %x) {
+; CHECK-LABEL: @test_snan_quiet_bit1(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 4194304
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 4194304
+  ret i32 %masked
+}
+
+define i32 @test_snan_quiet_bit2(float nofpclass(sub norm inf qnan) %x) {
+; CHECK-LABEL: @test_snan_quiet_bit2(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 2097152
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 2097152
+  ret i32 %masked
+}
+
+define i32 @test_qnan_quiet_bit1(float nofpclass(sub norm inf snan) %x) {
+; CHECK-LABEL: @test_qnan_quiet_bit1(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 4194304
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 4194304
+  ret i32 %masked
+}
+
+define i32 @test_qnan_quiet_bit2(float nofpclass(sub norm inf snan) %x) {
+; CHECK-LABEL: @test_qnan_quiet_bit2(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 2097152
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 2097152
+  ret i32 %masked
+}
+
+define i16 @test_simplify_mask(i32 %ui, float %x) {
+; CHECK-LABEL: @test_simplify_mask(
+; CHECK-NEXT:    [[CONV:%.*]] = uitofp i32 [[UI:%.*]] to float
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ogt float [[CONV]], [[X:%.*]]
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_ELSE:%.*]], label [[IF_END:%.*]]
+; CHECK:       if.end:
+; CHECK-NEXT:    ret i16 31744
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i16 0
+;
+  %conv = uitofp i32 %ui to float
+  %cmp = fcmp olt float %x, %conv
+  br i1 %cmp, label %if.else, label %if.end
+
+if.end:
+  %cast = bitcast float %conv to i32
+  %shr = lshr i32 %cast, 16
+  %trunc = trunc i32 %shr to i16
+  %and = and i16 %trunc, -32768
+  %or = or disjoint i16 %and, 31744
+  ret i16 %or
+
+if.else:
+  ret i16 0
+}
+
+; TODO: %cmp always evaluates to false
+
+define i1 @test_simplify_icmp2(double %x) {
+; CHECK-LABEL: @test_simplify_icmp2(
+; CHECK-NEXT:    [[ABS:%.*]] = tail call double @llvm.fabs.f64(double [[X:%.*]])
+; CHECK-NEXT:    [[COND:%.*]] = fcmp oeq double [[ABS]], 0x7FF0000000000000
+; CHECK-NEXT:    br i1 [[COND]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    [[CAST:%.*]] = bitcast double [[X]] to i64
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i64 [[CAST]], 3458764513820540928
+; CHECK-NEXT:    ret i1 [[CMP]]
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i1 false
+;
+  %ab...
[truncated]

llvmbot · 2024-05-14T08:36:52Z

@llvm/pr-subscribers-llvm-analysis

Author: Yingwei Zheng (dtcxzyw)

Changes

This patch relands #86409.

I mistakenly thought that Known.makeNegative() clears the sign bit of Known.Zero. This patch fixes the assertion failure by explicitly clearing the sign bit.

Patch is 21.00 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/92084.diff

6 Files Affected:

(modified) llvm/include/llvm/IR/PatternMatch.h (+1-1)
(modified) llvm/lib/Analysis/ValueTracking.cpp (+36)
(modified) llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll (+7-7)
(modified) llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll (+6-6)
(modified) llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll (+5-5)
(modified) llvm/test/Transforms/InstCombine/known-bits.ll (+284)

diff --git a/llvm/include/llvm/IR/PatternMatch.h b/llvm/include/llvm/IR/PatternMatch.h
index 171ddab977dea..0d6d86cb47e67 100644
--- a/llvm/include/llvm/IR/PatternMatch.h
+++ b/llvm/include/llvm/IR/PatternMatch.h
@@ -1904,7 +1904,7 @@ template <typename Op_t> struct ElementWiseBitCast_match {
   ElementWiseBitCast_match(const Op_t &OpMatch) : Op(OpMatch) {}
 
   template <typename OpTy> bool match(OpTy *V) {
-    BitCastInst *I = dyn_cast<BitCastInst>(V);
+    auto *I = dyn_cast<BitCastInst>(V);
     if (!I)
       return false;
     Type *SrcType = I->getSrcTy();
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 375385aca7a39..99b621461fcdc 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -1118,6 +1118,42 @@ static void computeKnownBitsFromOperator(const Operator *I,
       break;
     }
 
+    const Value *V;
+    // Handle bitcast from floating point to integer.
+    if (match(I, m_ElementWiseBitCast(m_Value(V))) &&
+        V->getType()->isFPOrFPVectorTy()) {
+      Type *FPType = V->getType()->getScalarType();
+      KnownFPClass Result = computeKnownFPClass(V, fcAllFlags, Depth + 1, Q);
+      FPClassTest FPClasses = Result.KnownFPClasses;
+
+      if (Result.isKnownNever(fcNormal | fcSubnormal | fcNan)) {
+        Known.Zero.setAllBits();
+        Known.One.setAllBits();
+
+        if (FPClasses & fcInf)
+          Known = Known.intersectWith(KnownBits::makeConstant(
+              APFloat::getInf(FPType->getFltSemantics()).bitcastToAPInt()));
+
+        if (FPClasses & fcZero)
+          Known = Known.intersectWith(KnownBits::makeConstant(
+              APInt::getZero(FPType->getScalarSizeInBits())));
+
+        Known.Zero.clearSignBit();
+        Known.One.clearSignBit();
+      }
+
+      if (Result.SignBit) {
+        if (*Result.SignBit)
+          Known.makeNegative();
+        else
+          Known.makeNonNegative();
+      }
+
+      assert(!Known.hasConflict() && "Bits known to be one AND zero?");
+
+      break;
+    }
+
     // Handle cast from vector integer type to scalar or vector integer.
     auto *SrcVecTy = dyn_cast<FixedVectorType>(SrcTy);
     if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() ||
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
index c4bd4bc126f73..5db25a59d33fc 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll
@@ -2216,7 +2216,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_sitofp(float %x, i32 %y)
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2304,7 +2304,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_uitofp(float %x, i32 %y)
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2353,7 +2353,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_uitofp_i256(float %x, i2
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2376,7 +2376,7 @@ define float @test_pow_afn_nnan_ninf_f32_known_integral_sitofp_i256(float %x, i2
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
@@ -2399,7 +2399,7 @@ define <2 x float> @test_pow_afn_nnan_ninf_v2f32_known_integral_sitofp(<2 x floa
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <2 x float> [[X]] to <2 x i32>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i32> [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x float> [[__EXP2]] to <2 x i32>
-; CHECK-NEXT:    [[TMP4:%.*]] = or <2 x i32> [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint <2 x i32> [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <2 x float>
 ; CHECK-NEXT:    ret <2 x float> [[TMP5]]
 ;
@@ -2448,7 +2448,7 @@ define <2 x float> @test_pow_afn_nnan_ninf_v2f32_known_integral_uitofp(<2 x floa
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <2 x float> [[X]] to <2 x i32>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i32> [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x float> [[__EXP2]] to <2 x i32>
-; CHECK-NEXT:    [[TMP4:%.*]] = or <2 x i32> [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint <2 x i32> [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <2 x float>
 ; CHECK-NEXT:    ret <2 x float> [[TMP5]]
 ;
@@ -2560,7 +2560,7 @@ define float @test_pow_afn_f32_nnan_ninf__y_known_integral_trunc(float %x, float
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP4:%.*]] = or i32 [[__POW_SIGN]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32 [[TMP4]] to float
 ; CHECK-NEXT:    ret float [[TMP5]]
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll
index 8ddaf243db92c..e298226ee7ccd 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pown.ll
@@ -680,7 +680,7 @@ define float @test_pown_afn_nnan_ninf_f32(float %x, i32 %y) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP2:%.*]] = or i32 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32 [[TMP2]] to float
 ; CHECK-NEXT:    ret float [[TMP3]]
 ;
@@ -703,7 +703,7 @@ define <2 x float> @test_pown_afn_nnan_ninf_v2f32(<2 x float> %x, <2 x i32> %y)
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <2 x float> [[X]] to <2 x i32>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i32> [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <2 x float> [[__EXP2]] to <2 x i32>
-; CHECK-NEXT:    [[TMP2:%.*]] = or <2 x i32> [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint <2 x i32> [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x i32> [[TMP2]] to <2 x float>
 ; CHECK-NEXT:    ret <2 x float> [[TMP3]]
 ;
@@ -772,7 +772,7 @@ define half @test_pown_afn_nnan_ninf_f16(half %x, i32 %y) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast half [[X]] to i16
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i16 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast half [[__EXP2]] to i16
-; CHECK-NEXT:    [[TMP2:%.*]] = or i16 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i16 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i16 [[TMP2]] to half
 ; CHECK-NEXT:    ret half [[TMP3]]
 ;
@@ -795,7 +795,7 @@ define <2 x half> @test_pown_afn_nnan_ninf_v2f16(<2 x half> %x, <2 x i32> %y) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <2 x half> [[X]] to <2 x i16>
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and <2 x i16> [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <2 x half> [[__EXP2]] to <2 x i16>
-; CHECK-NEXT:    [[TMP2:%.*]] = or <2 x i16> [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint <2 x i16> [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <2 x i16> [[TMP2]] to <2 x half>
 ; CHECK-NEXT:    ret <2 x half> [[TMP3]]
 ;
@@ -829,7 +829,7 @@ define float @test_pown_fast_f32_strictfp(float %x, i32 %y) #1 {
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP2:%.*]] = or i32 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32 [[TMP2]] to float
 ; CHECK-NEXT:    ret float [[TMP3]]
 ;
@@ -1075,7 +1075,7 @@ define float @test_pown_afn_ninf_nnan_f32__x_known_positive(float nofpclass(ninf
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast float [[X]] to i32
 ; CHECK-NEXT:    [[__POW_SIGN:%.*]] = and i32 [[__YEVEN]], [[TMP0]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float [[__EXP2]] to i32
-; CHECK-NEXT:    [[TMP2:%.*]] = or i32 [[__POW_SIGN]], [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = or disjoint i32 [[__POW_SIGN]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32 [[TMP2]] to float
 ; CHECK-NEXT:    ret float [[TMP3]]
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll b/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
index 204c8140d3f17..54ca33401ccf4 100644
--- a/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
+++ b/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
@@ -360,7 +360,7 @@ declare half @_Z4pownDhi(half, i32)
 ; GCN-NATIVE: %0 = bitcast half %x to i16
 ; GCN-NATIVE: %__pow_sign = and i16 %__yeven, %0
 ; GCN-NATIVE: %1 = bitcast half %__exp2 to i16
-; GCN-NATIVE: %2 = or i16 %__pow_sign, %1
+; GCN-NATIVE: %2 = or disjoint i16 %__pow_sign, %1
 ; GCN-NATIVE: %3 = bitcast i16 %2 to half
 define half @test_pown_f16(half %x, i32 %y) {
 entry:
@@ -378,7 +378,7 @@ declare float @_Z4pownfi(float, i32)
 ; GCN: %[[r0:.*]] = bitcast float %tmp to i32
 ; GCN: %__pow_sign = and i32 %[[r0]], -2147483648
 ; GCN: %[[r1:.*]] = bitcast float %__exp2 to i32
-; GCN: %[[r2:.*]] = or i32 %__pow_sign, %[[r1]]
+; GCN: %[[r2:.*]] = or disjoint i32 %__pow_sign, %[[r1]]
 ; GCN: store i32 %[[r2]], ptr addrspace(1) %a, align 4
 define amdgpu_kernel void @test_pow(ptr addrspace(1) nocapture %a) {
 entry:
@@ -414,7 +414,7 @@ entry:
 ; GCN: %[[r0:.*]] = bitcast float %tmp to i32
 ; GCN: %__pow_sign = and i32 %__yeven, %[[r0]]
 ; GCN: %[[r1:.*]] = bitcast float %__exp2 to i32
-; GCN: %[[r2:.*]] = or i32 %__pow_sign, %[[r1]]
+; GCN: %[[r2:.*]] = or disjoint i32 %__pow_sign, %[[r1]]
 ; GCN: store i32 %[[r2]], ptr addrspace(1) %a, align 4
 define amdgpu_kernel void @test_pown(ptr addrspace(1) nocapture %a) {
 entry:
@@ -438,7 +438,7 @@ declare <2 x half> @_Z3powDv2_DhS_(<2 x half>, <2 x half>)
 ; GCN: %1 = bitcast half %x to i16
 ; GCN: %__pow_sign = and i16 %1, -32768
 ; GCN: %2 = bitcast half %__exp2 to i16
-; GCN: %3 = or i16 %__pow_sign, %2
+; GCN: %3 = or disjoint i16 %__pow_sign, %2
 ; GCN: %4 = bitcast i16 %3 to half
 define half @test_pow_fast_f16__y_13(half %x) {
   %powr = tail call fast half @_Z3powDhDh(half %x, half 13.0)
@@ -453,7 +453,7 @@ define half @test_pow_fast_f16__y_13(half %x) {
 ; GCN: %1 = bitcast <2 x half> %x to <2 x i16>
 ; GCN: %__pow_sign = and <2 x i16> %1, <i16 -32768, i16 -32768>
 ; GCN: %2 = bitcast <2 x half> %__exp2 to <2 x i16>
-; GCN: %3 = or <2 x i16> %__pow_sign, %2
+; GCN: %3 = or disjoint <2 x i16> %__pow_sign, %2
 ; GCN: %4 = bitcast <2 x i16> %3 to <2 x half>
 define <2 x half> @test_pow_fast_v2f16__y_13(<2 x half> %x) {
   %powr = tail call fast <2 x half> @_Z3powDv2_DhS_(<2 x half> %x, <2 x half> <half 13.0, half 13.0>)
diff --git a/llvm/test/Transforms/InstCombine/known-bits.ll b/llvm/test/Transforms/InstCombine/known-bits.ll
index 8b4249b2c25a9..7a38020500517 100644
--- a/llvm/test/Transforms/InstCombine/known-bits.ll
+++ b/llvm/test/Transforms/InstCombine/known-bits.ll
@@ -1374,5 +1374,289 @@ define i8 @nonzero_reduce_xor_vscale_odd(<vscale x 3 x i8> %xx) {
   ret i8 %r
 }
 
+define i1 @test_sign_pos(float %x) {
+; CHECK-LABEL: @test_sign_pos(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call float @llvm.fabs.f32(float %x)
+  %y = bitcast float %fabs to i32
+  %sign = icmp sgt i32 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_pos_half(half %x) {
+; CHECK-LABEL: @test_sign_pos_half(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call half @llvm.fabs.f16(half %x)
+  %y = bitcast half %fabs to i16
+  %sign = icmp sgt i16 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_pos_half_non_elementwise(<2 x half> %x) {
+; CHECK-LABEL: @test_sign_pos_half_non_elementwise(
+; CHECK-NEXT:    [[FABS:%.*]] = call <2 x half> @llvm.fabs.v2f16(<2 x half> [[X:%.*]])
+; CHECK-NEXT:    [[Y:%.*]] = bitcast <2 x half> [[FABS]] to i32
+; CHECK-NEXT:    [[SIGN:%.*]] = icmp sgt i32 [[Y]], -1
+; CHECK-NEXT:    ret i1 [[SIGN]]
+;
+  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %x)
+  %y = bitcast <2 x half> %fabs to i32
+  %sign = icmp sgt i32 %y, -1
+  ret i1 %sign
+}
+
+define i1 @test_sign_neg(float %x) {
+; CHECK-LABEL: @test_sign_neg(
+; CHECK-NEXT:    ret i1 true
+;
+  %fabs = call float @llvm.fabs.f32(float %x)
+  %fnabs = fneg float %fabs
+  %y = bitcast float %fnabs to i32
+  %sign = icmp slt i32 %y, 0
+  ret i1 %sign
+}
+
+define <2 x i1> @test_sign_pos_vec(<2 x float> %x) {
+; CHECK-LABEL: @test_sign_pos_vec(
+; CHECK-NEXT:    ret <2 x i1> zeroinitializer
+;
+  %fabs = call <2 x float> @llvm.fabs.v2f32(<2 x float> %x)
+  %y = bitcast <2 x float> %fabs to <2 x i32>
+  %sign = icmp slt <2 x i32> %y, zeroinitializer
+  ret <2 x i1> %sign
+}
+
+define i32 @test_inf_only(float nofpclass(nan sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_only(
+; CHECK-NEXT:    ret i32 2139095040
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2147483647
+  ret i32 %and
+}
+
+define i16 @test_inf_only_bfloat(bfloat nofpclass(nan sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_only_bfloat(
+; CHECK-NEXT:    ret i16 32640
+;
+  %y = bitcast bfloat %x to i16
+  %and = and i16 %y, 32767
+  ret i16 %and
+}
+
+define i128 @test_inf_only_ppc_fp128(ppc_fp128 nofpclass(nan sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_only_ppc_fp128(
+; CHECK-NEXT:    ret i128 9218868437227405312
+;
+  %y = bitcast ppc_fp128 %x to i128
+  %and = and i128 %y, 170141183460469231731687303715884105727
+  ret i128 %and
+}
+
+define i32 @test_zero_only(float nofpclass(nan sub norm inf) %x) {
+; CHECK-LABEL: @test_zero_only(
+; CHECK-NEXT:    ret i32 0
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2147483647
+  ret i32 %and
+}
+
+define i80 @test_zero_only_non_ieee(x86_fp80 nofpclass(nan sub norm inf) %x) {
+; CHECK-LABEL: @test_zero_only_non_ieee(
+; CHECK-NEXT:    ret i80 0
+;
+  %y = bitcast x86_fp80 %x to i80
+  %and = and i80 %y, 604462909807314587353087
+  ret i80 %and
+}
+
+define i32 @test_inf_nan_only(float nofpclass(sub norm zero) %x) {
+; CHECK-LABEL: @test_inf_nan_only(
+; CHECK-NEXT:    [[Y:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[Y]], 2130706432
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2130706432
+  ret i32 %and
+}
+
+define i32 @test_sub_zero_only(float nofpclass(nan norm inf) %x) {
+; CHECK-LABEL: @test_sub_zero_only(
+; CHECK-NEXT:    [[Y:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[Y]], 2130706432
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 2130706432
+  ret i32 %and
+}
+
+define i32 @test_inf_zero_only(float nofpclass(nan norm sub) %x) {
+; CHECK-LABEL: @test_inf_zero_only(
+; CHECK-NEXT:    [[Y:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[AND:%.*]] = and i32 [[Y]], 8388608
+; CHECK-NEXT:    ret i32 [[AND]]
+;
+  %y = bitcast float %x to i32
+  %and = and i32 %y, 16777215
+  ret i32 %and
+}
+
+; Make sure that the signbit is cleared.
+define i32 @test_ninf_only(double %x) {
+; CHECK-LABEL: @test_ninf_only(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp oeq double [[X:%.*]], 0xFFF0000000000000
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    ret i32 0
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i32 0
+;
+  %cmp = fcmp oeq double %x, 0xFFF0000000000000
+  br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+  %cast = bitcast double %x to i64
+  %trunc = trunc i64 %cast to i32
+  ret i32 %trunc
+
+if.else:
+  ret i32 0
+}
+
+define i1 @test_simplify_icmp(i32 %x) {
+; CHECK-LABEL: @test_simplify_icmp(
+; CHECK-NEXT:    ret i1 false
+;
+  %cast1 = uitofp i32 %x to double
+  %cast2 = bitcast double %cast1 to i64
+  %mask = and i64 %cast2, -140737488355328
+  %cmp = icmp eq i64 %mask, -1970324836974592
+  ret i1 %cmp
+}
+
+define i32 @test_snan_quiet_bit1(float nofpclass(sub norm inf qnan) %x) {
+; CHECK-LABEL: @test_snan_quiet_bit1(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 4194304
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 4194304
+  ret i32 %masked
+}
+
+define i32 @test_snan_quiet_bit2(float nofpclass(sub norm inf qnan) %x) {
+; CHECK-LABEL: @test_snan_quiet_bit2(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 2097152
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 2097152
+  ret i32 %masked
+}
+
+define i32 @test_qnan_quiet_bit1(float nofpclass(sub norm inf snan) %x) {
+; CHECK-LABEL: @test_qnan_quiet_bit1(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 4194304
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 4194304
+  ret i32 %masked
+}
+
+define i32 @test_qnan_quiet_bit2(float nofpclass(sub norm inf snan) %x) {
+; CHECK-LABEL: @test_qnan_quiet_bit2(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast float [[X:%.*]] to i32
+; CHECK-NEXT:    [[MASKED:%.*]] = and i32 [[BITS]], 2097152
+; CHECK-NEXT:    ret i32 [[MASKED]]
+;
+  %bits = bitcast float %x to i32
+  %masked = and i32 %bits, 2097152
+  ret i32 %masked
+}
+
+define i16 @test_simplify_mask(i32 %ui, float %x) {
+; CHECK-LABEL: @test_simplify_mask(
+; CHECK-NEXT:    [[CONV:%.*]] = uitofp i32 [[UI:%.*]] to float
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ogt float [[CONV]], [[X:%.*]]
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_ELSE:%.*]], label [[IF_END:%.*]]
+; CHECK:       if.end:
+; CHECK-NEXT:    ret i16 31744
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i16 0
+;
+  %conv = uitofp i32 %ui to float
+  %cmp = fcmp olt float %x, %conv
+  br i1 %cmp, label %if.else, label %if.end
+
+if.end:
+  %cast = bitcast float %conv to i32
+  %shr = lshr i32 %cast, 16
+  %trunc = trunc i32 %shr to i16
+  %and = and i16 %trunc, -32768
+  %or = or disjoint i16 %and, 31744
+  ret i16 %or
+
+if.else:
+  ret i16 0
+}
+
+; TODO: %cmp always evaluates to false
+
+define i1 @test_simplify_icmp2(double %x) {
+; CHECK-LABEL: @test_simplify_icmp2(
+; CHECK-NEXT:    [[ABS:%.*]] = tail call double @llvm.fabs.f64(double [[X:%.*]])
+; CHECK-NEXT:    [[COND:%.*]] = fcmp oeq double [[ABS]], 0x7FF0000000000000
+; CHECK-NEXT:    br i1 [[COND]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    [[CAST:%.*]] = bitcast double [[X]] to i64
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i64 [[CAST]], 3458764513820540928
+; CHECK-NEXT:    ret i1 [[CMP]]
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i1 false
+;
+  %ab...
[truncated]

llvm/lib/Analysis/ValueTracking.cpp

…m#92084) This patch relands llvm#86409. I mistakenly thought that `Known.makeNegative()` clears the sign bit of `Known.Zero`. This patch fixes the assertion failure by explicitly clearing the sign bit.

dyung · 2024-05-15T05:15:42Z

Hi @dtcxzyw, one of our internal tests started failing and I bisected it back to this change you made. I have put the repro in #92217, can you take a look?

dtcxzyw · 2024-05-15T05:37:28Z

Hi @dtcxzyw, one of our internal tests started failing and I bisected it back to this change you made. I have put the repro in #92217, can you take a look?

Thank you! I will have a look :)

mikaelholmen · 2024-05-16T05:48:44Z

Hi @dtcxzyw
The following starts crashing with this patch:

opt -passes="instcombine" bbi-95530.ll -o /dev/null

It fails with

opt: ../lib/Analysis/ValueTracking.cpp:1152: void computeKnownBitsFromOperator(const llvm::Operator *, const llvm::APInt &, llvm::KnownBits &, unsigned int, const llvm::SimplifyQuery &): Assertion `!Known.hasConflict() && "Bits known to be one AND zero?"' failed.

bbi-95530.ll.gz

dtcxzyw · 2024-05-16T06:04:02Z

Hi @dtcxzyw The following starts crashing with this patch:

opt -passes="instcombine" bbi-95530.ll -o /dev/null

It fails with

opt: ../lib/Analysis/ValueTracking.cpp:1152: void computeKnownBitsFromOperator(const llvm::Operator *, const llvm::APInt &, llvm::KnownBits &, unsigned int, const llvm::SimplifyQuery &): Assertion `!Known.hasConflict() && "Bits known to be one AND zero?"' failed.

bbi-95530.ll.gz

I will take a look. If necessary, feel free to revert this commit :)

dtcxzyw · 2024-05-16T06:27:47Z

Hi @dtcxzyw The following starts crashing with this patch:

opt -passes="instcombine" bbi-95530.ll -o /dev/null

It fails with

opt: ../lib/Analysis/ValueTracking.cpp:1152: void computeKnownBitsFromOperator(const llvm::Operator *, const llvm::APInt &, llvm::KnownBits &, unsigned int, const llvm::SimplifyQuery &): Assertion `!Known.hasConflict() && "Bits known to be one AND zero?"' failed.

bbi-95530.ll.gz

It is weird that computeKnownFPClass returns (none). I will post a fix.

BTW, I don't think we need to revert this patch.

dtcxzyw · 2024-05-16T06:31:54Z

Reduced reproducer:

define i64 @test(double %x) {
  %cmp = fcmp uno double %x, 0.000000e+00
  br i1 %cmp, label %if.then1, label %if.else

if.then1:
  br i1 %cmp, label %if.else, label %if.then2

if.then2:
  %cast = bitcast double %x to i64
  %and = and i64 %cast, 1
  ret i64 %and

if.else:
  ret i64 0
}

dtcxzyw · 2024-05-16T06:38:37Z

@nikic Do you have a plan that treating the use as a poison value when the knownbit of it has conflicts?

jayfoad · 2024-05-16T09:01:29Z

Reduced reproducer:

define i64 @test(double %x) {
  %cmp = fcmp uno double %x, 0.000000e+00
  br i1 %cmp, label %if.then1, label %if.else

if.then1:
  br i1 %cmp, label %if.else, label %if.then2

if.then2:
  %cast = bitcast double %x to i64
  %and = and i64 %cast, 1
  ret i64 %and

if.else:
  ret i64 0
}

So if.then2 is dead code. It seems perfectly natural to me that you can get conflicting KnownBits info in dead code, and we should handle it instead of asserting. Treating it as poison sounds good to me.

nikic · 2024-05-16T09:28:11Z

@nikic Do you have a plan that treating the use as a poison value when the knownbit of it has conflicts?

The plan is to remove all of these assertions and fold conflict to poison. Someone just needs to implement it :)

…rns fcNone (#92355) Fixes #92084 (comment).

mikaelholmen · 2024-05-16T12:18:53Z

Thanks for the quick fix @dtcxzyw !

…m#92084) This patch relands llvm#86409. I mistakenly thought that `Known.makeNegative()` clears the sign bit of `Known.Zero`. This patch fixes the assertion failure by explicitly clearing the sign bit.

…rns fcNone (llvm#92355) Fixes llvm#92084 (comment).

dtcxzyw added 2 commits May 14, 2024 16:04

Revert "Revert "[ValueTracking] Compute knownbits from known fp class…

f9cf582

…es (llvm#86409)"" This reverts commit 2e165a2.

[ValueTracking] Fix assertion failure.

88d4f85

dtcxzyw requested review from mstorsjo and arsenm May 14, 2024 08:36

dtcxzyw requested a review from nikic as a code owner May 14, 2024 08:36

llvmbot added backend:AMDGPU llvm:ir llvm:analysis llvm:transforms labels May 14, 2024

arsenm approved these changes May 14, 2024

View reviewed changes

llvm/lib/Analysis/ValueTracking.cpp Outdated Show resolved Hide resolved

[ValueTracking] Remove the blank line. NFC.

1fc82bd

dtcxzyw merged commit 3aae916 into llvm:main May 14, 2024
3 of 4 checks passed

dtcxzyw deleted the fix-pr86409 branch May 14, 2024 10:10

dtcxzyw mentioned this pull request May 16, 2024

[ValueTracking] Fix assertion failure when computeKnownFPClass returns fcNone #92355

Merged

dtcxzyw added a commit that referenced this pull request May 16, 2024

[ValueTracking] Fix assertion failure when computeKnownFPClass retu…

46bc54f

…rns fcNone (#92355) Fixes #92084 (comment).

mub-at-arm pushed a commit to mub-at-arm/llvm-project that referenced this pull request May 16, 2024

[ValueTracking] Fix assertion failure when computeKnownFPClass retu…

d28b48f

…rns fcNone (llvm#92355) Fixes llvm#92084 (comment).

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Reland "[ValueTracking] Compute knownbits from known fp classes" #92084

Reland "[ValueTracking] Compute knownbits from known fp classes" #92084

dtcxzyw commented May 14, 2024

llvmbot commented May 14, 2024 •

edited

llvmbot commented May 14, 2024

llvmbot commented May 14, 2024

dyung commented May 15, 2024

dtcxzyw commented May 15, 2024

mikaelholmen commented May 16, 2024

dtcxzyw commented May 16, 2024

dtcxzyw commented May 16, 2024

dtcxzyw commented May 16, 2024

dtcxzyw commented May 16, 2024

jayfoad commented May 16, 2024

nikic commented May 16, 2024

mikaelholmen commented May 16, 2024

Reland "[ValueTracking] Compute knownbits from known fp classes" #92084

Reland "[ValueTracking] Compute knownbits from known fp classes" #92084

Conversation

dtcxzyw commented May 14, 2024

llvmbot commented May 14, 2024 • edited

llvmbot commented May 14, 2024

llvmbot commented May 14, 2024

dyung commented May 15, 2024

dtcxzyw commented May 15, 2024

mikaelholmen commented May 16, 2024

dtcxzyw commented May 16, 2024

dtcxzyw commented May 16, 2024

dtcxzyw commented May 16, 2024

dtcxzyw commented May 16, 2024

jayfoad commented May 16, 2024

nikic commented May 16, 2024

mikaelholmen commented May 16, 2024

llvmbot commented May 14, 2024 •

edited