[llvm] 86f0399 - [InstCombine] Fold expression using basic properties of floor and ceiling function (#107107)
via llvm-commits
llvm-commits at lists.llvm.org
Sun Sep 15 03:25:04 PDT 2024
Author: c8ef
Date: 2024-09-15T14:25:00+04:00
New Revision: 86f0399c1fdec80c34e22821607696a485857421
URL: https://github.com/llvm/llvm-project/commit/86f0399c1fdec80c34e22821607696a485857421
DIFF: https://github.com/llvm/llvm-project/commit/86f0399c1fdec80c34e22821607696a485857421.diff
LOG: [InstCombine] Fold expression using basic properties of floor and ceiling function (#107107)
alive2: ~~https://alive2.llvm.org/ce/z/Ag3Ki7~~
https://alive2.llvm.org/ce/z/ywP5t2
related: #76438
This patch adds the following foldings: `floor(x) <= x --> true` and `x
<= ceil(x) --> true`. We leverage the properties of these math functions
and ensure there is no floating point input of `nan`.
---------
Co-authored-by: Yingwei Zheng <dtcxzyw at qq.com>
Added:
llvm/test/Transforms/InstCombine/fp-floor-ceil.ll
Modified:
llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index 8e8d472a5df1d3..5cdfeada7f0aa2 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -8178,6 +8178,75 @@ static Instruction *foldFCmpFSubIntoFCmp(FCmpInst &I, Instruction *LHSI,
return nullptr;
}
+static Instruction *foldFCmpWithFloorAndCeil(FCmpInst &I,
+ InstCombinerImpl &IC) {
+ Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
+ Type *OpType = LHS->getType();
+ CmpInst::Predicate Pred = I.getPredicate();
+
+ bool FloorX = match(LHS, m_Intrinsic<Intrinsic::floor>(m_Specific(RHS)));
+ bool CeilX = match(LHS, m_Intrinsic<Intrinsic::ceil>(m_Specific(RHS)));
+
+ if (!FloorX && !CeilX) {
+ if ((FloorX = match(RHS, m_Intrinsic<Intrinsic::floor>(m_Specific(LHS)))) ||
+ (CeilX = match(RHS, m_Intrinsic<Intrinsic::ceil>(m_Specific(LHS))))) {
+ std::swap(LHS, RHS);
+ Pred = I.getSwappedPredicate();
+ }
+ }
+
+ switch (Pred) {
+ case FCmpInst::FCMP_OLE:
+ // fcmp ole floor(x), x => fcmp ord x, 0
+ if (FloorX)
+ return new FCmpInst(FCmpInst::FCMP_ORD, RHS, ConstantFP::getZero(OpType),
+ "", &I);
+ break;
+ case FCmpInst::FCMP_OGT:
+ // fcmp ogt floor(x), x => false
+ if (FloorX)
+ return IC.replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
+ break;
+ case FCmpInst::FCMP_OGE:
+ // fcmp oge ceil(x), x => fcmp ord x, 0
+ if (CeilX)
+ return new FCmpInst(FCmpInst::FCMP_ORD, RHS, ConstantFP::getZero(OpType),
+ "", &I);
+ break;
+ case FCmpInst::FCMP_OLT:
+ // fcmp olt ceil(x), x => false
+ if (CeilX)
+ return IC.replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
+ break;
+ case FCmpInst::FCMP_ULE:
+ // fcmp ule floor(x), x => true
+ if (FloorX)
+ return IC.replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
+ break;
+ case FCmpInst::FCMP_UGT:
+ // fcmp ugt floor(x), x => fcmp uno x, 0
+ if (FloorX)
+ return new FCmpInst(FCmpInst::FCMP_UNO, RHS, ConstantFP::getZero(OpType),
+ "", &I);
+ break;
+ case FCmpInst::FCMP_UGE:
+ // fcmp uge ceil(x), x => true
+ if (CeilX)
+ return IC.replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
+ break;
+ case FCmpInst::FCMP_ULT:
+ // fcmp ult ceil(x), x => fcmp uno x, 0
+ if (CeilX)
+ return new FCmpInst(FCmpInst::FCMP_UNO, RHS, ConstantFP::getZero(OpType),
+ "", &I);
+ break;
+ default:
+ break;
+ }
+
+ return nullptr;
+}
+
Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
bool Changed = false;
@@ -8382,6 +8451,9 @@ Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
if (Instruction *R = foldSqrtWithFcmpZero(I, *this))
return R;
+ if (Instruction *R = foldFCmpWithFloorAndCeil(I, *this))
+ return R;
+
if (match(Op0, m_FNeg(m_Value(X)))) {
// fcmp pred (fneg X), C --> fcmp swap(pred) X, -C
Constant *C;
diff --git a/llvm/test/Transforms/InstCombine/fp-floor-ceil.ll b/llvm/test/Transforms/InstCombine/fp-floor-ceil.ll
new file mode 100644
index 00000000000000..c90b7ee11310cb
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/fp-floor-ceil.ll
@@ -0,0 +1,260 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+
+define i1 @floor_x_ole(float %x) {
+; CHECK-LABEL: @floor_x_ole(
+; CHECK-NEXT: [[RET:%.*]] = fcmp ninf ord float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp ninf ole float %floor, %x
+ ret i1 %ret
+}
+
+define i1 @floor_x_ule(float %x) {
+; CHECK-LABEL: @floor_x_ule(
+; CHECK-NEXT: ret i1 true
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp ule float %floor, %x
+ ret i1 %ret
+}
+
+define i1 @floor_x_ogt(float %x) {
+; CHECK-LABEL: @floor_x_ogt(
+; CHECK-NEXT: ret i1 false
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp ogt float %floor, %x
+ ret i1 %ret
+}
+
+define i1 @floor_x_ugt(float %x) {
+; CHECK-LABEL: @floor_x_ugt(
+; CHECK-NEXT: [[RET:%.*]] = fcmp ninf uno float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp ninf ugt float %floor, %x
+ ret i1 %ret
+}
+
+define i1 @x_floor_oge(float %x) {
+; CHECK-LABEL: @x_floor_oge(
+; CHECK-NEXT: [[RET:%.*]] = fcmp ninf ord float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp ninf oge float %x, %floor
+ ret i1 %ret
+}
+
+define i1 @x_floor_uge(float %x) {
+; CHECK-LABEL: @x_floor_uge(
+; CHECK-NEXT: ret i1 true
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp uge float %x, %floor
+ ret i1 %ret
+}
+
+define i1 @x_floor_olt(float %x) {
+; CHECK-LABEL: @x_floor_olt(
+; CHECK-NEXT: ret i1 false
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp olt float %x, %floor
+ ret i1 %ret
+}
+
+define i1 @x_floor_ult(float %x) {
+; CHECK-LABEL: @x_floor_ult(
+; CHECK-NEXT: [[RET:%.*]] = fcmp ninf uno float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp ninf ult float %x, %floor
+ ret i1 %ret
+}
+
+define <2 x i1> @x_floor_olt_vec(<2 x float> %x) {
+; CHECK-LABEL: @x_floor_olt_vec(
+; CHECK-NEXT: ret <2 x i1> zeroinitializer
+;
+ %floor = call <2 x float> @llvm.floor.f32(<2 x float> %x)
+ %ret = fcmp olt <2 x float> %x, %floor
+ ret <2 x i1> %ret
+}
+
+define i1 @x_floor_ole_neg(float %x) {
+; CHECK-LABEL: @x_floor_ole_neg(
+; CHECK-NEXT: [[FLOOR:%.*]] = call float @llvm.floor.f32(float [[X:%.*]])
+; CHECK-NEXT: [[RET:%.*]] = fcmp ole float [[X]], [[FLOOR]]
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp ole float %x, %floor
+ ret i1 %ret
+}
+
+define i1 @x_floor_ogt_neg(float %x) {
+; CHECK-LABEL: @x_floor_ogt_neg(
+; CHECK-NEXT: [[FLOOR:%.*]] = call float @llvm.floor.f32(float [[X:%.*]])
+; CHECK-NEXT: [[RET:%.*]] = fcmp ogt float [[X]], [[FLOOR]]
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp ogt float %x, %floor
+ ret i1 %ret
+}
+
+define i1 @x_floor_ueq_neg(float %x) {
+; CHECK-LABEL: @x_floor_ueq_neg(
+; CHECK-NEXT: [[FLOOR:%.*]] = call float @llvm.floor.f32(float [[X:%.*]])
+; CHECK-NEXT: [[RET:%.*]] = fcmp ueq float [[X]], [[FLOOR]]
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp ueq float %x, %floor
+ ret i1 %ret
+}
+
+define i1 @x_floor_une_neg(float %x) {
+; CHECK-LABEL: @x_floor_une_neg(
+; CHECK-NEXT: [[FLOOR:%.*]] = call float @llvm.floor.f32(float [[X:%.*]])
+; CHECK-NEXT: [[RET:%.*]] = fcmp une float [[X]], [[FLOOR]]
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %floor = call float @llvm.floor.f32(float %x)
+ %ret = fcmp une float %x, %floor
+ ret i1 %ret
+}
+
+define i1 @ceil_x_oge(float %x) {
+; CHECK-LABEL: @ceil_x_oge(
+; CHECK-NEXT: [[RET:%.*]] = fcmp ninf ord float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp ninf oge float %ceil, %x
+ ret i1 %ret
+}
+
+define i1 @ceil_x_uge(float %x) {
+; CHECK-LABEL: @ceil_x_uge(
+; CHECK-NEXT: ret i1 true
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp uge float %ceil, %x
+ ret i1 %ret
+}
+
+define i1 @ceil_x_olt(float %x) {
+; CHECK-LABEL: @ceil_x_olt(
+; CHECK-NEXT: ret i1 false
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp olt float %ceil, %x
+ ret i1 %ret
+}
+
+define i1 @ceil_x_ult(float %x) {
+; CHECK-LABEL: @ceil_x_ult(
+; CHECK-NEXT: [[RET:%.*]] = fcmp ninf uno float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp ninf ult float %ceil, %x
+ ret i1 %ret
+}
+
+define i1 @x_ceil_ole(float %x) {
+; CHECK-LABEL: @x_ceil_ole(
+; CHECK-NEXT: [[RET:%.*]] = fcmp ninf ord float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp ninf ole float %x, %ceil
+ ret i1 %ret
+}
+
+define i1 @x_ceil_ule(float %x) {
+; CHECK-LABEL: @x_ceil_ule(
+; CHECK-NEXT: ret i1 true
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp ule float %x, %ceil
+ ret i1 %ret
+}
+
+define i1 @x_ceil_ogt(float %x) {
+; CHECK-LABEL: @x_ceil_ogt(
+; CHECK-NEXT: ret i1 false
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp ogt float %x, %ceil
+ ret i1 %ret
+}
+
+define i1 @x_ceil_ugt(float %x) {
+; CHECK-LABEL: @x_ceil_ugt(
+; CHECK-NEXT: [[RET:%.*]] = fcmp ninf uno float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp ninf ugt float %x, %ceil
+ ret i1 %ret
+}
+
+define <2 x i1> @x_ceil_ogt_vec(<2 x float> %x) {
+; CHECK-LABEL: @x_ceil_ogt_vec(
+; CHECK-NEXT: ret <2 x i1> zeroinitializer
+;
+ %ceil = call <2 x float> @llvm.ceil.f32(<2 x float> %x)
+ %ret = fcmp ogt <2 x float> %x, %ceil
+ ret <2 x i1> %ret
+}
+
+define i1 @x_ceil_oge_neg(float %x) {
+; CHECK-LABEL: @x_ceil_oge_neg(
+; CHECK-NEXT: [[CEIL:%.*]] = call float @llvm.ceil.f32(float [[X:%.*]])
+; CHECK-NEXT: [[RET:%.*]] = fcmp oge float [[X]], [[CEIL]]
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp oge float %x, %ceil
+ ret i1 %ret
+}
+
+define i1 @x_ceil_olt_neg(float %x) {
+; CHECK-LABEL: @x_ceil_olt_neg(
+; CHECK-NEXT: [[CEIL:%.*]] = call float @llvm.ceil.f32(float [[X:%.*]])
+; CHECK-NEXT: [[RET:%.*]] = fcmp olt float [[X]], [[CEIL]]
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp olt float %x, %ceil
+ ret i1 %ret
+}
+
+define i1 @x_ceil_oeq_neg(float %x) {
+; CHECK-LABEL: @x_ceil_oeq_neg(
+; CHECK-NEXT: [[CEIL:%.*]] = call float @llvm.ceil.f32(float [[X:%.*]])
+; CHECK-NEXT: [[RET:%.*]] = fcmp oeq float [[X]], [[CEIL]]
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp oeq float %x, %ceil
+ ret i1 %ret
+}
+
+define i1 @x_ceil_one_neg(float %x) {
+; CHECK-LABEL: @x_ceil_one_neg(
+; CHECK-NEXT: [[CEIL:%.*]] = call float @llvm.ceil.f32(float [[X:%.*]])
+; CHECK-NEXT: [[RET:%.*]] = fcmp one float [[X]], [[CEIL]]
+; CHECK-NEXT: ret i1 [[RET]]
+;
+ %ceil = call float @llvm.ceil.f32(float %x)
+ %ret = fcmp one float %x, %ceil
+ ret i1 %ret
+}
More information about the llvm-commits
mailing list