[InstSimplify] Simplify fcmp implied by dominating fcmp #161090
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
llvmbot
added
backend:AMDGPU
llvm:instcombine
|
@llvm/pr-subscribers-llvm-analysis @llvm/pr-subscribers-backend-amdgpu Author: Yingwei Zheng (dtcxzyw) ChangesThis patch simplifies an fcmp into true/false if it is implied by a dominating fcmp.
Note: It doesn't fix #70985, as the second fcmp in the motivating case is not dominated by the edge. We may need to adjust JumpThreading to handle this case. Comptime impact (~+0.1%): https://llvm-compile-time-tracker.com/compare.php?from=a728f213c863e4dd19f8969a417148d2951323c0&to=8ca70404fb0d66a824f39d83050ac38e2f1b25b9&stat=instructions:u Full diff: https://github.com/llvm/llvm-project/pull/161090.diff 5 Files Affected:
diff --git a/llvm/lib/Analysis/InstructionSimplify.cpp b/llvm/lib/Analysis/InstructionSimplify.cpp
index 07f4a8e5c889e..0d978d4da125e 100644
--- a/llvm/lib/Analysis/InstructionSimplify.cpp
+++ b/llvm/lib/Analysis/InstructionSimplify.cpp
@@ -4164,6 +4164,10 @@ static Value *simplifyFCmpInst(CmpPredicate Pred, Value *LHS, Value *RHS,
return ConstantInt::get(RetTy, Pred == CmpInst::FCMP_UNO);
}
+ if (std::optional<bool> Res =
+ isImpliedByDomCondition(Pred, LHS, RHS, Q.CxtI, Q.DL))
+ return ConstantInt::getBool(RetTy, *Res);
+
const APFloat *C = nullptr;
match(RHS, m_APFloatAllowPoison(C));
std::optional<KnownFPClass> FullKnownClassLHS;
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 6f11b250cf21f..4d049fc356273 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -39,6 +39,7 @@
#include "llvm/IR/Attributes.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/Constant.h"
+#include "llvm/IR/ConstantFPRange.h"
#include "llvm/IR/ConstantRange.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DerivedTypes.h"
@@ -9447,6 +9448,69 @@ isImpliedCondICmps(CmpPredicate LPred, const Value *L0, const Value *L1,
return std::nullopt;
}
+/// Return true if LHS implies RHS (expanded to its components as "R0 RPred R1")
+/// is true. Return false if LHS implies RHS is false. Otherwise, return
+/// std::nullopt if we can't infer anything.
+static std::optional<bool>
+isImpliedCondFCmps(FCmpInst::Predicate LPred, const Value *L0, const Value *L1,
+ FCmpInst::Predicate RPred, const Value *R0, const Value *R1,
+ const DataLayout &DL, bool LHSIsTrue) {
+ // The rest of the logic assumes the LHS condition is true. If that's not the
+ // case, invert the predicate to make it so.
+ if (!LHSIsTrue)
+ LPred = FCmpInst::getInversePredicate(LPred);
+
+ // We can have non-canonical operands, so try to normalize any common operand
+ // to L0/R0.
+ if (L0 == R1) {
+ std::swap(R0, R1);
+ RPred = FCmpInst::getSwappedPredicate(RPred);
+ }
+ if (R0 == L1) {
+ std::swap(L0, L1);
+ LPred = FCmpInst::getSwappedPredicate(LPred);
+ }
+ if (L1 == R1) {
+ // If we have L0 == R0 and L1 == R1, then make L1/R1 the constants.
+ if (L0 != R0 || match(L0, m_ImmConstant())) {
+ std::swap(L0, L1);
+ LPred = ICmpInst::getSwappedCmpPredicate(LPred);
+ std::swap(R0, R1);
+ RPred = ICmpInst::getSwappedCmpPredicate(RPred);
+ }
+ }
+
+ // Can we infer anything when the two compares have matching operands?
+ if (L0 == R0 && L1 == R1) {
+ if ((LPred & RPred) == LPred)
+ return true;
+ if ((LPred & ~RPred) == LPred)
+ return false;
+ }
+
+ // See if we can infer anything if operand-0 matches and we have at least one
+ // constant.
+ const APFloat *L1C, *R1C;
+ if (L0 == R0 && match(L1, m_APFloat(L1C)) && match(R1, m_APFloat(R1C))) {
+ if (std::optional<ConstantFPRange> DomCR =
+ ConstantFPRange::makeExactFCmpRegion(LPred, *L1C)) {
+ if (std::optional<ConstantFPRange> ImpliedCR =
+ ConstantFPRange::makeExactFCmpRegion(RPred, *R1C)) {
+ if (ImpliedCR->contains(*DomCR))
+ return true;
+ }
+ if (std::optional<ConstantFPRange> ImpliedCR =
+ ConstantFPRange::makeExactFCmpRegion(
+ FCmpInst::getInversePredicate(RPred), *R1C)) {
+ if (ImpliedCR->contains(*DomCR))
+ return false;
+ }
+ }
+ }
+
+ return std::nullopt;
+}
+
/// Return true if LHS implies RHS is true. Return false if LHS implies RHS is
/// false. Otherwise, return std::nullopt if we can't infer anything. We
/// expect the RHS to be an icmp and the LHS to be an 'and', 'or', or a 'select'
@@ -9502,15 +9566,24 @@ llvm::isImpliedCondition(const Value *LHS, CmpPredicate RHSPred,
LHSIsTrue = !LHSIsTrue;
// Both LHS and RHS are icmps.
- if (const auto *LHSCmp = dyn_cast<ICmpInst>(LHS))
- return isImpliedCondICmps(LHSCmp->getCmpPredicate(), LHSCmp->getOperand(0),
- LHSCmp->getOperand(1), RHSPred, RHSOp0, RHSOp1,
- DL, LHSIsTrue);
- const Value *V;
- if (match(LHS, m_NUWTrunc(m_Value(V))))
- return isImpliedCondICmps(CmpInst::ICMP_NE, V,
- ConstantInt::get(V->getType(), 0), RHSPred,
- RHSOp0, RHSOp1, DL, LHSIsTrue);
+ if (RHSOp0->getType()->getScalarType()->isIntOrPtrTy()) {
+ if (const auto *LHSCmp = dyn_cast<ICmpInst>(LHS))
+ return isImpliedCondICmps(LHSCmp->getCmpPredicate(),
+ LHSCmp->getOperand(0), LHSCmp->getOperand(1),
+ RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue);
+ const Value *V;
+ if (match(LHS, m_NUWTrunc(m_Value(V))))
+ return isImpliedCondICmps(CmpInst::ICMP_NE, V,
+ ConstantInt::get(V->getType(), 0), RHSPred,
+ RHSOp0, RHSOp1, DL, LHSIsTrue);
+ } else {
+ assert(RHSOp0->getType()->isFPOrFPVectorTy() &&
+ "Expected floating point type only!");
+ if (const auto *LHSCmp = dyn_cast<FCmpInst>(LHS))
+ return isImpliedCondFCmps(LHSCmp->getPredicate(), LHSCmp->getOperand(0),
+ LHSCmp->getOperand(1), RHSPred, RHSOp0, RHSOp1,
+ DL, LHSIsTrue);
+ }
/// The LHS should be an 'or', 'and', or a 'select' instruction. We expect
/// the RHS to be an icmp.
@@ -9547,6 +9620,13 @@ std::optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS,
return InvertRHS ? !*Implied : *Implied;
return std::nullopt;
}
+ if (const FCmpInst *RHSCmp = dyn_cast<FCmpInst>(RHS)) {
+ if (auto Implied = isImpliedCondition(
+ LHS, RHSCmp->getPredicate(), RHSCmp->getOperand(0),
+ RHSCmp->getOperand(1), DL, LHSIsTrue, Depth))
+ return InvertRHS ? !*Implied : *Implied;
+ return std::nullopt;
+ }
const Value *V;
if (match(RHS, m_NUWTrunc(m_Value(V)))) {
diff --git a/llvm/test/CodeGen/AMDGPU/sgpr-copy.ll b/llvm/test/CodeGen/AMDGPU/sgpr-copy.ll
index c82b3410a0c42..5bc9cdb5de527 100644
--- a/llvm/test/CodeGen/AMDGPU/sgpr-copy.ll
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-copy.ll
@@ -256,7 +256,7 @@ endif: ; preds = %else, %if
define amdgpu_kernel void @copy1(ptr addrspace(1) %out, ptr addrspace(1) %in0) {
entry:
%tmp = load float, ptr addrspace(1) %in0
- %tmp1 = fcmp oeq float %tmp, 0.000000e+00
+ %tmp1 = fcmp one float %tmp, 0.000000e+00
br i1 %tmp1, label %if0, label %endif
if0: ; preds = %entry
diff --git a/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll b/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll
index 7f3276608c5da..0ccaa9c393654 100644
--- a/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll
+++ b/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll
@@ -172,10 +172,8 @@ define float @clamp_negative_wrong_const(float %x) {
; Like @clamp_test_1 but both are min
define float @clamp_negative_same_op(float %x) {
; CHECK-LABEL: @clamp_negative_same_op(
-; CHECK-NEXT: [[INNER_CMP_INV:%.*]] = fcmp fast oge float [[X:%.*]], 2.550000e+02
-; CHECK-NEXT: [[INNER_SEL:%.*]] = select nnan ninf i1 [[INNER_CMP_INV]], float 2.550000e+02, float [[X]]
-; CHECK-NEXT: [[OUTER_CMP:%.*]] = fcmp fast ult float [[X]], 1.000000e+00
-; CHECK-NEXT: [[R:%.*]] = select i1 [[OUTER_CMP]], float [[INNER_SEL]], float 1.000000e+00
+; CHECK-NEXT: [[OUTER_CMP_INV:%.*]] = fcmp fast oge float [[X:%.*]], 1.000000e+00
+; CHECK-NEXT: [[R:%.*]] = select nnan ninf i1 [[OUTER_CMP_INV]], float 1.000000e+00, float [[X]]
; CHECK-NEXT: ret float [[R]]
;
%inner_cmp = fcmp fast ult float %x, 255.0
diff --git a/llvm/test/Transforms/InstSimplify/domcondition.ll b/llvm/test/Transforms/InstSimplify/domcondition.ll
index 43be5de6bf95d..2893bb155bbe7 100644
--- a/llvm/test/Transforms/InstSimplify/domcondition.ll
+++ b/llvm/test/Transforms/InstSimplify/domcondition.ll
@@ -278,3 +278,210 @@ end:
}
declare void @foo(i32)
+
+
+define i1 @simplify_fcmp_implied_by_dom_cond_range_true(float %x) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_range_true(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 true
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, 0.0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp olt float %x, 1.0
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_in_else_implied_by_dom_cond_range_true(float %x) {
+; CHECK-LABEL: @simplify_fcmp_in_else_implied_by_dom_cond_range_true(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], 1.000000e+00
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 true
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 true
+;
+ %cmp = fcmp olt float %x, 1.0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ ret i1 true
+
+if.else:
+ %cmp2 = fcmp uge float %x, 0.5
+ ret i1 %cmp2
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_range_false(float %x) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_range_false(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 false
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, 0.0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp ogt float %x, 1.0
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_pred_true(float %x, float %y) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_pred_true(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 true
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, %y
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp ole float %x, %y
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_pred_false(float %x, float %y) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_pred_false(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 false
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, %y
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp ogt float %x, %y
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_pred_commuted(float %x, float %y) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_pred_commuted(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 true
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, %y
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp oge float %y, %x
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+; Negative tests
+
+define i1 @simplify_fcmp_implied_by_dom_cond_wrong_range(float %x) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_wrong_range(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp olt float [[X]], -1.000000e+00
+; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, 0.0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp olt float %x, -1.0
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_range_mismatched_operand(float %x, float %y) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_range_mismatched_operand(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp olt float [[Y:%.*]], 1.000000e+00
+; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, 0.0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp olt float %y, 1.0
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_wrong_pred(float %x, float %y) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_wrong_pred(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp ole float [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp olt float [[X]], [[Y]]
+; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp ole float %x, %y
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp olt float %x, %y
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_pred_mismatched_operand(float %x, float %y, float %z) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_pred_mismatched_operand(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp ole float [[X]], [[Z:%.*]]
+; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, %y
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp ole float %x, %z
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
|
|
@llvm/pr-subscribers-llvm-transforms Author: Yingwei Zheng (dtcxzyw) ChangesThis patch simplifies an fcmp into true/false if it is implied by a dominating fcmp.
Note: It doesn't fix #70985, as the second fcmp in the motivating case is not dominated by the edge. We may need to adjust JumpThreading to handle this case. Comptime impact (~+0.1%): https://llvm-compile-time-tracker.com/compare.php?from=a728f213c863e4dd19f8969a417148d2951323c0&to=8ca70404fb0d66a824f39d83050ac38e2f1b25b9&stat=instructions:u Full diff: https://github.com/llvm/llvm-project/pull/161090.diff 5 Files Affected:
diff --git a/llvm/lib/Analysis/InstructionSimplify.cpp b/llvm/lib/Analysis/InstructionSimplify.cpp
index 07f4a8e5c889e..0d978d4da125e 100644
--- a/llvm/lib/Analysis/InstructionSimplify.cpp
+++ b/llvm/lib/Analysis/InstructionSimplify.cpp
@@ -4164,6 +4164,10 @@ static Value *simplifyFCmpInst(CmpPredicate Pred, Value *LHS, Value *RHS,
return ConstantInt::get(RetTy, Pred == CmpInst::FCMP_UNO);
}
+ if (std::optional<bool> Res =
+ isImpliedByDomCondition(Pred, LHS, RHS, Q.CxtI, Q.DL))
+ return ConstantInt::getBool(RetTy, *Res);
+
const APFloat *C = nullptr;
match(RHS, m_APFloatAllowPoison(C));
std::optional<KnownFPClass> FullKnownClassLHS;
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 6f11b250cf21f..4d049fc356273 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -39,6 +39,7 @@
#include "llvm/IR/Attributes.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/Constant.h"
+#include "llvm/IR/ConstantFPRange.h"
#include "llvm/IR/ConstantRange.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DerivedTypes.h"
@@ -9447,6 +9448,69 @@ isImpliedCondICmps(CmpPredicate LPred, const Value *L0, const Value *L1,
return std::nullopt;
}
+/// Return true if LHS implies RHS (expanded to its components as "R0 RPred R1")
+/// is true. Return false if LHS implies RHS is false. Otherwise, return
+/// std::nullopt if we can't infer anything.
+static std::optional<bool>
+isImpliedCondFCmps(FCmpInst::Predicate LPred, const Value *L0, const Value *L1,
+ FCmpInst::Predicate RPred, const Value *R0, const Value *R1,
+ const DataLayout &DL, bool LHSIsTrue) {
+ // The rest of the logic assumes the LHS condition is true. If that's not the
+ // case, invert the predicate to make it so.
+ if (!LHSIsTrue)
+ LPred = FCmpInst::getInversePredicate(LPred);
+
+ // We can have non-canonical operands, so try to normalize any common operand
+ // to L0/R0.
+ if (L0 == R1) {
+ std::swap(R0, R1);
+ RPred = FCmpInst::getSwappedPredicate(RPred);
+ }
+ if (R0 == L1) {
+ std::swap(L0, L1);
+ LPred = FCmpInst::getSwappedPredicate(LPred);
+ }
+ if (L1 == R1) {
+ // If we have L0 == R0 and L1 == R1, then make L1/R1 the constants.
+ if (L0 != R0 || match(L0, m_ImmConstant())) {
+ std::swap(L0, L1);
+ LPred = ICmpInst::getSwappedCmpPredicate(LPred);
+ std::swap(R0, R1);
+ RPred = ICmpInst::getSwappedCmpPredicate(RPred);
+ }
+ }
+
+ // Can we infer anything when the two compares have matching operands?
+ if (L0 == R0 && L1 == R1) {
+ if ((LPred & RPred) == LPred)
+ return true;
+ if ((LPred & ~RPred) == LPred)
+ return false;
+ }
+
+ // See if we can infer anything if operand-0 matches and we have at least one
+ // constant.
+ const APFloat *L1C, *R1C;
+ if (L0 == R0 && match(L1, m_APFloat(L1C)) && match(R1, m_APFloat(R1C))) {
+ if (std::optional<ConstantFPRange> DomCR =
+ ConstantFPRange::makeExactFCmpRegion(LPred, *L1C)) {
+ if (std::optional<ConstantFPRange> ImpliedCR =
+ ConstantFPRange::makeExactFCmpRegion(RPred, *R1C)) {
+ if (ImpliedCR->contains(*DomCR))
+ return true;
+ }
+ if (std::optional<ConstantFPRange> ImpliedCR =
+ ConstantFPRange::makeExactFCmpRegion(
+ FCmpInst::getInversePredicate(RPred), *R1C)) {
+ if (ImpliedCR->contains(*DomCR))
+ return false;
+ }
+ }
+ }
+
+ return std::nullopt;
+}
+
/// Return true if LHS implies RHS is true. Return false if LHS implies RHS is
/// false. Otherwise, return std::nullopt if we can't infer anything. We
/// expect the RHS to be an icmp and the LHS to be an 'and', 'or', or a 'select'
@@ -9502,15 +9566,24 @@ llvm::isImpliedCondition(const Value *LHS, CmpPredicate RHSPred,
LHSIsTrue = !LHSIsTrue;
// Both LHS and RHS are icmps.
- if (const auto *LHSCmp = dyn_cast<ICmpInst>(LHS))
- return isImpliedCondICmps(LHSCmp->getCmpPredicate(), LHSCmp->getOperand(0),
- LHSCmp->getOperand(1), RHSPred, RHSOp0, RHSOp1,
- DL, LHSIsTrue);
- const Value *V;
- if (match(LHS, m_NUWTrunc(m_Value(V))))
- return isImpliedCondICmps(CmpInst::ICMP_NE, V,
- ConstantInt::get(V->getType(), 0), RHSPred,
- RHSOp0, RHSOp1, DL, LHSIsTrue);
+ if (RHSOp0->getType()->getScalarType()->isIntOrPtrTy()) {
+ if (const auto *LHSCmp = dyn_cast<ICmpInst>(LHS))
+ return isImpliedCondICmps(LHSCmp->getCmpPredicate(),
+ LHSCmp->getOperand(0), LHSCmp->getOperand(1),
+ RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue);
+ const Value *V;
+ if (match(LHS, m_NUWTrunc(m_Value(V))))
+ return isImpliedCondICmps(CmpInst::ICMP_NE, V,
+ ConstantInt::get(V->getType(), 0), RHSPred,
+ RHSOp0, RHSOp1, DL, LHSIsTrue);
+ } else {
+ assert(RHSOp0->getType()->isFPOrFPVectorTy() &&
+ "Expected floating point type only!");
+ if (const auto *LHSCmp = dyn_cast<FCmpInst>(LHS))
+ return isImpliedCondFCmps(LHSCmp->getPredicate(), LHSCmp->getOperand(0),
+ LHSCmp->getOperand(1), RHSPred, RHSOp0, RHSOp1,
+ DL, LHSIsTrue);
+ }
/// The LHS should be an 'or', 'and', or a 'select' instruction. We expect
/// the RHS to be an icmp.
@@ -9547,6 +9620,13 @@ std::optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS,
return InvertRHS ? !*Implied : *Implied;
return std::nullopt;
}
+ if (const FCmpInst *RHSCmp = dyn_cast<FCmpInst>(RHS)) {
+ if (auto Implied = isImpliedCondition(
+ LHS, RHSCmp->getPredicate(), RHSCmp->getOperand(0),
+ RHSCmp->getOperand(1), DL, LHSIsTrue, Depth))
+ return InvertRHS ? !*Implied : *Implied;
+ return std::nullopt;
+ }
const Value *V;
if (match(RHS, m_NUWTrunc(m_Value(V)))) {
diff --git a/llvm/test/CodeGen/AMDGPU/sgpr-copy.ll b/llvm/test/CodeGen/AMDGPU/sgpr-copy.ll
index c82b3410a0c42..5bc9cdb5de527 100644
--- a/llvm/test/CodeGen/AMDGPU/sgpr-copy.ll
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-copy.ll
@@ -256,7 +256,7 @@ endif: ; preds = %else, %if
define amdgpu_kernel void @copy1(ptr addrspace(1) %out, ptr addrspace(1) %in0) {
entry:
%tmp = load float, ptr addrspace(1) %in0
- %tmp1 = fcmp oeq float %tmp, 0.000000e+00
+ %tmp1 = fcmp one float %tmp, 0.000000e+00
br i1 %tmp1, label %if0, label %endif
if0: ; preds = %entry
diff --git a/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll b/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll
index 7f3276608c5da..0ccaa9c393654 100644
--- a/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll
+++ b/llvm/test/Transforms/InstCombine/clamp-to-minmax.ll
@@ -172,10 +172,8 @@ define float @clamp_negative_wrong_const(float %x) {
; Like @clamp_test_1 but both are min
define float @clamp_negative_same_op(float %x) {
; CHECK-LABEL: @clamp_negative_same_op(
-; CHECK-NEXT: [[INNER_CMP_INV:%.*]] = fcmp fast oge float [[X:%.*]], 2.550000e+02
-; CHECK-NEXT: [[INNER_SEL:%.*]] = select nnan ninf i1 [[INNER_CMP_INV]], float 2.550000e+02, float [[X]]
-; CHECK-NEXT: [[OUTER_CMP:%.*]] = fcmp fast ult float [[X]], 1.000000e+00
-; CHECK-NEXT: [[R:%.*]] = select i1 [[OUTER_CMP]], float [[INNER_SEL]], float 1.000000e+00
+; CHECK-NEXT: [[OUTER_CMP_INV:%.*]] = fcmp fast oge float [[X:%.*]], 1.000000e+00
+; CHECK-NEXT: [[R:%.*]] = select nnan ninf i1 [[OUTER_CMP_INV]], float 1.000000e+00, float [[X]]
; CHECK-NEXT: ret float [[R]]
;
%inner_cmp = fcmp fast ult float %x, 255.0
diff --git a/llvm/test/Transforms/InstSimplify/domcondition.ll b/llvm/test/Transforms/InstSimplify/domcondition.ll
index 43be5de6bf95d..2893bb155bbe7 100644
--- a/llvm/test/Transforms/InstSimplify/domcondition.ll
+++ b/llvm/test/Transforms/InstSimplify/domcondition.ll
@@ -278,3 +278,210 @@ end:
}
declare void @foo(i32)
+
+
+define i1 @simplify_fcmp_implied_by_dom_cond_range_true(float %x) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_range_true(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 true
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, 0.0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp olt float %x, 1.0
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_in_else_implied_by_dom_cond_range_true(float %x) {
+; CHECK-LABEL: @simplify_fcmp_in_else_implied_by_dom_cond_range_true(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], 1.000000e+00
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 true
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 true
+;
+ %cmp = fcmp olt float %x, 1.0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ ret i1 true
+
+if.else:
+ %cmp2 = fcmp uge float %x, 0.5
+ ret i1 %cmp2
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_range_false(float %x) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_range_false(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 false
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, 0.0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp ogt float %x, 1.0
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_pred_true(float %x, float %y) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_pred_true(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 true
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, %y
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp ole float %x, %y
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_pred_false(float %x, float %y) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_pred_false(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 false
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, %y
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp ogt float %x, %y
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_pred_commuted(float %x, float %y) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_pred_commuted(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: ret i1 true
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, %y
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp oge float %y, %x
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+; Negative tests
+
+define i1 @simplify_fcmp_implied_by_dom_cond_wrong_range(float %x) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_wrong_range(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp olt float [[X]], -1.000000e+00
+; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, 0.0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp olt float %x, -1.0
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_range_mismatched_operand(float %x, float %y) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_range_mismatched_operand(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp olt float [[Y:%.*]], 1.000000e+00
+; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, 0.0
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp olt float %y, 1.0
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_wrong_pred(float %x, float %y) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_wrong_pred(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp ole float [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp olt float [[X]], [[Y]]
+; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp ole float %x, %y
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp olt float %x, %y
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
+
+define i1 @simplify_fcmp_implied_by_dom_cond_pred_mismatched_operand(float %x, float %y, float %z) {
+; CHECK-LABEL: @simplify_fcmp_implied_by_dom_cond_pred_mismatched_operand(
+; CHECK-NEXT: [[CMP:%.*]] = fcmp olt float [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK: if.then:
+; CHECK-NEXT: [[CMP2:%.*]] = fcmp ole float [[X]], [[Z:%.*]]
+; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK: if.else:
+; CHECK-NEXT: ret i1 false
+;
+ %cmp = fcmp olt float %x, %y
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+ %cmp2 = fcmp ole float %x, %z
+ ret i1 %cmp2
+
+if.else:
+ ret i1 false
+}
|
dtcxzyw
commented
Sep 28, 2025
Rajveer100
reviewed
Sep 29, 2025
Generally, only one component is involved in one patch. Although this patch doesn't resolve the original issue, it still removes many redundant fcmps in real-world code. |
Rajveer100
approved these changes
Sep 30, 2025
There was a problem hiding this comment.
LG once llvm/test/CodeGen/AMDGPU/sgpr-copy.ll is addressed.
nikic
approved these changes
Oct 4, 2025
There was a problem hiding this comment.
This look ok. I'm not a fan of isImpliedByDomCondition() because it's a hack that only supports dominating conditions from a directly preceding branch -- we're going to need ConstantRangeFP support in one of the lattice propagating passes to cover the general case. But I guess it does work for some cases...
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/190/builds/28528 Here is the relevant piece of the build log for the reference |
aokblast
pushed a commit
to aokblast/llvm-project
that referenced
this pull request
Oct 6, 2025
This patch simplifies an fcmp into true/false if it is implied by a dominating fcmp. As an initial support, it only handles two cases: + `fcmp pred1, X, Y -> fcmp pred2, X, Y`: use set operations. + `fcmp pred1, X, C1 -> fcmp pred2, X, C2`: use `ConstantFPRange` and set operations. Note: It doesn't fix llvm#70985, as the second fcmp in the motivating case is not dominated by the edge. We may need to adjust JumpThreading to handle this case. Comptime impact (~+0.1%): https://llvm-compile-time-tracker.com/compare.php?from=a728f213c863e4dd19f8969a417148d2951323c0&to=8ca70404fb0d66a824f39d83050ac38e2f1b25b9&stat=instructions:u IR diff: dtcxzyw/llvm-opt-benchmark#2848
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This patch simplifies an fcmp into true/false if it is implied by a dominating fcmp.
As an initial support, it only handles two cases:
fcmp pred1, X, Y -> fcmp pred2, X, Y: use set operations.fcmp pred1, X, C1 -> fcmp pred2, X, C2: useConstantFPRangeand set operations.Note: It doesn't fix #70985, as the second fcmp in the motivating case is not dominated by the edge. We may need to adjust JumpThreading to handle this case.
Comptime impact (~+0.1%): https://llvm-compile-time-tracker.com/compare.php?from=a728f213c863e4dd19f8969a417148d2951323c0&to=8ca70404fb0d66a824f39d83050ac38e2f1b25b9&stat=instructions:u
IR diff: dtcxzyw/llvm-opt-benchmark#2848