[llvm] 9e587ce - [SCCP] Simplify [us]cmp(X, Y) into X - Y (#144717)
via llvm-commits
llvm-commits at lists.llvm.org
Sun Jul 20 00:01:47 PDT 2025
Author: Yingwei Zheng
Date: 2025-07-20T15:01:44+08:00
New Revision: 9e587ce6f0da91ee71dd6920dff0930b291169bb
URL: https://github.com/llvm/llvm-project/commit/9e587ce6f0da91ee71dd6920dff0930b291169bb
DIFF: https://github.com/llvm/llvm-project/commit/9e587ce6f0da91ee71dd6920dff0930b291169bb.diff
LOG: [SCCP] Simplify [us]cmp(X, Y) into X - Y (#144717)
If the difference between [us]cmp's operands is not greater than 1, we
can simplify it into `X - Y`.
Alive2: https://alive2.llvm.org/ce/z/JS55so
llvm-opt-benchmark diff:
https://github.com/dtcxzyw/llvm-opt-benchmark/pull/2464/files
Added:
llvm/test/Transforms/SCCP/uscmp.ll
Modified:
llvm/lib/Transforms/Utils/SCCPSolver.cpp
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/Utils/SCCPSolver.cpp b/llvm/lib/Transforms/Utils/SCCPSolver.cpp
index 586874f57ae73..b9292af093bd0 100644
--- a/llvm/lib/Transforms/Utils/SCCPSolver.cpp
+++ b/llvm/lib/Transforms/Utils/SCCPSolver.cpp
@@ -19,7 +19,9 @@
#include "llvm/Analysis/ValueLattice.h"
#include "llvm/Analysis/ValueLatticeUtils.h"
#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstVisitor.h"
+#include "llvm/IR/NoFolder.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/Debug.h"
@@ -245,11 +247,43 @@ static Value *simplifyInstruction(SCCPSolver &Solver,
const APInt *RHSC;
// Remove masking operations.
if (match(&Inst, m_And(m_Value(X), m_LowBitMask(RHSC)))) {
- ConstantRange LRange = GetRange(Inst.getOperand(0));
+ ConstantRange LRange = GetRange(X);
if (LRange.getUnsignedMax().ule(*RHSC))
return X;
}
+ // Check if we can simplify [us]cmp(X, Y) to X - Y.
+ if (auto *Cmp = dyn_cast<CmpIntrinsic>(&Inst)) {
+ Value *LHS = Cmp->getOperand(0);
+ Value *RHS = Cmp->getOperand(1);
+ unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
+ // Bail out on 1-bit comparisons.
+ if (BitWidth == 1)
+ return nullptr;
+ ConstantRange LRange = GetRange(LHS);
+ if (LRange.isSizeLargerThan(3))
+ return nullptr;
+ ConstantRange RRange = GetRange(RHS);
+ if (RRange.isSizeLargerThan(3))
+ return nullptr;
+ ConstantRange RHSLower = RRange.sub(APInt(BitWidth, 1));
+ ConstantRange RHSUpper = RRange.add(APInt(BitWidth, 1));
+ ICmpInst::Predicate Pred =
+ Cmp->isSigned() ? CmpInst::ICMP_SLE : CmpInst::ICMP_ULE;
+ if (!RHSLower.icmp(Pred, LRange) || !LRange.icmp(Pred, RHSUpper))
+ return nullptr;
+
+ IRBuilder<NoFolder> Builder(&Inst);
+ Value *Sub = Builder.CreateSub(LHS, RHS, Inst.getName(), /*HasNUW=*/false,
+ /*HasNSW=*/Cmp->isSigned());
+ InsertedValues.insert(Sub);
+ if (Sub->getType() != Inst.getType()) {
+ Sub = Builder.CreateSExtOrTrunc(Sub, Inst.getType());
+ InsertedValues.insert(Sub);
+ }
+ return Sub;
+ }
+
return nullptr;
}
diff --git a/llvm/test/Transforms/SCCP/uscmp.ll b/llvm/test/Transforms/SCCP/uscmp.ll
new file mode 100644
index 0000000000000..d010c0632c809
--- /dev/null
+++ b/llvm/test/Transforms/SCCP/uscmp.ll
@@ -0,0 +1,185 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -passes=sccp -S < %s | FileCheck %s
+
+define i32 @scmp_to_sub(i32 range(i32 -1, 2) %a) {
+; CHECK-LABEL: define i32 @scmp_to_sub(
+; CHECK-SAME: i32 range(i32 -1, 2) [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP:%.*]] = sub nsw i32 [[A]], 0
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %scmp = call i32 @llvm.scmp(i32 %a, i32 0)
+ ret i32 %scmp
+}
+
+define i32 @scmp_zext_to_sub(i1 %a, i1 %b) {
+; CHECK-LABEL: define i32 @scmp_zext_to_sub(
+; CHECK-SAME: i1 [[A:%.*]], i1 [[B:%.*]]) {
+; CHECK-NEXT: [[ZEXT_A:%.*]] = zext i1 [[A]] to i32
+; CHECK-NEXT: [[ZEXT_B:%.*]] = zext i1 [[B]] to i32
+; CHECK-NEXT: [[SCMP:%.*]] = sub nsw i32 [[ZEXT_A]], [[ZEXT_B]]
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %zext_a = zext i1 %a to i32
+ %zext_b = zext i1 %b to i32
+ %scmp = call i32 @llvm.scmp(i32 %zext_a, i32 %zext_b)
+ ret i32 %scmp
+}
+
+define i8 @scmp_to_sub_trunc(i32 range(i32 -1, 2) %a) {
+; CHECK-LABEL: define i8 @scmp_to_sub_trunc(
+; CHECK-SAME: i32 range(i32 -1, 2) [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP1:%.*]] = sub nsw i32 [[A]], 0
+; CHECK-NEXT: [[SCMP:%.*]] = trunc i32 [[SCMP1]] to i8
+; CHECK-NEXT: ret i8 [[SCMP]]
+;
+ %scmp = call i8 @llvm.scmp(i32 %a, i32 0)
+ ret i8 %scmp
+}
+
+define i64 @scmp_to_sub_sext(i32 range(i32 -1, 2) %a) {
+; CHECK-LABEL: define i64 @scmp_to_sub_sext(
+; CHECK-SAME: i32 range(i32 -1, 2) [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP1:%.*]] = sub nsw i32 [[A]], 0
+; CHECK-NEXT: [[SCMP:%.*]] = sext i32 [[SCMP1]] to i64
+; CHECK-NEXT: ret i64 [[SCMP]]
+;
+ %scmp = call i64 @llvm.scmp(i32 %a, i32 0)
+ ret i64 %scmp
+}
+
+define i32 @scmp_to_sub_small_range(i32 range(i32 -1, 1) %a) {
+; CHECK-LABEL: define i32 @scmp_to_sub_small_range(
+; CHECK-SAME: i32 range(i32 -1, 1) [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP:%.*]] = sub nsw i32 [[A]], 0
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %scmp = call i32 @llvm.scmp(i32 %a, i32 0)
+ ret i32 %scmp
+}
+
+define i32 @ucmp_to_sub(i32 range(i32 0, 3) %a) {
+; CHECK-LABEL: define i32 @ucmp_to_sub(
+; CHECK-SAME: i32 range(i32 0, 3) [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP:%.*]] = sub i32 [[A]], 1
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %ucmp = call i32 @llvm.ucmp(i32 %a, i32 1)
+ ret i32 %ucmp
+}
+
+define i8 @ucmp_to_sub_trunc(i32 range(i32 0, 3) %a) {
+; CHECK-LABEL: define i8 @ucmp_to_sub_trunc(
+; CHECK-SAME: i32 range(i32 0, 3) [[A:%.*]]) {
+; CHECK-NEXT: [[UCMP1:%.*]] = sub i32 [[A]], 1
+; CHECK-NEXT: [[UCMP:%.*]] = trunc i32 [[UCMP1]] to i8
+; CHECK-NEXT: ret i8 [[UCMP]]
+;
+ %ucmp = call i8 @llvm.ucmp(i32 %a, i32 1)
+ ret i8 %ucmp
+}
+
+define i64 @ucmp_to_sub_sext(i32 range(i32 0, 3) %a) {
+; CHECK-LABEL: define i64 @ucmp_to_sub_sext(
+; CHECK-SAME: i32 range(i32 0, 3) [[A:%.*]]) {
+; CHECK-NEXT: [[UCMP1:%.*]] = sub i32 [[A]], 1
+; CHECK-NEXT: [[UCMP:%.*]] = sext i32 [[UCMP1]] to i64
+; CHECK-NEXT: ret i64 [[UCMP]]
+;
+ %ucmp = call i64 @llvm.ucmp(i32 %a, i32 1)
+ ret i64 %ucmp
+}
+
+; TODO: we can fold this into %a.
+define i32 @ucmp_to_sub_small_range(i32 range(i32 0, 2) %a) {
+; CHECK-LABEL: define i32 @ucmp_to_sub_small_range(
+; CHECK-SAME: i32 range(i32 0, 2) [[A:%.*]]) {
+; CHECK-NEXT: [[UCMP:%.*]] = call i32 @llvm.ucmp.i32.i32(i32 [[A]], i32 0)
+; CHECK-NEXT: ret i32 [[UCMP]]
+;
+ %ucmp = call i32 @llvm.ucmp(i32 %a, i32 0)
+ ret i32 %ucmp
+}
+
+define i32 @scmp_to_sub_large_range(i32 range(i32 -1, 3) %a) {
+; CHECK-LABEL: define i32 @scmp_to_sub_large_range(
+; CHECK-SAME: i32 range(i32 -1, 3) [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP:%.*]] = call i32 @llvm.scmp.i32.i32(i32 [[A]], i32 0)
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %scmp = call i32 @llvm.scmp(i32 %a, i32 0)
+ ret i32 %scmp
+}
+
+define i32 @ucmp_to_sub_large_range(i32 range(i32 -1, 3) %a) {
+; CHECK-LABEL: define i32 @ucmp_to_sub_large_range(
+; CHECK-SAME: i32 range(i32 -1, 3) [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP:%.*]] = call i32 @llvm.ucmp.i32.i32(i32 [[A]], i32 0)
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %ucmp = call i32 @llvm.ucmp(i32 %a, i32 0)
+ ret i32 %ucmp
+}
+
+define i32 @scmp_to_sub_wrap(i8 range(i8 127, -126) %a) {
+; CHECK-LABEL: define i32 @scmp_to_sub_wrap(
+; CHECK-SAME: i8 range(i8 127, -126) [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP:%.*]] = call i32 @llvm.scmp.i32.i8(i8 [[A]], i8 -128)
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %scmp = call i32 @llvm.scmp(i8 %a, i8 -128)
+ ret i32 %scmp
+}
+
+define i32 @ucmp_to_sub_wrap(i8 range(i8 -1, 2) %a) {
+; CHECK-LABEL: define i32 @ucmp_to_sub_wrap(
+; CHECK-SAME: i8 range(i8 -1, 2) [[A:%.*]]) {
+; CHECK-NEXT: [[UCMP:%.*]] = call i32 @llvm.ucmp.i32.i8(i8 [[A]], i8 0)
+; CHECK-NEXT: ret i32 [[UCMP]]
+;
+ %ucmp = call i32 @llvm.ucmp(i8 %a, i8 0)
+ ret i32 %ucmp
+}
+
+; It is incorrect to convert a ucmp into sub when the input type is i1.
+define i32 @ucmp_to_sub_i1_rhs_const(i1 %a) {
+; CHECK-LABEL: define i32 @ucmp_to_sub_i1_rhs_const(
+; CHECK-SAME: i1 [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP:%.*]] = call i32 @llvm.ucmp.i32.i1(i1 [[A]], i1 false)
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %ucmp = call i32 @llvm.ucmp(i1 %a, i1 false)
+ ret i32 %ucmp
+}
+
+; It is incorrect to convert a ucmp into sub when the input type is i1.
+define i32 @ucmp_to_sub_i1_lhs_const(i1 %a) {
+; CHECK-LABEL: define i32 @ucmp_to_sub_i1_lhs_const(
+; CHECK-SAME: i1 [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP:%.*]] = call i32 @llvm.ucmp.i32.i1(i1 false, i1 [[A]])
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %ucmp = call i32 @llvm.ucmp(i1 false, i1 %a)
+ ret i32 %ucmp
+}
+
+; It is incorrect to convert a ucmp into sub when the input type is i1.
+define i32 @ucmp_to_sub_i1(i1 %a, i1 %b) {
+; CHECK-LABEL: define i32 @ucmp_to_sub_i1(
+; CHECK-SAME: i1 [[A:%.*]], i1 [[B:%.*]]) {
+; CHECK-NEXT: [[SCMP:%.*]] = call i32 @llvm.ucmp.i32.i1(i1 [[A]], i1 [[B]])
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %ucmp = call i32 @llvm.ucmp(i1 %a, i1 %b)
+ ret i32 %ucmp
+}
+
+; It is incorrect to convert a scmp into sub when the input type is i1.
+define i32 @scmp_to_sub_i1_rhs_const(i1 %a) {
+; CHECK-LABEL: define i32 @scmp_to_sub_i1_rhs_const(
+; CHECK-SAME: i1 [[A:%.*]]) {
+; CHECK-NEXT: [[SCMP:%.*]] = call i32 @llvm.scmp.i32.i1(i1 [[A]], i1 false)
+; CHECK-NEXT: ret i32 [[SCMP]]
+;
+ %scmp = call i32 @llvm.scmp(i1 %a, i1 false)
+ ret i32 %scmp
+}
More information about the llvm-commits
mailing list