Skip to content

Commit 23f0f06

Browse files
dc03dc03
committed
[InstCombine] Fold icmps comparing uadd_sat with a constant
This patch is a continuation of D154206. It introduces a fold for the operation "uadd_sat(X, C) pred C2" where "C" and "C2" are constants. The fold is: uadd_sat(X, C) pred C2 => (X >= ~C) || ((X + C) pred C2) -> when (UINT_MAX pred C2) is true => (X < ~C) && ((X + C) pred C2) -> when (UINT_MAX pred C2) is false This patch also generalizes the fold to work with any saturating intrinsic as long as the saturating value is known. Proofs: https://alive2.llvm.org/ce/z/wWeirP Differential Revision: https://reviews.llvm.org/D154565
1 parent ed91084 commit 23f0f06

File tree

3 files changed

+69
-53
lines changed

3 files changed

+69
-53
lines changed

‎llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

Lines changed: 51 additions & 24 deletions
Original file line numberDiff line numberDiff line change
@@ -3580,48 +3580,73 @@ Instruction *InstCombinerImpl::foldICmpBinOpWithConstant(ICmpInst &Cmp,
35803580
}
35813581

35823582
static Instruction *
3583-
foldICmpUSubSatWithConstant(ICmpInst::Predicate Pred, IntrinsicInst *II,
3584-
const APInt &C, InstCombiner::BuilderTy &Builder) {
3583+
foldICmpUSubSatOrUAddSatWithConstant(ICmpInst::Predicate Pred,
3584+
SaturatingInst *II, const APInt &C,
3585+
InstCombiner::BuilderTy &Builder) {
35853586
// This transform may end up producing more than one instruction for the
35863587
// intrinsic, so limit it to one user of the intrinsic.
35873588
if (!II->hasOneUse())
35883589
return nullptr;
35893590

3590-
// Let Y = usub_sat(X, C) pred C2
3591-
// => Y = (X < C ? 0 : (X - C)) pred C2
3592-
// => Y = (X < C) ? (0 pred C2) : ((X - C) pred C2)
3591+
// Let Y = [add/sub]_sat(X, C) pred C2
3592+
// SatVal = The saturating value for the operation
3593+
// WillWrap = Whether or not the operation will underflow / overflow
3594+
// => Y = (WillWrap ? SatVal : (X binop C)) pred C2
3595+
// => Y = WillWrap ? (SatVal pred C2) : ((X binop C) pred C2)
35933596
//
3594-
// When (0 pred C2) is true, then
3595-
// Y = (X < C) ? true : ((X - C) pred C2)
3596-
// => Y = (X < C) || ((X - C) pred C2)
3597+
// When (SatVal pred C2) is true, then
3598+
// Y = WillWrap ? true : ((X binop C) pred C2)
3599+
// => Y = WillWrap || ((X binop C) pred C2)
35973600
// else
3598-
// Y = (X < C) ? false : ((X - C) pred C2)
3599-
// => Y = !(X < C) && ((X - C) pred C2)
3600-
// => Y = (X >= C) && ((X - C) pred C2)
3601+
// Y = WillWrap ? false : ((X binop C) pred C2)
3602+
// => Y = !WillWrap ? ((X binop C) pred C2) : false
3603+
// => Y = !WillWrap && ((X binop C) pred C2)
36013604
Value *Op0 = II->getOperand(0);
36023605
Value *Op1 = II->getOperand(1);
36033606

3604-
// Check (0 pred C2)
3605-
auto [NewPred, LogicalOp] =
3606-
ICmpInst::compare(APInt::getZero(C.getBitWidth()), C, Pred)
3607-
? std::make_pair(ICmpInst::ICMP_ULT, Instruction::BinaryOps::Or)
3608-
: std::make_pair(ICmpInst::ICMP_UGE, Instruction::BinaryOps::And);
3609-
36103607
const APInt *COp1;
3611-
// This transform only works when the usub_sat has an integral constant or
3608+
// This transform only works when the intrinsic has an integral constant or
36123609
// splat vector as the second operand.
36133610
if (!match(Op1, m_APInt(COp1)))
36143611
return nullptr;
36153612

3616-
ConstantRange C1 = ConstantRange::makeExactICmpRegion(NewPred, *COp1);
3617-
// Convert '(X - C) pred C2' into 'X pred C2' shifted by C.
3613+
APInt SatVal;
3614+
switch (II->getIntrinsicID()) {
3615+
default:
3616+
llvm_unreachable(
3617+
"This function only works with usub_sat and uadd_sat for now!");
3618+
case Intrinsic::uadd_sat:
3619+
SatVal = APInt::getAllOnes(C.getBitWidth());
3620+
break;
3621+
case Intrinsic::usub_sat:
3622+
SatVal = APInt::getZero(C.getBitWidth());
3623+
break;
3624+
}
3625+
3626+
// Check (SatVal pred C2)
3627+
bool SatValCheck = ICmpInst::compare(SatVal, C, Pred);
3628+
3629+
// !WillWrap.
3630+
ConstantRange C1 = ConstantRange::makeExactNoWrapRegion(
3631+
II->getBinaryOp(), *COp1, II->getNoWrapKind());
3632+
3633+
// WillWrap.
3634+
if (SatValCheck)
3635+
C1 = C1.inverse();
3636+
36183637
ConstantRange C2 = ConstantRange::makeExactICmpRegion(Pred, C);
3619-
C2 = C2.add(*COp1);
3638+
if (II->getBinaryOp() == Instruction::Add)
3639+
C2 = C2.sub(*COp1);
3640+
else
3641+
C2 = C2.add(*COp1);
3642+
3643+
Instruction::BinaryOps CombiningOp =
3644+
SatValCheck ? Instruction::BinaryOps::Or : Instruction::BinaryOps::And;
36203645

36213646
std::optional<ConstantRange> Combination;
3622-
if (LogicalOp == Instruction::BinaryOps::Or)
3647+
if (CombiningOp == Instruction::BinaryOps::Or)
36233648
Combination = C1.exactUnionWith(C2);
3624-
else /* LogicalOp == Instruction::BinaryOps::And */
3649+
else /* CombiningOp == Instruction::BinaryOps::And */
36253650
Combination = C1.exactIntersectWith(C2);
36263651

36273652
if (!Combination)
@@ -3649,8 +3674,10 @@ Instruction *InstCombinerImpl::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
36493674
switch (II->getIntrinsicID()) {
36503675
default:
36513676
break;
3677+
case Intrinsic::uadd_sat:
36523678
case Intrinsic::usub_sat:
3653-
if (auto *Folded = foldICmpUSubSatWithConstant(Pred, II, C, Builder))
3679+
if (auto *Folded = foldICmpUSubSatOrUAddSatWithConstant(
3680+
Pred, cast<SaturatingInst>(II), C, Builder))
36543681
return Folded;
36553682
break;
36563683
}

‎llvm/test/Transforms/InstCombine/icmp-uadd-sat.ll

Lines changed: 16 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -10,8 +10,7 @@
1010
define i1 @icmp_eq_basic(i8 %arg) {
1111
; CHECK-LABEL: define i1 @icmp_eq_basic
1212
; CHECK-SAME: (i8 [[ARG:%.*]]) {
13-
; CHECK-NEXT: [[ADD:%.*]] = call i8 @llvm.uadd.sat.i8(i8 [[ARG]], i8 2)
14-
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i8 [[ADD]], 5
13+
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i8 [[ARG]], 3
1514
; CHECK-NEXT: ret i1 [[CMP]]
1615
;
1716
%add = call i8 @llvm.uadd.sat.i8(i8 %arg, i8 2)
@@ -22,8 +21,7 @@ define i1 @icmp_eq_basic(i8 %arg) {
2221
define i1 @icmp_ne_basic(i16 %arg) {
2322
; CHECK-LABEL: define i1 @icmp_ne_basic
2423
; CHECK-SAME: (i16 [[ARG:%.*]]) {
25-
; CHECK-NEXT: [[ADD:%.*]] = call i16 @llvm.uadd.sat.i16(i16 [[ARG]], i16 8)
26-
; CHECK-NEXT: [[CMP:%.*]] = icmp ne i16 [[ADD]], 9
24+
; CHECK-NEXT: [[CMP:%.*]] = icmp ne i16 [[ARG]], 1
2725
; CHECK-NEXT: ret i1 [[CMP]]
2826
;
2927
%add = call i16 @llvm.uadd.sat.i16(i16 %arg, i16 8)
@@ -34,8 +32,7 @@ define i1 @icmp_ne_basic(i16 %arg) {
3432
define i1 @icmp_ule_basic(i32 %arg) {
3533
; CHECK-LABEL: define i1 @icmp_ule_basic
3634
; CHECK-SAME: (i32 [[ARG:%.*]]) {
37-
; CHECK-NEXT: [[ADD:%.*]] = call i32 @llvm.uadd.sat.i32(i32 [[ARG]], i32 2)
38-
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i32 [[ADD]], 4
35+
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i32 [[ARG]], 2
3936
; CHECK-NEXT: ret i1 [[CMP]]
4037
;
4138
%add = call i32 @llvm.uadd.sat.i32(i32 %arg, i32 2)
@@ -46,8 +43,7 @@ define i1 @icmp_ule_basic(i32 %arg) {
4643
define i1 @icmp_ult_basic(i64 %arg) {
4744
; CHECK-LABEL: define i1 @icmp_ult_basic
4845
; CHECK-SAME: (i64 [[ARG:%.*]]) {
49-
; CHECK-NEXT: [[ADD:%.*]] = call i64 @llvm.uadd.sat.i64(i64 [[ARG]], i64 5)
50-
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i64 [[ADD]], 20
46+
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i64 [[ARG]], 15
5147
; CHECK-NEXT: ret i1 [[CMP]]
5248
;
5349
%add = call i64 @llvm.uadd.sat.i64(i64 %arg, i64 5)
@@ -58,8 +54,7 @@ define i1 @icmp_ult_basic(i64 %arg) {
5854
define i1 @icmp_uge_basic(i8 %arg) {
5955
; CHECK-LABEL: define i1 @icmp_uge_basic
6056
; CHECK-SAME: (i8 [[ARG:%.*]]) {
61-
; CHECK-NEXT: [[ADD:%.*]] = call i8 @llvm.uadd.sat.i8(i8 [[ARG]], i8 4)
62-
; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i8 [[ADD]], 7
57+
; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i8 [[ARG]], 3
6358
; CHECK-NEXT: ret i1 [[CMP]]
6459
;
6560
%add = call i8 @llvm.uadd.sat.i8(i8 %arg, i8 4)
@@ -70,8 +65,7 @@ define i1 @icmp_uge_basic(i8 %arg) {
7065
define i1 @icmp_ugt_basic(i16 %arg) {
7166
; CHECK-LABEL: define i1 @icmp_ugt_basic
7267
; CHECK-SAME: (i16 [[ARG:%.*]]) {
73-
; CHECK-NEXT: [[ADD:%.*]] = call i16 @llvm.uadd.sat.i16(i16 [[ARG]], i16 1)
74-
; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i16 [[ADD]], 3
68+
; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i16 [[ARG]], 2
7569
; CHECK-NEXT: ret i1 [[CMP]]
7670
;
7771
%add = call i16 @llvm.uadd.sat.i16(i16 %arg, i16 1)
@@ -82,8 +76,7 @@ define i1 @icmp_ugt_basic(i16 %arg) {
8276
define i1 @icmp_sle_basic(i32 %arg) {
8377
; CHECK-LABEL: define i1 @icmp_sle_basic
8478
; CHECK-SAME: (i32 [[ARG:%.*]]) {
85-
; CHECK-NEXT: [[ADD:%.*]] = call i32 @llvm.uadd.sat.i32(i32 [[ARG]], i32 10)
86-
; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[ADD]], 9
79+
; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i32 [[ARG]], 2147483637
8780
; CHECK-NEXT: ret i1 [[CMP]]
8881
;
8982
%add = call i32 @llvm.uadd.sat.i32(i32 %arg, i32 10)
@@ -94,8 +87,7 @@ define i1 @icmp_sle_basic(i32 %arg) {
9487
define i1 @icmp_slt_basic(i64 %arg) {
9588
; CHECK-LABEL: define i1 @icmp_slt_basic
9689
; CHECK-SAME: (i64 [[ARG:%.*]]) {
97-
; CHECK-NEXT: [[ADD:%.*]] = call i64 @llvm.uadd.sat.i64(i64 [[ARG]], i64 24)
98-
; CHECK-NEXT: [[CMP:%.*]] = icmp slt i64 [[ADD]], 5
90+
; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i64 [[ARG]], 9223372036854775783
9991
; CHECK-NEXT: ret i1 [[CMP]]
10092
;
10193
%add = call i64 @llvm.uadd.sat.i64(i64 %arg, i64 24)
@@ -106,8 +98,8 @@ define i1 @icmp_slt_basic(i64 %arg) {
10698
define i1 @icmp_sge_basic(i8 %arg) {
10799
; CHECK-LABEL: define i1 @icmp_sge_basic
108100
; CHECK-SAME: (i8 [[ARG:%.*]]) {
109-
; CHECK-NEXT: [[ADD:%.*]] = call i8 @llvm.uadd.sat.i8(i8 [[ARG]], i8 1)
110-
; CHECK-NEXT: [[CMP:%.*]] = icmp sgt i8 [[ADD]], 3
101+
; CHECK-NEXT: [[TMP1:%.*]] = add i8 [[ARG]], -3
102+
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i8 [[TMP1]], 124
111103
; CHECK-NEXT: ret i1 [[CMP]]
112104
;
113105
%add = call i8 @llvm.uadd.sat.i8(i8 %arg, i8 1)
@@ -118,8 +110,8 @@ define i1 @icmp_sge_basic(i8 %arg) {
118110
define i1 @icmp_sgt_basic(i16 %arg) {
119111
; CHECK-LABEL: define i1 @icmp_sgt_basic
120112
; CHECK-SAME: (i16 [[ARG:%.*]]) {
121-
; CHECK-NEXT: [[ADD:%.*]] = call i16 @llvm.uadd.sat.i16(i16 [[ARG]], i16 2)
122-
; CHECK-NEXT: [[CMP:%.*]] = icmp sgt i16 [[ADD]], 5
113+
; CHECK-NEXT: [[TMP1:%.*]] = add i16 [[ARG]], -4
114+
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i16 [[TMP1]], 32762
123115
; CHECK-NEXT: ret i1 [[CMP]]
124116
;
125117
%add = call i16 @llvm.uadd.sat.i16(i16 %arg, i16 2)
@@ -150,8 +142,7 @@ define i1 @icmp_eq_multiuse(i8 %arg) {
150142
define <2 x i1> @icmp_eq_vector_equal(<2 x i8> %arg) {
151143
; CHECK-LABEL: define <2 x i1> @icmp_eq_vector_equal
152144
; CHECK-SAME: (<2 x i8> [[ARG:%.*]]) {
153-
; CHECK-NEXT: [[ADD:%.*]] = call <2 x i8> @llvm.uadd.sat.v2i8(<2 x i8> [[ARG]], <2 x i8> <i8 2, i8 2>)
154-
; CHECK-NEXT: [[CMP:%.*]] = icmp eq <2 x i8> [[ADD]], <i8 5, i8 5>
145+
; CHECK-NEXT: [[CMP:%.*]] = icmp eq <2 x i8> [[ARG]], <i8 3, i8 3>
155146
; CHECK-NEXT: ret <2 x i1> [[CMP]]
156147
;
157148
%add = call <2 x i8> @llvm.uadd.sat.v2i8(<2 x i8> %arg, <2 x i8> <i8 2, i8 2>)
@@ -174,8 +165,7 @@ define <2 x i1> @icmp_eq_vector_unequal(<2 x i8> %arg) {
174165
define <2 x i1> @icmp_ne_vector_equal(<2 x i16> %arg) {
175166
; CHECK-LABEL: define <2 x i1> @icmp_ne_vector_equal
176167
; CHECK-SAME: (<2 x i16> [[ARG:%.*]]) {
177-
; CHECK-NEXT: [[ADD:%.*]] = call <2 x i16> @llvm.uadd.sat.v2i16(<2 x i16> [[ARG]], <2 x i16> <i16 3, i16 3>)
178-
; CHECK-NEXT: [[CMP:%.*]] = icmp ne <2 x i16> [[ADD]], <i16 5, i16 5>
168+
; CHECK-NEXT: [[CMP:%.*]] = icmp ne <2 x i16> [[ARG]], <i16 2, i16 2>
179169
; CHECK-NEXT: ret <2 x i1> [[CMP]]
180170
;
181171
%add = call <2 x i16> @llvm.uadd.sat.v2i16(<2 x i16> %arg, <2 x i16> <i16 3, i16 3>)
@@ -198,8 +188,7 @@ define <2 x i1> @icmp_ne_vector_unequal(<2 x i16> %arg) {
198188
define <2 x i1> @icmp_ule_vector_equal(<2 x i32> %arg) {
199189
; CHECK-LABEL: define <2 x i1> @icmp_ule_vector_equal
200190
; CHECK-SAME: (<2 x i32> [[ARG:%.*]]) {
201-
; CHECK-NEXT: [[ADD:%.*]] = call <2 x i32> @llvm.uadd.sat.v2i32(<2 x i32> [[ARG]], <2 x i32> <i32 3, i32 3>)
202-
; CHECK-NEXT: [[CMP:%.*]] = icmp ult <2 x i32> [[ADD]], <i32 5, i32 5>
191+
; CHECK-NEXT: [[CMP:%.*]] = icmp ult <2 x i32> [[ARG]], <i32 2, i32 2>
203192
; CHECK-NEXT: ret <2 x i1> [[CMP]]
204193
;
205194
%add = call <2 x i32> @llvm.uadd.sat.v2i32(<2 x i32> %arg, <2 x i32> <i32 3, i32 3>)
@@ -222,8 +211,7 @@ define <2 x i1> @icmp_ule_vector_unequal(<2 x i32> %arg) {
222211
define <2 x i1> @icmp_sgt_vector_equal(<2 x i64> %arg) {
223212
; CHECK-LABEL: define <2 x i1> @icmp_sgt_vector_equal
224213
; CHECK-SAME: (<2 x i64> [[ARG:%.*]]) {
225-
; CHECK-NEXT: [[ADD:%.*]] = call <2 x i64> @llvm.uadd.sat.v2i64(<2 x i64> [[ARG]], <2 x i64> <i64 409623, i64 409623>)
226-
; CHECK-NEXT: [[CMP:%.*]] = icmp sgt <2 x i64> [[ADD]], <i64 1234, i64 1234>
214+
; CHECK-NEXT: [[CMP:%.*]] = icmp ult <2 x i64> [[ARG]], <i64 9223372036854366185, i64 9223372036854366185>
227215
; CHECK-NEXT: ret <2 x i1> [[CMP]]
228216
;
229217
%add = call <2 x i64> @llvm.uadd.sat.v2i64(<2 x i64> %arg, <2 x i64> <i64 409623, i64 409623>)

‎llvm/test/Transforms/InstCombine/icmp-usub-sat.ll

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,8 @@
11
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
22
; RUN: opt < %s -passes=instcombine -S | FileCheck %s
33

4-
; Tests for InstCombineCompares.cpp::foldICmpUSubSatWithConstant
4+
; Tests for InstCombineCompares.cpp::foldICmpUSubSatOrUAddSatWithConstant
5+
; - usub_sat case
56
; https://github.com/llvm/llvm-project/issues/58342
67

78
; ==============================================================================

0 commit comments

Comments
 (0)