[SCCP] Don't allow undef ranges when performing operations #93163
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
When performing some range operation (e.g. and) on a constant range that includes undef, we currently just ignore the undef value, which is obviously incorrect. Instead, we can do one of two things: * Say that the result range also includes undef. * Treat undef as a full range. This patch goes with the second approach -- I'd expect it to be a bit better overall, e.g. it allows preserving the fact that a zext of a range with undef isn't a full range.
|
@llvm/pr-subscribers-llvm-transforms Author: Nikita Popov (nikic) ChangesWhen performing some range operation (e.g. and) on a constant range that includes undef, we currently just ignore the undef value, which is obviously incorrect. Instead, we can do one of two things:
This patch goes with the second approach -- I'd expect it to be a bit better overall, e.g. it allows preserving the fact that a zext of a range with undef isn't a full range. Fixes #93096. Full diff: https://github.com/llvm/llvm-project/pull/93163.diff 3 Files Affected:
diff --git a/llvm/lib/Transforms/Utils/SCCPSolver.cpp b/llvm/lib/Transforms/Utils/SCCPSolver.cpp
index ce40e8b31b767..4f36bac11e34b 100644
--- a/llvm/lib/Transforms/Utils/SCCPSolver.cpp
+++ b/llvm/lib/Transforms/Utils/SCCPSolver.cpp
@@ -43,7 +43,7 @@ static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() {
}
static ConstantRange getConstantRange(const ValueLatticeElement &LV, Type *Ty,
- bool UndefAllowed = true) {
+ bool UndefAllowed) {
assert(Ty->isIntOrIntVectorTy() && "Should be int or int vector");
if (LV.isConstantRange(UndefAllowed))
return LV.getConstantRange();
@@ -1297,7 +1297,8 @@ void SCCPInstVisitor::visitCastInst(CastInst &I) {
if (I.getDestTy()->isIntegerTy() && I.getSrcTy()->isIntOrIntVectorTy()) {
auto &LV = getValueState(&I);
- ConstantRange OpRange = getConstantRange(OpSt, I.getSrcTy());
+ ConstantRange OpRange =
+ getConstantRange(OpSt, I.getSrcTy(), /*UndefAllowed=*/false);
Type *DestTy = I.getDestTy();
// Vectors where all elements have the same known constant range are treated
@@ -1329,8 +1330,8 @@ void SCCPInstVisitor::handleExtractOfWithOverflow(ExtractValueInst &EVI,
return; // Wait to resolve.
Type *Ty = LHS->getType();
- ConstantRange LR = getConstantRange(L, Ty);
- ConstantRange RR = getConstantRange(R, Ty);
+ ConstantRange LR = getConstantRange(L, Ty, /*UndefAllowed=*/false);
+ ConstantRange RR = getConstantRange(R, Ty, /*UndefAllowed=*/false);
if (Idx == 0) {
ConstantRange Res = LR.binaryOp(WO->getBinaryOp(), RR);
mergeInValue(&EVI, ValueLatticeElement::getRange(Res));
@@ -1534,8 +1535,10 @@ void SCCPInstVisitor::visitBinaryOperator(Instruction &I) {
return markOverdefined(&I);
// Try to simplify to a constant range.
- ConstantRange A = getConstantRange(V1State, I.getType());
- ConstantRange B = getConstantRange(V2State, I.getType());
+ ConstantRange A =
+ getConstantRange(V1State, I.getType(), /*UndefAllowed=*/false);
+ ConstantRange B =
+ getConstantRange(V2State, I.getType(), /*UndefAllowed=*/false);
auto *BO = cast<BinaryOperator>(&I);
ConstantRange R = ConstantRange::getEmpty(I.getType()->getScalarSizeInBits());
@@ -1818,7 +1821,8 @@ void SCCPInstVisitor::handleCallResult(CallBase &CB) {
// Combine range info for the original value with the new range from the
// condition.
- auto CopyOfCR = getConstantRange(CopyOfVal, CopyOf->getType());
+ auto CopyOfCR = getConstantRange(CopyOfVal, CopyOf->getType(),
+ /*UndefAllowed=*/true);
auto NewCR = ImposedCR.intersectWith(CopyOfCR);
// If the existing information is != x, do not use the information from
// a chained predicate, as the != x information is more likely to be
@@ -1863,7 +1867,8 @@ void SCCPInstVisitor::handleCallResult(CallBase &CB) {
const ValueLatticeElement &State = getValueState(Op);
if (State.isUnknownOrUndef())
return;
- OpRanges.push_back(getConstantRange(State, Op->getType()));
+ OpRanges.push_back(
+ getConstantRange(State, Op->getType(), /*UndefAllowed=*/false));
}
ConstantRange Result =
diff --git a/llvm/test/Transforms/SCCP/ip-add-range-to-call.ll b/llvm/test/Transforms/SCCP/ip-add-range-to-call.ll
index c24c554102ddf..91efbcc4ee382 100644
--- a/llvm/test/Transforms/SCCP/ip-add-range-to-call.ll
+++ b/llvm/test/Transforms/SCCP/ip-add-range-to-call.ll
@@ -159,7 +159,7 @@ exit:
}
define i32 @caller5() {
-; CHECK-LABEL: define range(i32 200, 401) i32 @caller5() {
+; CHECK-LABEL: define i32 @caller5() {
; CHECK-NEXT: [[C1:%.*]] = call i32 @callee5(i32 10, i32 100)
; CHECK-NEXT: [[C2:%.*]] = call i32 @callee5(i32 20, i32 200)
; CHECK-NEXT: [[A:%.*]] = add i32 [[C1]], [[C2]]
diff --git a/llvm/test/Transforms/SCCP/range-with-undef.ll b/llvm/test/Transforms/SCCP/range-with-undef.ll
index 444b47df55697..9b8d415171140 100644
--- a/llvm/test/Transforms/SCCP/range-with-undef.ll
+++ b/llvm/test/Transforms/SCCP/range-with-undef.ll
@@ -2,7 +2,6 @@
; RUN: opt -S -passes=ipsccp < %s | FileCheck %s
; Make sure that constant ranges including undef are propagated correctly.
-; FIXME: All of the following are currently miscompiled.
define i8 @test_binop(i1 %cond, i8 %a) {
; CHECK-LABEL: define i8 @test_binop(
@@ -15,7 +14,7 @@ define i8 @test_binop(i1 %cond, i8 %a) {
; CHECK: [[JOIN]]:
; CHECK-NEXT: [[PHI:%.*]] = phi i16 [ undef, %[[ENTRY]] ], [ [[A_EXT]], %[[IF]] ]
; CHECK-NEXT: [[AND:%.*]] = and i16 [[PHI]], -1
-; CHECK-NEXT: [[TRUNC:%.*]] = trunc nuw i16 [[AND]] to i8
+; CHECK-NEXT: [[TRUNC:%.*]] = trunc i16 [[AND]] to i8
; CHECK-NEXT: ret i8 [[TRUNC]]
;
entry:
@@ -43,7 +42,7 @@ define i8 @test_cast(i1 %cond, i8 %a) {
; CHECK: [[JOIN]]:
; CHECK-NEXT: [[PHI:%.*]] = phi i16 [ undef, %[[ENTRY]] ], [ [[A_EXT]], %[[IF]] ]
; CHECK-NEXT: [[ZEXT:%.*]] = zext i16 [[PHI]] to i32
-; CHECK-NEXT: [[TRUNC:%.*]] = trunc nuw i32 [[ZEXT]] to i8
+; CHECK-NEXT: [[TRUNC:%.*]] = trunc i32 [[ZEXT]] to i8
; CHECK-NEXT: ret i8 [[TRUNC]]
;
entry:
@@ -61,7 +60,7 @@ join:
}
define i8 @test_intrin(i1 %cond, i8 %a) {
-; CHECK-LABEL: define range(i8 42, 0) i8 @test_intrin(
+; CHECK-LABEL: define i8 @test_intrin(
; CHECK-SAME: i1 [[COND:%.*]], i8 [[A:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*]]:
; CHECK-NEXT: br i1 [[COND]], label %[[IF:.*]], label %[[JOIN:.*]]
@@ -71,7 +70,7 @@ define i8 @test_intrin(i1 %cond, i8 %a) {
; CHECK: [[JOIN]]:
; CHECK-NEXT: [[PHI:%.*]] = phi i16 [ undef, %[[ENTRY]] ], [ [[A_EXT]], %[[IF]] ]
; CHECK-NEXT: [[UMAX:%.*]] = call i16 @llvm.umax.i16(i16 [[PHI]], i16 42)
-; CHECK-NEXT: [[TRUNC:%.*]] = trunc nuw i16 [[UMAX]] to i8
+; CHECK-NEXT: [[TRUNC:%.*]] = trunc i16 [[UMAX]] to i8
; CHECK-NEXT: ret i8 [[TRUNC]]
;
entry:
@@ -89,7 +88,7 @@ join:
}
define i9 @test_with_overflow(i1 %cond, i8 %a) {
-; CHECK-LABEL: define range(i9 1, -255) i9 @test_with_overflow(
+; CHECK-LABEL: define i9 @test_with_overflow(
; CHECK-SAME: i1 [[COND:%.*]], i8 [[A:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*]]:
; CHECK-NEXT: br i1 [[COND]], label %[[IF:.*]], label %[[JOIN:.*]]
@@ -100,7 +99,7 @@ define i9 @test_with_overflow(i1 %cond, i8 %a) {
; CHECK-NEXT: [[PHI:%.*]] = phi i16 [ undef, %[[ENTRY]] ], [ [[A_EXT]], %[[IF]] ]
; CHECK-NEXT: [[WO:%.*]] = call { i16, i1 } @llvm.uadd.with.overflow.i16(i16 [[PHI]], i16 1)
; CHECK-NEXT: [[ADD:%.*]] = extractvalue { i16, i1 } [[WO]], 0
-; CHECK-NEXT: [[TRUNC:%.*]] = trunc nuw i16 [[ADD]] to i9
+; CHECK-NEXT: [[TRUNC:%.*]] = trunc i16 [[ADD]] to i9
; CHECK-NEXT: ret i9 [[TRUNC]]
;
entry:
|
|
@llvm/pr-subscribers-function-specialization Author: Nikita Popov (nikic) ChangesWhen performing some range operation (e.g. and) on a constant range that includes undef, we currently just ignore the undef value, which is obviously incorrect. Instead, we can do one of two things:
This patch goes with the second approach -- I'd expect it to be a bit better overall, e.g. it allows preserving the fact that a zext of a range with undef isn't a full range. Fixes #93096. Full diff: https://github.com/llvm/llvm-project/pull/93163.diff 3 Files Affected:
diff --git a/llvm/lib/Transforms/Utils/SCCPSolver.cpp b/llvm/lib/Transforms/Utils/SCCPSolver.cpp
index ce40e8b31b767..4f36bac11e34b 100644
--- a/llvm/lib/Transforms/Utils/SCCPSolver.cpp
+++ b/llvm/lib/Transforms/Utils/SCCPSolver.cpp
@@ -43,7 +43,7 @@ static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() {
}
static ConstantRange getConstantRange(const ValueLatticeElement &LV, Type *Ty,
- bool UndefAllowed = true) {
+ bool UndefAllowed) {
assert(Ty->isIntOrIntVectorTy() && "Should be int or int vector");
if (LV.isConstantRange(UndefAllowed))
return LV.getConstantRange();
@@ -1297,7 +1297,8 @@ void SCCPInstVisitor::visitCastInst(CastInst &I) {
if (I.getDestTy()->isIntegerTy() && I.getSrcTy()->isIntOrIntVectorTy()) {
auto &LV = getValueState(&I);
- ConstantRange OpRange = getConstantRange(OpSt, I.getSrcTy());
+ ConstantRange OpRange =
+ getConstantRange(OpSt, I.getSrcTy(), /*UndefAllowed=*/false);
Type *DestTy = I.getDestTy();
// Vectors where all elements have the same known constant range are treated
@@ -1329,8 +1330,8 @@ void SCCPInstVisitor::handleExtractOfWithOverflow(ExtractValueInst &EVI,
return; // Wait to resolve.
Type *Ty = LHS->getType();
- ConstantRange LR = getConstantRange(L, Ty);
- ConstantRange RR = getConstantRange(R, Ty);
+ ConstantRange LR = getConstantRange(L, Ty, /*UndefAllowed=*/false);
+ ConstantRange RR = getConstantRange(R, Ty, /*UndefAllowed=*/false);
if (Idx == 0) {
ConstantRange Res = LR.binaryOp(WO->getBinaryOp(), RR);
mergeInValue(&EVI, ValueLatticeElement::getRange(Res));
@@ -1534,8 +1535,10 @@ void SCCPInstVisitor::visitBinaryOperator(Instruction &I) {
return markOverdefined(&I);
// Try to simplify to a constant range.
- ConstantRange A = getConstantRange(V1State, I.getType());
- ConstantRange B = getConstantRange(V2State, I.getType());
+ ConstantRange A =
+ getConstantRange(V1State, I.getType(), /*UndefAllowed=*/false);
+ ConstantRange B =
+ getConstantRange(V2State, I.getType(), /*UndefAllowed=*/false);
auto *BO = cast<BinaryOperator>(&I);
ConstantRange R = ConstantRange::getEmpty(I.getType()->getScalarSizeInBits());
@@ -1818,7 +1821,8 @@ void SCCPInstVisitor::handleCallResult(CallBase &CB) {
// Combine range info for the original value with the new range from the
// condition.
- auto CopyOfCR = getConstantRange(CopyOfVal, CopyOf->getType());
+ auto CopyOfCR = getConstantRange(CopyOfVal, CopyOf->getType(),
+ /*UndefAllowed=*/true);
auto NewCR = ImposedCR.intersectWith(CopyOfCR);
// If the existing information is != x, do not use the information from
// a chained predicate, as the != x information is more likely to be
@@ -1863,7 +1867,8 @@ void SCCPInstVisitor::handleCallResult(CallBase &CB) {
const ValueLatticeElement &State = getValueState(Op);
if (State.isUnknownOrUndef())
return;
- OpRanges.push_back(getConstantRange(State, Op->getType()));
+ OpRanges.push_back(
+ getConstantRange(State, Op->getType(), /*UndefAllowed=*/false));
}
ConstantRange Result =
diff --git a/llvm/test/Transforms/SCCP/ip-add-range-to-call.ll b/llvm/test/Transforms/SCCP/ip-add-range-to-call.ll
index c24c554102ddf..91efbcc4ee382 100644
--- a/llvm/test/Transforms/SCCP/ip-add-range-to-call.ll
+++ b/llvm/test/Transforms/SCCP/ip-add-range-to-call.ll
@@ -159,7 +159,7 @@ exit:
}
define i32 @caller5() {
-; CHECK-LABEL: define range(i32 200, 401) i32 @caller5() {
+; CHECK-LABEL: define i32 @caller5() {
; CHECK-NEXT: [[C1:%.*]] = call i32 @callee5(i32 10, i32 100)
; CHECK-NEXT: [[C2:%.*]] = call i32 @callee5(i32 20, i32 200)
; CHECK-NEXT: [[A:%.*]] = add i32 [[C1]], [[C2]]
diff --git a/llvm/test/Transforms/SCCP/range-with-undef.ll b/llvm/test/Transforms/SCCP/range-with-undef.ll
index 444b47df55697..9b8d415171140 100644
--- a/llvm/test/Transforms/SCCP/range-with-undef.ll
+++ b/llvm/test/Transforms/SCCP/range-with-undef.ll
@@ -2,7 +2,6 @@
; RUN: opt -S -passes=ipsccp < %s | FileCheck %s
; Make sure that constant ranges including undef are propagated correctly.
-; FIXME: All of the following are currently miscompiled.
define i8 @test_binop(i1 %cond, i8 %a) {
; CHECK-LABEL: define i8 @test_binop(
@@ -15,7 +14,7 @@ define i8 @test_binop(i1 %cond, i8 %a) {
; CHECK: [[JOIN]]:
; CHECK-NEXT: [[PHI:%.*]] = phi i16 [ undef, %[[ENTRY]] ], [ [[A_EXT]], %[[IF]] ]
; CHECK-NEXT: [[AND:%.*]] = and i16 [[PHI]], -1
-; CHECK-NEXT: [[TRUNC:%.*]] = trunc nuw i16 [[AND]] to i8
+; CHECK-NEXT: [[TRUNC:%.*]] = trunc i16 [[AND]] to i8
; CHECK-NEXT: ret i8 [[TRUNC]]
;
entry:
@@ -43,7 +42,7 @@ define i8 @test_cast(i1 %cond, i8 %a) {
; CHECK: [[JOIN]]:
; CHECK-NEXT: [[PHI:%.*]] = phi i16 [ undef, %[[ENTRY]] ], [ [[A_EXT]], %[[IF]] ]
; CHECK-NEXT: [[ZEXT:%.*]] = zext i16 [[PHI]] to i32
-; CHECK-NEXT: [[TRUNC:%.*]] = trunc nuw i32 [[ZEXT]] to i8
+; CHECK-NEXT: [[TRUNC:%.*]] = trunc i32 [[ZEXT]] to i8
; CHECK-NEXT: ret i8 [[TRUNC]]
;
entry:
@@ -61,7 +60,7 @@ join:
}
define i8 @test_intrin(i1 %cond, i8 %a) {
-; CHECK-LABEL: define range(i8 42, 0) i8 @test_intrin(
+; CHECK-LABEL: define i8 @test_intrin(
; CHECK-SAME: i1 [[COND:%.*]], i8 [[A:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*]]:
; CHECK-NEXT: br i1 [[COND]], label %[[IF:.*]], label %[[JOIN:.*]]
@@ -71,7 +70,7 @@ define i8 @test_intrin(i1 %cond, i8 %a) {
; CHECK: [[JOIN]]:
; CHECK-NEXT: [[PHI:%.*]] = phi i16 [ undef, %[[ENTRY]] ], [ [[A_EXT]], %[[IF]] ]
; CHECK-NEXT: [[UMAX:%.*]] = call i16 @llvm.umax.i16(i16 [[PHI]], i16 42)
-; CHECK-NEXT: [[TRUNC:%.*]] = trunc nuw i16 [[UMAX]] to i8
+; CHECK-NEXT: [[TRUNC:%.*]] = trunc i16 [[UMAX]] to i8
; CHECK-NEXT: ret i8 [[TRUNC]]
;
entry:
@@ -89,7 +88,7 @@ join:
}
define i9 @test_with_overflow(i1 %cond, i8 %a) {
-; CHECK-LABEL: define range(i9 1, -255) i9 @test_with_overflow(
+; CHECK-LABEL: define i9 @test_with_overflow(
; CHECK-SAME: i1 [[COND:%.*]], i8 [[A:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*]]:
; CHECK-NEXT: br i1 [[COND]], label %[[IF:.*]], label %[[JOIN:.*]]
@@ -100,7 +99,7 @@ define i9 @test_with_overflow(i1 %cond, i8 %a) {
; CHECK-NEXT: [[PHI:%.*]] = phi i16 [ undef, %[[ENTRY]] ], [ [[A_EXT]], %[[IF]] ]
; CHECK-NEXT: [[WO:%.*]] = call { i16, i1 } @llvm.uadd.with.overflow.i16(i16 [[PHI]], i16 1)
; CHECK-NEXT: [[ADD:%.*]] = extractvalue { i16, i1 } [[WO]], 0
-; CHECK-NEXT: [[TRUNC:%.*]] = trunc nuw i16 [[ADD]] to i9
+; CHECK-NEXT: [[TRUNC:%.*]] = trunc i16 [[ADD]] to i9
; CHECK-NEXT: ret i9 [[TRUNC]]
;
entry:
|
fhahn
reviewed
May 23, 2024
| ConstantRange B = getConstantRange(V2State, I.getType()); | ||
| ConstantRange A = | ||
| getConstantRange(V1State, I.getType(), /*UndefAllowed=*/false); | ||
| ConstantRange B = |
There was a problem hiding this comment.
I think the code that uses the constant range information from ValueLatticeElements already should be handling the range-may-include-undef case properly (i.e. not use it when refining instructions, only when using it to replace with a constant).
But it looks like the code below (mergeInValue call) simply drops the may-include-undef bit from the operands. If we propagate the bit to the result, things should be fine?
- mergeInValue(&I, ValueLatticeElement::getRange(R));
+ mergeInValue(&I, ValueLatticeElement::getRange(
+ R, V1State.isConstantRangeIncludingUndef() ||
+ V2State.isConstantRangeIncludingUndef()));
(IIRC the may-inlcude-undef bit was explicitly added to account for cases such as the AND test cases and I was pretty sure this was covered, but perhaps this was a place that got missed)
There was a problem hiding this comment.
If the above works out, it may make sense to make the MayIncludeUndef argument for ValueLatticeElement::getRange() required or flip the default?
There was a problem hiding this comment.
This is the alternative I mentioned in the patch description, and actually what I implemented first. But I think that this will usually produce a worse result than treating undef as a full range.
For example, if you have zext i8 ([a, b] | undef) to i16 we can either produce i16 [0, 255] (this patch) or i16 ([a, b] | undef), where the latter will be interprted as i16 full-range in places that don't allow undef. So we can either produce a range that is usable everywhere but is larger, or a range that is only usable in allow-undef contexts but is smaller. As a lot of folds in SCCP don't allow undef, I figured the first variant is preferable.
There was a problem hiding this comment.
For the record, this was my initial patch: nikic@94ed59e
There was a problem hiding this comment.
Worth noting that the "treat undef as full range" approach is also what we do in LVI.
There was a problem hiding this comment.
Ah yes that makes sense, thanks!
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
When performing some range operation (e.g. and) on a constant range that includes undef, we currently just ignore the undef value, which is obviously incorrect. Instead, we can do one of two things:
This patch goes with the second approach -- I'd expect it to be a bit better overall, e.g. it allows preserving the fact that a zext of a range with undef isn't a full range.
Fixes #93096.