[LVI] Support no constant range of cast value in getEdgeValueLocal. by andjo403 · Pull Request #157614 · llvm/llvm-project

andjo403 · 2025-09-09T05:20:00Z

proof: https://alive2.llvm.org/ce/z/8emkHY

…geValueLocal. (NFC)

llvmbot · 2025-09-09T05:20:31Z

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-llvm-analysis

Author: Andreas Jonson (andjo403)

Changes

proof: https://alive2.llvm.org/ce/z/8emkHY

Full diff: https://github.com/llvm/llvm-project/pull/157614.diff

2 Files Affected:

(modified) llvm/lib/Analysis/LazyValueInfo.cpp (+18)
(modified) llvm/test/Transforms/CorrelatedValuePropagation/range.ll (+144)

diff --git a/llvm/lib/Analysis/LazyValueInfo.cpp b/llvm/lib/Analysis/LazyValueInfo.cpp
index c7b0ca97a8e43..cbb9ed3b12934 100644
--- a/llvm/lib/Analysis/LazyValueInfo.cpp
+++ b/llvm/lib/Analysis/LazyValueInfo.cpp
@@ -1493,6 +1493,24 @@ LazyValueInfoImpl::getEdgeValueLocal(Value *Val, BasicBlock *BBFrom,
             //   br %Condition, label %then, label %else
             APInt ConditionVal(1, isTrueDest ? 1 : 0);
             Result = constantFoldUser(Usr, Condition, ConditionVal, DL);
+          } else if (isa<CastInst>(Usr) || isa<FreezeInst>(Usr)) {
+            ValueLatticeElement OpLatticeVal =
+                *getValueFromCondition(Usr->getOperand(0), Condition,
+                                       isTrueDest, /*UseBlockValue*/ false);
+
+            if (isa<FreezeInst>(Usr) || !OpLatticeVal.isConstantRange())
+              return OpLatticeVal;
+
+            const unsigned ResultBitWidth =
+                Usr->getType()->getScalarSizeInBits();
+            if (auto *Trunc = dyn_cast<TruncInst>(Usr))
+              return ValueLatticeElement::getRange(
+                  OpLatticeVal.getConstantRange().truncate(
+                      ResultBitWidth, Trunc->getNoWrapKind()));
+
+            return ValueLatticeElement::getRange(
+                OpLatticeVal.getConstantRange().castOp(
+                    cast<CastInst>(Usr)->getOpcode(), ResultBitWidth));
           } else {
             // If one of Val's operand has an inferred value, we may be able to
             // infer the value of Val.
diff --git a/llvm/test/Transforms/CorrelatedValuePropagation/range.ll b/llvm/test/Transforms/CorrelatedValuePropagation/range.ll
index 8d787fd5586ef..fd6722cfc1671 100644
--- a/llvm/test/Transforms/CorrelatedValuePropagation/range.ll
+++ b/llvm/test/Transforms/CorrelatedValuePropagation/range.ll
@@ -1235,6 +1235,150 @@ define i1 @neg_icmp_eq_range_call() {
   ret i1 %cmp
 }
 
+define i16 @return_range_for_edge_value_zext(i8 %a) {
+; CHECK-LABEL: define range(i16 0, 98) i16 @return_range_for_edge_value_zext(
+; CHECK-SAME: i8 [[A:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[B:%.*]] = zext i8 [[A]] to i16
+; CHECK-NEXT:    br label %[[DISPATCH:.*]]
+; CHECK:       [[DISPATCH]]:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i8 [[A]], 98
+; CHECK-NEXT:    br i1 [[CMP]], label %[[TARGET:.*]], label %[[DISPATCH]]
+; CHECK:       [[TARGET]]:
+; CHECK-NEXT:    ret i16 [[B]]
+;
+entry:
+  %b = zext i8 %a to i16
+  br label %dispatch
+
+dispatch:
+  %cmp = icmp ult i8 %a, 98
+  br i1 %cmp, label %target, label %dispatch
+
+target:
+  ret i16 %b
+}
+
+define i16 @return_range_for_edge_value_sext(i8 %a) {
+; CHECK-LABEL: define range(i16 0, 98) i16 @return_range_for_edge_value_sext(
+; CHECK-SAME: i8 [[A:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[B:%.*]] = sext i8 [[A]] to i16
+; CHECK-NEXT:    br label %[[DISPATCH:.*]]
+; CHECK:       [[DISPATCH]]:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i8 [[A]], 98
+; CHECK-NEXT:    br i1 [[CMP]], label %[[TARGET:.*]], label %[[DISPATCH]]
+; CHECK:       [[TARGET]]:
+; CHECK-NEXT:    ret i16 [[B]]
+;
+entry:
+  %b = sext i8 %a to i16
+  br label %dispatch
+
+dispatch:
+  %cmp = icmp ult i8 %a, 98
+  br i1 %cmp, label %target, label %dispatch
+
+target:
+  ret i16 %b
+}
+
+define i8 @return_range_for_edge_value_trunc(i16 %a) {
+; CHECK-LABEL: define range(i8 0, 98) i8 @return_range_for_edge_value_trunc(
+; CHECK-SAME: i16 [[A:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[B:%.*]] = trunc i16 [[A]] to i8
+; CHECK-NEXT:    br label %[[DISPATCH:.*]]
+; CHECK:       [[DISPATCH]]:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i16 [[A]], 98
+; CHECK-NEXT:    br i1 [[CMP]], label %[[TARGET:.*]], label %[[DISPATCH]]
+; CHECK:       [[TARGET]]:
+; CHECK-NEXT:    ret i8 [[B]]
+;
+entry:
+  %b = trunc i16 %a to i8
+  br label %dispatch
+
+dispatch:
+  %cmp = icmp ult i16 %a, 98
+  br i1 %cmp, label %target, label %dispatch
+
+target:
+  ret i8 %b
+}
+
+define i8 @neg_return_range_for_edge_value_trunc(i16 %a) {
+; CHECK-LABEL: define i8 @neg_return_range_for_edge_value_trunc(
+; CHECK-SAME: i16 [[A:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[B:%.*]] = trunc i16 [[A]] to i8
+; CHECK-NEXT:    br label %[[DISPATCH:.*]]
+; CHECK:       [[DISPATCH]]:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i16 [[A]], 200
+; CHECK-NEXT:    br i1 [[CMP]], label %[[TARGET:.*]], label %[[DISPATCH]]
+; CHECK:       [[TARGET]]:
+; CHECK-NEXT:    ret i8 [[B]]
+;
+entry:
+  %b = trunc i16 %a to i8
+  br label %dispatch
+
+dispatch:
+  %cmp = icmp ugt i16 %a, 200
+  br i1 %cmp, label %target, label %dispatch
+
+target:
+  ret i8 %b
+}
+
+define i8 @return_range_for_edge_value_trunc_nuw(i16 %a) {
+; CHECK-LABEL: define range(i8 -55, 0) i8 @return_range_for_edge_value_trunc_nuw(
+; CHECK-SAME: i16 [[A:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[B:%.*]] = trunc nuw i16 [[A]] to i8
+; CHECK-NEXT:    br label %[[DISPATCH:.*]]
+; CHECK:       [[DISPATCH]]:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i16 [[A]], 200
+; CHECK-NEXT:    br i1 [[CMP]], label %[[TARGET:.*]], label %[[DISPATCH]]
+; CHECK:       [[TARGET]]:
+; CHECK-NEXT:    ret i8 [[B]]
+;
+entry:
+  %b = trunc nuw i16 %a to i8
+  br label %dispatch
+
+dispatch:
+  %cmp = icmp ugt i16 %a, 200
+  br i1 %cmp, label %target, label %dispatch
+
+target:
+  ret i8 %b
+}
+
+define i8 @return_range_for_edge_value_freeze(i8 %a) {
+; CHECK-LABEL: define range(i8 0, 98) i8 @return_range_for_edge_value_freeze(
+; CHECK-SAME: i8 [[A:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[B:%.*]] = freeze i8 [[A]]
+; CHECK-NEXT:    br label %[[DISPATCH:.*]]
+; CHECK:       [[DISPATCH]]:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i8 [[A]], 98
+; CHECK-NEXT:    br i1 [[CMP]], label %[[TARGET:.*]], label %[[DISPATCH]]
+; CHECK:       [[TARGET]]:
+; CHECK-NEXT:    ret i8 [[B]]
+;
+entry:
+  %b = freeze i8 %a
+  br label %dispatch
+
+dispatch:
+  %cmp = icmp ult i8 %a, 98
+  br i1 %cmp, label %target, label %dispatch
+
+target:
+  ret i8 %b
+}
+
 declare i16 @llvm.ctlz.i16(i16, i1)
 declare i16 @llvm.cttz.i16(i16, i1)
 declare i16 @llvm.ctpop.i16(i16)

andjo403 · 2025-09-09T05:51:55Z

I needs to look in to why there is failures in dtcxzyw/llvm-opt-benchmark#2768

…lueLocal.

andjo403 · 2025-09-09T06:21:01Z

limited the cast to trunc, zext and sext as that is what I have tested

dtcxzyw · 2025-09-09T14:34:00Z

+  br label %dispatch
+
+dispatch:
+  %cmp = icmp ult i8 %a, 98


InstCombine canonicalizes all uses of %a to %b. Can you put freeze-related changes into a separate PR?

good catch there was no changes for freeze when I did a benchmark run I only added it as it was listed in isOperationFoldable. but is dropped now

dtcxzyw

Changes w/o freeze look good.

…tEdgeValueLocal.

nikic

LGTM

…e in getEdgeValueLocal.

Closes #161367. In #157614, we ignored cases where OpLatticeVal might be a constant or notconstant. Directly returning the result causes a type mismatch. I apologize for the oversight in the previous code review. This patch applies the cast op to constants. For notconstant value lattices, I'd leave it as a todo (it is similar to the constant case, except for trunc without nsw/nuw).

… (#161410) Closes llvm/llvm-project#161367. In llvm/llvm-project#157614, we ignored cases where OpLatticeVal might be a constant or notconstant. Directly returning the result causes a type mismatch. I apologize for the oversight in the previous code review. This patch applies the cast op to constants. For notconstant value lattices, I'd leave it as a todo (it is similar to the constant case, except for trunc without nsw/nuw).

Closes llvm#161367. In llvm#157614, we ignored cases where OpLatticeVal might be a constant or notconstant. Directly returning the result causes a type mismatch. I apologize for the oversight in the previous code review. This patch applies the cast op to constants. For notconstant value lattices, I'd leave it as a todo (it is similar to the constant case, except for trunc without nsw/nuw).

Closes llvm/llvm-project#161367. In llvm/llvm-project#157614, we ignored cases where OpLatticeVal might be a constant or notconstant. Directly returning the result causes a type mismatch. I apologize for the oversight in the previous code review. This patch applies the cast op to constants. For notconstant value lattices, I'd leave it as a todo (it is similar to the constant case, except for trunc without nsw/nuw).

[LVI] Add test of no constant range of cast and freeze value in getEd…

3b6e08d

…geValueLocal. (NFC)

andjo403 requested review from dtcxzyw and nikic September 9, 2025 05:20

llvmbot added llvm:analysis Includes value tracking, cost tables and constant folding llvm:transforms labels Sep 9, 2025

andjo403 mentioned this pull request Sep 9, 2025

Task submission dtcxzyw/llvm-opt-benchmark#1312

Open

zyw-bot mentioned this pull request Sep 9, 2025

pre-commit: PR157614 dtcxzyw/llvm-opt-benchmark#2768

Closed

andjo403 force-pushed the NonConstEdgeValueCastFreeze branch from 98aa2b2 to 3c4ac99 Compare September 9, 2025 06:01

[LVI] Support no constant range of cast and freeze value in getEdgeVa…

7cce3e7

…lueLocal.

andjo403 force-pushed the NonConstEdgeValueCastFreeze branch from 3c4ac99 to 7cce3e7 Compare September 9, 2025 06:20

zyw-bot mentioned this pull request Sep 9, 2025

pre-commit: PR157614 dtcxzyw/llvm-opt-benchmark#2769

Closed

dtcxzyw reviewed Sep 9, 2025

View reviewed changes

dtcxzyw mentioned this pull request Sep 9, 2025

Fuzz PR157614 dtcxzyw/llvm-fuzz-service#133

Closed

dtcxzyw approved these changes Sep 9, 2025

View reviewed changes

fixup! [LVI] Support no constant range of cast and freeze value in ge…

d75a408

…tEdgeValueLocal.

nikic approved these changes Sep 9, 2025

View reviewed changes

Comment thread llvm/test/Transforms/CorrelatedValuePropagation/range.ll Outdated

Comment thread llvm/test/Transforms/JumpThreading/branch-debug-info2.ll Outdated

andjo403 changed the title ~~[LVI] Support no constant range of cast and freeze value in getEdgeValueLocal.~~ [LVI] Support no constant range of cast value in getEdgeValueLocal. Sep 9, 2025

fixup! fixup! [LVI] Support no constant range of cast and freeze valu…

46b3556

…e in getEdgeValueLocal.

andjo403 merged commit 2701a22 into llvm:main Sep 10, 2025
9 checks passed

andjo403 deleted the NonConstEdgeValueCastFreeze branch September 10, 2025 16:19

dtcxzyw mentioned this pull request Sep 30, 2025

[LVI] Handle constant value lattice in getEdgeValueLocal #161410

Merged

aemerson mentioned this pull request Oct 16, 2025

[clang-reorder-fields] Check for flexible array member (#160262) #163697

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[LVI] Support no constant range of cast value in getEdgeValueLocal.#157614

[LVI] Support no constant range of cast value in getEdgeValueLocal.#157614
andjo403 merged 4 commits into
llvm:mainfrom
andjo403:NonConstEdgeValueCastFreeze

andjo403 commented Sep 9, 2025

Uh oh!

llvmbot commented Sep 9, 2025 •

edited

Loading

Uh oh!

andjo403 commented Sep 9, 2025 •

edited

Loading

Uh oh!

andjo403 commented Sep 9, 2025

Uh oh!

dtcxzyw Sep 9, 2025

Uh oh!

andjo403 Sep 9, 2025

Uh oh!

dtcxzyw left a comment

Uh oh!

nikic left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

andjo403 commented Sep 9, 2025

Uh oh!

llvmbot commented Sep 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

andjo403 commented Sep 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

andjo403 commented Sep 9, 2025

Uh oh!

dtcxzyw Sep 9, 2025

Choose a reason for hiding this comment

Uh oh!

andjo403 Sep 9, 2025

Choose a reason for hiding this comment

Uh oh!

dtcxzyw left a comment

Choose a reason for hiding this comment

Uh oh!

nikic left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

llvmbot commented Sep 9, 2025 •

edited

Loading

andjo403 commented Sep 9, 2025 •

edited

Loading