[SCEV] Collect guard info for ICMP NE w/o constants. #160500

fhahn · 2025-09-24T11:06:50Z

When collecting information from loop guards, use UMax(1, %b - %a) for ICMP NE %a, %b, if neither are constant.

This improves results in some cases, and will be even more useful together with

https://alive2.llvm.org/ce/z/YyBvoT

llvmbot · 2025-09-24T11:07:25Z

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-llvm-analysis

Author: Florian Hahn (fhahn)

Changes

When collecting information from loop guards, use UMax(1, %b - %a) for ICMP NE %a, %b, if neither are constant.

This improves results in some cases, and will be even more useful together with

https://alive2.llvm.org/ce/z/YyBvoT

Full diff: https://github.com/llvm/llvm-project/pull/160500.diff

4 Files Affected:

(modified) llvm/lib/Analysis/ScalarEvolution.cpp (+10)
(modified) llvm/test/Analysis/ScalarEvolution/ptrtoint.ll (+5-5)
(modified) llvm/test/Transforms/IndVarSimplify/pointer-loop-guards.ll (+1-1)
(modified) llvm/test/Transforms/LoopUnroll/scevunroll.ll (+1-2)

diff --git a/llvm/lib/Analysis/ScalarEvolution.cpp b/llvm/lib/Analysis/ScalarEvolution.cpp
index b08399b381f34..d9f14bbceeffd 100644
--- a/llvm/lib/Analysis/ScalarEvolution.cpp
+++ b/llvm/lib/Analysis/ScalarEvolution.cpp
@@ -15866,6 +15866,16 @@ void ScalarEvolution::LoopGuards::collectFromBlock(
           const SCEV *OneAlignedUp =
               DividesBy ? GetNextSCEVDividesByDivisor(One, DividesBy) : One;
           To = SE.getUMaxExpr(FromRewritten, OneAlignedUp);
+        } else {
+          if (LHS->getType()->isPointerTy()) {
+            LHS = SE.getLosslessPtrToIntExpr(LHS);
+            RHS = SE.getLosslessPtrToIntExpr(RHS);
+            if (isa<SCEVCouldNotCompute>(LHS) || isa<SCEVCouldNotCompute>(RHS))
+              break;
+          }
+          From = SE.getMinusSCEV(RHS, LHS);
+          FromRewritten = From;
+          To = SE.getUMaxExpr(FromRewritten, SE.getOne(From->getType()));
         }
         break;
       default:
diff --git a/llvm/test/Analysis/ScalarEvolution/ptrtoint.ll b/llvm/test/Analysis/ScalarEvolution/ptrtoint.ll
index e784d25385980..6db2f21352c49 100644
--- a/llvm/test/Analysis/ScalarEvolution/ptrtoint.ll
+++ b/llvm/test/Analysis/ScalarEvolution/ptrtoint.ll
@@ -382,7 +382,7 @@ define void @pr46786_c26_char(ptr %arg, ptr %arg1, ptr %arg2) {
 ; X64-NEXT:    %i9 = ptrtoint ptr %i7 to i64
 ; X64-NEXT:    --> {(ptrtoint ptr %arg to i64),+,1}<nuw><%bb6> U: full-set S: full-set Exits: (-1 + (ptrtoint ptr %arg1 to i64)) LoopDispositions: { %bb6: Computable }
 ; X64-NEXT:    %i10 = sub i64 %i9, %i4
-; X64-NEXT:    --> {0,+,1}<nuw><%bb6> U: full-set S: full-set Exits: (-1 + (-1 * (ptrtoint ptr %arg to i64)) + (ptrtoint ptr %arg1 to i64)) LoopDispositions: { %bb6: Computable }
+; X64-NEXT:    --> {0,+,1}<nuw><%bb6> U: [0,-1) S: [0,-1) Exits: (-1 + (-1 * (ptrtoint ptr %arg to i64)) + (ptrtoint ptr %arg1 to i64)) LoopDispositions: { %bb6: Computable }
 ; X64-NEXT:    %i11 = getelementptr inbounds i8, ptr %arg2, i64 %i10
 ; X64-NEXT:    --> {%arg2,+,1}<nw><%bb6> U: full-set S: full-set Exits: (-1 + (-1 * (ptrtoint ptr %arg to i64)) + (ptrtoint ptr %arg1 to i64) + %arg2) LoopDispositions: { %bb6: Computable }
 ; X64-NEXT:    %i12 = load i8, ptr %i11, align 1
@@ -393,7 +393,7 @@ define void @pr46786_c26_char(ptr %arg, ptr %arg1, ptr %arg2) {
 ; X64-NEXT:    --> {(1 + %arg),+,1}<nuw><%bb6> U: full-set S: full-set Exits: ((-1 * (ptrtoint ptr %arg to i64)) + (ptrtoint ptr %arg1 to i64) + %arg) LoopDispositions: { %bb6: Computable }
 ; X64-NEXT:  Determining loop execution counts for: @pr46786_c26_char
 ; X64-NEXT:  Loop %bb6: backedge-taken count is (-1 + (-1 * (ptrtoint ptr %arg to i64)) + (ptrtoint ptr %arg1 to i64))
-; X64-NEXT:  Loop %bb6: constant max backedge-taken count is i64 -1
+; X64-NEXT:  Loop %bb6: constant max backedge-taken count is i64 -2
 ; X64-NEXT:  Loop %bb6: symbolic max backedge-taken count is (-1 + (-1 * (ptrtoint ptr %arg to i64)) + (ptrtoint ptr %arg1 to i64))
 ; X64-NEXT:  Loop %bb6: Trip multiple is 1
 ;
@@ -406,9 +406,9 @@ define void @pr46786_c26_char(ptr %arg, ptr %arg1, ptr %arg2) {
 ; X32-NEXT:    %i8 = load i8, ptr %i7, align 1
 ; X32-NEXT:    --> %i8 U: full-set S: full-set Exits: <<Unknown>> LoopDispositions: { %bb6: Variant }
 ; X32-NEXT:    %i9 = ptrtoint ptr %i7 to i64
-; X32-NEXT:    --> {(zext i32 (ptrtoint ptr %arg to i32) to i64),+,1}<nuw><%bb6> U: [0,8589934591) S: [0,8589934591) Exits: ((zext i32 (-1 + (-1 * (ptrtoint ptr %arg to i32)) + (ptrtoint ptr %arg1 to i32)) to i64) + (zext i32 (ptrtoint ptr %arg to i32) to i64)) LoopDispositions: { %bb6: Computable }
+; X32-NEXT:    --> {(zext i32 (ptrtoint ptr %arg to i32) to i64),+,1}<nuw><%bb6> U: [0,8589934590) S: [0,8589934590) Exits: ((zext i32 (-1 + (-1 * (ptrtoint ptr %arg to i32)) + (ptrtoint ptr %arg1 to i32)) to i64) + (zext i32 (ptrtoint ptr %arg to i32) to i64)) LoopDispositions: { %bb6: Computable }
 ; X32-NEXT:    %i10 = sub i64 %i9, %i4
-; X32-NEXT:    --> {0,+,1}<nuw><%bb6> U: [0,4294967296) S: [0,4294967296) Exits: (zext i32 (-1 + (-1 * (ptrtoint ptr %arg to i32)) + (ptrtoint ptr %arg1 to i32)) to i64) LoopDispositions: { %bb6: Computable }
+; X32-NEXT:    --> {0,+,1}<nuw><%bb6> U: [0,4294967295) S: [0,4294967295) Exits: (zext i32 (-1 + (-1 * (ptrtoint ptr %arg to i32)) + (ptrtoint ptr %arg1 to i32)) to i64) LoopDispositions: { %bb6: Computable }
 ; X32-NEXT:    %i11 = getelementptr inbounds i8, ptr %arg2, i64 %i10
 ; X32-NEXT:    --> {%arg2,+,1}<%bb6> U: full-set S: full-set Exits: (-1 + (-1 * (ptrtoint ptr %arg to i32)) + (ptrtoint ptr %arg1 to i32) + %arg2) LoopDispositions: { %bb6: Computable }
 ; X32-NEXT:    %i12 = load i8, ptr %i11, align 1
@@ -419,7 +419,7 @@ define void @pr46786_c26_char(ptr %arg, ptr %arg1, ptr %arg2) {
 ; X32-NEXT:    --> {(1 + %arg),+,1}<nuw><%bb6> U: full-set S: full-set Exits: ((-1 * (ptrtoint ptr %arg to i32)) + (ptrtoint ptr %arg1 to i32) + %arg) LoopDispositions: { %bb6: Computable }
 ; X32-NEXT:  Determining loop execution counts for: @pr46786_c26_char
 ; X32-NEXT:  Loop %bb6: backedge-taken count is (-1 + (-1 * (ptrtoint ptr %arg to i32)) + (ptrtoint ptr %arg1 to i32))
-; X32-NEXT:  Loop %bb6: constant max backedge-taken count is i32 -1
+; X32-NEXT:  Loop %bb6: constant max backedge-taken count is i32 -2
 ; X32-NEXT:  Loop %bb6: symbolic max backedge-taken count is (-1 + (-1 * (ptrtoint ptr %arg to i32)) + (ptrtoint ptr %arg1 to i32))
 ; X32-NEXT:  Loop %bb6: Trip multiple is 1
 ;
diff --git a/llvm/test/Transforms/IndVarSimplify/pointer-loop-guards.ll b/llvm/test/Transforms/IndVarSimplify/pointer-loop-guards.ll
index 89b132e981899..9371fe24be05d 100644
--- a/llvm/test/Transforms/IndVarSimplify/pointer-loop-guards.ll
+++ b/llvm/test/Transforms/IndVarSimplify/pointer-loop-guards.ll
@@ -18,7 +18,7 @@ define i64 @test_ptr_compare_guard(ptr %start, ptr %end) {
 ; CHECK-NEXT:    br i1 [[C_1]], label %[[LOOP_LATCH]], label %[[EXIT_LOOPEXIT:.*]]
 ; CHECK:       [[LOOP_LATCH]]:
 ; CHECK-NEXT:    [[PTR_IV_NEXT]] = getelementptr i8, ptr [[PTR_IV]], i64 1
-; CHECK-NEXT:    [[I64_IV_NEXT]] = add i64 [[I64_IV]], 1
+; CHECK-NEXT:    [[I64_IV_NEXT]] = add nuw i64 [[I64_IV]], 1
 ; CHECK-NEXT:    [[C_2:%.*]] = icmp eq ptr [[PTR_IV_NEXT]], [[END]]
 ; CHECK-NEXT:    br i1 [[C_2]], label %[[EXIT_LOOPEXIT]], label %[[LOOP_HEADER]]
 ; CHECK:       [[EXIT_LOOPEXIT]]:
diff --git a/llvm/test/Transforms/LoopUnroll/scevunroll.ll b/llvm/test/Transforms/LoopUnroll/scevunroll.ll
index fa55eab062198..bc63f79296a30 100644
--- a/llvm/test/Transforms/LoopUnroll/scevunroll.ll
+++ b/llvm/test/Transforms/LoopUnroll/scevunroll.ll
@@ -465,8 +465,7 @@ define void @peel_int_eq_condition(i32 %start) {
 ; CHECK-NEXT:    br label [[LOOP:%.*]]
 ; CHECK:       loop:
 ; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[IV_NEXT_PEEL]], [[ENTRY_PEEL_NEWPH]] ], [ [[IV_NEXT:%.*]], [[LOOP_LATCH:%.*]] ]
-; CHECK-NEXT:    [[C_0:%.*]] = icmp eq i32 [[IV]], [[START]]
-; CHECK-NEXT:    br i1 [[C_0]], label [[IF_THEN:%.*]], label [[LOOP_LATCH]]
+; CHECK-NEXT:    br i1 false, label [[IF_THEN:%.*]], label [[LOOP_LATCH]]
 ; CHECK:       if.then:
 ; CHECK-NEXT:    call void @fn(i32 [[IV]])
 ; CHECK-NEXT:    br label [[LOOP_LATCH]]

github-actions · 2025-09-24T20:24:41Z

✅ With the latest revision this PR passed the C/C++ code formatter.

fhahn · 2025-09-25T17:31:03Z

IR corpus results: https://github.com/dtcxzyw/llvm-opt-benchmark/pull/2859/files

fhahn · 2025-10-01T16:36:20Z

ping :)

nikic · 2025-10-07T15:14:32Z

llvm/lib/Analysis/ScalarEvolution.cpp

+            if (isa<SCEVCouldNotCompute>(LHS) || isa<SCEVCouldNotCompute>(RHS))
+              break;
+          }
+          From = SE.getMinusSCEV(RHS, LHS);


Hm, wouldn't we have to add this for both RHS - LHS and LHS - RHS? Otherwise the guard will or will not work depending on whether you write RHS != LHS or LHS != RHS.

Yep, added support for both variants and added a test with swapped operands, thanks

Additional test coverage for #160500.

When collecting information from loop guards, use UMax(1, %b - %a) for ICMP NE %a, %b, if neither are constant. This improves results in some cases, and will be even more useful together with * llvm#160012 * llvm#159942 https://alive2.llvm.org/ce/z/YyBvoT

fhahn

ping :)

Additional test coverage for llvm/llvm-project#160500.

nikic · 2025-10-13T10:58:01Z

llvm/lib/Analysis/ScalarEvolution.cpp

+          auto AddSubRewrite = [&](const SCEV *A, const SCEV *B) {
+            const SCEV *Sub = SE.getMinusSCEV(A, B);
+            AddRewrite(Sub, Sub,
+                       SE.getUMaxExpr(Sub, SE.getOne(From->getType())));


Should use GetNextSCEVDividesByDivisor here, similar to the code path above?

I think we would need to get the next SCEV based on both the multiple of LHS & RHS, which would be a slightly larger change, perhaps better done as follow-up?

Put up as follow-up: #163250

(can also fold in here if you prefer it in one go)

Follow-up to llvm#160500 to preserve divisibiltiy info when creating the UMax.

fhahn requested review from efriedma-quic, nikic and preames September 24, 2025 11:06

llvmbot added llvm:analysis Includes value tracking, cost tables and constant folding llvm:transforms labels Sep 24, 2025

fhahn mentioned this pull request Sep 24, 2025

Task submission dtcxzyw/llvm-opt-benchmark#1312

Open

This was referenced Sep 24, 2025

pre-commit: PR160500 dtcxzyw/llvm-opt-benchmark#2853

Open

pre-commit: PR160500 dtcxzyw/llvm-opt-benchmark#2855

Closed

pre-commit: PR160500 dtcxzyw/llvm-opt-benchmark#2856

Closed

pre-commit: PR160500 dtcxzyw/llvm-opt-benchmark#2857

Closed

fhahn force-pushed the scev-guards-icmp-ne-diff branch from 2492c48 to 40e2e7b Compare September 24, 2025 20:18

zyw-bot mentioned this pull request Sep 24, 2025

pre-commit: PR160500 dtcxzyw/llvm-opt-benchmark#2859

Closed

fhahn force-pushed the scev-guards-icmp-ne-diff branch from 40e2e7b to b7b6455 Compare September 25, 2025 17:23

fhahn force-pushed the scev-guards-icmp-ne-diff branch from b7b6455 to e89632a Compare October 1, 2025 16:37

nikic reviewed Oct 7, 2025

View reviewed changes

fhahn force-pushed the scev-guards-icmp-ne-diff branch from e89632a to 9947da6 Compare October 8, 2025 12:28

fhahn added a commit that referenced this pull request Oct 13, 2025

[SCEV] Add test with ptrtoint guards and their order swapped.

735ee5c

Additional test coverage for #160500.

fhahn added 2 commits October 13, 2025 11:28

!fixup add both

7ede0fa

fhahn force-pushed the scev-guards-icmp-ne-diff branch from 9947da6 to 7ede0fa Compare October 13, 2025 10:29

fhahn commented Oct 13, 2025

View reviewed changes

llvm-sync bot pushed a commit to arm/arm-toolchain that referenced this pull request Oct 13, 2025

Automerge: [SCEV] Add test with ptrtoint guards and their order swapped.

2269ea7

Additional test coverage for llvm/llvm-project#160500.

nikic reviewed Oct 13, 2025

View reviewed changes

fhahn added a commit to fhahn/llvm-project that referenced this pull request Oct 13, 2025

[SCEV] Preserve divisor info when adding guard info for ICMP_NE via Sub.

b0ccf0c

Follow-up to llvm#160500 to preserve divisibiltiy info when creating the UMax.

fhahn mentioned this pull request Oct 13, 2025

[SCEV] Preserve divisor info when adding guard info for ICMP_NE via Sub. #163250

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[SCEV] Collect guard info for ICMP NE w/o constants. #160500

[SCEV] Collect guard info for ICMP NE w/o constants. #160500

fhahn commented Sep 24, 2025

Uh oh!

llvmbot commented Sep 24, 2025 •

edited

Loading

Uh oh!

github-actions bot commented Sep 24, 2025 •

edited

Loading

Uh oh!

fhahn commented Sep 25, 2025

Uh oh!

fhahn commented Oct 1, 2025

Uh oh!

nikic Oct 7, 2025

Uh oh!

fhahn Oct 8, 2025

Uh oh!

fhahn left a comment

Uh oh!

nikic Oct 13, 2025

Uh oh!

fhahn Oct 13, 2025

Uh oh!

fhahn Oct 13, 2025

Uh oh!

fhahn Oct 13, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

[SCEV] Collect guard info for ICMP NE w/o constants. #160500

Are you sure you want to change the base?

[SCEV] Collect guard info for ICMP NE w/o constants. #160500

Conversation

fhahn commented Sep 24, 2025

Uh oh!

llvmbot commented Sep 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Sep 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

fhahn commented Sep 25, 2025

Uh oh!

fhahn commented Oct 1, 2025

Uh oh!

nikic Oct 7, 2025

Choose a reason for hiding this comment

Uh oh!

fhahn Oct 8, 2025

Choose a reason for hiding this comment

Uh oh!

fhahn left a comment

Choose a reason for hiding this comment

Uh oh!

nikic Oct 13, 2025

Choose a reason for hiding this comment

Uh oh!

fhahn Oct 13, 2025

Choose a reason for hiding this comment

Uh oh!

fhahn Oct 13, 2025

Choose a reason for hiding this comment

Uh oh!

fhahn Oct 13, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

llvmbot commented Sep 24, 2025 •

edited

Loading

github-actions bot commented Sep 24, 2025 •

edited

Loading