[llvm] [AggressiveInstCombine] Ignore debug instructions when load combining (PR #70200)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: None (mikaelholmen)

<details>
<summary>Changes</summary>

We previously included debug instructions when counting instructions when
looking for loads to combine. This meant that the presence of debug
instructions could affect optimization, as shown in the updated testcase.

This fixes #<!-- -->69925.


---
Full diff: https://github.com/llvm/llvm-project/pull/70200.diff


2 Files Affected:

- (modified) llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp (+10-1) 
- (added) llvm/test/Transforms/AggressiveInstCombine/AArch64/combine_ignore_debug.ll (+51) 


``````````diff
diff --git a/llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp b/llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp
index a55d01645f10eb8..72f55e237ca9151 100644
--- a/llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp
+++ b/llvm/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp
@@ -701,12 +701,21 @@ static bool foldLoadsRecursive(Value *V, LoadOps &LOps, const DataLayout &DL,
       Loc = Loc.getWithNewSize(LOps.LoadSize);
   } else
     Loc = MemoryLocation::get(End);
+
+  // Ignore debug info (and other "AssumeLike" intrinsics) so that's not counted
+  // against MaxInstrsToScan. Otherwise debug info could affect codegen.
+  auto IsAssumeLikeIntr = [](const Instruction &I) {
+    if (auto *II = dyn_cast<IntrinsicInst>(&I))
+      return II->isAssumeLikeIntrinsic();
+    return false;
+  };
   unsigned NumScanned = 0;
   for (Instruction &Inst :
        make_range(Start->getIterator(), End->getIterator())) {
     if (Inst.mayWriteToMemory() && isModSet(AA.getModRefInfo(&Inst, Loc)))
       return false;
-    if (++NumScanned > MaxInstrsToScan)
+
+    if (!IsAssumeLikeIntr(Inst) && ++NumScanned > MaxInstrsToScan)
       return false;
   }
 
diff --git a/llvm/test/Transforms/AggressiveInstCombine/AArch64/combine_ignore_debug.ll b/llvm/test/Transforms/AggressiveInstCombine/AArch64/combine_ignore_debug.ll
new file mode 100644
index 000000000000000..68455a1f9074ecb
--- /dev/null
+++ b/llvm/test/Transforms/AggressiveInstCombine/AArch64/combine_ignore_debug.ll
@@ -0,0 +1,51 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 3
+; RUN: opt -mtriple aarch64 -aggressive-instcombine-max-scan-instrs=1 -passes="aggressive-instcombine" -S < %s | FileCheck %s -check-prefix DBG
+; RUN: opt -strip-debug -mtriple aarch64 -aggressive-instcombine-max-scan-instrs=1 -passes="aggressive-instcombine" -S < %s | FileCheck %s -check-prefix NODBG
+
+; The DBG and NODBG cases should be the same. I.e. we should optimize the DBG
+; case too even if there is a dbg.value.
+
+target datalayout = "E"
+
+%s = type { i16, i16 }
+
+ at e = global %s zeroinitializer, align 1
+ at l = global %s zeroinitializer, align 1
+
+define void @test() {
+; DBG-LABEL: define void @test() {
+; DBG-NEXT:  entry:
+; DBG-NEXT:    [[L1:%.*]] = load i32, ptr @e, align 1
+; DBG-NEXT:    call void @llvm.dbg.value(metadata i32 undef, metadata [[META3:![0-9]+]], metadata !DIExpression()), !dbg [[DBG5:![0-9]+]]
+; DBG-NEXT:    store i32 [[L1]], ptr @l, align 1
+; DBG-NEXT:    ret void
+;
+; NODBG-LABEL: define void @test() {
+; NODBG-NEXT:  entry:
+; NODBG-NEXT:    [[L1:%.*]] = load i32, ptr @e, align 1
+; NODBG-NEXT:    store i32 [[L1]], ptr @l, align 1
+; NODBG-NEXT:    ret void
+;
+entry:
+  %l1 = load i16, ptr @e, align 1
+  call void @llvm.dbg.value(metadata i32 undef, metadata !3, metadata !DIExpression()), !dbg !5
+  %l2 = load i16, ptr getelementptr inbounds (%s, ptr @e, i16 0, i32 1), align 1
+  %e2 = zext i16 %l2 to i32
+  %e1 = zext i16 %l1 to i32
+  %s1 = shl nuw i32 %e1, 16
+  %o1 = or i32 %s1, %e2
+  store i32 %o1, ptr @l, align 1
+  ret void
+}
+
+declare void @llvm.dbg.value(metadata, metadata, metadata)
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!2}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C11, file: !1)
+!1 = !DIFile(filename: "foo.c", directory: "/")
+!2 = !{i32 2, !"Debug Info Version", i32 3}
+!3 = !DILocalVariable(scope: !4)
+!4 = distinct !DISubprogram(unit: !0)
+!5 = !DILocation(scope: !4)

``````````

</details>


https://github.com/llvm/llvm-project/pull/70200