[DSE] Use alias analysis to identify no-op stores

[AlexMaclean]

No description provided.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Alex MacLean (AlexMaclean)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/122610.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp (+3-1)
• (modified) llvm/test/Transforms/DeadStoreElimination/noop-stores.ll (+11)

diff --git a/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp b/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp
index cae5b9c41a37f1..908b9110bd595d 100644
--- a/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp
+++ b/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp
@@ -2116,7 +2116,9 @@ struct DSEState {
       return true;
 
     if (auto *LoadI = dyn_cast<LoadInst>(Store->getOperand(0))) {
-      if (LoadI->getPointerOperand() == Store->getOperand(1)) {
+      if (LoadI->getPointerOperand() == Store->getOperand(1) ||
+          AA.isMustAlias(MemoryLocation::get(LoadI),
+                         MemoryLocation::get(Store))) {
         // Get the defining access for the load.
         auto *LoadAccess = MSSA.getMemoryAccess(LoadI)->getDefiningAccess();
         // Fast path: the defining accesses are the same.
diff --git a/llvm/test/Transforms/DeadStoreElimination/noop-stores.ll b/llvm/test/Transforms/DeadStoreElimination/noop-stores.ll
index 9fc20d76da5eb4..5bedc1ebe8281e 100644
--- a/llvm/test/Transforms/DeadStoreElimination/noop-stores.ll
+++ b/llvm/test/Transforms/DeadStoreElimination/noop-stores.ll
@@ -1152,3 +1152,14 @@ if.else:
 end:
   ret void
 }
+
+define i32 @test_use_alias_analysis(ptr %Q) {
+; CHECK-LABEL: @test_use_alias_analysis(
+; CHECK-NEXT:    [[A:%.*]] = load i32, ptr [[Q:%.*]], align 4
+; CHECK-NEXT:    ret i32 [[A]]
+;
+  %a = load i32, ptr %Q
+  %Q2 = addrspacecast ptr %Q to ptr addrspace(5)
+  store i32 %a, ptr addrspace(5) %Q2
+  ret i32 %a
+}

[nikic]

Compile-time: http://llvm-compile-time-tracker.com/compare.php?from=bfe93aedcc7d393c2697e66d6569baffb701ba6f&to=d41dfc385f0a808507612043cf76924d721aa325&stat=instructions:u

[fhahn]

Compile-time: http://llvm-compile-time-tracker.com/compare.php?from=bfe93aedcc7d393c2697e66d6569baffb701ba6f&to=d41dfc385f0a808507612043cf76924d721aa325&stat=instructions:u

Hmm seems quite high for no changes. @AlexMaclean do you have any data on how many additional stores can be removed with this change on larger workloads?

[justinfargnoli]

Confirming my understanding of the problem: checking that the address spaces are not equal is insufficient because address spaces may not be disjoint, correct?