[Mlir-commits] [mlir] [MLIR][LLVM] Fix non-deterministic alloca order in SROA under LLVM_REVERSE_ITERATION (PR #192087)

[Mehdi Amini]

https://github.com/joker-eph updated https://github.com/llvm/llvm-project/pull/192087

>From c28268546488889e22e77812a4afa581cf1fa8df Mon Sep 17 00:00:00 2001
From: Mehdi Amini <joker.eph at gmail.com>
Date: Tue, 14 Apr 2026 09:47:31 -0700
Subject: [PATCH] [MLIR][LLVM] Fix non-deterministic alloca order in SROA under
 LLVM_REVERSE_ITERATION

AllocaOp::destructure iterated over usedIndices (SmallPtrSet) whose
order depends on pointer values, causing allocas for destructured
subslots to be emitted in a non-deterministic order when
LLVM_REVERSE_ITERATION is enabled. Sort indices ascending by integer
value before creating allocas to guarantee a stable output order.
Update four test cases in sroa-intrinsics.mlir whose CHECK patterns
relied on the old non-deterministic ordering.

Assisted-by: Claude Code
Co-Authored-By: Claude Sonnet 4.6 <noreply at anthropic.com>
---
 mlir/lib/Dialect/LLVMIR/IR/LLVMMemorySlot.cpp | 8 +++++++-
 mlir/test/Dialect/LLVMIR/sroa-intrinsics.mlir | 8 ++++----
 2 files changed, 11 insertions(+), 5 deletions(-)

diff --git a/mlir/lib/Dialect/LLVMIR/IR/LLVMMemorySlot.cpp b/mlir/lib/Dialect/LLVMIR/IR/LLVMMemorySlot.cpp
index 8f3c0dac026a7..66d6a592a93df 100644
--- a/mlir/lib/Dialect/LLVMIR/IR/LLVMMemorySlot.cpp
+++ b/mlir/lib/Dialect/LLVMIR/IR/LLVMMemorySlot.cpp
@@ -87,7 +87,13 @@ DenseMap<Attribute, MemorySlot> LLVM::AllocaOp::destructure(
 
   auto destructurableType = cast<DestructurableTypeInterface>(getElemType());
   DenseMap<Attribute, MemorySlot> slotMap;
-  for (Attribute index : usedIndices) {
+  // Iterate subelements in their original type order to produce allocas in a
+  // deterministic, readable order (matching appearance in the source type).
+  Type i32 = IntegerType::get(getContext(), 32);
+  for (size_t i = 0; i < slot.subelementTypes.size(); i++) {
+    Attribute index = IntegerAttr::get(i32, i);
+    if (!usedIndices.contains(index))
+      continue;
     Type elemType = destructurableType.getTypeAtIndex(index);
     assert(elemType && "used index must exist");
     auto subAlloca = LLVM::AllocaOp::create(
diff --git a/mlir/test/Dialect/LLVMIR/sroa-intrinsics.mlir b/mlir/test/Dialect/LLVMIR/sroa-intrinsics.mlir
index 6dc8a97884ee1..3253077a76077 100644
--- a/mlir/test/Dialect/LLVMIR/sroa-intrinsics.mlir
+++ b/mlir/test/Dialect/LLVMIR/sroa-intrinsics.mlir
@@ -277,8 +277,8 @@ llvm.func @invalid_indirect_memset_inline() -> i32 {
 // CHECK-LABEL: llvm.func @memset_double_use
 llvm.func @memset_double_use() -> i32 {
   // CHECK: %[[ALLOCA_LEN:.*]] = llvm.mlir.constant(1 : i32) : i32
-  // CHECK: %[[ALLOCA_FLOAT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x f32
   // CHECK: %[[ALLOCA_INT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x i32
+  // CHECK: %[[ALLOCA_FLOAT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x f32
   // CHECK: %[[MEMSET_VALUE:.*]] = llvm.mlir.constant(42 : i8) : i8
   %0 = llvm.mlir.constant(1 : i32) : i32
   %1 = llvm.alloca %0 x !llvm.struct<"foo", (i32, f32)> {alignment = 8 : i64} : (i32) -> !llvm.ptr
@@ -309,8 +309,8 @@ llvm.func @memset_double_use() -> i32 {
 // CHECK-LABEL: llvm.func @memset_inline_double_use
 llvm.func @memset_inline_double_use() -> i32 {
   // CHECK: %[[ALLOCA_LEN:.*]] = llvm.mlir.constant(1 : i32) : i32
-  // CHECK: %[[ALLOCA_FLOAT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x f32
   // CHECK: %[[ALLOCA_INT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x i32
+  // CHECK: %[[ALLOCA_FLOAT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x f32
   // CHECK: %[[MEMSET_VALUE:.*]] = llvm.mlir.constant(42 : i8) : i8
   %0 = llvm.mlir.constant(1 : i32) : i32
   %1 = llvm.alloca %0 x !llvm.struct<"foo", (i32, f32)> {alignment = 8 : i64} : (i32) -> !llvm.ptr
@@ -398,8 +398,8 @@ llvm.func @memset_inline_considers_alignment() -> i32 {
 // CHECK-LABEL: llvm.func @memset_considers_packing
 llvm.func @memset_considers_packing() -> i32 {
   // CHECK: %[[ALLOCA_LEN:.*]] = llvm.mlir.constant(1 : i32) : i32
-  // CHECK: %[[ALLOCA_FLOAT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x f32
   // CHECK: %[[ALLOCA_INT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x i32
+  // CHECK: %[[ALLOCA_FLOAT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x f32
   // CHECK: %[[MEMSET_VALUE:.*]] = llvm.mlir.constant(42 : i8) : i8
   %0 = llvm.mlir.constant(1 : i32) : i32
   %1 = llvm.alloca %0 x !llvm.struct<"foo", packed (i8, i32, f32)> {alignment = 8 : i64} : (i32) -> !llvm.ptr
@@ -430,8 +430,8 @@ llvm.func @memset_considers_packing() -> i32 {
 // CHECK-LABEL: llvm.func @memset_inline_considers_packing
 llvm.func @memset_inline_considers_packing() -> i32 {
   // CHECK: %[[ALLOCA_LEN:.*]] = llvm.mlir.constant(1 : i32) : i32
-  // CHECK: %[[ALLOCA_FLOAT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x f32
   // CHECK: %[[ALLOCA_INT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x i32
+  // CHECK: %[[ALLOCA_FLOAT:.*]] = llvm.alloca %[[ALLOCA_LEN]] x f32
   // CHECK: %[[MEMSET_VALUE:.*]] = llvm.mlir.constant(42 : i8) : i8
   %0 = llvm.mlir.constant(1 : i32) : i32
   %1 = llvm.alloca %0 x !llvm.struct<"foo", packed (i8, i32, f32)> {alignment = 8 : i64} : (i32) -> !llvm.ptr