[mlir][vector] Propagate alignment from vector to llvm dialects. #153482
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
amd-eochoalo
force-pushed
the
eochoa/2025-08-13/alignment-flow
branch
from
88979e9 to
514cde7
Compare
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
amd-eochoalo
force-pushed
the
eochoa/2025-08-13/alignment-flow
branch
from
514cde7 to
3b48a3a
Compare
amd-eochoalo
force-pushed
the
eochoa/2025-08-13/alignment-flow
branch
from
3b48a3a to
69c6fdf
Compare
amd-eochoalo
requested review from
banach-space,
dcaballe and
nicolasvasilache
as code owners
|
@llvm/pr-subscribers-mlir Author: Erick Ochoa Lopez (amd-eochoalo) ChangesAllows alignment to be propagated correctly from vector to LLVM dialect operations. Full diff: https://github.com/llvm/llvm-project/pull/153482.diff 2 Files Affected:
diff --git a/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp b/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp
index afc3d1b12ac0d..7d29750ddcf39 100644
--- a/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp
+++ b/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp
@@ -299,8 +299,9 @@ class VectorGatherOpConversion
}
// Resolve alignment.
- unsigned align;
- if (failed(getVectorToLLVMAlignment(*this->getTypeConverter(), vType,
+ unsigned align = gather.getAlignment().value_or(0);
+ if (!align &&
+ failed(getVectorToLLVMAlignment(*this->getTypeConverter(), vType,
memRefType, align, useVectorAlignment)))
return rewriter.notifyMatchFailure(gather, "could not resolve alignment");
@@ -354,8 +355,9 @@ class VectorScatterOpConversion
}
// Resolve alignment.
- unsigned align;
- if (failed(getVectorToLLVMAlignment(*this->getTypeConverter(), vType,
+ unsigned align = scatter.getAlignment().value_or(0);
+ if (!align &&
+ failed(getVectorToLLVMAlignment(*this->getTypeConverter(), vType,
memRefType, align, useVectorAlignment)))
return rewriter.notifyMatchFailure(scatter,
"could not resolve alignment");
@@ -399,8 +401,14 @@ class VectorExpandLoadOpConversion
Value ptr = getStridedElementPtr(rewriter, loc, memRefType,
adaptor.getBase(), adaptor.getIndices());
+ // From:
+ // https://llvm.org/docs/LangRef.html#llvm-masked-expandload-intrinsics
+ // The pointer alignment defaults to 1.
+ uint64_t alignment = expand.getAlignment().value_or(1);
+
rewriter.replaceOpWithNewOp<LLVM::masked_expandload>(
- expand, vtype, ptr, adaptor.getMask(), adaptor.getPassThru());
+ expand, vtype, ptr, adaptor.getMask(), adaptor.getPassThru(),
+ alignment);
return success();
}
};
@@ -421,8 +429,13 @@ class VectorCompressStoreOpConversion
Value ptr = getStridedElementPtr(rewriter, loc, memRefType,
adaptor.getBase(), adaptor.getIndices());
+ // From:
+ // https://llvm.org/docs/LangRef.html#llvm-masked-compressstore-intrinsics
+ // The pointer alignment defaults to 1.
+ uint64_t alignment = compress.getAlignment().value_or(1);
+
rewriter.replaceOpWithNewOp<LLVM::masked_compressstore>(
- compress, adaptor.getValueToStore(), ptr, adaptor.getMask());
+ compress, adaptor.getValueToStore(), ptr, adaptor.getMask(), alignment);
return success();
}
};
diff --git a/mlir/test/Conversion/VectorToLLVM/vector-to-llvm-interface.mlir b/mlir/test/Conversion/VectorToLLVM/vector-to-llvm-interface.mlir
index 9b57b1b6fb4c7..5973c2ba2cbd0 100644
--- a/mlir/test/Conversion/VectorToLLVM/vector-to-llvm-interface.mlir
+++ b/mlir/test/Conversion/VectorToLLVM/vector-to-llvm-interface.mlir
@@ -2042,6 +2042,16 @@ func.func @gather_1d_from_2d_scalable(%arg0: memref<4x?xf32>, %arg1: vector<[4]x
// -----
+func.func @gather_with_alignment(%arg0: memref<?xf32>, %arg1: vector<3xi32>, %arg2: vector<3xi1>, %arg3: vector<3xf32>, %0: index) -> vector<3xf32> {
+ %1 = vector.gather %arg0[%0][%arg1], %arg2, %arg3 {alignment = 8} : memref<?xf32>, vector<3xi32>, vector<3xi1>, vector<3xf32> into vector<3xf32>
+ return %1 : vector<3xf32>
+}
+
+// CHECK-LABEL: func @gather_with_alignment
+// CHECK: llvm.intr.masked.gather %{{.*}}, %{{.*}}, %{{.*}} {alignment = 8 : i32} : (vector<3x!llvm.ptr>, vector<3xi1>, vector<3xf32>) -> vector<3xf32>
+
+// -----
+
//===----------------------------------------------------------------------===//
// vector.scatter
//===----------------------------------------------------------------------===//
@@ -2118,6 +2128,17 @@ func.func @scatter_1d_into_2d_scalable(%arg0: memref<4x?xf32>, %arg1: vector<[4]
// CHECK: %[[P:.*]] = llvm.getelementptr %[[B]][%{{.*}}] : (!llvm.ptr, vector<[4]xi32>) -> vector<[4]x!llvm.ptr>, f32
// CHECK: llvm.intr.masked.scatter %{{.*}}, %[[P]], %{{.*}} {alignment = 4 : i32} : vector<[4]xf32>, vector<[4]xi1> into vector<[4]x!llvm.ptr>
+// -----
+
+func.func @scatter_with_alignment(%arg0: memref<?xf32>, %arg1: vector<3xi32>, %arg2: vector<3xi1>, %arg3: vector<3xf32>, %0: index) {
+ vector.scatter %arg0[%0][%arg1], %arg2, %arg3 { alignment = 8 } : memref<?xf32>, vector<3xi32>, vector<3xi1>, vector<3xf32>
+ return
+}
+
+// CHECK-LABEL: func @scatter_with_alignment
+// CHECK: llvm.intr.masked.scatter %{{.*}}, %{{.*}}, %{{.*}} {alignment = 8 : i32} : vector<3xf32>, vector<3xi1> into vector<3x!llvm.ptr>
+
+
// -----
//===----------------------------------------------------------------------===//
@@ -2149,6 +2170,15 @@ func.func @expand_load_op_index(%arg0: memref<?xindex>, %arg1: vector<11xi1>, %a
// -----
+func.func @expand_load_op_with_alignment(%arg0: memref<?xindex>, %arg1: vector<11xi1>, %arg2: vector<11xindex>, %c0: index) -> vector<11xindex> {
+ %0 = vector.expandload %arg0[%c0], %arg1, %arg2 { alignment = 8 } : memref<?xindex>, vector<11xi1>, vector<11xindex> into vector<11xindex>
+ return %0 : vector<11xindex>
+}
+// CHECK-LABEL: func @expand_load_op_with_alignment
+// CHECK: %{{.*}} = "llvm.intr.masked.expandload"(%{{.*}}, %{{.*}}, %{{.*}}) <{arg_attrs = [{llvm.align = 8 : i64}, {}, {}]}> : (!llvm.ptr, vector<11xi1>, vector<11xi64>) -> vector<11xi64>
+
+// -----
+
//===----------------------------------------------------------------------===//
// vector.compressstore
//===----------------------------------------------------------------------===//
@@ -2177,6 +2207,15 @@ func.func @compress_store_op_index(%arg0: memref<?xindex>, %arg1: vector<11xi1>,
// -----
+func.func @compress_store_op_with_alignment(%arg0: memref<?xindex>, %arg1: vector<11xi1>, %arg2: vector<11xindex>, %c0: index) {
+ vector.compressstore %arg0[%c0], %arg1, %arg2 { alignment = 8 } : memref<?xindex>, vector<11xi1>, vector<11xindex>
+ return
+}
+// CHECK-LABEL: func @compress_store_op_with_alignment
+// CHECK: "llvm.intr.masked.compressstore"(%{{.*}}, %{{.*}}, %{{.*}}) <{arg_attrs = [{}, {llvm.align = 8 : i64}, {}]}> : (vector<11xi64>, !llvm.ptr, vector<11xi1>) -> ()
+
+// -----
+
//===----------------------------------------------------------------------===//
// vector.splat
//===----------------------------------------------------------------------===//
|
Comment on lines
+302
to
+304
| unsigned align = gather.getAlignment().value_or(0); | ||
| if (!align && | ||
| failed(getVectorToLLVMAlignment(*this->getTypeConverter(), vType, |
There was a problem hiding this comment.
Note that this means that explicit alignment takes priority over e.g. --convert-vector-to-llvm='use-vector-alignment=1. This feels like the right design decision, but we should make sure that it is documented (perhaps just add a comment here?) and tested (e.g. in "use-vector-alignment.mlir").
Similar comment for scatter. Thanks!
There was a problem hiding this comment.
Thanks for the review @banach-space! See here for the changes:
There was a problem hiding this comment.
Thanks! Sorry I didn't replay earlier, I was on a sick leave.
Alignment can be specified either explicitly via attributes in vector operations or via the option to use-vector-alignment=<N> in the --convert-vector-to-llvm pass. The explicit attribute takes precedent over the option used as input to the pass.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Allows alignment to be propagated correctly from vector to LLVM dialect operations.