[mlir][amd] fix LLVM::InsertValueOp::create failure to disambiguate

[makslevental]

fixes #149879 (comment)

Note this happens because ADL can't disambiguate between mlir::DenseI64ArrayAttr and llvm::ArrayRef<int64_t> for the value 0 which I guess is equal to nullptr on some (most?) systems.

Note, this only occurs with the value 0.

[llvmbot]

@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-backend-amdgpu

Author: Maksim Levental (makslevental)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/150605.diff

1 Files Affected:

• (modified) mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp (+2-2)

diff --git a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
index bc0d9bf2ea771..b6f6167d1dfb3 100644
--- a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
+++ b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
@@ -232,8 +232,8 @@ struct FatRawBufferCastLowering
     Value result = MemRefDescriptor::poison(
         rewriter, loc,
         getTypeConverter()->convertType(op.getResult().getType()));
-    result = LLVM::InsertValueOp::create(rewriter, loc, result, fatPtr,
-                                         kAllocatedPtrPosInMemRefDescriptor);
+    SmallVector<int64_t> pos{kAllocatedPtrPosInMemRefDescriptor};
+    result = LLVM::InsertValueOp::create(rewriter, loc, result, fatPtr, pos);
     result = LLVM::InsertValueOp::create(rewriter, loc, result, fatPtr,
                                          kAlignedPtrPosInMemRefDescriptor);
     result = LLVM::InsertValueOp::create(rewriter, loc, result, offset,

[llvmbot]

@llvm/pr-subscribers-mlir-gpu

Author: Maksim Levental (makslevental)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/150605.diff

1 Files Affected:

• (modified) mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp (+2-2)

diff --git a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
index bc0d9bf2ea771..b6f6167d1dfb3 100644
--- a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
+++ b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
@@ -232,8 +232,8 @@ struct FatRawBufferCastLowering
     Value result = MemRefDescriptor::poison(
         rewriter, loc,
         getTypeConverter()->convertType(op.getResult().getType()));
-    result = LLVM::InsertValueOp::create(rewriter, loc, result, fatPtr,
-                                         kAllocatedPtrPosInMemRefDescriptor);
+    SmallVector<int64_t> pos{kAllocatedPtrPosInMemRefDescriptor};
+    result = LLVM::InsertValueOp::create(rewriter, loc, result, fatPtr, pos);
     result = LLVM::InsertValueOp::create(rewriter, loc, result, fatPtr,
                                          kAlignedPtrPosInMemRefDescriptor);
     result = LLVM::InsertValueOp::create(rewriter, loc, result, offset,

[makslevental]

I don't have a windows machine so I'm gonna wait for CI to complete just to be sure...

[kuhar]

You should be able to do ArrayRef{kAllocated...}

[makslevental]

you have have to do ArrayRef<int64_t>{...} and I thought this was version was clearer but yea you're right.

[DavidSpickett]

The 2 stage build will take ages, but this fixed https://lab.llvm.org/buildbot/#/builders/207/builds/4343 as well so we know it works. Thanks!

A bit confused why you only had to change the first one, but if it works it works.

[makslevental]

A bit confused why you only had to change the first one, but if it works it works.

It's because the ADL failure is due to 0 == nullptr and both ArrayRef, DenseI64ArrayAttr are "reference" types. For the rest of the callsites the "pos" arg is non-zero and so ADL works and correctly selects the ArrayRef<int64_t> position overload.

[DavidSpickett]

Oh right there's 3 different constants being used, I read them as all being the same.