Integrate LLVM at llvm/llvm-project@8955e285e1

[ScottTodd]

Advance to llvm/llvm-project@8955e285e1

• Still carrying a revert of llvm/llvm-project@fa06668
• Dropped revert of llvm/llvm-project@bbd4af5
• Dropped local patch of llvm/llvm-project@4f80508

[ScottTodd]

Putting this on hold for a little bit. Full debug chats here on Discord.

Issues encountered:

CUDA Translation failed for operation: builtin.unrealized_conversion_cast

Logs: https://github.com/iree-org/iree/actions/runs/10113405122/job/27970714347?pr=18010#step:9:55

<unknown>:0: error: LLVM Translation failed for operation: builtin.unrealized_conversion_cast
<unknown>:0: note: see current operation: %71 = "builtin.unrealized_conversion_cast"(%70) : (!llvm.array<1 x vector<1xf32>>) -> vector<1x1xf32>
model.mlir:7:10: error: failed to translate the MLIR LLVM dialect to the native llvm::Module

Repro: this input file and iree-compile model.mlir --iree-hal-target-backends=cuda -o /tmp/model_gpu_cuda.vmfb

Culprit: llvm/llvm-project@f6431f0

Removing LLVM::FMAOp from the target.addIllegalOp<...> block in mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp fixes compilation. I suggest debugging around

iree/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp

Lines 133 to 151 in 1ee68bd

	RewritePatternSet llvmPatterns(&getContext());
	populateLowerHALInterfaceOp(llvmPatterns);
	populateLLVMConversionPatterns(&getContext(), llvmPatterns, converter);
	populateComplexToLLVMConversionPatterns(converter, llvmPatterns);
	populateMathToLLVMConversionPatterns(converter, llvmPatterns);
	memref::populateExpandStridedMetadataPatterns(llvmPatterns);
	populateFinalizeMemRefToLLVMConversionPatterns(converter, llvmPatterns);
	populateFuncToLLVMConversionPatterns(converter, llvmPatterns);
	cf::populateControlFlowToLLVMConversionPatterns(converter, llvmPatterns);
	arith::populateArithToLLVMConversionPatterns(converter, llvmPatterns);
	populateVectorToLLVMConversionPatterns(converter, llvmPatterns);
	populateGpuToNVVMConversionPatterns(converter, llvmPatterns);
	populateNVGPUToNVVMConversionPatterns(converter, llvmPatterns);
	populateGpuWMMAToNVVMConversionPatterns(converter, llvmPatterns);
	LLVMConversionTarget target(getContext());
	configureGpuToNVVMConversionLegality(target);
	if (failed(applyPartialConversion(m, target, std::move(llvmPatterns)))) {
	signalPassFailure();
	}

Perhaps some of the patterns there are overlapping or running in the wrong order (change pattern benefit?)

---

iree/compiler/Codegen/SPIRV/test/vectorize_load_store.mlir.test

Logs: https://github.com/iree-org/iree/actions/runs/10113405115/job/27969720942?pr=18010#step:5:1720

+ iree-opt --split-input-file '--pass-pipeline=builtin.module(func.func(iree-spirv-vectorize-load-store, canonicalize, cse))' --mlir-print-local-scope /work/compiler/src/iree/compiler/Codegen/SPIRV/test/vectorize_load_store.mlir
+ FileCheck /work/compiler/src/iree/compiler/Codegen/SPIRV/test/vectorize_load_store.mlir
iree-opt: /work/third_party/llvm-project/llvm/include/llvm/Support/Casting.h:566: decltype(auto) llvm::cast(const From &) [To = mlir::VectorType, From = mlir::Type]: Assertion `isa<To>(Val) && "cast<Ty>() argument of incompatible type!"' failed.

Might relate to llvm/llvm-project@bbd4af5

---

iree/compiler/Codegen/LLVMGPU/test/set_transform_strategy_pad.mlir.test

lit test failure

---

iree/compiler/Dialect/Stream/Transforms/test/refine_usage.mlir.test

lit test failure

[ScottTodd]

Started a smaller integrate at #18016