Murali Vijayaraghavan

The original issue https://github.com/iree-org/iree/issues/9726#issue-1296754705 by @dcaballe says the debug info has to be stripped before generating LLVM code. The current flag `-iree-llvm-debug-symbols=false` already removes debug info from LLVM, right? So...

That sounds like a useful tool indeed. And it should exist in MLIR (not IREE), right? On a somewhat related note, it looks like, unlike for LLVM, there's no serialized...

Does the comment in https://github.com/ucb-bar/berkeley-hardfloat/blob/master/src/main/scala/DivSqrtRecFN_small.scala suffice?

The best way to split this would be to keep only bedrock2 language and semantics in this repo. RISC-V semantics should be completely orthogonal to bedrock2 semantics (for instance bedrock2...

I agree with Andres' overall splitting strategy. However, I still think it is a good idea to do the splitting that I mentioned, namely the compiler to machine code and...

Looks like the flag does zilch. Let me see if the `reassoc` attribute is generated.

I was planning to file a bug in MLIR

Alloca gets optimized away in llvm's sroa pass. It remains in our generated code On Fri, Oct 14, 2022, 11:14 Han-Chung Wang ***@***.***> wrote: > I thought the issue with...

The allocation size solely depends on the (single dimension) reduction tiling size. On Fri, Oct 14, 2022, 12:13 Han-Chung Wang ***@***.***> wrote: > Yes, I agree with you. We should...

@MaheshRavishankar had explained why alloca was being generated earlier. From what I understand, bufferization can reuse input buffers only if there's a chain of use-defs from inputs to outputs. If...