[llvm] [NVPTX] Decide .ftz on the whole denormal mode, not just the output half (PR #222645)

[Ashutosh Kumar Singh via llvm-commits]

Ashutosh0x wrote:

Confirming I have read the LLVM AI Tool Use Policy, the Code-Review Policy and Practices, and the Developer Policy.

This change was written with AI assistance, and I have noted it in the PR description and added an `Assisted-by:` trailer to the commit. I have read and reviewed the entire patch and I am accountable for it.

Two things I did specifically so that this is a good use of reviewer time rather than a plausible-looking patch: the `.ftz` semantics come from the PTX ISA and LangRef passages quoted in the description, not from a model's recollection; and every row of the before/after table is measured, by building `llc` twice from this tree — once with the patch, once with it stashed — and diffing the emitted PTX. Happy to answer questions on any part of it.


https://github.com/llvm/llvm-project/pull/222645