Okay, thanks.

Sorry for the delay in replying.

Hi @asmok-g I will be leaving it reverted, I think the case I was interested in is that just generally broken :)

The patch is certainly triggering on IR though, twice. I will look into it.

Er, I can't read X86 machine code. And I also don't consider a 4K line IR file as a reduced reproducer....

I'm preparing a reduced, verified case.

Okay, great, thanks.

Do those figures include using wasm-opt after linking?

No worries!

Right, thanks. My next question was going to be whether you're building libc++ :) - thanks!

Hmmm, would you be able to send post the .ll file?

Hi @bgraur Thanks for the reproducer, but I'm unable to reproduce the issue on my x64 box... What platform were you targeting? Could you please confirm that the reproducer still demonstrates the bug?

Okay, thanks!

Any update on that reproducer @bgraur?

Taking D141926 as a base, this transform brings the total code size of my local benchmarks down from 7982KB to 7970KB.

How is code size affected?

Using wasi-sdk-19, the total code size of my suite is 7985 KB. With this patch it's reduced slightly to 7982 KB.

Modified, and now using, isKnownNeverZeroFloat.

Ping?

• Added check for no signed zeros.
• Fixed up vector support.
• Added arm tests.

Hopefully the FP clamping and min/max idiom misses have now been fixed in D142093, D142481, D142535, D143106 and D143256.

Can you preserve the match with nsz?

Thanks for doing this.

Added SETNE as well.

Thanks for the reproducer, I'll revert for now.

Added setugt and setult support.

Added support for checking against SETLT and SETGT nodes too. X86 and Arm test changes look rather noisey...

Thanks, extracted the bitwidth calculation into APFloatBase.

Rebased and added comment.

Avoiding integer comparison warning.

Using semanticsMaxExponent, so hopefully correct now...

Removed non-poison input restriction for fptosi.

If you want to add some more, the initial set is in: https://github.com/llvm/llvm-project/commit/714286f9e641209411609deaf80dd865aa2198c5

Okay, great. Thanks for clarification on both fronts. I'm just about to commit some tests.

it should be sufficient to have the intermediate integer width be one more than the minimum required type width, so "%i = fptoui half %x to i17".

Now only checking for poison/undef for the signed case.

Added extra preserved analyses, though omitting MemorySSA for now.

Thanks for the quick review! I will be porting hardware-loops and mve-tail-predication as well.

Fixed typo.