Thanks, @rjmccall . I'm sorry I don't have much time on this patch recently. I will update it later.

Use -mextend-small-integers=<arg> instead of boolean option

In D124435#3474130, @skan wrote:

Should we update the clang/docs/ReleaseNotes.rst for this?

add one missing comment

Address comments

In D124435#3474130, @skan wrote:

Should we update the clang/docs/ReleaseNotes.rst for this?

fix bug

I think you need to add a new ConservativeExtend kind to ABIArgInfo and make X86_64ABIInfo::classifyArgumentType (and classifyReturnType?) use that instead of Extend.

In D122963#3460934, @rjmccall wrote:

If people think that we'll reliably get the right behavior by the frontend adding sext and zext to call sites but not to function arguments, I can walk you through how to make that change in the frontend.

What we can't do in the frontend is just switch to unconditionally not using sext and zext at all on x86_64, because we need to maintain compatibility with old versions of clang. But we can introduce a flag that makes clang generate incompatible code if someone wants to opt in to it for performance.

According to our internal tests, the performance of most of the tests is not affected. The most affected test performance dropped by %0.76. Average performance dropped 0.25%. This matches our expectations

If there is no better solution, I hope this patch merge in. This is a bug that needs to be fixed.

rebase

Address Phoebe's comments: remove 32-bit target check.

Ping? This bug will cause correctness issues, we need to fix it. I think it is possible to fix in the backend in the current situation.
Is there any other better method, @jyknight @lebedev.ri

It seems we all agree we should extend the parameter both in caller and in callee. This is also the behavior of gcc.
For now, the front-end doesn't have an accurate attribute that describes what we're expecting right now and doesn't know whether the function is external linked. By default clang will do the extension only in the caller. This forced us to do some workarounds on the backend.
We also thought about setting these attributes correctly in the IPO, but is there any possibility that IPO won't do in some cases?
D112942 introduces a new parameter attribute NoExt might help us. But this still means that by default, we will do extension in both caller and callee. And only in some optimization cases, we will do extension in caller or callee. Still not the ideal method that lebedev.ri expected.
Besides, I think that as an ABI issue, the correctness should be guaranteed first, optimization is not always guaranteed.
@lebedev.ri, is there any better approach to fix this issue?

In D122963#3435619, @lebedev.ri wrote:

Still not okay with it. The fix is to fix whatever is emitting the attribute, not completely ignore it.

Address Shenchen's comment

Address Yuanke's comments: do the extension only if the function isn't local linked.

Thanks for you review.

Can you take a review again and +1 if no problem? I missed changing some testcases. Thanks.

Fix check-all fail.

Thanks for you review. The order of operand is changed in final commit: The addend of FMA builtin is moved from the first to the third.

Address comments.

1. Remove redundant option checkRemove redundant option check.
2. Reformat.

Distinguish 0.0 and -0.0 in FMA.

Adress comments

Adress comments.

Adress comments.

Address comments

Thanks for reviewing this patch!

Address pengfei's comments.

What is the usage model you’re trying to enable? This has been broken for a long time. Why is it important now?

In D100091#2707521, @craig.topper wrote:

Does adding HasX87 predicates to the instructions cause the conversions in my example to fail?

In D100091#2677235, @craig.topper wrote:

In D100091#2675996, @LiuChen3 wrote:

In D100091#2675968, @craig.topper wrote:

What about cases that use x87 to do float/double conversions? I think it’s i64 conversions in 32-bit mode. But maybe other cases. Does gcc disable those with -mno-80387?

You are right! GCC can do the conversions: https://godbolt.org/z/87ez838oc .
With this patch, llvm will fail in ISEL. I want to make another patch to do this.

I'm more concerned about these cases that don't use long double and aren't missed constant folding. https://godbolt.org/z/qGv54ef34

Fix format issue.

1. Rebase.
2. Adding _mm512_i32loscatter_epi64 and _mm512_mask_i32loscatter_epi64.

Thanks for your review. Hope this patch won't cause too many ABI issues in the future.

Address Craig's comments

Address Simon's comments

In D100368#2685189, @RKSimon wrote:

Add _mm512_i32loscatter_epi64 and _mm512_mask_i32loscatter_epi64 for completeness?