[AArch64][CodeGen] Fix wrong operand order when creating vcmla intrin…

[ellishg]

…sic (llvm#65278)

[ellishg]

We've discovered a floating point miscompilation in XCode 16 when using -ffp-contract=fast. I found this patch that fixes the issue.

@fhahn can we get this merged into the release branch?

[ellishg]

Oh, and the patch did not cherry-pick cleanly so I had to fix the tests to make them pass on this branch.

[drodriguez]

@swift-ci please test llvm

[drodriguez]

@swift-ci please test

[drodriguez]

(I think the LLVM tests will fail in some cases, but hopefully not in the specific tests that the commit changes)

[drodriguez]

About the LLVM Test failures:

I left a comment about ExtractAPI/enum.c last week in #8987 (comment)

For the Driver/ ones is probably because #8668 is in stable/20230725, but not in swift/release/6.0 (has been like that a couple of months)

[ellishg]

Friendly ping @fhahn @ravikandhadai

[ravikandhadai]

cc @Gerolf-Apple

[fhahn]

Unfortunately I think it is too late at this point in the release to take this change.

[drodriguez]

Should it be applied to stable/20230725 at least so it is used by a future Swift 6.0.x release? What's the process to get it in the next release before stable/20240723 is used?

[ellishg]

This is a critical codegen fix. If it isn't picked, some floating point operations will simply give the wrong result. Also, the non-test changes are pretty simple and safe in my opinion.

[Gerolf-Apple]

Looks good to me.

[drodriguez]

@Gerolf-Apple are you able to merge into this branch for us? Is this the right branch to merge at this point?

[Gerolf-Apple]

We will need to get this on a newer branch.