x86: Implement fma intrinsic

[BradleyWood]

This PR implements Math.fma intrinsics on x86 with vfmadd instructions.

Issue: #7474

[BradleyWood]

Requires fma, extensions.

Here are performance numbers,

Benchmark	Compared to Baseline	Compared to Hotspot
FMABench.benchFloatFMA	552x	1.19x
FMABench.benchDoubleFMA	1482x	1.16x

[BradleyWood]

FYI @0xdaryl, @JamesKingdon

@hzongaro Would you mind reviewing?

[JamesKingdon]

Thanks Brad, what a result!

[hzongaro]

Thanks for pulling this together so quickly! I just have a few comments and questions.

[BradleyWood]

See the force-push

[hzongaro]

I think the changes look good. I will wait for @0xdaryl to review before running tests.

[0xdaryl]

I think the logic behind choosing the best instruction format to use is sound. I just have a few issues with the approach behind that.

[BradleyWood]

@hzongaro The only thing I changed is comments, but was forced to rebase and fix conflicts. I think everything is ready to run testing.

[hzongaro]

@0xdaryl, may I ask you to confirm that your comments were all resolved to your satisfaction?

[hzongaro]

Jenkins test sanity.functional,sanity.openjdk xlinux,win,xmac jdk11,jdk17,jdk21

[hzongaro]

The many jvmti test failures in JDK 11 sanity.functional xlinux and Windows appear to be unrelated to this change. I see the same failures with another pull request test run.

The failures running cmdlineTester_jfr_0 for JDK 17 sanity.functional Windows and JDK 21 sanity.functional Windows are due to known issue #21181.

The failures in JDK running jdk_util_1 for JDK 21 sanity.openjdk x86-64 macOS appears to be known issue #17270.

I will rerun sanity.functional testing for JDK 11 x86-64 macOS

Jenkins test sanity.functional xmac jdk11

[hzongaro]

@BradleyWood, it looks there are merge conflicts that you will need to resolve before I will be able to rerun the JDK 11 xmac testing.

[BradleyWood]

@hzongaro Fixed conflicts

[hzongaro]

Jenkins test sanity.functional xmac jdk11

[BradleyWood]

@hzongaro Should I double-deliver on monday if no issues arise by then?

[hzongaro]

Should I double-deliver on monday if no issues arise by then?

I this is a safe change. I would be OK with it if there are no objections.

[BradleyWood]

Is this is a safe change.

I think its on the safer side as far as intrinsics go. The only concern I have is that the operands which may be loaded directly from memory could come in any of many possible combinations. I did my best to test this whilst I was working on it buy disabling a few optimizations and placing constants (such as 1f) in place of variables to force memory loads at specific spots.

On the other hand, this change has survived 4 or 5 days of builds/tests without issues, provides significant performance benefits, and has a workaround option to disable the intrinsic if needed.

FYI, @pshipton

[hzongaro]

I this is a safe change.

Sorry for that typo! I meant to say, "I think this is a safe change." I agree with @BradleyWood's analysis.

[pshipton]

Getting it in today would be good, before we start any milestone builds.