[AMDGPU] Added hot-block-rematerialize pass

[adam-yang]

Note: Before this is ready for review, I intend to remove AMDGPUMirSyncDependency and AMDGPUMirDivergenceAnalysis. They were basically forks of the IR versions of the passes from before the upstream MIR versions were available.

I have the code path that uses MachineUniformityInfo and the divergence related unit test passes with it, but I'd like to do more testing before deleting the other ones.

[github-actions]

✅ With the latest revision this PR passed the undef deprecator.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[dstutt]

Added Jay and Carl fyi (still needs cleanup)

[perlfu]

Since this is only a draft and MirDivergence parts are going away I only I had a quick look.
I wonder if parts like AMDGPULatencyTracker should be full analysis passes, rather than just utilities?

[jayfoad]

Need to #include <cmath> for this. But perhaps it would be better to use an integer-only power operation (not sure if we already have one anywhere in LLVM).

[jayfoad]

Is this intended to be run just before pre-RA MachineScheduler?

Update: I see the pass currently requires SSA form, so it would have to be run earlier than that. Where would you put it in the pass pipeline?

[dstutt]

We've got some work in the pipeline to allow pre-RA machinescheduler to run in SSA form as well, so that might be ok.
Also, this might work well with some other work we're considering to add spilling as a separate pass before RA (but in SSA form).

[jayfoad]

Hi @adam-yang, I've done some testing with this patch and noticed that options -amdgpu-remat-enable-hot-block-remat-aggressive and -amdgpu-remat-enable-sub-exp-remat seem to cause many crashes in the compiler. Are you interested in test cases for these? Or are these options not expected to be working yet?

[adam-yang]

Hi @adam-yang, I've done some testing with this patch and noticed that options -amdgpu-remat-enable-hot-block-remat-aggressive and -amdgpu-remat-enable-sub-exp-remat seem to cause many crashes in the compiler. Are you interested in test cases for these? Or are these options not expected to be working yet?

Yes could you send them to me please.

[jayfoad]

OK, here's a somewhat reduced test case for a crash with -amdgpu-remat-enable-sub-exp-remat. The verifier reports "Bad machine code: Found PHI instruction after non-PHI".
r1.txt

[jayfoad]

Here's one for -amdgpu-remat-enable-hot-block-remat-aggressive which fails an assertion "Invalid RC for virtual register": r2.txt

[adam-yang]

Here's one for -amdgpu-remat-enable-hot-block-remat-aggressive which fails an assertion "Invalid RC for virtual register": r2.txt

This revealed some incorrect assumptions in this PR. Lanes are assumed to be 32-bit, but new 16-bit subregs have been added so lanes are now 16-bit. I'll have to go through the change to fix some more issues related to this.