[llvm] [AArch64] Define high bits of FPR and GPR registers (take 2) (PR #114827)

[via llvm-commits]

zeroomega wrote:

> Thanks for letting me know. From what I can quickly see there is only a time difference for the `runtime` tests. The `clang`, `polly`, `llvm` and `lld` tests run in about the same time. @PiJoules the build you linked to was interrupted while building the `check-clang` target, which for the other builds pointed to by @zeroomega wasn't any slower. Perhaps I'm missing something here though. I do see various other patches in the blamelist, has it been confirmed that this patch is the cause of the increased testing time?

The patch is confirmed through bisecting. 

https://ci.chromium.org/b/8730624321112150225 builds e52238b59f250aef5dc0925866d0308305a19dbf and the test time looks normal.

https://ci.chromium.org/b/8730630997932921601 builds c1c68baf7e0fcaef1f4ee86b527210f1391b55f6 and the test time regressed. c1c68baf7e0fcaef1f4ee86b527210f1391b55f6 (https://github.com/llvm/llvm-project/pull/114827) is 1 commit after e52238b59f250aef5dc0925866d0308305a19dbf


For stage2 builders, I didn't do a bisecting since they uses LTO and will take too much time. But
https://ci.chromium.org/ui/p/fuchsia/builders/prod/clang-linux-arm64/b8730794949594129265/overview is from 2 days ago (of course after this patch was landed). It took 2.5hrs for building clang and runtimes. and 2.1hrs for running all the tests. 
https://ci.chromium.org/ui/p/fuchsia/builders/prod/clang-linux-arm64/b8731348039888781153/overview is from 11/13 (before this patch was landed). It took 43mins to build clang and runtimes and 18 minutes to run the tests. If you feel it is necessary to perform a before/after stage2 build test on build time, though we feel it is unnecessary since the bisecting on the stage1 build already shows it clearly,, we can do that, it will take about 6 hours. 

I suspect the patch causes a general performance regression of binaries built by clang. Because the stage1 clang build is using the host toolchian, so the build time wasn't affected. But clang, lld and runtimes tests are using the freshly built clang, they are affected. The stage2 builder is built using the stage1 toolchain (which has the performance bug), so the build time also regressed. 

Let us know if you need additional information.


https://github.com/llvm/llvm-project/pull/114827