Takes a long time to compile ARM32 Android libart-disassembler

[kongy]

It takes more than 20 minutes to link libart-assembler, it used to take less than 1 minute. This is reproducible on commit f142f8a as well as ToT build.

Reproducer: https://drive.google.com/file/d/112_Y7rUIjrG0o6AFHJq_FIENsY3lRlJc/view?usp=sharing

[llvmbot]

@llvm/issue-subscribers-backend-arm

Author: Yi Kong (kongy)

This is reproducible on commit f142f8a as well as ToT build.

Reproducer: https://drive.google.com/file/d/112_Y7rUIjrG0o6AFHJq_FIENsY3lRlJc/view?usp=sharing

[pirama-arumuga-nainar]

cc: @smithp35

[smithp35]

I don't think that this is a LLD problem. The vast majority of the time is taken in LTO. I've uploaded the output of --time-trace

If I'm reading it correctly, the time is being spent in __cxx_global_var_init in libvixl.a(decoder-aarch64.o) with instruction selection taking a large amount of time.

This is way beyond what I can efficiently investigate. The decoder-aarch64.o takes a minute to compile on its own without LTO (via llc), but it is possible that with LTO and or the LTO passes that causes it to blowup somewhere.

libart-disassembler.so.time-trace.txt
I had to rename the time-trace to .txt to get Github to accept it.

[kongy]

I was able to reduce this to reproduce without LTO.

Command:

Clang++ -mthumb -fno-exceptions -msoft-float -march=armv8-a -mfloat-abi=softfp -mfpu=neon-fp-armv8 -target armv7a-linux-androideabi31 -std=gnu++20 -O2 -w -S -o /dev/null decoder-aarch64.i

And the source file:

https://drive.google.com/file/d/1W0oWZ7Xw-DPwyf0E_fvuu7YEq3h9H-n6/view?usp=sharing

[pirama-arumuga-nainar]

https://github.com/Linaro/vixl/blob/main/src/aarch64/decoder-constants-aarch64.h#L51 is the global variable that's contributing to the slow compile.

[kongy]

Bisection points to change 8e8d259.

CC: @nikic

[nikic]

From a quick look, the time is spent in PerformSTORECombine -> CombineBaseUpdate -> hasPredecessorHelper.