How to make when developing machine function pass ?

Project Infrastructure LLVM Dev List Archives
#1

Hi

I am modifying X86RetpolineThunks.cpp.

X86RetpolineThunks.cpp 's location is llvm-src/lib/Target/X86.

Which target should I use , next time use clang test.c , I can see the difference .

I found that “make llc” doesn’t work.

And either “make LLVMX86CodeGen” doen’t work.

“make clang” waste a lot of time, even -j64.

Every time modifying the file of machine function pass ,
I have to wait for 1 min.

It drive me crazy.

Which “target” should use to “make” when developing machine function pass ?

Or is there other way to develop machine function pass?

Thank you.

#2

One minute to recompile and relink one file is too long. There are ways to make incremental development significantly faster, and there have been a few threads about this on llvm-dev. This one came to mind:
https://groups.google.com/g/llvm-dev/c/ur-2aX_KvfM/m/FFRlZdZECAAJ

If you search around, you can find some of the standard recommendations:

  • Use ninja instead of make
  • Use lld or gold instead of ld.bfd
  • Adjust debug info settings (-gmlt, split dwarf, or disable it)
  • Use external tablegen (probably not your problem)

Good luck!

#3

The google groups link seems to work for me, but is broken when I am not logged into my Google account. >_> I mean this thread titled “debug build busts memory”:
http://llvm.1065342.n5.nabble.com/llvm-dev-debug-build-busts-memory-td133229.html#a133237

#4

Gold linker really speed up.
You help me a lot.
Thank you !!

I would keep going trying ninja~

Reid Kleckner <rnk@google.com> 於 2020年3月11日 週三 上午3:33寫道：