search

espressif / llvm-project

Fork of LLVM with Xtensa specific patches. To be upstreamed.
Other
226 stars 20 forks source link

indirect calls are always used (LLVM-321) #90

Open yamt opened 4 months ago

yamt commented 4 months ago

currently, -mlongcalls is always effectively on. i had to patch llvm for my use case. cf. https://github.com/yamt/llvm-project/commit/3f04244a4fd94952a3f208b7d1ee3afef2da1e3a it would be nicer to have a way to use direct calls. ideally, the backend can automatically choose direct calls when the call distance is not too long, i guess.

related: https://github.com/bytecodealliance/wasm-micro-runtime/pull/3141 https://github.com/bytecodealliance/wasm-micro-runtime/pull/3142

yamt commented 4 months ago

for my use case, it's ok to manually adding attributes at IR level. eg. xtensa-short-call. mips has something along the line: https://clang.llvm.org/docs/AttributeReference.html#short-call-near

gerekon commented 4 months ago

@yamt Thanks for pointing this out. We are working on direct calls support using callN instruction.

cc @sstefan1

sstefan1 commented 4 months ago

for my use case, it's ok to manually adding attributes at IR level. eg. xtensa-short-call. mips has something along the line: https://clang.llvm.org/docs/AttributeReference.html#short-call-near

We're working on implementing the support for direct calls, the same way gcc does it. That is a bit more complex. In the mean time we can implement short-call attributes for xtensa.

dongsheng28849455 commented 2 months ago

hi, @sstefan1, can I ask the progress for this feature?

sstefan1 commented 2 months ago

hi, @sstefan1, can I ask the progress for this feature?

There hasn't been any real progress since that coment. I'll write here when I have some updates.

dongsheng28849455 commented 1 month ago

hi, @sstefan1, can I ask the progress for this feature?

There hasn't been any real progress since that coment. I'll write here when I have some updates.

is that fixed by https://github.com/espressif/llvm-project/releases/tag/esp-17.0.1_20240408?

yamt commented 1 month ago

hi, @sstefan1, can I ask the progress for this feature?

There hasn't been any real progress since that coment. I'll write here when I have some updates.

is that fixed by https://github.com/espressif/llvm-project/releases/tag/esp-17.0.1_20240408?

the short-call attribute is already in the releases and it seems working as far as i tested. the "ideal" solution (ie. somehow automatically use short/long calls) is not there yet. i guess he was talking about the latter.

gerekon commented 1 month ago

hi, @sstefan1, can I ask the progress for this feature?

There hasn't been any real progress since that coment. I'll write here when I have some updates.

is that fixed by https://github.com/espressif/llvm-project/releases/tag/esp-17.0.1_20240408?

the short-call attribute is already in the releases and it seems working as far as i tested. the "ideal" solution (ie. somehow automatically use short/long calls) is not there yet. i guess he was talking about the latter.

Yes, exactly. Automatic use of short/long calls is in progress. Just for information: We do not expect it to be ready in the 1st half of this year.

andyross commented 2 weeks ago

Just bumped into this myself. Is there no support for at least changing the default e.g. ("-mno-longcalls")? The range of a CALLx instruction is +/- 512k, which is quite a bit larger than the .text segment of many/most firmware images in the wild. Right now the conservative default is a pretty severe performance regression, and it's needless for a lot of applications.

andreisfr commented 2 weeks ago

@andyross , thank you for the question. Currently we work on implementing CALLX support in the LLD.