LTO doesn't store -mcmodel= information in the bitcode files

[petrhosek]

Bugzilla Link	33306
Resolution	FIXED
Resolved on	Sep 24, 2018 11:01
Version	trunk
OS	All
CC	@cmtice,@frobtech

Extended Description

LTO doesn't store -mcmodel= information in the bitcode files as it does for many other codegen options and this has to be passed explicitly when invoking linker through the -mllvm -code-model= flag.

[petrhosek]

assigned to @cmtice

[llvmbot]

Interesting. Do you have a repro where this matters, Petr? Thanks in advance!

[petrhosek]

Yes, in Fuchsia on AArch64 -mcmodel=kernel is different from -mcmodel=small, concretely it moves the thread pointer from TPIDR_EL0 to TPIDR_EL1, see https://reviews.llvm.org/D31624. We're trying to enable LTO for our kernel and we hit this issue.

[frobtech]

-fPIE (and presumably -fPIC and all the related variants) seems to be in the same boat. But I can't even figure out how to force that with an lld -mllvm option as we do with -mllvm -code-model=kernel. Should we file that as a separate bug or consider it part of this one?

I think it's a general bug if there's anything on the individual TU command line that's lost in LTO so that you can't just use the exact same command lines with the sole addition of -flto for your compile and link steps.

[frobtech]

Hmm, looks like PIC/PIE is encoded in IR metadata.
But LTO codegen seems to be ignoring it in the Zircon (Fuchsia) kernel build.

[cmtice]

I believe https://reviews.llvm.org/D52322 and https://reviews.llvm.org/D52323 fix this problem. They definitely fix the issue of passing the mcmodel information to LTO.