compiler_rt: Update Windows ABI for float<->int conversion routines #12071
Conversation
topolarity
force-pushed
the
windows-abi-change
branch
from
1b50749 to
c8d2f5f
Compare
|
Great work - really appreciate your help with the LLVM stuff lately 🙏 |
There was a problem hiding this comment.
I believe the failure is due to codegen/llvm.zig emitting explicit calls to compiler_rt functions in the cases where LLVM will fail to properly do so - meaning that this function signature must be adjusted in codegen.cpp and codegen/llvm.zig to match.
Are you up for those changes too? Let me know if you want some help.
Happy to lend a hand! These are fun bugs to chase for now 🐞 |
Very good point - should have thought of our explicitly-lowered calls. Let me take a stab and see how far I can get. I'll reach out if it gets hairy. |
topolarity
force-pushed
the
windows-abi-change
branch
from
c8d2f5f to
1529092
Compare
Starting with LLVM 14, the Libcalls to these functions are now lowered using a Vec(2, u64) instead of the standard ABI for i128 integers, so our compiler-rt implementation needs to be updated to expose the same ABI on Windows.
This is just a cosmetic change. The goal is to keep the export logic relatively flat and centralized.
This change is the Zig counterpart to https://reviews.llvm.org/D110413 Since we lower some libcalls directly (just like clang does), we need to make sure that the ABI we call with matches the ABI of the compiler-rt we are providing (and also the ABI expected by LLVM). While I was at it, I noticed some flawed vector handling in the binary soft float ops in stage 1, so I shored up the logic a bit and expanded an existing test to cover the missing functionality.
This change is the Zig counterpart to https://reviews.llvm.org/D110413 Same as the prior commit, but for stage2
andrewrk
force-pushed
the
windows-abi-change
branch
from
1529092 to
2b52154
Compare
Resolves #12063
Starting with LLVM 14, the Libcalls to these functions are now lowered using a
Vector(2, u64)instead of the standard ABI for i128 integers on Windows, so our compiler-rt implementation needs to be updated to expose the same ABIRelevant LLVM revision: https://reviews.llvm.org/D110413