bjoernager/rust
bjoernager/rust
1
Fork 0
Code Issues Pull requests Activity

Remove the ThinLTO CU hack

Browse source 
This reverts #46722, commit e0ab5d5feb.

Since #111167, commit 10b69dde3f, we are
generating DWARF subprograms in a way that is meant to be more compatible
with LLVM's expectations, so hopefully we don't need this workaround
rewriting CUs anymore.
This commit is contained in:
Josh Stone 2023-05-04 17:20:49 -07:00
parent ce042889f7
commit c836c24994
3 changed files with 0 additions and 112 deletions
Download patch file Download diff file Expand all files Collapse all files

49
compiler/rustc_codegen_llvm/src/back/lto.rs
View file

@ -25,7 +25,6 @@ use std::fs::File;
use std::io;
use std::iter;
use std::path::Path;
use std::ptr;
use std::slice;
use std::sync::Arc;
@ -709,17 +708,6 @@ pub unsafe fn optimize_thin_module(
let llmod = module.module_llvm.llmod();
save_temp_bitcode(cgcx, &module, "thin-lto-input");
// Before we do much else find the "main" `DICompileUnit` that we'll be
// using below. If we find more than one though then rustc has changed
// in a way we're not ready for, so generate an ICE by returning
// an error.
let mut cu1 = ptr::null_mut();
let mut cu2 = ptr::null_mut();
llvm::LLVMRustThinLTOGetDICompileUnit(llmod, &mut cu1, &mut cu2);
if !cu2.is_null() {
return Err(write::llvm_err(&diag_handler, LlvmError::MultipleSourceDiCompileUnit));
}
// Up next comes the per-module local analyses that we do for Thin LTO.
// Each of these functions is basically copied from the LLVM
// implementation and then tailored to suit this implementation. Ideally
@ -766,43 +754,6 @@ pub unsafe fn optimize_thin_module(
save_temp_bitcode(cgcx, &module, "thin-lto-after-import");
}
// Ok now this is a bit unfortunate. This is also something you won't
// find upstream in LLVM's ThinLTO passes! This is a hack for now to
// work around bugs in LLVM.
//
// First discovered in #45511 it was found that as part of ThinLTO
// importing passes LLVM will import `DICompileUnit` metadata
// information across modules. This means that we'll be working with one
// LLVM module that has multiple `DICompileUnit` instances in it (a
// bunch of `llvm.dbg.cu` members). Unfortunately there's a number of
// bugs in LLVM's backend which generates invalid DWARF in a situation
// like this:
//
// https://bugs.llvm.org/show_bug.cgi?id=35212
// https://bugs.llvm.org/show_bug.cgi?id=35562
//
// While the first bug there is fixed the second ended up causing #46346
// which was basically a resurgence of #45511 after LLVM's bug 35212 was
// fixed.
//
// This function below is a huge hack around this problem. The function
// below is defined in `PassWrapper.cpp` and will basically "merge"
// all `DICompileUnit` instances in a module. Basically it'll take all
// the objects, rewrite all pointers of `DISubprogram` to point to the
// first `DICompileUnit`, and then delete all the other units.
//
// This is probably mangling to the debug info slightly (but hopefully
// not too much) but for now at least gets LLVM to emit valid DWARF (or
// so it appears). Hopefully we can remove this once upstream bugs are
// fixed in LLVM.
{
let _timer = cgcx
.prof
.generic_activity_with_arg("LLVM_thin_lto_patch_debuginfo", thin_module.name());
llvm::LLVMRustThinLTOPatchDICompileUnit(llmod, cu1);
save_temp_bitcode(cgcx, &module, "thin-lto-after-patch");
}
// Alright now that we've done everything related to the ThinLTO
// analysis it's time to run some optimizations! Here we use the same
// `run_pass_manager` as the "fat" LTO above except that we tell it to