[RISCV] Code model other than medium does not work with lto

[kxxt]

I tried this code:

// hello.rs
fn main() {
    println!("Hello, world!");
}

// command on riscv64
rustc -C code-model=large -C lto hello.rs
// command on x86_64
rustc --target riscv64gc-unknown-linux-gnu -C code-model=large -C lto hello.rs

I expected to see this happen: code compiles without error

Instead, this happened: failed to load bitcode of module "addr2line-*"

on riscv64, archrv rust:

$ rustc -C code-model=large -C lto hello.rs
warning: linking module flags 'Code Model': IDs have conflicting values in '' and 'hello.8b1675587fc60370-cgu.0'

error: failed to load bitcode of module "addr2line-4a5a0cf02a84c587.addr2line.d0fbcb35ed0a8d23-cgu.0.rcgu.o": 

error: aborting due to 1 previous error; 1 warning emitted

On x86_64, latest stable from rustup:

rustc --target riscv64gc-unknown-linux-gnu -C code-model=large -C lto hello.rs
warning: linking module flags 'Code Model': IDs have conflicting values in '' and 'hello.eea8546ba018bd24-cgu.0'

error: failed to load bitcode of module "addr2line-28336428312034c6.addr2line.134ee917f0cf37e1-cgu.0.rcgu.o": 

error: aborting due to 1 previous error; 1 warning emitted

On x86_64, Nightly:

rustc +nightly --target riscv64gc-unknown-linux-gnu -C code-model=large -C lto hello.rs
warning: linking module flags 'Code Model': IDs have conflicting values: 'i32 3' from , and 'i32 4' from hello.4a3c48060548947b-cgu.0

error: failed to load bitcode of module "addr2line-62ef9fdb0627a182.addr2line.746071221f6a6099-cgu.0.rcgu.o": 

error: aborting due to 1 previous error; 1 warning emitted

Meta

rustc --version --verbose:

rustc 1.85.1 (4eb161250 2025-03-15) (Arch Linux rust 1:1.85.1-1)
binary: rustc
commit-hash: 4eb161250e340c8f48f66e2b929ef4a5bed7c181
commit-date: 2025-03-15
host: riscv64gc-unknown-linux-gnu
release: 1.85.1
LLVM version: 19.1.7

rustc 1.86.0 (05f9846f8 2025-03-31)
binary: rustc
commit-hash: 05f9846f893b09a1be1fc8560e33fc3c815cfecb
commit-date: 2025-03-31
host: x86_64-unknown-linux-gnu
release: 1.86.0
LLVM version: 19.1.7

rustc 1.88.0-nightly (5e17a2a91 2025-04-05)
binary: rustc
commit-hash: 5e17a2a91dd7dbefd8b4a1087c2e42257457deeb
commit-date: 2025-04-05
host: x86_64-unknown-linux-gnu
release: 1.88.0-nightly
LLVM version: 20.1.2

This is specific to RISC-V. The same code works fine for x86_64 and aarch64 target. e.g.

rustc +nightly --target aarch64-unknown-linux-gnu -C code-model=large -C lto hello.rs -C linker=/bin/aarch64-linux-gnu-gcc 

[kxxt]

@rustbot label +O-riscv +A-LTO +A-LLVM +S-has-mcve

[kxxt]

rustc +1.42 --target riscv64gc-unknown-linux-gnu -C code-model=large -C lto hello.rs -C linker=/bin/riscv64-linux-gnu-gcc works. This is a regression.

@rustbot label +regression-from-stable-to-stable

[kxxt]

This regression is first introduced in 1.52.0.

[kxxt]

cargo-bisect-rustc:

********************************************************************************
Regression in nightly-2021-03-15
********************************************************************************

fetching https://static.rust-lang.org/dist/2021-03-14/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2021-03-14: 40 B / 40 B [=============================================================================================================================================================================] 100.00 % 1.29 MB/s converted 2021-03-14 to acca818928654807ed3bc1ce0e97df118f8716c8
fetching https://static.rust-lang.org/dist/2021-03-15/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2021-03-15: 40 B / 40 B [=============================================================================================================================================================================] 100.00 % 1.61 MB/s converted 2021-03-15 to d6eaea1c8860adb5302d2fbaad409e36585ab217
looking for regression commit between 2021-03-14 and 2021-03-15
fetching (via remote github) commits from max(acca818928654807ed3bc1ce0e97df118f8716c8, 2021-03-12) to d6eaea1c8860adb5302d2fbaad409e36585ab217
ending github query because we found starting sha: acca818928654807ed3bc1ce0e97df118f8716c8
get_commits_between returning commits, len: 8
  commit[0] 2021-03-13: Auto merge of #83064 - cjgillot:fhash, r=jackh726
  commit[1] 2021-03-14: Auto merge of #83105 - JohnTitor:rollup-tqpm8pb, r=JohnTitor
  commit[2] 2021-03-14: Auto merge of #83028 - GuillaumeGomez:prevent-js-error-if-no-filter, r=Nemo157
  commit[3] 2021-03-14: Auto merge of #83094 - GuillaumeGomez:crates-js-location, r=Nemo157
  commit[4] 2021-03-14: Auto merge of #83044 - kubo39:set-llvm-code-model, r=nikic
  commit[5] 2021-03-14: Auto merge of #83082 - cjgillot:defkey-ii, r=oli-obk
  commit[6] 2021-03-14: Auto merge of #82399 - petrochenkov:modin2, r=Aaron1011
  commit[7] 2021-03-14: Auto merge of #83062 - JohnTitor:improve-reassign-err, r=davidtwco
ERROR: no CI builds available between acca818928654807ed3bc1ce0e97df118f8716c8 and d6eaea1c8860adb5302d2fbaad409e36585ab217 within last 167 days

So it is probably this one: #83044

[kxxt]

Probably related to https://github.com/rust-lang/rust/blob/master/compiler/rustc_target/src/spec/targets/riscv64gc_unknown_linux_gnu.rs#L18

riscv is among the architectures with target_options.code_model set.

[apiraino]

#83044 was a while ago, unsure if something changed for the riscv64gc-unknown-linux-gnu compile target.

cc: riscv maintainers in case they have an opinion (sorry for the mass mention)

@kito-cheng
@michaelmaitland
@robin-randhawa-sifive
@topperc

[saethlin]

The code model for that target has always been "medium".

Rust 1.52.0 contains an LLVM upgrade. I bet the "regression" here is actually a fix from llvm/llvm-project#32653

[kxxt]

#83044 probably only reveals this bug instead of causing it.

warning: linking module flags 'Code Model': IDs have conflicting values: 'i32 3' from , and 'i32 4' from hello.4a3c48060548947b-cgu.0

There is an unnamed module carrying the medium code model, which is causing this bug. Any idea on how to find out what it is? The large code model from codegen options seems not applied to it.

[bjorn3]

Could be the allocator shim:

rust/compiler/rustc_codegen_llvm/src/lib.rs

Lines 116 to 122 in c6c1796

	let module_llvm = ModuleLlvm::new_metadata(tcx, module_name);
	let cx =
	SimpleCx::new(module_llvm.llmod(), &module_llvm.llcx, tcx.data_layout.pointer_size);
	unsafe {
	allocator::codegen(tcx, cx, module_name, kind, alloc_error_handler_kind);
	}
	module_llvm

[kxxt]

I think I get it. When code_model is set in TargetOptions(e.g. riscv64gc-unknown-linux-gnu sets it to Medium),
the generated LLVM bitcode embeded in fat lto object in libadler2-20d8a55a46d89958.rlib contains !2 = !{i32 1, !"Code Model", i32 3}, which then causes the above error when using -C code-model=large option because the prebuilt libadler2 rlib is using Medium code model and LLVM 12 starts to forbid that.

However, when code_model is not set in TargetOptions(e.g. x86_64), the generated LLVM bitcode embeded in fat lto object in libadler2-*.rlib uses the default code model and would not contain the Code Model option. But when we manually set -C code-model=whatever-not-the-default later when building hello.rs, we are still linking modules with different Code Models together. It is just that LLVM is not complaining about it because the prebuilt rlibs does not have Code Model set in their module flags, silently triggering undefined behavior on x86_64.

So I think #83044 is an incomplete fix and thus unfortunately using non-default code model with LTO is not possible in Rust without building std and all dependencies from source again. Session::code_model() should be updated to always return Some(DEFAULT_LLVM_CODE_MODEL_FOR_TARGET) in default case instead of returning None.

[kxxt]

I see that there is a None code model.

rust/compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp

Line 223 in c6c1796

None,

If not setting code model does not mean using the default code model and instead bitcode without code model makes sense. I think we should compile target rust sysroot without specifying the code model. Thoughts?