Uh oh!
There was an error while loading. Please reload this page.
Update toolchain to latest nightly (Nov 13) - #460
Conversation
- We now get warnings about unnecessary parentheses around `dyn` types. - Spelling of `sf.name.prefer_remapped_unconditionaly` was fixed (now ends in `lly`). - `Pointer::into_parts` has been replaced with `Pointer::prov_and_relative_offset`.
- `BuilderMethods::dynamic_alloca` was removed.
- `TargetDataLayout::pointer_{size,align}` were changed from fields to
methods.
- `AddressSpace::DATA` was renamed `AddressSpace::ZERO`.- `run_fat_lto`, `optimize_fat`, and `autodiff` were merged into `run_and_optimize_fat_lto`, and the parameters were changed. - `run_thin_lto` parameters were changed. - `codegen` parameters were changed. - `LtoModuleCodegen` was removed. - Minor error message changes: - Some line number changes in rustc. - Some error message wording tweaks. The LTO changes occurred in rust-lang/rust PR 143388 and PR 144062.
- `tcx.get_attrs_unchecked(...)` was replaced with `tcx.get_all_attrs(...)`. - `run_link` has been removed. - `BuilderMethods::tail_call` was added. It's currently `todo!()`. - Adjust expected test outputs for very minor error message formatting changes. - And update to glam 0.30.8.
This reverts commit 42dc34a.
schell
commented
Nov 27, 2025
@dvdplm is this PR ready to go or are you looking to collab on a fix for the ICE you encountered? |
dvdplm
commented
Nov 28, 2025
It is definitely not ready to go, and the ICE is still there. :/ I think the issue stems from the recent work to support C variadics in FFI, but that's speculation on my part (and a moving target). This PR should really be a draft, possibly as the starting point for further investigation. |
schell
commented
Dec 1, 2025
Hrm, I can't set the PR as draft, can you @dvdplm? |
| cgcx: &CodegenContext<Self>, | ||
| // FIXME(bjorn3): Limit LTO exports to these symbols | ||
| _exported_symbols_for_lto: &[String], | ||
| _each_linked_rlib_for_lto: &[PathBuf], // njn: ? |
There was a problem hiding this comment.
This is the set of rlibs that will be linked in during thin LTO. It is empty for thin local LTO. FWIW as of a couple of nightlies ago, if you return false from CodegenBackend::thin_lto_supported, thin local LTO will be unconditionally disabled (and -Clto=thin will be downgraded to no LTO) allowing you to replace run_thin_lto and optimize_thin with unreachable!().
| o.env = self.env.to_string().into(); | ||
| o.vendor = self.vendor.clone().into(); | ||
| // FIXME: this is problematic I believe. Used to be a `Cow` but is now an enum. | ||
| o.env = rustc_target::spec::Env::Unspecified; |
There was a problem hiding this comment.
You can use Env::Other(self.env.to_string().into()).
bjorn3
commented
Feb 19, 2026
FWIW I want to remove
Indeed. Due to array decay in C, |
fee1-dead
commented
Mar 3, 2026
@bjorn3 did you mean |
bjorn3
commented
Mar 4, 2026
With the exception of the one for the main shim, all |
Based on @nnethercote's work in #420, but with a more recent toolchain.
The toolchain update from
nightly-2025-06-30tonightly-2025-11-13uncovered a slew of minor incompatibilities and a major(?) ABI issue:rustc_target::callconv::PassModehandling in function ABIs changed upstream. The bulk of that work happened in #144529 with the introduction of the new internal#[rustc_pass_indirectly_in_non_rustic_abis]attribute.This change causes an ICE when we run the compiletests from
rust-gpuwith, e.g.cargo run -p compiletests --no-default-features --features "use-installed-tools" typed_buffer.rs -- --target-env vulkan1.1,vulkan1.2,vulkan1.3,vulkan1.4,spv1.3. The assertion we hit when compilingcoreis here:The assertion fires in our query hook (abi.rs:104) when calling the default provider's
fn_abi_of_instance, which now runs the new sanity check from PR #144529.rust-gpuconvertsextern "C"toextern "unadjusted"(line 47 in abi.rs), then attempts to force all arguments toPassMode::Directviareadjust_fn_abi()(line 68). However, PR #144529 introduced a new invariant that requires certain types marked with#[rustc_pass_indirectly_in_non_rustic_abis]to be passed indirectly, even in unadjusted ABIs.Example ICE output (Click to view)
I have a suspicion that this change is involved, but I have to admit that making
rust-gpuwork with a locally builtrustchas been quite challenging!The other changes in this PR are simple adjustments and tweaks to accommodate for upstream changes. I suspect that some of my hack-fixes here do not actually work, but it's hard to tell with the ICE in the way.