Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
rustc: SIMD types use pointers in Rust's ABI
This commit changes the ABI of SIMD types in the "Rust" ABI to unconditionally be passed via pointers instead of being passed as immediates. This should fix a longstanding issue, rust-lang#44367, where SIMD-using programs ended up showing very odd behavior at runtime because the ABI between functions was mismatched. As a bit of a recap, this is sort of an LLVM bug and sort of an LLVM feature (today's behavior). LLVM will generate code for a function solely looking at the function it's generating, including calls to other functions. Let's then say you've got something that looks like: ```llvm define void @foo() { ; no target features enabled call void @bar(<i64 x 4> zeroinitializer) ret void } define void @bar(<i64 x 4>) #0 { ; enables the AVX feature ... } ``` LLVM will codegen the call to `bar` *without* using AVX registers becauase `foo` doesn't have access to these registers. Instead it's generated with emulation that uses two 128-bit registers. The `bar` function, on the other hand, will expect its argument in an AVX register (as it has AVX enabled). This means we've got a codegen problem! Comments on rust-lang#44367 have some more contexutal information but the crux of the issue is that if we want SIMD to work in general we'll need to ensure that whenever a function calls another they ABI of the arguments being passed is in agreement. One possible solution to this would be to insert "shim functions" where whenever a `target_feature` mismatch is detected the compiler inserts a shim function where you pass arguments via memory to the shim and then the shim loads the values and calls the target function (where the shim and the target have the same target features enabled). This unfortunately is quite nontrivial to implement in rustc today (especially when accounting for function pointers and such). This commit takes a different solution, *always* passing SIMD arguments through memory instead of passing as immediates. This strategy solves the problem at the LLVM layer because the ABI between two functions never uses SIMD registers. This also shouldn't be a hit to performance because SIMD performance is thought to often rely on inlining anyway, where a `call` instruction, even if using SIMD registers, would be disastrous to performance regardless. LLVM should then be more than capable of fixing all our memory usage to use registers instead after enough inlining has been performed. Note that there's a few caveats to this commit though: * The "platform intrinsic" ABI is omitted from "always pass via memory". This ABI is used to define intrinsics like `simd_shuffle4` where LLVM and rustc need to have the arguments as an immediate. * Additionally this commit does *not* fix the `extern` ("C") ABI. This means that the bug in rust-lang#44367 can still happen when using non-Rust-ABI functions. My hope is that before stabilization we can ban and/or warn about SIMD types in these functions (as AFAIK there's not much motivation to belong there anyway), but I'll leave that for a later commit and if this is merged I'll file a follow-up issue. All in all this... Closes rust-lang#44367
- Loading branch information