Wasm32: Rust should not spill v128 function params into shadow stack

You are allowed to pass core::arch::wasm32::v128 around even when the v128 target feature is disabled. In that case we have to pass v128 values by-ref. To prevent ABI mismatches when one crate disables v128 (libcore has it disabled) while other crates have it enabled, we have to unconditionally pass v128 by-ref.

github.com/rust-lang/rust

compiler/rustc_target/src/callconv/mod.rs

1f8e04d34

    819. BackendRepr::SimdVector { .. } => {


    820.     // This is a fun case! The gist of what this is doing is


    821.     // that we want callers and callees to always agree on the


    822.     // ABI of how they pass SIMD arguments. If we were to *not*


    823.     // make these arguments indirect then they'd be immediates


    824.     // in LLVM, which means that they'd used whatever the


    825.     // appropriate ABI is for the callee and the caller. That


    826.     // means, for example, if the caller doesn't have AVX


    827.     // enabled but the callee does, then passing an AVX argument


    828.     // across this boundary would cause corrupt data to show up.


    829.     //


    830.     // This problem is fixed by unconditionally passing SIMD


    831.     // arguments through memory between callers and callees


    832.     // which should get them all to agree on ABI regardless of


    833.     // target feature sets. Some more information about this


    834.     // issue can be found in #44367.


    835.     //


    836.     // We *could* do better in some cases, e.g. on x86_64 targets where SSE2 is


    837.     // required. However, it turns out that that makes LLVM worse at optimizing this


    838.     // code, so we pass things indirectly even there. See #139029 for more on that.

Dit bestand is afgekapt. origineel weergeven

It shouldn't matter too much in the general case though as vendor intrinsics are always inlined and you generally want to inline any simd helpers too.