On Sun, Nov 26, 2017 at 01:40:06PM +0000, Stuart Henderson wrote: > On 2017/11/26 14:00, Sebastien Marie wrote: > > Hi, > > > > The following diff makes i386 to be compiled without --enable-rust-simd. > > simd is "Single instruction, multiple data" (aka MMX, SSE, SSE2...). > > > > rustc when compiled with external LLVM doesn't export target_features as > > compilation variable, making firefox building to fail. > > > > I workarounded the problem on amd64 by manually exported SSE2 symbol > > (always present on amd64). But for i386, same hack isn't possible (as > > not all i386 processors have such extensions). So it is preferable to > > not enable simd on i386. > > > > I tested the diff with rustc 1.22.1 and firefox still build well (but no > > runtime test). > > Is it actually needed? I can't imagine anything that can actually *run* > firefox (RAM-wise) that doesn't at least have SSE. >
Regarding firefox, I agree that SSE2 should be available on any processor that will use firefox. But the same hack for i386 isn't possible. It is a compiler change and I am unsure about controlled effects on rust programs. But it is possible to force simd crate to not check the target_feature, and always assumes SSE2 to be present. But it will be a new (unmergeable) patch on firefox and landry@ disagree. > Due to the serious lack of registers on i386 (especially with PIE) I think > it needs all the help it can get... My diff on rustc doesn't change the way the compiler uses LLVM. With external LLVM (opposed to using the embedded one), rustc has no way to get the list of target_feature that LLVM part has used to generate the code. So my hack on rustc is to always announce "SSE2" as present for amd64. For me, dropping --enable-rust-simd on i386 isn't problem for compilation. The impact will be performance wise at runtime for decoding pages (simd is used by encoding_rs). -- Sebastien Marie
