Rework benchmarks to make it easier to get assembly.

[josephlr]

This change:

• Moves the benchmarks from mod.rs to buffer.rs
• Passes a &[u8] to test::black_box for both benchmarks
• Move the inner loop we benchmark into an #[inline(never)] function
• Includes instructions for getting the ASM for a specific benchmark

This should hopefully reduce the variance of these benchmarks and make it easier to figure out if we are emitting the assembly or IR we expect for a particular implementation.

Signed-off-by: Joe Richey joerichey@google.com

[josephlr]

This PR came about because I discovered the amazing cargo-show-asm tool, and I wanted to figure out a way to make it easy to use this tool with our benchmarks.

After installing the tool, we can run cargo asm --bench buffer --release buffer::p384::bench_getrandom::inner. Then after some minor cleanups, we get:

buffer::p384::bench_getrandom::inner:pushrbxsubrsp,64 ; Zero the bufferxorpsxmm0,xmm0movaps xmmword ptr [rsp+48],xmm0movaps xmmword ptr [rsp+32],xmm0movaps xmmword ptr [rsp+16],xmm0 ; Call the funtionlearbx,[rsp+16]movesi,48movrdi,rbxcall qword ptr [rip+ getrandom::imp::getrandom_inner@GOTPCREL] ; Check for errortesteax,eaxjne .LBB17_1 ; test::black_box(slice);mov qword ptr [rsp],rbxmov qword ptr [rsp+8],48movrax,rspaddrsp,64poprbxret

We can see the effect of using getrandom_uninit:

buffer::p384::bench_getrandom_uninit::inner:pushrbxsubrsp,64learbx,[rsp+16]movesi,48movrdi,rbxcall qword ptr [rip+ getrandom::imp::getrandom_inner@GOTPCREL]testeax,eaxjne .LBB18_1mov qword ptr [rsp],rbxmov qword ptr [rsp+8],48movrax,rspaddrsp,64poprbxret

As the benchmarks are compiled as separate crates, we can see the effect of inlining. Removing the #[inline] from getrandom_uninit() gives:

buffer::p384::bench_getrandom_uninit::inner:pushrbxsubrsp,80learbx,[rsp+16]learsi,[rsp+32]movedx,48movrdi,rbxcall qword ptr [rip+ getrandom::getrandom_uninit@GOTPCREL]movrax, qword ptr [rsp+16]testrax,raxje .LBB18_1movrcx, qword ptr [rsp+24]mov qword ptr [rsp+16],raxmov qword ptr [rsp+24],rcxaddrsp,80poprbxret

We can also see that passing the entire array to test::black_box creates an additional copy:

buffer::p384::bench_getrandom_uninit::inner:subrsp,104leardi,[rsp+56]movesi,48call qword ptr [rip+ getrandom::imp::getrandom_inner@GOTPCREL]testeax,eaxjne .LBB18_1 ; 48 byte copymovupsxmm0, xmmword ptr [rsp+56]movupsxmm1, xmmword ptr [rsp+72]movupsxmm2, xmmword ptr [rsp+88]movaps xmmword ptr [rsp+32],xmm2movaps xmmword ptr [rsp+16],xmm1movaps xmmword ptr [rsp],xmm0movrax,rspaddrsp,104ret

@briansmith this relates to #291 (comment) about how the type you pass to black_box affects the code generation, so when comparing implementations, we should be sure to use the same types here.

[briansmith]

No major objections from me.

• Passes a &[u8] to test::black_box for both benchmarks

I think most users don't really want a &[u8] but rather want a [u8; N] or even a stronger type (see the getrandom_array discussion). The getrandom benchmark does produce an [u8; N] but the getrandom_uninit benchmark never does. It is good that you got rid of the extra copy caused by passing the entire array to black_box but it would be better to pass a &[u8; N] instead. In reality the assume_init() call that would be required is almost definitely going to get optimized away; however, according to bug reports from last year that wasn't always the case.