MMX Locals
Move MMX registers to local array, so clang can optimise a bit better, i.e.:
- Keep them in host's registers
- Assume a clean slate/set of values (rather than having to preserve state across funcs)
Where this happens
run in codegen/optimizations/mmx_locals.rs.
How it works
Same idea as x87 locals. So this page is kinda a stub, mkay?
MMX registers (mm0 to mm7) normally live in ctx.fpu.mmx.
Other functions might read them, so clang must keep every write.
When it's safe, we move them to a local uint64_t mmx[8].
Then clang can keep them in CPU registers and drop writes nothing reads.
It's safe when:
mmregister isn't a param- called functions don't need it.
- no mid-assembly hooks (<= kinda overkill, can do syncs, but alas, not needed rn)
Assumption
We assume callers don't read the mm values a function leaves.
This may be specific to MSVC. I've not verified it across a variety of games, so tread caution.
Consider this 'experimental', nya~ 🐱
Before and after
sub_4023C0 from Shadow the Hedgehog.
// movq mm0, [eax]
ctx.fpu.mmx[0] = X86_LOAD_U64(ctx.eax.u32);
// pavgb mm0, [ebx]
uint64_t dst_val_4203511 = MMX_REG(0);
uint64_t src_val_4203511 = X86_MMX_LOAD(ctx.ebx.u32);
MMX_REG(0) = X86_PAVGB(dst_val_4203511, src_val_4203511);
// movq [ecx], mm0
X86_STORE_U64(ctx.ecx.u32, ctx.fpu.mmx[0]);
uint64_t mmx[8]; // <- at the top of the function
// movq mm0, [eax]
mmx[0] = X86_LOAD_U64(ctx.eax.u32);
// pavgb mm0, [ebx]
uint64_t dst_val_4203511 = mmx[0];
uint64_t src_val_4203511 = X86_MMX_LOAD(ctx.ebx.u32);
mmx[0] = X86_PAVGB(dst_val_4203511, src_val_4203511);
// movq [ecx], mm0
X86_STORE_U64(ctx.ecx.u32, mmx[0]);