The FPU unroller (+90, -85) is checked into CVS. Had some last minute issues with 64 bit memory address emitting while loading constants. That has been worked around by using the following.
#define md_load_const_float_32(inst,reg,mem) \
do { \
amd64_push_reg((inst), MD_REG_PC); \
amd64_mov_reg_imm_size((inst), MD_REG_PC, (mem), 8); \
amd64_fld_membase((inst), (MD_REG_PC), (0), 0); \
amd64_pop_reg((inst), MD_REG_PC);\
}while(0)
Technically speaking the following code should've just worked.
#define md_load_const_float_32(inst,reg,mem) \ amd64_fld((inst), (reg), (mem))
But the amd64-codegen.h does not actually accept 64 bit pointers in there. Which is really strange, because if it doesn't, I don't understand how Mono has a working AMD64 jit (same amd64-codegen.h in both places). The direct amd64_fld would be a lot faster than an indirect membase read. There are more small things left to do in there, like SSE/2/3 opcodes for speed. For example there is a convenient cvttsd2si to convert floats and doubles to integers without too much mucking around. Also I plan to implement trig functions directly using the FPU sometime in the future. Here's an old patch of how it could be done for x86/amd64.
The float benchmark has gone up 3x approximately. From a November snapshot of the engine giving 540 on floats, it has gone up into giving 1671. I think that a further 50-60 should come in as soon as I go SSE on this codebase. And for the first time ever, my entire pnetmark has gone beyond 10,000 - 10257.
And now I have no excuse for not hacking libjit ;)
--Eat flaming death, you minicomputer mongrels