Reload pass ignores constraints. Why?
Wed Apr 15 16:55:00 GMT 2009
Ian Lance Taylor schrieb:
> Georg-Johann Lay <firstname.lastname@example.org> writes:
>>> that--if LEGITIMATE_CONSTANT_P accepts 0xffffff--then you need to ensure
>>> that your movsi insn will handle 0xffffff directly, without using any
>>> pseudo-registers when can_create_pseudo_p returns false.
>> That works, of course. But I must admit that I prefer to express what
>> is going on in terms of algebra, i.e. in terms of RTL instead of
>> acting as if the core could handle the constant and just printng out
>> some asm sequencs. movsi expands constants that cannot be loaded in
>> one machine instruction to a movesi_insn and an arithmetic insn, and
>> movsi_insn therefore allows only constants that are easy to load.
> You shouldn't print out some asm sequence, you should make movsi a
> define_expand which emits a sequence of insns which do not require new
> pseudo-registers. See, e.g., mips_move_integer which is called
> (indirectly) by the mov<mode> define_expand in mips.md.
Up to now, I see the following stategies:
i) Expand to a single insn that matched by *movsi_insn:
Bad, because not optimal and the insn output will print several
asm statements en bloc (the movsi expander does not print
anything, of course)
ii) Expand into MOV+LSHIFTRT and deny the resulting const in
*movsi_insn predicate and condition:
Bad, because it crashes reload (as we saw above)
iii) Expand into MOV+LSHIFTRT and allow the resulting const in
CSE et. al. will reconstruct the original constant and
replace MOV+LSHIFTRT with a single SET:
Bad: expanding was in vain and we Goto i)
iv) Expand into MOV+LSHIFTRT and remove the const from
Bad, because constant will end up in constant pool.
v) Expand into MOV+ADD sequence. Works(?), but
Bad: Code bloat of 100% compared to optimum.
Ok, I could catch it in a peep2...
The difference is that in contrast to lshiftrt the add
can handle the required addition without need of reloads.
vi) Expand into MOV+LSHIFTRT and allow the constand only if
Bad: runs into ICE
I didn't follow this path any further. Looks too hacky.
The movMM expa
>> What I do not understand is that a MOV/ADD sequence (which covers
>> large constants) works on RTL level, whereas MOV/SHIFTRT (which is
>> more efficient in some cases) shreds global alloc. Other strategies
>> could be MOV/[AND|IOR|XOR|BSWAP...] which won't work either, though.
> I don't know exactly what is going on. But it is most likely just a
> coincidence that it is failing when using SHIFTRT. There is probably
> some way to make it fail in other ways as well.
Would state it like this: If the movMM expander expands the move into
several insns, each insn must be able to handle an alternative (which
reload might select) without needing a reload.
More information about the Gcc-help