https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92140
--- Comment #18 from Segher Boessenkool <segher at gcc dot gnu.org> ---
(In reply to Uroš Bizjak from comment #15)
> Is it possible to lift the limitation from the combine pass, where the
> combine tries to split the insn, but expects exactly two new insn patterns
> to be generated. From the docs:
>
> --q--
> When the combiner phase tries to split an insn pattern, it is always
> the case that the pattern is _not_ matched by any 'define_insn'. The
> combiner pass first tries to split a single 'set' expression and then
> the same 'set' expression inside a 'parallel', but followed by a
> 'clobber' of a pseudo-reg to use as a scratch register. In these cases,
> the combiner expects exactly two new insn patterns to be generated. It
> will verify that these patterns match some 'define_insn' definitions, so
> you need not do this test in the 'define_split' (of course, there is no
> point in writing a 'define_split' that will never produce insns that
> match).
> --/q--
>
> Many define_and_split insns in i386.md that use can_create_pseudo_p have
> this unwanted condition in their insn condition just because of the
> mentioned limitation.
What unwanted condition? What limitation?
From combine.c:
/* If we were combining three insns and the result is a simple SET
with no ASM_OPERANDS that wasn't recognized, try to split it into two
insns. There are two ways to do this. It can be split using a
machine-specific method (like when you have an addition of a large
constant) or by combine in the function find_split_point. */
/* See if the MD file can split NEWPAT. If it can't, see if letting it
use I2DEST as a scratch register will help. In the latter case,
convert I2DEST to the mode of the source of NEWPAT if we can. */
/* ??? Reusing i2dest without resetting the reg_stat entry for it
(temporarily, until we are committed to this instruction
combination) does not work: for example, any call to nonzero_bits
on the register (from a splitter in the MD file, for example)
will get the old information, which is invalid.
Since nowadays we can create registers during combine just fine,
we should just create a new one here, not reuse i2dest. */
Exactly two new insns... Exactly one is also okay. I'll fix the docs.
