On Wed, 23 Sep 2020, Ilya Leoshkevich via Gcc wrote:
> Hi,
>
> "Defining How to Split Instructions" in gccint states the following:
>
> The preparation-statements are similar to those statements that are
> specified for define_expand ... Unlike those in define_expand, however,
> these statements must not generate any new pseudo-registers.
>
> I see that there is code that does this anyway, e.g. "Split
> calculate_pic_address into pic_load_addr_* and a move." in arm.md.
>
> Is this restriction still valid today? Is there a reason we can't
> introduce new pseudos in a splitter before LRA?
There are passes splitting insns before register allocation.
The pattern after the quoted comment 6557 gates its call to
gen_reg_rtx conditional by can_create_pseudo_p (), thus is
valid.
IIUC the restriction is still in place, but the gccint section
you quote should be updated to the effect of "must not generate
any new pseudo-registers *after register allocation has started,
@xref{How-to-conditionalize-pre-post-reload}*" with the xref
referring to a new section to-be-written from pieces given by
grep reload_completed and can_create_pseudo_p.
brgds, H-P
