Jeff Law <law@redhat.com> writes:
On 09/18/14 04:07, Richard Sandiford wrote:
This series is a cleaned-up version of:
https://gcc.gnu.org/ml/gcc/2014-03/msg00163.html
The underlying problem is that the semantics of subregs depend on the
word size. You can't have a subreg for byte 2 of a 4-byte word, say,
but you can have a subreg for word 2 of a 4-word value (as well as lowpart
subregs of that word, etc.). This causes problems when an architecture has
wider-than-word registers, since the addressability of a word can then depend
on which register class is used.
The register allocators need to fix up cases where a subreg turns out to
be invalid for a particular class. This is really an extension of what
we need to do for CANNOT_CHANGE_MODE_CLASS.
Tested on x86_64-linux-gnu, powerpc64-linux-gnu and aarch64_be-elf.
I thought we fixed these problems long ago with the change to subreg_byte?!?
No, that was fixing something else. (I'm just about old enough to remember
that too!) The problem here is that (say):
(subreg:SI (reg:DI X) 4)
is independently addressable on little-endian AArch32 if X assigned
to a GPR, but not if X is assigned to a vector register. We need
to allow these kinds of subreg on pseudos in order to decompose multiword
arithmetic. It's then up to the RA to realise that a reload would be
needed if X were assigned to a vector register, since the upper half
of a vector register cannot be independently accessed.
Note that you could write this example even with the old word-style offsets
and IIRC the effect would have been the same.