Matthew Fortune <[email protected]> writes:
> Hi Catherine/Richard,
>
> I think there may be some impact on register move costs by introducing
> this class.
Yeah, I was worried about that too. I'm going to do some code comparison
tests for SE and MIPS16 to see what happens.
> Is it worth having mips_canonicalize_move_class return M16_REGS for
> M16_STORE_REGS to reduce the effect on costings? Given the extra
> register is only $0 then this would seem mostly acceptable albeit
> slightly strange.
This will happen automatically, since M16_STORE_REGS is just M16_REGS
without $16. (M16_STORE_REGS doesn't include $0, since that's handled
as a constant instead.)
However:
Index: config/mips/mips.c
===================================================================
--- config/mips/mips.c (revision 209307)
+++ config/mips/mips.c (working copy)
@@ -648,7 +648,7 @@ static mips_one_only_stub *mips16_set_fcsr_stub;
/* Index R is the smallest register class that contains register R. */
const enum reg_class mips_regno_to_class[FIRST_PSEUDO_REGISTER] = {
- LEA_REGS, LEA_REGS, M16_REGS, V1_REG,
+ M16_STORE_REGS, LEA_REGS, M16_REGS, V1_REG,
M16_REGS, M16_REGS, M16_REGS, M16_REGS,
LEA_REGS, LEA_REGS, LEA_REGS, LEA_REGS,
LEA_REGS, LEA_REGS, LEA_REGS, LEA_REGS,
this bit isn't right. We should leave entry 0 alone and change the
M16_REGS entries for every register except $16.
Thanks,
Richard
