On 07/03/13 11:42, Richard Sandiford wrote:
Graham Stott <graham.st...@btinternet.com> writes:
Hi Richard,
This patch attemps to elimate the TEQ instruction div DIV/MOD instructions
if possible (i.e the numerator is known to be non-zero)
I have introduced and seperated UNSPEC UNSPEC_SET_HILO_NOTRAP
which is generation by a peephole2 when the trap is known not to be required.
The peephole's work by checking for a REG_EQUAL or/REG_EQUIV on the
instruction which sets the numerator for the DIV/MOD this doesn't catch all
possible cases but does catch the common cases where ths numerator is
set set in a insn immediately preceeding the DIV/MOD.
Nice idea, but TBH, I think I'd prefer to expose the trap code at expand time
and let the rtl optimisers try to remove it. I can't remember any correctness
reason why that wouldn't work. I think the only reason we don't do it yet is
because (before gimple) there was a preference for keeping the division as
a unit during early rtl optimisations.
Yea, we ought to be reasonably good at detecting non-zero ranges in VRP
and eliminating the trap. Doing it that way may also expose weaknesses
in VRP which obviously we'd like to fix.
jeff
