On 08/27/2012 10:36 AM, Richard Sandiford wrote:
Sandra Loosemore<san...@codesourcery.com> writes:
On 08/19/2012 11:22 AM, Richard Sandiford wrote:
Not sure whether a peephole is the right choice here. In practice,
I'd imagine these opportunities would only come from a DImode move of
$0 into a doubleword register, so we could simply emit the pattern in
mips_split_doubleword_move.
That would also allow us to use it for plain HI and LO. It wasn't
obvious from the patch why it was restricted to the DSP extension
registers.
Please also add a scan-assembler test.
How is this version of the fix?
Just to say that I've not forgotten about this. I'd still like to
remove the !TARGET_64BIT and ISA_HAS_DSP_MULT tests, because the
idea isn't specific to either.
Also, reviewing the patch made me realise that it might be better
to keep the move intact and simply use "mult" in the output code.
That's my fault for suggesting the wrong thing, though, so I was
hoping to find time this weekend to try it myself. The testsuite
stuff ended up taking up all the available time instead.
Richard,
Have you had time to think about this some more? I am not sure I can
guess how you'd like me to fix this patch now without some more specific
review and/or suggestions about where the optimization should happen and
what cases it should be extended to detect in addition to the dsp
accumulator multiplies.
-Sandra
