On Feb 18, 2015, at 3:32 PM, Maxim Kuvyrkov <maxim.kuvyr...@linaro.org> wrote:
> First of all, applauses! I realize how difficult it was to reduce this > problem. > > Your patch looks OK to me, but I can't shake off feeling that it will > pessimize cases when d8 is not used afterwards. In particular, your patch > makes it impossible to use same register for output (operand 0) and inputs > (operands 1 and 2). > > Did you consider using SCRATCHes instead of re-using operand 0 with early > clobber like in the attached [untested] patch? If I got it all correct, > register allocator will get more freedom in deciding which register to use > for negated shift temporary, while still allowing reusing register from > operand 0 for one of the inputs. There is a typo in the patch I sent (mode for last match_scratch should be QI, not DI). Corrected patch attached. -- Maxim Kuvyrkov www.linaro.org
bz1149.patch
Description: Binary data
