Hi Modi, > The zero extract now matching against other modes would generate a test + > branch rather > than the combined instruction which led to the code size regression. I've > updated the patch > so that tbnz etc. matches GPI and that brings code size down to <0.2% in > spec2017 and <0.4% in spec2006.
That's looking better indeed. I notice there are still differences, eg. tbz/tbnz counts are significantly different in perlbench, with ~350 missed cases overall (mostly tbz reg, #7). There are also more uses of uxtw, ubfiz, sbfiz - for example I see cases like this in namd: 42c7dc: 13007400 sbfx w0, w0, #0, #30 42c7e0: 937c7c00 sbfiz x0, x0, #4, #32 So it would be a good idea to check any benchmarks where there is still a non-trivial codesize difference. You can get a quick idea what is happening by grepping for instructions like this: grep -c sbfiz out1.txt out2.txt out1.txt:872 out2.txt:934 grep -c tbnz out1.txt out2.txt out1.txt:5189 out2.txt:4989 > Can you send me the necessary documents to make that happen? Thanks! That's something you need to sort out with the fsf. There is a mailing list for this: ass...@gnu.org. Cheers, Wilco
