On Tue, 24 Jul 2012, Richard Henderson wrote:
> On 07/21/2012 06:10 AM, Oleg Endo wrote: > > I think on SH the cost test in lshift_cheap_p with > > gen_rtx_ASHIFT (word_mode, const1_rtx, reg), speed_p); > > > > will always 'fail', because of sh.c (shiftcosts): > > /* There is no pattern for constant first operand. */ > > if (CONST_INT_P (XEXP (x, 0))) > > return MAX_COST; > > > > On SH3 / SH4* / SH2A there is a dynamic shift that does "reg << reg" or > > "reg >> reg", which is not that expensive actually. However, the > > constant "1" must be loaded into a register first. I'm currently trying > > to brush up the shift code in SH a little bit and could add some things > > to handle the "const << reg" case. > > Having to load the "1" into a register is something common to all targets. IIUC, no, not for testing a single bit and/or several bits from bit 0 up to the numbered bit. See cris *btst and m68k (several unnamed patterns with about the same contents, setting cc0, calling output_btst) r188179 m68k.md:666 (evil patterns?) > I shouldn't think that SH should need to do anything special. Indeed, that > conditional that you quote above looks very wrong. I believe you should > simply recurse (or let the rtx_cost driver recurse) to find the cost of > arguments to the shift. > > > r~ >
