2011/9/16 Georg-Johann Lay <a...@gjlay.de>: > Georg-Johann Lay schrieb: >> Denis Chertykov schrieb: >>> 2011/9/12 Georg-Johann Lay <a...@gjlay.de>: >>>> This patch introduces patterns for multiply-add and multiply-sub. >>>> >>>> On the enhanced core, these operations can be performed with the product >>>> in R0; >>>> there is no need to MOVW it out of that register. The code is smaller and >>>> faster and has lower register pressure. >>>> >>>> Tested without regressions. >>>> >>>> Ok to commit? >>> Ok. >>> >>> Denis. >> >> This is the second part to fix this PR; it introduced multiply-add/-sub for >> QImode and one insn for HI = sign_extend (QI << 1). >> >> With this patch PR50358 is fixed up to some corner cases. >> >> The insns with CONST_INT split the load of the constant after reload. >> avr_rtx_costs describes these costs, but it would be advantageous to do the >> split pre-reload because IRA/reload could reuse constants. >> >> The trouble is that reload_in_progress is false in IRA and therefore the >> patterns match in IRA, so here is the same trouble I faced in the patch for >> widening multiply where a function like avr_gate_split1() was regarded as to >> hackish because it tested the pass-number to help out. Without such a >> function >> in the insn condition, the insn matches in IRA and a register is re-replaced >> with CONST_INT again, leading to crash in split2 because of gen_reg_rtx or >> because of !reload_completed in the insn condition. >> >> So this patch comes without reusing constants. >> >> Besides that, the "Write as one pattern" changes gather two insns and write >> them down as one; that's no functional change, it's just about using >> iterators >> to reduce lines of code. The order of insns changes, but that does not >> matter >> here. I didn't make an extra patch for that. > > Split the two patches > > Ok to install? >
Please, commit. Denis.
