From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jeffrey A Law To: Joern Rennecke Cc: m.hayes@elec.canterbury.ac.nz (Michael Hayes), gcc@gcc.gnu.org, amylaar@cygnus.com Subject: Re: Autoincrement examples Date: Tue, 30 Nov 1999 23:37:00 -0000 Message-ID: <373.942951315@upchuck> References: <199911181532.PAA09078@phal.cygnus.co.uk> X-SW-Source: 1999-11n/msg00491.html Message-ID: <19991130233700.F437EhxadlC_KNHv_DWijdt352iPapMYkBhCF_ta7bI@z> In message < 199911181532.PAA09078@phal.cygnus.co.uk >you write: > However, we are not supposed to generate autoincrement before flow. The > point of doin this in regmove is that we alerady got constraint information > there - using autoincrement instead of three-address-adds is much more > useful when you don't actually have a three-address add to start with. It seems to me that we can get this information during flow/life analysis just as easily as in regmove.c The interactions between loop & autoinc are clearly something we need to deal with. Both of you have noted the fact that loop opts tend to disguise autoinc opportunities. But I believe we can attack the problems separately -- ie, we can improve autoinc detection separately from making the loop optimizer generate code that is more autoinc friendly. The biggest obstacle in doing that is deciding precisely what sequences we want to detect for autoinc opts and which ones we're going to depend on loop generating autoinc friendly code. jeff