Richard Guenther <rguent...@suse.de> wrote on 02/08/2011 01:33:49 PM: > > On Tue, 2 Aug 2011, Ira Rosen wrote: > > > > > > + /* Now we have matched the statement pattern > > > + > > > + rhs1 = (T1)x; > > > + rhs2 = (T1)y; > > > + op_result = rhs1 OP rhs2; > > > + lhs = (T2)op_result; > > > > Just a note that the patch I proposed for the vectorizer ( > > http://gcc.gnu.org/ml/gcc-patches/2011-07/msg01472.html) also handles > > constants, multiple statements (i.e., op_result doesn't have to be promoted > > itself, but the sequence needs to end up with a promotion), and also it may > > use an intermediate type for OP. The tests in my patch don't match the > > pattern this patch detects. > > Ok, I only looked at the description of your patch, not the patch itself. > > The patch already handles constant 2nd operands. > > It shouldn't be difficult to handle multiple statements here, either by > instead of the above match only > > op_result = rhs1 OP rhs2; > lhs = (T2)op_result; > > and thus allow iteration on the promoted/demoted operation operands > or by collecting all defs first. > > Do you handle arbitrary def trees or only a linear chain as suggested > by > > + S2 x_T = (TYPE) x_t; > + S3 res0_T = op (x_T, C0); > + S4 res1_T = op (res0_T, C1); > + S5 ... = () res1_T; - type demotion > > ? Thus, do you handle res1_T = op (res0_T, res2_T) with a possibly > different TYPE in its def?
Only linear chains. But it doesn't seem too complicated to only check if res2_T is a result of a type promotion. Thanks, Ira > The case of > > op_result = rhs1 OP CST; > lhs = (T2)op_result; > > is probably always profitable to demote to > > rhs1' = (T2)rhs1; > lhs = rhs1' OP (T2)CST; > > and "iterating" that should be simple (handling two variable > operands will probably get a bit convoluted). > > Thanks, > Richard.
