http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46801
Richard Guenther <rguenth at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Target Milestone|--- |4.6.0 --- Comment #11 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-01-30 14:36:37 UTC --- SRA does Pack9.Copy (struct pack9__r2 * const x, struct pack9__r2 * const y) { + integer t$i2; const struct pack9__r2 t; integer D.2584; integer D.2583; <bb 2>: t = *y_1(D); - D.2583_2 = t.i2; + t$i2_9 = y_1(D)->i2; + D.2583_2 = t$i2_9; D.2584_3 = y_1(D)->i2; D.2584_4 = D.2584_3; if (D.2583_2 != D.2584_4) @@ -206,6 +58,7 @@ <bb 4>: *x_5(D) = t; + x_5(D)->i2 = t$i2_9; return; } thus eliminates 't' and makes D.2583_2 and D.2584_3 redundant (and VN figure that out and remove the if stmt). Value-numbering does not see that in <bb 2>: t = *y_1(D); <bb 3>: D.2584_2 = t.i2; D.2585_3 = y_1(D)->i2; the two loads are the same (because it enters (only) the non-rewritten ops into the hashtable). See PR52054.