On Wed, 2011-07-20 at 11:03 +0200, Richard Guenther wrote: > I wonder if the code below triggered at all as since we expand from > SSA we no longer see the larger trees in-place but you have to > look them up via SSA defs using get_gimple_for_ssa_name (or the > helper get_def_for_expr). So I expect that the check for a MULT_EXPR > offset only will trigger if the memory reference was still in the > form of a variable ARRAY_REF (then get_inner_reference will produce > an offset of the form i * sizeof (element)). I think what we want > to catch here is also the case of *(ptr + i) which will not show > up as ARRAY_REF but as MEM[ptr + 0] where ptr is an SSA name > with a defining statement doing ptr' + i * sizeof (element). > > Richard.
Confirmed. The compiler is pretty smart about reconstructing ARRAY_REFs out of *(p + i) if p is known to point to an array, but if p is a scalar pointer with unknown target the code doesn't kick in. I'll cobble up a version that tests for both patterns and get some new numbers. Thanks, Bill
