[Bug tree-optimization/50444] [4.6/4.7 Regression] -ftree-sra ignores alignment

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50444

--- Comment #16 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-01-20 15:46:40 UTC ---
(In reply to comment #15)
> Created attachment 26395 [details]
> other candidate patch
> 
> I'm testing the following patch instead, which avoids changing access types
> for all-scalar across-link propagations (we're going to create proper V_C_Es
> later).  I also remove the fancy code that tries to avoid adding V_C_Es,
> it looks it will cause more trouble than missed-optimizations.
> 
> That way we completely avoid needing to care for alignment at that particular
> places.  Whether the aggregate copy across-link propagation is affected in
> a similar way remains to be seen.
> 
> I'll see if I run into the same issue as you and investigate that.

gcc.dg/torture/pr47228.c shows that we rely on the build-ref-for-model
path in sra_modify_assign as we scalarize

struct S4
{
  unsigned f0:24;
} __attribute__((__packed__));

to unsigned int : 24, which is of different size, so we refuse to
VIEW_CONVERT the SImode register to BLKmode.  I'm not entirely sure
what's the best cause of action here, but certainly either detecting
the size-mismatch issue at analysis phase or papering over the issue
with build-ref-for-model (which might not always suceed?!).

Other FAILs this patch causes are

Running target unix/
FAIL: gcc.dg/tree-ssa/pr45144.c scan-tree-dump optimized " =
VIEW_CONVERT_EXPR<u
nsigned int>\\(a\\);"

Running target unix/-m32
FAIL: gcc.dg/torture/pr45678-2.c  -Os  execution test
FAIL: gcc.dg/tree-ssa/pr45144.c scan-tree-dump optimized " =
VIEW_CONVERT_EXPR<u
nsigned int>\\(a\\);"

Running target unix/
FAIL: 20_util/hash/chi2_quality.cc execution test
FAIL: 23_containers/forward_list/capacity/resize_size.cc execution test
FAIL: 23_containers/forward_list/modifiers/2.cc execution test
FAIL: 23_containers/list/operations/3.cc execution test
FAIL: 23_containers/list/operations/3_c++0x.cc execution test
FAIL: ext/pb_ds/example/basic_map.cc execution test
FAIL: ext/pb_ds/example/basic_multiset.cc execution test
FAIL: ext/pb_ds/example/basic_set.cc execution test
FAIL: ext/pb_ds/example/erase_if.cc execution test
FAIL: ext/pb_ds/example/tree_intervals.cc execution test
FAIL: ext/pb_ds/example/tree_join.cc execution test
FAIL: ext/pb_ds/example/tree_order_statistics.cc execution test
FAIL: ext/pb_ds/regression/associative_containers.cc execution test
FAIL: ext/pb_ds/regression/tree_map_rand.cc execution test
FAIL: ext/pb_ds/regression/tree_set_rand.cc execution test

Running target unix//-m32
FAIL: 23_containers/list/operations/3_c++0x.cc execution test
FAIL: 25_algorithms/nth_element/2.cc execution test

I'm going to test the two parts of the patch separately now.