The following is the truly minimal fix for the middle-end issue
with SRA and memcpy folding interaction. I've tried more variants
that "make sense" but as they all end up folding slightly more
memcpy calls than before we run into optimization testcase
regressions in places that look for __builtin_memcpy and do
not deal with aggregate copies as MEM[..] = MEM[..];
Bootstrap and regtest running on x86_64-unknown-linux-gnu.
GCC 9 is the time we can try dealing with such fallout, it just
doesn't feel like the correct time to do now...
Richard.
2017-11-27 Richard Biener <rguent...@suse.de>
PR middle-end/83141
* gimple-fold.c (gimple_fold_builtin_memory_op): For aggregate
copies generated from memcpy use a character array as reference
type.
Index: gcc/gimple-fold.c
===================================================================
--- gcc/gimple-fold.c (revision 255172)
+++ gcc/gimple-fold.c (working copy)
@@ -1038,9 +1038,25 @@ gimple_fold_builtin_memory_op (gimple_st
gimple_assign_set_lhs (new_stmt, srcvar);
gimple_set_vuse (new_stmt, gimple_vuse (stmt));
gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
+ new_stmt = gimple_build_assign (destvar, srcvar);
+ goto set_vop_and_replace;
}
}
- new_stmt = gimple_build_assign (destvar, srcvar);
+
+ /* We get an aggregate copy. Use an unsigned char[] type to
+ perform the copying to preserve padding and to avoid any issues
+ with TREE_ADDRESSABLE types or float modes behavior on copying. */
+ desttype = build_array_type_nelts (unsigned_char_type_node,
+ tree_to_uhwi (len));
+ srctype = desttype;
+ if (src_align > TYPE_ALIGN (srctype))
+ srctype = build_aligned_type (srctype, src_align);
+ if (dest_align > TYPE_ALIGN (desttype))
+ desttype = build_aligned_type (desttype, dest_align);
+ new_stmt
+ = gimple_build_assign (fold_build2 (MEM_REF, desttype, dest, off0),
+ fold_build2 (MEM_REF, srctype, src, off0));
+set_vop_and_replace:
gimple_set_vuse (new_stmt, gimple_vuse (stmt));
gimple_set_vdef (new_stmt, gimple_vdef (stmt));
if (gimple_vdef (new_stmt)
