https://gcc.gnu.org/bugzilla/show_bug.cgi?id=49330
--- Comment #24 from Richard Biener <rguenth at gcc dot gnu.org> ---
On GCC testcases one large group of MEMs only disambiguated through
base_alias_check is disambiguations agains DSEs group_info->base_mem
which is BLKmode mems based on some "base" pointer. This base_mem
lacks a MEM_EXPR but I think it shouldn't be difficult to add one,
like with (completely lacking sanity testing):
diff --git a/gcc/dse.c b/gcc/dse.c
index 389c52d4284..098c77165de 100644
--- a/gcc/dse.c
+++ b/gcc/dse.c
@@ -1097,6 +1097,7 @@ canon_address (rtx mem,
{
machine_mode address_mode = get_address_mode (mem);
rtx mem_address = XEXP (mem, 0);
+ tree mem_expr = MEM_EXPR (mem);
rtx expanded_address, address;
int expanded;
@@ -1165,6 +1166,9 @@ canon_address (rtx mem,
&& const_or_frame_p (address))
{
group_info *group = get_group_info (address);
+ if (!MEM_EXPR (group->base_mem)
+ && mem_expr)
+ set_mem_expr (group->base_mem, get_base_address (mem_expr));
if (dump_file && (dump_flags & TDF_DETAILS))
{
btw, the disambiguations like
(mem/c:SI (symbol_ref:DI ("g") [flags 0x2] <var_decl 0x7fe01f78a510 g>) [1 g+0
S4 A32])
vs. (mem:DI (pre_dec:DI (reg/f:DI 7 sp)) [0 S8 A8])
are handled through REG_BASE_VALUE which assigns 'sp' (address:DI -1).
I believe we should be working towards adding proper MEM_EXPRs to more
places and simply make find_base_term more conservative which means
simplifying the PLUS/MINUS cases.
