https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64882
--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> --- (In reply to Jakub Jelinek from comment #7) > But what about vsib_address_operand, address_mpx_no_base_operand, > address_mpx_no_index_operand? These *do* check modes, the problematic predicate is defined as a special predicate (AKA "define_special_predicate") > Even address_operand is used directly: > [(prefetch (match_operand:P 0 "address_operand" "p") > (const_int 1) > (const_int 2))] Hm... don't know the reason, but effectively address_operand is a special predicate, too. > The comment above address_operand is definitely confusing, because when it > is used directly in the *.md (various ports), then the mode of the address > is passed to it, the comment above address_operand is confusing: > "Return 1 if OP is a valid memory address for a memory reference of mode > MODE." > and memory_address_p it calls wants address of the MEM, not the address. FYI, testcase tweaked to be 32-bit clean and still exhibit the issue with -m64: With the patch from Comment #5? It works for me with patched gcc (BTW: I think this testcase should be put into torture directory).
