https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69671
--- Comment #5 from Kirill Yukhin <kyukhin at gcc dot gnu.org> ---
(In reply to ktkachov from comment #3)
> CC'ing Kirill for AVX512 opinion
I suppose that there's something wrong w/ MD patterns.
E.g. for example provided pattern is:
;; /export/users/kyukhin/gcc/git/gcc/gcc/config/i386/sse.md: 9199
(define_insn ("avx512vl_truncatev4siv4qi2_mask")
[
(set (match_operand:V16QI 0 ("register_operand") ("=v"))
(vec_concat:V16QI (vec_merge:V4QI (truncate:V4QI
(match_operand:V4SI 1 ("register_operand") ("v")))
(vec_select:V4QI (match_operand:V16QI 2
("vector_move_operand") ("0C"))
(parallel [
(const_int 0 [0])
(const_int 1 [0x1])
(const_int 2 [0x2])
(const_int 3 [0x3])
]))
(match_operand:QI 3 ("register_operand") ("Yk")))
(const_vector:V12QI [
(const_int 0 [0])
(const_int 0 [0])
(const_int 0 [0])
(const_int 0 [0])
(const_int 0 [0])
(const_int 0 [0])
(const_int 0 [0])
(const_int 0 [0])
(const_int 0 [0])
(const_int 0 [0])
(const_int 0 [0])
(const_int 0 [0])
Right now I think that 2nd operand predicate is not correct.
It should be const0_rtx (of corresponding mode) or duplicate of operand 0
(result actually)
This is whats contstraint tells.
However predicate says simply that operand is either const0_rtx or
nonimmediate: no connection with operand 0.
