]> git.ipfire.org Git - thirdparty/gcc.git/commit
i386: correct integer division modeling in znver.md
authorAlexander Monakov <amonakov@ispras.ru>
Mon, 31 Oct 2022 14:35:57 +0000 (17:35 +0300)
committerAlexander Monakov <amonakov@ispras.ru>
Tue, 1 Nov 2022 12:20:57 +0000 (15:20 +0300)
commit5cee5f94000ee5eabce9b223c44c7923c1c69f61
tree628e3f4e76dc49e92652b88160826ebdc6e01bef
parent0122faae30fe1ad1dfa8c69f3d3f0428b996b600
i386: correct integer division modeling in znver.md

In znver.md, division instructions have descriptions like

(define_insn_reservation "znver1_idiv_DI" 41
                        (and (eq_attr "cpu" "znver1,znver2")
                             (and (eq_attr "type" "idiv")
                                  (and (eq_attr "mode" "DI")
                                       (eq_attr "memory" "none"))))
                        "znver1-double,znver1-ieu2*41")

which says that DImode idiv has latency 41 (which is correct) and that
it occupies 2nd integer execution unit for 41 consecutive cycles, but
that is not correct:

1) the division instruction is partially pipelined, and has throughput
   1/14, not 1/41;

2) for the most part it occupies a separate division unit, not the
   general arithmetic unit.

Evidently, interaction of such 41-cycle paths with the rest of
reservations causes a combinatorial explosion in the automaton.

Fix this by modeling the integer division unit properly, and correcting
reservations to use the measured reciprocal throughput of those
instructions (available from uops.info). A similar correction for
floating-point divisions is left for a followup patch.

Top 5 znver table sizes, before:

68692 r znver1_ieu_check
68692 r znver1_ieu_transitions
99792 r znver1_ieu_min_issue_delay
428108 r znver1_fp_min_issue_delay
856216 r znver1_fp_transitions

After:

1454 r znver1_ieu_translate
1454 r znver1_translate
2304 r znver1_ieu_transitions
428108 r znver1_fp_min_issue_delay
856216 r znver1_fp_transitions

gcc/ChangeLog:

PR target/87832
* config/i386/znver.md (znver1_idiv): New automaton.
(znver1-idiv): New unit.
(znver1_idiv_DI): Correct unit and cycles in the reservation.
(znver1_idiv_SI): Ditto.
(znver1_idiv_HI): Ditto.
(znver1_idiv_QI): Ditto.
(znver1_idiv_mem_DI): Ditto.
(znver1_idiv_mem_SI): Ditto.
(znver1_idiv_mem_HI): Ditto.
(znver1_idiv_mem_QI): Ditto.
(znver3_idiv_DI): Ditto.
(znver3_idiv_SI): Ditto.
(znver3_idiv_HI): Ditto.
(znver3_idiv_QI): Ditto.
(znver3_idiv_mem_DI): Ditto.
(znver3_idiv_mem_SI): Ditto.
(znver3_idiv_mem_HI): Ditto.
(znver3_idiv_mem_QI): Ditto.
gcc/config/i386/znver.md