Re: [PATCH] Improve RTL ifcvt for empty else_bb (PR rtl-optimization/80491)

[Jeff Law]

On 04/25/2017 01:31 PM, Jakub Jelinek wrote:

Hi!

The following patch is a partial fix for PR80491, an improvement for
if-conversion if there is empty else_bb.

What we do right now is in that case we only look at the immediately
preceeding (non-debug/non-note) instruction before cond_earliest
and if it is not the set of x, we just turn it into attempt to
use previous value of x rather than whatever it has been initialized to.

On the testcases from the PR, we have:
(insn 7 4 8 2 (set (reg:DI 97 [ _15+8 ])
         (const_int 0 [0])) 81 {*movdi_internal}
      (nil))
(insn 8 7 9 2 (set (reg:DI 104 [ a_11(D)->low ])
         (mem:DI (reg/v/f:DI 101 [ a ]) [2 a_11(D)->low+0 S8 A64])) 81 
{*movdi_internal}
      (nil))
(insn 9 8 10 2 (set (reg:DI 105 [ b_12(D)->low ])
         (mem:DI (reg/v/f:DI 102 [ b ]) [2 b_12(D)->low+0 S8 A64])) 81 
{*movdi_internal}
      (nil))
(insn 10 9 11 2 (parallel [
             (set (reg:CCC 17 flags)
                 (compare:CCC (plus:DI (reg:DI 104 [ a_11(D)->low ])
                         (reg:DI 105 [ b_12(D)->low ]))
                     (reg:DI 104 [ a_11(D)->low ])))
             (set (reg:DI 103)
                 (plus:DI (reg:DI 104 [ a_11(D)->low ])
                     (reg:DI 105 [ b_12(D)->low ])))
         ]) 316 {*adddi3_cc_overflow_1}
      (expr_list:REG_DEAD (reg:DI 105 [ b_12(D)->low ])
         (expr_list:REG_DEAD (reg:DI 104 [ a_11(D)->low ])
             (nil))))
(jump_insn 11 10 14 2 (set (pc)
         (if_then_else (ltu (reg:CCC 17 flags)
                 (const_int 0 [0]))
             (label_ref 14)
             (pc))) 617 {*jcc_1}
      (expr_list:REG_DEAD (reg:CCC 17 flags)
         (int_list:REG_BR_PROB 4 (nil)))
  -> 14)
(code_label 14 11 35 3 3 (nil) [1 uses])
(note 35 14 15 3 [bb 3] NOTE_INSN_BASIC_BLOCK)
(insn 15 35 16 3 (set (reg:DI 97 [ _15+8 ])
         (const_int 1 [0x1])) 81 {*movdi_internal}
      (nil))
(code_label 16 15 36 4 2 (nil) [0 uses])
If insn 7 would come after insn 9 (which doesn't change the behavior, as
insn 8 and insn 9 don't clobber pseudo 97 and const_int 0 is constant),
we'd turn that into a setcc pattern, but otherwise we fail.

This patch let us search for x's setter earlier in the bb.
During testing I found that modified_in_p/modified_in_between_p don't
actually take into account calls that could change MEMs, so the patch
handles that too.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

During bootstrap/regtest it seems to trigger over 10000 times, but
my statistics collection only noted cases where it found an earlier setter
and the noce_process_if_block has been successful, has not attempted to
check if insn_b/set_b was NULL whether it would otherwise fail (but even
in cases where it doesn't fail perhaps with the patch it can generate
better code).

2017-04-25  Jakub Jelinek  <ja...@redhat.com>

        PR rtl-optimization/80491
        * ifcvt.c (noce_process_if_block): When looking for x setter
        with missing else_bb, don't check only the insn right before
        cond_earliest, but look for the last insn that x is modified in
        within the same bb.

OK.

jeff