[PATCH] update edge profile info in nvptx.c

[Tom de Vries]

On 07/20/2017 03:04 PM, Tom de Vries wrote:
> On 07/13/2017 06:53 PM, Cesar Philippidis wrote:
>> Similarly, for nvptx vector reductions, when it comes time to initialize
>> the reduction variable, the nvptx BE constructs a branch so that only
>> vector lanes 1 to vector_length-1 are initialized the the default value
>> for a given reduction type, where vector lane 0 retains the original
>> value of the reduction variable. For similar reason to the gang and
>> worker reductions, I set the probability of the new edge introduced for
>> the vector reduction to even.
>>
> 
> Hi,
> 
> The problem that you describe in abstract term looks like this concretely:
> ....
> (gdb) call debug_bb_n (4)
> ;; basic block 4, loop depth 0, freq 662, maybe hot
> ;;  prev block 3, next block 16, flags: (VISITED)
> ;;  pred:       3 [always (guessed)]  (FALLTHRU,EXECUTABLE)
> # VUSE <.MEM_61>
> # PT = nonlocal unit-escaped null
> _18 = MEM[(const struct .omp_data_t.33D.1518 &).omp_data_i_9(D)
>            clique 1 base 1].s2D.1519;
> # VUSE <.MEMD.1540>
> # USE = anything
> _72 = GOACC_DIM_POS (2);
> if (_72 != 0)
>    goto <bb 16>; [100.00%] [count: INV]
> else
>    goto <bb 17>; [INV] [count: INV]
> ;;  succ:       16 [always]  (TRUE_VALUE)
> ;;              17 (FALSE_VALUE)
> ...
> 
> The edge to bb16 has probability 100%. The edge to bb17 has no 
> probability set.
> 

Hi,


I.

the patch below fixes the probabilities on the outgoing edges, setting 
them to even:
...
(gdb) call debug_bb_n (4)
;; basic block 4, loop depth 0, freq 662, maybe hot
;;  prev block 3, next block 16, flags: (VISITED)
;;  pred:       3 [always (guessed)]  (FALLTHRU,EXECUTABLE)
# VUSE <.MEM_61>
# PT = nonlocal unit-escaped null
_18 = MEM[(const struct .omp_data_t.33D.1518 &).omp_data_i_9(D) clique 1 
base 1].s2D.1519;
# VUSE <.MEMD.1540>
# USE = anything
_72 = GOACC_DIM_POS (2);
if (_72 != 0)
   goto <bb 16>; [50.00%] [count: INV]
else
   goto <bb 17>; [50.00%] [count: INV]
;;  succ:       16 [50.0% (adjusted)]  (TRUE_VALUE)
;;              17 [50.0% (adjusted)]  (FALSE_VALUE)
...


II.

The quality is 'adjusted'. [ Even() first calls always() which has 
quality precise, and then applies scale(), which downgrades the quality 
from 'precise' to 'adjusted'. ]

The reason for that is explained in this comment AFAIU:
...
    Named probabilities except for never/always are assumed to be
    statically guessed and thus not necessarily accurate.
...

When I look at the definitions of 'adjusted' and 'precise':
...
   /* Profile was originally based on feedback but it was adjusted
      by code duplicating optimization.  It may not precisely reflect the
      particular code path.  */
   profile_adjusted = 2,
   /* Profile was read from profile feedback or determined by accurate
      static method.  */
   profile_precise = 3
...

I wonder: there seem to be two situations in which 'precise' is possible:
- Profile was read from profile feedback
- Profile was determined by accurate static method
But there is only one situation where 'adjusted' is possible:
- Profile was originally based on feedback but it was adjusted by code
   duplicating optimization.
I can imagine as well that we originally have a static method giving 
precise information, and that this information is downgraded by a code 
duplication optimization.

So, should this be instead:
...
   /* Profile was originally based on feedback or accurate static method,
      but it was adjusted by code duplicating optimization.  It may not
      precisely reflect the particular code path.  */
   profile_adjusted = 2,
...
?


III.

In this particular case, we insert a conditional jump based on a special 
machine register which gives us the knowledge that both the true and 
false edge are equally likely (in fact, both are executed, with 
different warp enabling mask).

So I think the most accurate representation would be 50/50 'precise', 
but that does not fit with the assumption above.

I'll commit the patch below for now.

But I'm curious if you have any comments.

Thanks,
- Tom

diff --git a/gcc/config/nvptx/nvptx.c b/gcc/config/nvptx/nvptx.c
index 6314653..2c427fa 100644
--- a/gcc/config/nvptx/nvptx.c
+++ b/gcc/config/nvptx/nvptx.c
@@ -5149,6 +5149,7 @@ nvptx_goacc_reduction_init (gcall *call)

        /* Fixup flags from call_bb to init_bb.  */
        init_edge->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE;
+      init_edge->probability = profile_probability::even ();

        /* Set the initialization stmts.  */
        gimple_seq init_seq = NULL;
@@ -5164,6 +5165,7 @@ nvptx_goacc_reduction_init (gcall *call)

        /* Create false edge from call_bb to dst_bb.  */
        edge nop_edge = make_edge (call_bb, dst_bb, EDGE_FALSE_VALUE);
+      nop_edge->probability = profile_probability::even ();

        /* Create phi node in dst block.  */
        gphi *phi = create_phi_node (lhs, dst_bb);