Hi,
On Thu, 11 Jun 2015 12:57:34 +0200, Jakub Jelinek wrote:
>
> On Thu, Jun 11, 2015 at 12:38:40PM +0200, Bernd Edlinger wrote:
>> On Thu, 11 Jun 2015 10:02:03, Jakub Jelinek wrote:
>>>
>>> IMHO the
>>> #if 0
>>> #endif
>>> stuff doesn't belong to the patch.
>>>
>>
>> I just wanted to leave a hint, how I debugged this function, and how
>> to assess the performance of the decision that is taken here.
>
> What I usually do in these cases is something like:
> FILE *f = fopen ("/tmp/mylogfile", "a");
> fprintf (f, "%s %d ...\n", main_input_filename ? main_input_filename : "-",
> (int) BITS_PER_WORD, ...);
> fclose (f);
> and do full bootstrap/regtest (usually both x86_64-linux and i686-linux)
> with it, then look at the log file.
> But I keep those for myself, don't keep them even as comments.
> In this case, you could post the hack as incremental patch for interested
> folks to test on their architecture, but I'm not convinced we want to keep
> it in the source, whether #if 0 or in a comment.
>
I am not too sure about it either.
But I think, it is quite helpful data, however I am even tempted
to add the name of the current function, and the pass we are in at the moment,
but I have no idea how to grab that information...
> So, for a full bootstrap/regtest, how many log messages do you get, and are
> they always resolved conservatively (i.e. if unsure the offset is ok, return
> 1)?
>
In stage 2 of the build (with all languages) I get:
2930 messages of the form
*** frame can trap: offset=16, size=8, low_bound=-3152, high_bound=0
74 messages of the form
*** sp can trap: offset=112, size=4, low_bound=-144, high_bound=112
202 messages of the from
*** argp can trap: offset=16, size=8, low_bound=-56, high_bound=16
10 messages of the form
*** fp can trap: offset=40, size=4, low_bound=-264, high_bound=24
My patch does not change the handling of frame_pointer_rtx,
except that it avoids a possible integer overflow in "adj_offset + size - 1>= 0"
so these 2930 suppressed optimizations were already introduced by Eric's patch.
I think that is probably a new effect, that [FP+x] is now used more
often than before to access values at [ARGP+x]. I have not tried, but
maybe it would be possible to use the crtl->args.size, here too, to get more
optimistic upper bounds on the argument sizes.
So all in all my patch changed 286 times the return value of
rtx_addr_can_trap_p_1
in the whole pass 2.
But OTOH there are millions of times, where the rtx_addr_can_trap_p_1
returns 0, which is rtx can not trap.
Bernd.
