Re: [avr-gcc-list] avr_rtx_cost for DImode values?

[Georg-Johann Lay]

Am 03/18/2014 06:33 PM, schrieb Senthil Kumar Selvaraj:
> Hi,
>
> I see that the avr_rtx_costs_1 function returns relative costs for a
> variety of RTL codes, but only for non-DImode operands. For DImode,
> it returns false, and gcc uses its idea of the cost plus the costs of
> operands (found by recursing into the RTL expression) instead.
>
> Is there a reason why this has not been done?
>
> I was trying to debug compilation timeouts in the latest trunk test
> results (http://gcc.gnu.org/ml/gcc-testresults/2014-03/msg01239.html)
> - specifically
> WARNING: program timed out.
> FAIL: gcc.c-torture/execute/20041011-1.c compilation,  -O1

I cannot reproduce this all 22 tests from execute.exp=20041011-1.c take around 
6 seconds (including test suite overhead).

> and found that expmed.c:choose_mult_variant takes a lot of time to
> finish if a DImode value of -513 is provided. From my initial
> analysis, it looks like it has to do with cost values being off for
> DImode RTL operations (making MULT in avr_rtx_costs_1 to return
> twice the cost of SImode for DImode fixed the timeout). Has
> someone seen or fixed something like this before?

No. Even if the choice of a particular MUL strategy is not optimal, it should 
not hang the compiler.  I suspect you are using a 64-bit host? (I am on 32 bit 
Linux).

Johann