On Fri, May 13, 2011 at 6:00 PM, Nathan Froyd <froy...@codesourcery.com> wrote:
> On 05/13/2011 11:26 AM, Richard Guenther wrote:
>> On Fri, May 13, 2011 at 5:01 PM, Nathan Froyd <froy...@codesourcery.com>
>> wrote:
>>> On 05/13/2011 10:52 AM, William J. Schmidt wrote:
>>>> This patch addresses PR46728, which notes that pow and powi need to be
>>>> lowered in tree phases to restore lost FMA opportunities and expose
>>>> vectorization opportunities.
>>>>
>>>> +struct gimple_opt_pass pass_lower_pow =
>>>> +{
>>>> + {
>>>> + GIMPLE_PASS,
>>>> + "lower_pow", /* name */
>>>> + NULL, /* gate */
>>>
>>> Please make this controlled by an option; this pass doesn't need to be run
>>> all
>>> the time.
>>>
>>> IMHO, the pass shouldn't run at anything less than -O3, but that's for other
>>> people to decide.
>>
>> It was run unconditionally before, so unless we preserve the code at
>> expansion time we have to do it here.
>
> We were doing it unconditionally before because we were calling it through
> folding and expansion, both of which only fired if we were expanding pow
> calls; now we're groveling over the whole IR to look for optimization
> opportunities that, let's be honest, the vast majority of code is never going
> to care about. The whole point of the patch is "to restore lost FMA
> opportunities and expose vectorization opportunities". The first reason
> *might* be good justification for running it at -O2, but the second reason
> calls for -O3 or at the very least flag_tree_vectorize.
The patch of course does more, like expanding to power series.
> However, I don't think anybody's going to notice/care if -O1 stopped folding
> pow calls; could we at least add a flag_expensive_optimizations && optimize
> gate?
The pass doesn't look very expensive and the IL walk could be
shared with cse_sincos. Note that I agree that the transforms
probably should never have been applied at -O0 (we IMHO do too
much builtin folding at -O0 which can be surprising - like
transforming pow (x, 0.5) to sqrt (x)). So I wouldn't worry about
losing the -O0 transforms at all.
Richard.
> -Nathan
>
>
