On Tue, Oct 17, 2023 at 2:39 PM Hanke Zhang <hkzhang...@gmail.com> wrote:
>
> Hi Richard
> I get it, thank you again.
>
> And I got another problem, so I'd like ask it by the way. Can the left
> shift of the induction variable in a loop be optimized as a constant?
> Like the code below:
>
> int ans = 0;
> int width = rand() % 16;
> for (int j = 0; j < width; j++)
> ans += 1 << (j + width)
>
> into:
>
> int width = rand() % 16;
> ans = (1 << (2 * width) - (1 << width));
>
> I came across a more complex version of that and found that gcc
> doesn't seem to handle it, so wanted to write a pass myself to
> optimize it.
>
> I got two questions here. Does GCC have such optimizations? If I want
> to do my own optimization, where should I put it? Put it behind the
> pass_iv_optimize?
GCC has the final value replacement pass (pass_scev_cprop) doing these
kind of transforms. Since 'ans' does not have an affine evolution this
case would need to be pattern matched (there are some existing pattern
matchings in the pass).
> Thanks
> Hanke Zhang
>
> Richard Biener <richard.guent...@gmail.com> 于2023年10月17日周二 20:00写道：
> >
> > On Tue, Oct 17, 2023 at 1:54 PM Hanke Zhang <hkzhang...@gmail.com> wrote:
> > >
> > > Richard Biener <richard.guent...@gmail.com> 于2023年10月17日周二 17:26写道：
> > > >
> > > > On Thu, Oct 12, 2023 at 2:18 PM Hanke Zhang via Gcc <gcc@gcc.gnu.org>
> > > > wrote:
> > > > >
> > > > > Hi, I'm recently working on vectorization of GCC. I'm stuck in a small
> > > > > problem and would like to ask for advice.
> > > > >
> > > > > For example, for the following code:
> > > > >
> > > > > int main() {
> > > > > int size = 1000;
> > > > > int *foo = malloc(sizeof(int) * size);
> > > > > int c1 = rand(), t1 = rand();
> > > > >
> > > > > for (int i = 0; i < size; i++) {
> > > > > if (foo[i] & c1) {
> > > > > foo[i] = t1;
> > > > > }
> > > > > }
> > > > >
> > > > > // prevents the loop above from being optimized
> > > > > for (int i = 0; i < size; i++) {
> > > > > printf("%d", foo[i]);
> > > > > }
> > > > > }
> > > > >
> > > > > First of all, the if statement block in the loop will be converted to
> > > > > a MASK_STORE through if-conversion optimization. But after
> > > > > tree-vector, it will still become a branched form. The part of the
> > > > > final disassembly structure probably looks like below(Using IDA to do
> > > > > this), and you can see that there is still such a branch 'if ( !_ZF )'
> > > > > in it, which will lead to low efficiency.
> > > > >
> > > > > do
> > > > > {
> > > > > while ( 1 )
> > > > > {
> > > > > __asm
> > > > > {
> > > > > vpand ymm0, ymm2, ymmword ptr [rax]
> > > > > vpcmpeqd ymm0, ymm0, ymm1
> > > > > vpcmpeqd ymm0, ymm0, ymm1
> > > > > vptest ymm0, ymm0
> > > > > }
> > > > > if ( !_ZF )
> > > > > break;
> > > > > _RAX += 8;
> > > > > if ( _RAX == v9 )
> > > > > goto LABEL_5;
> > > > > }
> > > > > __asm { vpmaskmovd ymmword ptr [rax], ymm0, ymm3 }
> > > > > _RAX += 8;
> > > > > }
> > > > > while ( _RAX != v9 );
> > > > >
> > > > > Why can't we just replace the vptest and if statement with some other
> > > > > instructions like vpblendvb so that it can be faster? Or is there a
> > > > > good way to do that?
> > > >
> > > > The branch is added by optimize_mask_stores after vectorization because
> > > > fully masked (disabled) masked stores can incur a quite heavy penalty on
> > > > some architectures when fault assists (read-only pages, but also COW
> > > > pages)
> > > > are ran into. All the microcode handling needs to possibly be carried
> > > > out
> > > > multiple times, for each such access to the same page. That can cause
> > > > a 1000x slowdown when you hit this case. Thus every masked store
> > > > is replaced by
> > > >
> > > > if (mask != 0)
> > > > masked_store ();
> > > >
> > > > and this is an optimization (which itself has a small cost).
> > > >
> > > > Richard.
> > >
> > > Yeah, I know that and I have seen the code of optimize_mask_store().
> > > And the main problem here is that when multiple MASK_STORE appear in
> > > the same loop, many branches will appear, resulting in a decrease in
> > > overall efficiency.
> > >
> > > And my original idea is that why can't we replace MASK_STORE with more
> > > effective SIMD instructions because icc can do much better in this
> > > case.
> >
> > ICC probably doesn't care for the case where foo[] isn't writable. In
> > fact for the case at hand we see it comes from malloc() which we
> > can assume to return writable memory I guess. That means if-conversion
> > can treat the unconditional read as a way to also allow to speculate
> > the write (with -fallow-store-data-races).
> >
> > Note there's currently no pointer analysis that tracks writability.
> >
> > > Then I give it up, because the ability to analyze vectorization
> > > of gcc is not as good as icc and my ability does not support me
> > > modifying this part of the code.
> > >
> > > Thanks very much for your reply.
> >
> > You're welcome.
> >
> > Richard.
> >
> > > >
> > > > >
> > > > > Thanks
> > > > > Hanke Zhang
