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
