On Mon, 5 Oct 2020, Jakub Jelinek wrote:
> Hi!
>
> Compiling the following testcase with -O2 -fopenmp:
> int a[10000][128];
>
> __attribute__((noipa)) void
> foo (void)
> {
> #pragma omp for simd schedule (simd: dynamic, 32) collapse(2)
> for (int i = 0; i < 10000; i++)
> for (int j = 0; j < 128; j++)
> a[i][j] += 3;
> }
>
> int
> main ()
> {
> for (int i = 0; i < 10000; i++)
> for (int j = 0; j < 128; j++)
> {
> asm volatile ("" : : "r" (&a[0][0]) : "memory");
> a[i][j] = i + j;
> }
> foo ();
> for (int i = 0; i < 10000; i++)
> for (int j = 0; j < 128; j++)
> if (a[i][j] != i + j + 3)
> __builtin_abort ();
> return 0;
> }
> doesn't seem result in the vectorization I was hoping to see.
> As has been changed recently, I'm only trying to vectorize now the
> innermost loop of the collapse with outer loops around it being normal
> scalar loops like those written in the source and with only omp simd
> it works fine, but for the combined constructs the current thread gets
> assigned some range of logical iterations, therefore I get a pair of
> in this case i and j starting values.
>
> At the end of ompexp I have:
> ...
> D.2106 = (unsigned int) D.2105;
> D.2107 = MIN_EXPR <D.2104, D.2106>;
> D.2103 = D.2107 + .iter.4;
> goto <bb 5>; [INV]
> ;; succ: 5
>
> ;; basic block 4, loop depth 2
> ;; pred: 5
> i = i.0;
> j = j.1;
> _1 = a[i][j];
> _2 = _1 + 3;
> a[i][j] = _2;
> .iter.4 = .iter.4 + 1;
> j.1 = j.1 + 1;
> ;; succ: 5
>
> ;; basic block 5, loop depth 2
> ;; pred: 4
> ;; 3
> ;; 7
> if (.iter.4 < D.2103)
> goto <bb 4>; [87.50%]
> else
> goto <bb 6>; [12.50%]
> ;; succ: 4
> ;; 6
>
> ;; basic block 6, loop depth 2
> ;; pred: 5
> i.0 = i.0 + 1;
> if (i.0 < 10000)
> goto <bb 7>; [87.50%]
> else
> goto <bb 8>; [12.50%]
> ;; succ: 8
> ;; 7
>
> ;; basic block 7, loop depth 2
> ;; pred: 6
> j.1 = 0;
> D.2108 = D.2099 - .iter.4;
> D.2109 = MIN_EXPR <D.2108, 128>;
> D.2103 = D.2109 + .iter.4;
> goto <bb 5>; [INV]
>
> I was really hoping bbs 4 and 5 would be one loop (the one I set safelen
> and force_vectorize etc. for) and that basic blocks 6 and 7 would be
> together with that inner loop another loop, but apparently loop discovery
> thinks it is just one loop.
> Any ideas what I'm doing wrong or is there any way how to make it two loops
> (that would also survive all the cfg cleanups until vectorization)?
The early CFG looks like we have a common header with two latches
so it boils down to how we disambiguate those in the end (we seem
to unify the latches via a forwarder). IIRC OMP lowering builds
loops itself, could it not do the appropriate disambiguation itself?
Richard.
> Essentially, in C I'm trying to have:
> int a[10000][128];
> void get_me_start_end (int *, int *);
> void
> foo (void)
> {
> int start, end, curend, i, j;
> get_me_start_end (&start, &end);
> i = start / 128;
> j = start % 128;
> curend = start + (end - start > 128 - j ? 128 - j : end - start);
> goto doit;
> for (i = 0; i < 10000; i++)
> {
> j = 0;
> curend = start + (end - start > 128 ? 128 : end - start);
> doit:;
> /* I'd use start < curend && j < 128 as condition here, but
> the vectorizer doesn't like that either. So I went to
> using a single IV. */
> for (; start < curend; start++, j++)
> a[i][j] += 3;
> }
> }
>
> This isn't vectorized with -O3 either for the same reason.
>
> Jakub
>
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Felix Imend
