[Bug tree-optimization/94092] Code size and performance degradations after -ftree-loop-distribute-patterns was enabled at -O[2s]+

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94092

--- Comment #12 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 3 Mar 2021, bina2374 at gmail dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94092
> 
> --- Comment #11 from Mel Chen <bina2374 at gmail dot com> ---
> (In reply to Mel Chen from comment #10)
> > (In reply to Richard Biener from comment #9)
> > > (In reply to Mel Chen from comment #8)
> > > > Sorry for using the bad example to describe the problem I am facing. Let me
> > > > clarify my question with a more precise example.
> > > > 
> > > > void array_mul(int N, int *C, short *A, short *B) {
> > > >   int i, j;
> > > >   for (i = 0; i < N; i++) {
> > > >     C[i] = 0; // Will be transformed to __builtin_memset
> > > >     for (j = 0; j < N; j++) {
> > > >       C[i] += (int)A[i * N + j] * (int)B[j];
> > > >     }
> > > >   }
> > > > }
> > > > 
> > > > If I compile the case with -O2 -fno-tree-loop-distribute-patterns, the store
> > > > operation 'C[i] = 0' can be eliminated by dead store elimination (dse3). But
> > > > without -fno-tree-loop-distribute-patterns, it will be transformed to memset
> > > > by loop distribution (ldist) because ldist executes before dse3. Finally the
> > > > memset will not be eliminated.
> > > > 
> > > > Another point is if there are other operations in the same level loop as the
> > > > store operation, is it really beneficial to do loop distribution and then
> > > > convert to builtin function?
> > > 
> > > Sure, it shows a cost modeling issue given that usually loop distribution
> > > merges partitions which touch the same memory stream (but IIRC maybe only
> > > for loads).  But more to the point we're missing to eliminate the dead store
> > > which should be appearant at least after PRE - LIM2 applied store motion
> > > but only PRE elides the resulting load of C[i].  Usually DCE and DSE come in
> > > pairs but after PRE we have DCE, CDDCE w/o accompaning DSE only with the
> > > next DSE only happening after loop distribution.
> > > 
> > > Which means we should eventually do
> > > 
> > > diff --git a/gcc/passes.def b/gcc/passes.def
> > > index e9ed3c7bc57..be3a9becde0 100644
> > > --- a/gcc/passes.def
> > > +++ b/gcc/passes.def
> > > @@ -254,6 +254,7 @@ along with GCC; see the file COPYING3.  If not see
> > >        NEXT_PASS (pass_sancov);
> > >        NEXT_PASS (pass_asan);
> > >        NEXT_PASS (pass_tsan);
> > > +      NEXT_PASS (pass_dse);
> > >        NEXT_PASS (pass_dce);
> > >        /* Pass group that runs when 1) enabled, 2) there are loops
> > >          in the function.  Make sure to run pass_fix_loops before
> > 
> > Yes, doing DSE before ldist is a simple and effective way.
> > This patch has been verified to be work on coremark. Not only improved
> > performance, but also code size.
> 
> The test case gcc.dg/tree-ssa/ldist-33.c is failure after I added DSE.
> 
> /* { dg-do compile { target size32plus } } */
> /* { dg-options "-O2 -ftree-loop-distribution -ftree-loop-distribute-patterns
> -fdump-tree-ldist-details" } */
> 
> #define N (1024)
> double a[N][N], b[N][N], c[N][N];
> 
> void
> foo (void)
> {
>   unsigned i, j, k;
> 
>   for (i = 0; i < N; ++i)
>     for (j = 0; j < N; ++j)
>       {
>         c[i][j] = 0.0;
>         for (k = 0; k < N; ++k)
>           c[i][j] += a[i][k] * b[k][j];
>       }
> }
> 
> /* { dg-final { scan-tree-dump "Loop nest . distributed: split to 1 loops and 1
> library" "ldist" } } */
> 
> It is similar to the example I showed earlier. DSE eliminated 'c[i][j] = 0.0'
> so no loop will be split. My question is how to handle this test case? Add
> -fno-tree-dse into dg-options, modify this test case, delete this test case, or
> others.

I'd say disable store-motion with a comment (-fno-tree-loop-im, we might
also finally add a separate switch to control the store-motion part of
GIMPLE invariant motion).  I think
the testcase should show we can recognize the memset in a non-innermost
loop (IIRC the above mimics what bwaves does).