Re: [PATCH] Fix PR36078, fallout from early loop unrolling

[Richard Guenther <rguenther at suse dot de>]

On Tue, 29 Apr 2008, Zdenek Dvorak wrote:

> Hi,
> 
> > This patch adds a missing SSA update after early complete loop unrolling.
> > Appearantly this is only needed if unrolling doesn't operate on
> > loop-closed SSA form, but it might also paper over a problem with not
> > doing so.
> 
> why does this help?  We already call update_ssa in
> try_unroll_loop_completely.

Hm, the difference with the (indeed extra) update_ssa is

--- a/t.f90.056t.cunrolli       2008-04-29 15:36:31.000000000 +0200
+++ b/t.f90.056t.cunrolli       2008-04-29 15:38:35.000000000 +0200
@@ -93,24 +93,6 @@
 Merging blocks 11 and 3
 Merging blocks 2 and 9
 Merging blocks 2 and 11
-
-
-Symbols to be put in SSA form
-
-{ PARM_NOALIAS.43 }
-
-
-SSA names to release after updating the SSA web
-
-PARM_NOALIAS.43_35 PARM_NOALIAS.43_265 PARM_NOALIAS.43_335 
-
-Incremental SSA update started at block: 0
-
-Number of blocks in CFG: 5
-Number of blocks to update: 4 ( 80%)
-
-
-
 foo (void (*<T2da>) (struct array1_real(kind=8) &, struct 
array2_real(kind=8) &) func, real(kind=8)[375] * p, logical(kind=4)[125] * 
eval)
 {
   void * D.1030;
@@ -490,7 +472,7 @@
   D.1057_341 = (*D.1052_340)[S.21_333];
   D.1058_342 = (integer(kind=8)) D.1057_341;
   # A.19_343 = VDEF <A.19_330>
-  # PARM_NOALIAS.43_344 = VDEF <PARM_NOALIAS.43_246>
+  # PARM_NOALIAS.43_344 = VDEF <PARM_NOALIAS.43_331>
   # SMT.51_345 = VDEF <SMT.51_332>
   (*D.1055_339)[S.21_333] = D.1058_342;
   S.21_346 = S.21_333 + 1;


as we run CFG cleanup after updating SSA form in 
try_unroll_loop_completely, some constants are propagated from PHIs
and the VOPs change during CFG cleanup requiring an SSA update
(cfg-cleanup vs. non-cfg-cleanup variant):

+
+<bb 8>:
+  # SMT.51_321 = PHI <SMT.51_253(2)>
+  # PARM_NOALIAS.43_322 = PHI <PARM_NOALIAS.43_246(2)>
+  # A.19_323 = PHI <A.19_245(2)>
+  # S.21_324 = PHI <0(2)>
+
+<bb 9>:
   # VUSE <atmp.18_263>
   D.1054_325 = atmp.18.data;
   D.1055_326 = (integer(kind=8)[2] *) D.1054_325;
   D.1052_327 = (integer(kind=4)[2] *) D.1051_32;
   # VUSE <A.16_257>
-  D.1057_328 = (*D.1052_327)[0];
+  D.1057_328 = (*D.1052_327)[S.21_324];
   D.1058_329 = (integer(kind=8)) D.1057_328;
-  # A.19_330 = VDEF <A.19_245>
-  # SMT.51_332 = VDEF <SMT.51_253>
-  (*D.1055_326)[0] = D.1058_329;
-  S.21_333 = 1;
+  # A.19_330 = VDEF <A.19_323>
+  # PARM_NOALIAS.43_331 = VDEF <PARM_NOALIAS.43_322>
+  # SMT.51_332 = VDEF <SMT.51_321>
+  (*D.1055_326)[S.21_324] = D.1058_329;
+  S.21_333 = S.21_324 + 1;

so with CFG cleanup we have

-  # A.19_330 = VDEF <A.19_245>
-  # SMT.51_332 = VDEF <SMT.51_253>
-  (*D.1055_326)[0] = D.1058_329;

while without we have

+  # A.19_330 = VDEF <A.19_323>
+  # PARM_NOALIAS.43_331 = VDEF <PARM_NOALIAS.43_322>
+  # SMT.51_332 = VDEF <SMT.51_321>
+  (*D.1055_326)[S.21_324] = D.1058_329;

so the use of PARM_NOALIAS.43_331 is no longer valid and an SSA
update is required.

This may be a latent general problem of course, which the patch
(bootstrapped and tested ok meanwhile) only papers over.

Richard.