[Bug tree-optimization/77975] [6/7 Regression] Missed optimization for some small constants

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77975

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
> Proved that loop 1 iterates 2 times using brute force.

so the question is why that doesn't work for the new form (and this is what we
should fix).  Because

static gphi *
get_base_for (struct loop *loop, tree x)
{
...
  init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
  next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));

  if (TREE_CODE (next) != SSA_NAME)
    return NULL;

Index: gcc/tree-ssa-loop-niter.c
===================================================================
--- gcc/tree-ssa-loop-niter.c   (revision 241148)
+++ gcc/tree-ssa-loop-niter.c   (working copy)
@@ -2545,13 +2551,11 @@ get_base_for (struct loop *loop, tree x)
   init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
   next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));

-  if (TREE_CODE (next) != SSA_NAME)
-    return NULL;
-
   if (!is_gimple_min_invariant (init))
     return NULL;

-  if (chain_of_csts_start (loop, next) != phi)
+  if (TREE_CODE (next) == SSA_NAME
+      && chain_of_csts_start (loop, next) != phi)
     return NULL;

   return phi;
@@ -2574,6 +2578,8 @@ get_val_for (tree x, tree base)

   if (!x)
     return base;
+  else if (is_gimple_min_invariant (x))
+    return x;

   stmt = SSA_NAME_DEF_STMT (x);
   if (gimple_code (stmt) == GIMPLE_PHI)


fixes it.  But I think it's not correct as it misses the intermediate stmt
evaluations.  Some more refactoring (passing in the PHI def to get_val_for)
is necessary I think.