(Short version of this email, is there a way to recalculate/rebuild virtual
phi nodes after modifying the CFG.)
I have a question about duplicating loops and virtual phi nodes.
I am trying to implement the following optimization as a pass:
Transform:
for (i = 0; i < n; i++) {
A[i] = A[i] + B[i];
C[i] = C[i-1] + D[i];
}
Into:
if (noalias between A&B, A&C, A&D)
for (i = 0; i < 100; i++)
A[i] = A[i] + B[i];
for (i = 0; i < 100; i++)
C[i] = C[i-1] + D[i];
else
for (i = 0; i < 100; i++) {
A[i] = A[i] + B[i];
C[i] = C[i-1] + D[i];
}
Right now the vectorizer sees that 'C[i] = C[i-1] + D[i];' cannot be
vectorized so it gives up and does not vectorize the loop. If we split
up the loop into two loops then the vector add with A[i] could be vectorized
even if the one with C[i] could not.
Currently I can introduce the first 'if' that checks for aliasing by
using loop_version() and that seems to work OK. (My actual compare
for aliasing is actually just an approximation for now.)
Where I am running into problems is with splitting up the single loop
under the noalias if condition into two sequential loops (which I then
intend to 'thin out' by removing one or the other set of instructions.
I am using slpeel_tree_duplicate_loop_to_edge_cfg() for that loop duplication
and while I get the CFG I want, the pass ends with verify_ssa failing due
to bad virtual/MEM PHI nodes. Perhaps there is a different function that
I should use duplicate the loop.
a.c: In function âfooâ:
a.c:2:5: error: PHI node with wrong VUSE on edge from BB 13
int foo(int *a, int *b, int *c, int *d, int n)
^~~
.MEM_40 = PHI <.MEM_15(D)(13), .MEM_34(9)>
expected .MEM_58
a.c:2:5: internal compiler error: verify_ssa failed
I have tried to fix up the PHI node by hand using SET_PHI_ARG_DEF but
have not had any luck. I was wondering if there was any kind of
'update all the phi nodes' function or just a 'update the virtual phi
nodes' function. The non-virtual PHI nodes seem to be OK, it is just
the virtual ones that seem wrong after I duplicate the loop into two
consecutive loops.
Steve Ellcey
[email protected]
