On Mon, 16 Feb 2015, Richard Biener wrote:
>
> Predictive commoning happens to re-use SSA names it released while
> there are still uses of them (oops), confusing the hell out of
> other code (expected). Fixed thus.
>
> Bootstrap and regtest running on x86_64-unknown-linux-gnu.
So I was wrong in that this doesn't fix PR65063 but it pointed at
a similar issue. The loop transform code doesn't handle the case
where we replace looparound PHIs and need a epilogue loop (thus
we use unrolling). The following patch disables unrolling in that
case.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
Richard.
2015-02-17 Richard Biener <rguent...@suse.de>
PR tree-optimization/65063
* tree-predcom.c (determine_unroll_factor): Return 1 if we
have replaced looparound PHIs.
* gcc.dg/pr65063.c: New testcase.
Index: gcc/tree-predcom.c
===================================================================
*** gcc/tree-predcom.c (revision 220755)
--- gcc/tree-predcom.c (working copy)
*************** determine_unroll_factor (vec<chain_p> ch
*** 1775,1783 ****
FOR_EACH_VEC_ELT (chains, i, chain)
{
! if (chain->type == CT_INVARIANT || chain->combined)
continue;
/* The best unroll factor for this chain is equal to the number of
temporary variables that we create for it. */
af = chain->length;
--- 1775,1794 ----
FOR_EACH_VEC_ELT (chains, i, chain)
{
! if (chain->type == CT_INVARIANT)
continue;
+ if (chain->combined)
+ {
+ /* For combined chains, we can't handle unrolling if we replace
+ looparound PHIs. */
+ dref a;
+ unsigned j;
+ for (j = 1; chain->refs.iterate (j, &a); j++)
+ if (gimple_code (a->stmt) == GIMPLE_PHI)
+ return 1;
+ }
+
/* The best unroll factor for this chain is equal to the number of
temporary variables that we create for it. */
af = chain->length;
Index: gcc/testsuite/gcc.dg/pr65063.c
===================================================================
*** gcc/testsuite/gcc.dg/pr65063.c (revision 0)
--- gcc/testsuite/gcc.dg/pr65063.c (working copy)
***************
*** 0 ****
--- 1,33 ----
+ /* { dg-do run } */
+ /* { dg-options "-O3 -fno-tree-loop-ivcanon -fno-tree-vectorize" } */
+
+ static int in[8][4];
+ static int out[4];
+ static const int check_result[] = {0, 16, 256, 4096};
+
+ static inline void foo ()
+ {
+ int sum;
+ int i, j, k;
+ for (k = 0; k < 4; k++)
+ {
+ sum = 1;
+ for (j = 0; j < 4; j++)
+ for (i = 0; i < 4; i++)
+ sum *= in[i + k][j];
+ out[k] = sum;
+ }
+ }
+
+ int main ()
+ {
+ int i, j, k;
+ for (i = 0; i < 8; i++)
+ for (j = 0; j < 4; j++)
+ in[i][j] = (i + 2) / 3;
+ foo ();
+ for (k = 0; k < 4; k++)
+ if (out[k] != check_result[k])
+ __builtin_abort ();
+ return 0;
+ }
