Hi, On Fri, 9 Dec 2016, Richard Biener wrote:
> On December 9, 2016 4:29:04 PM GMT+01:00, Michael Matz <m...@suse.de> wrote: > >Hi, > > > >if the induction variable on which we want to partition the loop > >iterations for loop splitting might overflow we would either need > >runtime > >tests checking if an overflow in fact does happen, or we can simply not > > > >split loops on such ones. I chose the latter. > > > >Fixes the testcase, and regstrapped without regressions on x86-64-linux > > > >(all languages+Ada). Okay for trunk? > > OK Actually there came another report in the same bug, which has a different but related cause. In the testcase the potential split-point is if (outer >= inner) where the final value of 'inner' is known constant in the _outer_ loop. But of course we have to consider the actual place of the use: the comparison statement which is inside the inner loop, not merely in the outer loop. Fixed with the amended patch. Regstrapping on x86-64-linux in progress (all langs+ada), but I know that it fixes both testcases and doesn't regress in gcc.dg. Okay for trunk if regstrapping succeeds? Ciao, Michael. PR tree-optimization/78725 * tree-ssa-loop-split.c (split_at_bb_p): Check for overflow and at correct use point. testsuite/ PR tree-optimization/78725 * gcc.dg/pr78725.c: New test. * gcc.dg/pr78725-2.c: New test. diff --git a/gcc/testsuite/gcc.dg/pr78725-2.c b/gcc/testsuite/gcc.dg/pr78725-2.c new file mode 100644 index 0000000..9de489e --- /dev/null +++ b/gcc/testsuite/gcc.dg/pr78725-2.c @@ -0,0 +1,19 @@ +/* { dg-do run } */ +/* { dg-options "-O3 -fsplit-loops" } */ + +int a, b, c; + +int main () +{ + int d; + for (; c < 1; c++) + for (d = 0; d < 3; d++) + for (b = 0; b < 1; b++) + if (c >= d) + a = 1; + + if (a != 1) + __builtin_abort (); + + return 0; +} diff --git a/gcc/testsuite/gcc.dg/pr78725.c b/gcc/testsuite/gcc.dg/pr78725.c new file mode 100644 index 0000000..ed95790 --- /dev/null +++ b/gcc/testsuite/gcc.dg/pr78725.c @@ -0,0 +1,19 @@ +/* { dg-do run } */ +/* { dg-options "-O3 -fsplit-loops" } */ + +int fn1 (int b, int c) +{ + return c < 0 || c > 31 ? 0 : b >> c; +} + +unsigned char d = 255; +int e, f; + +int main () +{ + for (; f < 2; f++) + e = fn1 (1, d++); + if (e != 1) + __builtin_abort (); + return 0; +} diff --git a/gcc/tree-ssa-loop-split.c b/gcc/tree-ssa-loop-split.c index dac68e6..84c0627 100644 --- a/gcc/tree-ssa-loop-split.c +++ b/gcc/tree-ssa-loop-split.c @@ -102,10 +102,11 @@ split_at_bb_p (struct loop *loop, basic_block bb, tree *border, affine_iv *iv) tree op0 = gimple_cond_lhs (stmt); tree op1 = gimple_cond_rhs (stmt); + struct loop *useloop = loop_containing_stmt (stmt); - if (!simple_iv (loop, loop, op0, iv, false)) + if (!simple_iv (loop, useloop, op0, iv, false)) return NULL_TREE; - if (!simple_iv (loop, loop, op1, &iv2, false)) + if (!simple_iv (loop, useloop, op1, &iv2, false)) return NULL_TREE; /* Make it so that the first argument of the condition is @@ -122,6 +123,8 @@ split_at_bb_p (struct loop *loop, basic_block bb, tree *border, affine_iv *iv) return NULL_TREE; if (!integer_zerop (iv2.step)) return NULL_TREE; + if (!iv->no_overflow) + return NULL_TREE; if (dump_file && (dump_flags & TDF_DETAILS)) {