Bootstrapped and tested on x86_64-unkown-linux-gnu, applied. Richard.
2017-09-26 Richard Biener <rguent...@suse.de> PR tree-optimization/82320 * tree-ssa-sccvn.c (set_ssa_val_to): Changing undef to undef isn't a change. * gcc.dg/torture/pr82320.c: New testcase. Index: gcc/tree-ssa-sccvn.c =================================================================== --- gcc/tree-ssa-sccvn.c (revision 253149) +++ gcc/tree-ssa-sccvn.c (working copy) @@ -3355,6 +3355,12 @@ set_ssa_val_to (tree from, tree to) if (currval != to && !operand_equal_p (currval, to, 0) + /* Different undefined SSA names are not actually different. See + PR82320 for a testcase were we'd otherwise not terminate iteration. */ + && !(TREE_CODE (currval) == SSA_NAME + && TREE_CODE (to) == SSA_NAME + && ssa_undefined_value_p (currval, false) + && ssa_undefined_value_p (to, false)) /* ??? For addresses involving volatile objects or types operand_equal_p does not reliably detect ADDR_EXPRs as equal. We know we are only getting invariant gimple addresses here, so can use Index: gcc/testsuite/gcc.dg/torture/pr82320.c =================================================================== --- gcc/testsuite/gcc.dg/torture/pr82320.c (nonexistent) +++ gcc/testsuite/gcc.dg/torture/pr82320.c (working copy) @@ -0,0 +1,39 @@ +/* { dg-do compile } */ + +void +ec (int n4, short int ea) +{ + if (1) + { + if (ea != 0) + { + int *c1 = (int *)&ea; + +nn: + for (;;) + ++*c1; + } + } + else + { + int *lq = &n4; + int *md; + int da; + + goto nn; + +r1: + md = lq; + for (da = 0; da < 1; ++da) + { +ig: + ++n4; + *md += n4; + } + } + + for (ea = 0; ea < 1; ++ea) + goto r1; + + goto ig; +}