I am testing the following (obvious) patch to avoid generating overlapping life-ranges for SSA names that occur in abnormal PHIs.
Bootstrap and regtest running on x86_64-unknown-linux-gnu. Richard. 2014-11-26 Richard Biener <rguent...@suse.de> PR middle-end/63738 * tree-data-ref.c (split_constant_offset_1): Do not follow SSA edges for SSA names with SSA_NAME_OCCURS_IN_ABNORMAL_PHI. * gcc.dg/torture/pr63738.c: New testcase. Index: gcc/tree-data-ref.c =================================================================== --- gcc/tree-data-ref.c (revision 218076) +++ gcc/tree-data-ref.c (working copy) @@ -674,6 +674,9 @@ split_constant_offset_1 (tree type, tree case SSA_NAME: { + if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op0)) + return false; + gimple def_stmt = SSA_NAME_DEF_STMT (op0); enum tree_code subcode; Index: gcc/testsuite/gcc.dg/torture/pr63738.c =================================================================== --- gcc/testsuite/gcc.dg/torture/pr63738.c (revision 0) +++ gcc/testsuite/gcc.dg/torture/pr63738.c (working copy) @@ -0,0 +1,27 @@ +/* { dg-do compile } */ + +#include <setjmp.h> + +struct longjmp_buffer { + jmp_buf buf; +}; + +void plouf(); + +extern long interprete() +{ + long * sp; + int i; + long *args; + int n; + + struct longjmp_buffer raise_buf; + _setjmp (raise_buf.buf); + + plouf(); + sp -= 4; + for (i = 0; i < n; i++) + args[i] = sp[10-i]; + plouf(); + return 0; +}