| Message ID | 8613782.AEIKBo5lNX@arcturus.home |
|---|---|
| State | New |
| Headers | show |
On Wed, Aug 31, 2016 at 8:46 PM, Eric Botcazou <ebotcazou@adacore.com> wrote: > Hi, > > the attached Ada testcase triggers a segfault in param_change_prob because the > parameter is VIEW_CONVERT_EXPR<INTEGER_CST> (or <SSA_NAME) for a variant) so > it fools the logic and walk_aliased_vdefs is called on a NULL second argument > as the call is to a pure function. The proposed fix is just to strip the VCE. > > Tested on x86_64-suse-linux, OK for the mainline? I think the fix is not enough if op is for example MEM[&"foo"] (no V_C_E) then get_base_address returns a STRING_CST and you run into exactly the same issue. So I think it's better to apply get_base_address here rather than only stripping VIEW_CONVERT_EXRPs. (beware of it returning NULL_TREE for WITH_SIZE_EXPRs, thus maybe strip those first as well) You could after all have COMPONENT_REF<VIEW_CONVERT_EXPR <INTEGER_CST>> as well ... Richard. > > 2016-08-31 Eric Botcazou <ebotcazou@adacore.com> > > * ipa-inline-analysis.c (param_change_prob): Strip VIEW_CONVERT_EXPR. > > > 2016-08-31 Eric Botcazou <ebotcazou@adacore.com> > > * gnat.dg/opt58.adb: New test. > * gnat.dg/opt58_pkg.ads: New helper. > > -- > Eric Botcazou
Index: ipa-inline-analysis.c =================================================================== --- ipa-inline-analysis.c (revision 239842) +++ ipa-inline-analysis.c (working copy) @@ -2185,9 +2185,13 @@ param_change_prob (gimple *stmt, int i) basic_block bb = gimple_bb (stmt); tree base; - /* Global invariants neve change. */ + if (TREE_CODE (op) == VIEW_CONVERT_EXPR) + op = TREE_OPERAND (op, 0); + + /* Global invariants never change. */ if (is_gimple_min_invariant (op)) return 0; + /* We would have to do non-trivial analysis to really work out what is the probability of value to change (i.e. when init statement is in a sibling loop of the call).