https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101373
--- Comment #4 from rguenther at suse dot de <rguenther at suse dot de> --- On Fri, 9 Jul 2021, ebotcazou at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101373 > > --- Comment #3 from Eric Botcazou <ebotcazou at gcc dot gnu.org> --- > > Eric, is it possible to write an equivalent testcase in Ada? > > procedure P is > > function Foo (J : Integer) return Integer; > pragma Machine_Attribute (Foo, "const"); > pragma Machine_Attribute (Foo, "noipa"); > > function Foo (J : Integer) return Integer is > begin > if J /= 0 then > raise Constraint_Error; > end if; > return 0; > end; > > function Bar (A : access Integer; N : Integer) return Integer; > pragma Machine_Attribute (Bar, "noipa"); > > function Bar (A : access Integer; N : Integer) return Integer is > Ret : Integer := 0; > begin > if N /= 0 then > Ret := Foo (N); > Ret := A.all; > end if; > Ret := Ret + A.all; > return Ret; > end; > > V : Integer; > pragma Volatile (V); > > begin > V := Bar (null, 1); > exception > when Constraint_Error => null; > end; > > compiled with -O2 -gnatp, but I cannot have the load hoisted in this case. The issue I can see is that the middle-end does not consider the Foo function 'const' (gimple_call_flags does not return ECF_CONST) which is likely because the decl isn't TREE_READONLY and for 'const' we nowhere lookup the actual attribute but handlers are expected to mimic c-family/c-attribs.c handling. With diff --git a/gcc/gimple.c b/gcc/gimple.c index 60a90667e4b..576b5df5cec 100644 --- a/gcc/gimple.c +++ b/gcc/gimple.c @@ -1470,7 +1470,11 @@ gimple_call_flags (const gimple *stmt) { tree decl = gimple_call_fndecl (stmt); if (decl) - flags = flags_from_decl_or_type (decl); + { + flags = flags_from_decl_or_type (decl); + if (lookup_attribute ("const", DECL_ATTRIBUTES (decl))) + flags |= ECF_CONST; + } flags |= flags_from_decl_or_type (gimple_call_fntype (stmt)); } I see the access hoisted (but as said, the Ada FE should set TREE_READONLY here). I first thought that non-call exceptions might be the issue (we wouldn't move A.all then), but non-call exceptions seem off.
