https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105329
Aldy Hernandez <aldyh at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |aldyh at gcc dot gnu.org
--- Comment #4 from Aldy Hernandez <aldyh at gcc dot gnu.org> ---
(In reply to Andrew Macleod from comment #3)
> p vr.dump(stderr)
> const size_type [1, 9223372036854775807]
>
> I also have a breakpoint in set_range_info for this name which hasn't been
> triggered. So either the set routine have been bypassed or perhaps
> inlining is setting this global value?
>
> Im still trying to figure out who and where has decided that __nleft_49 is
> [2, 0x7FFFFFFFFFFFFFFF] instead of [2, 0xFFFFFFFFFFFFFFFFFFFF]
Global ranges can also be set via duplicate_ssa_name_range_info, which is used
by the inliner.
It looks like [1, 9223372036854775807] was originally calculated in evrp for
__nleft_64:
=========== BB 45 ============
Imports: _27 __s_53(D)
Exports: _27 _34 __s_53(D) __nleft_64
_34 : _27(I) __s_53(D)(I)
__nleft_64 : _27(I) _34 __s_53(D)(I)
_27 char * [1B, -2B]
__s_53(D) const char * [0B, -3B]
Relational : (__s_53(D) < _26)
Relational : (_27 > __s_53(D))
<bb 45> :
_34 = _27 - __s_53(D);
__nleft_64 = (const size_type) _34;
if (_34 == 1)
goto <bb 46>; [34.00%]
else
goto <bb 47>; [66.00%]
_34 : long int [1, +INF]
__nleft_64 : long unsigned int [1, 9223372036854775807]
45->46 (T) _27 : char * [1B, -2B]
45->46 (T) _34 : long int [1, 1]
45->46 (T) __s_53(D) : const char * [0B, -3B]
45->46 (T) __nleft_64 : long unsigned int [1, 9223372036854775807]
45->47 (F) _27 : char * [1B, -2B]
45->47 (F) _34 : long int [2, +INF]
45->47 (F) __s_53(D) : const char * [0B, -3B]
45->47 (F) __nleft_64 : long unsigned int [1, 9223372036854775807]
This was later renamed to __nleft_40, and then copied by the IPA inliner to
__nleft_47 and ultimately to __nleft_49.