You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
…ork/lazy-method-loading
Note, this is a no-op when the LLDB setting
`plugin.typesystem.clang.experimental-redecl-completion` is not set (which is currently the default). So this patch has no effect unless the user explicitly opts into it.
The type-completion rework (aka redecl-completion) implemented in apple#8222 comes with a large performance penalty, since we now eagerly complete `RecordType`s. Completing a `RecordType` previously unconditionally resolved all member functions of said type. With redecl-completion completion, however, this meant we were now pulling in many more definitions than needed. Without redecl-completion, this isn't a problem, since importing method parameters is cheap (they are imported minimally), so we wouldn't notice that we always resolved all member functions.
This patch tries to load methods lazily when in redecl-completion mode. We do this by introducing a new `ExternalASTSource::FindExternalVisibleMethods` API which Clang calls when parsing a member access expression. The callback into LLDB will do a `FindFunctions` call for all methods with the method name we're looking for, filters out any mismatches, and lets Clang continue with its parsing. We still load following methods eagerly:
1. virtual functions: currently overrides are resolved in `CompleteRecordType`
2. operators: currently I couldn't find a point at which Clang can call into LLDB here to facilitate lazy loading
3. ctors/dtors: same reason as (2)
In our benchmark harness, we saw this patch bring down redecl-completion expression evaluation performance on-par with top-of-tree expression evaluation.
Extended Description
This is another try to get to the bottom of http://crbug.com/56471
hummer:delta thakis$ cat string16.ii
typedef long unsigned int size_t;
namespace std {
template
class basic_string {
private:
struct _Rep {
void _M_clone();
};
};
template
void basic_string<_CharT>::_Rep::_M_clone() {}
}
typedef unsigned short char16;
extern template class std::basic_string;
template class std::basic_string;
_M_clone becomes a private external symbol in gcc with -fvisibility=hidden:
hummer:delta thakis$ gcc -c string16.ii -fvisibility=hidden
hummer:delta thakis$ nm -m string16.o
0000000000000010 (__TEXT,__eh_frame) non-external EH_frame1
0000000000000000 (__TEXT,__text) private external __ZNSbItE4_Rep8_M_cloneEv
0000000000000030 (__TEXT,__eh_frame) private external __ZNSbItE4_Rep8_M_cloneEv.eh
(undefined) external ___gxx_personality_v0
It doesn't do that in clang:
hummer:delta thakis$ ~/src/llvm/Release+Asserts/bin/clang -c string16.ii -fvisibility=hidden
hummer:delta thakis$ nm -m string16.o
0000000000000010 (__TEXT,__eh_frame) non-external EH_frame0
0000000000000000 (__TEXT,__textcoal_nt) weak external __ZNSbItE4_Rep8_M_cloneEv
0000000000000028 (__TEXT,__eh_frame) weak external __ZNSbItE4_Rep8_M_cloneEv.eh
It works fine in clang if I move the method out of the containing inner struct or make the class not a template class.
The text was updated successfully, but these errors were encountered: