| Message ID | 20151201060240.GA31274@intrepid.com |
|---|---|
| State | New |
| Headers | show |
On Tue, Dec 1, 2015 at 7:02 AM, Gary Funck <gary@intrepid.com> wrote: > > Background > ---------- > > An overview email, describing the UPC-related changes is here: > https://gcc.gnu.org/ml/gcc-patches/2015-12/msg00005.html > > The GUPC branch is described here: > http://gcc.gnu.org/projects/gupc.html > > The UPC-related source code differences are summarized here: > http://gccupc.org/gupc-changes > > All languages (c, c++, fortran, go, lto, objc, obj-c++) have been > bootstrapped; no test suite regressions were introduced, > relative to the GCC trunk. > > If you are on the cc-list, your name was chosen either > because you are listed as a maintainer for the area that > applies to the patches described in this email, or you > were a frequent contributor of patches made to files listed > in this email. > > In the change log entries included in each patch, the directory > containing the affected files is listed, followed by the files. > When the patches are applied, the change log entries will be > distributed to the appropriate ChangeLog file. > > Overview > -------- > > UPC pointers-to-shared have an internal representation which is > defined as a 'struct' with three fields. Special logic is > needed in promote_mode() to handle this case. Errr - but how are 'struct's ever REFERENCE_TYPE or POINTER_TYPE? Richard. > 2015-11-30 Gary Funck <gary@intrepid.com> > > gcc/ > * explow.c (promote_mode): For UPC pointer-to-shared values, > return the mode of the UPC PTS representation type. > > Index: gcc/explow.c > =================================================================== > --- gcc/explow.c (.../trunk) (revision 231059) > +++ gcc/explow.c (.../branches/gupc) (revision 231080) > @@ -794,6 +794,8 @@ promote_mode (const_tree type ATTRIBUTE_ > case REFERENCE_TYPE: > case POINTER_TYPE: > *punsignedp = POINTERS_EXTEND_UNSIGNED; > + if (SHARED_TYPE_P (TREE_TYPE (type))) > + return TYPE_MODE (upc_pts_type_node); > return targetm.addr_space.address_mode > (TYPE_ADDR_SPACE (TREE_TYPE (type))); > break;
On Wed, Dec 02, 2015 at 01:10:44PM +0100, Richard Biener wrote: > > UPC pointers-to-shared have an internal representation which is > > defined as a 'struct' with three fields. Special logic is > > needed in promote_mode() to handle this case. > > Errr - but how are 'struct's ever REFERENCE_TYPE or POINTER_TYPE? Then obviously the FE should lower the pointers to RECORD_TYPEs then and present it to the middle-end as RECORD_TYPE. For debug info, the FEs lang flag or whatever on the side info should through some langhook propagate the info that some var is a UPC pointer to shared into the debug info, but the middle-end should treat it as any other RECORD_TYPE afterwards. > > 2015-11-30 Gary Funck <gary@intrepid.com> > > > > gcc/ > > * explow.c (promote_mode): For UPC pointer-to-shared values, > > return the mode of the UPC PTS representation type. > > > > Index: gcc/explow.c > > =================================================================== > > --- gcc/explow.c (.../trunk) (revision 231059) > > +++ gcc/explow.c (.../branches/gupc) (revision 231080) > > @@ -794,6 +794,8 @@ promote_mode (const_tree type ATTRIBUTE_ > > case REFERENCE_TYPE: > > case POINTER_TYPE: > > *punsignedp = POINTERS_EXTEND_UNSIGNED; > > + if (SHARED_TYPE_P (TREE_TYPE (type))) > > + return TYPE_MODE (upc_pts_type_node); > > return targetm.addr_space.address_mode > > (TYPE_ADDR_SPACE (TREE_TYPE (type))); > > break; Jakub
On 12/02/15 13:10:44, Richard Guenther wrote: > On Tue, Dec 1, 2015 at 7:02 AM, Gary Funck <gary@intrepid.com> wrote: > > UPC pointers-to-shared have an internal representation which is > > defined as a 'struct' with three fields. Special logic is > > needed in promote_mode() to handle this case. > > Errr - but how are 'struct's ever REFERENCE_TYPE or POINTER_TYPE? UPC pointers-to-shared types are pointer types, but they typically have a mode that is 2x the size of the normal pointer mode. Currently, 'upc_pts_type_node' is a global tree mode that describes the internal representation of a UPC pointer-to-shared. The code shown below takes the mode of 'upc_pts_type_node' (TIMode, typically on 64 bit targets) and returns that. As noted in an earlier reply, this mode could be saved in a global somewhere (similar to pointer_mode) which could be used to avoid the reference to 'upc_pts_type_node' below. > > +++ gcc/explow.c (.../branches/gupc) (revision 231080) > > @@ -794,6 +794,8 @@ promote_mode (const_tree type ATTRIBUTE_ > > case REFERENCE_TYPE: > > case POINTER_TYPE: > > *punsignedp = POINTERS_EXTEND_UNSIGNED; > > + if (SHARED_TYPE_P (TREE_TYPE (type))) > > + return TYPE_MODE (upc_pts_type_node); > > return targetm.addr_space.address_mode > > (TYPE_ADDR_SPACE (TREE_TYPE (type))); > > break; Asked in previous email: Can an address space be used for UPC pointers-to-shared? Yes, if TIMode were allowed, for example (on 64-bit targets), and that was the mode associated with th special UPC reserved address space number, it would take care of returning the proper mode for UPC pointers-to-shared types. Address spaces are insufficient however to handle the situations where the semantics of pointer arithmetic for UPC pointers-to-shared differs from those of regular "C" pointers, or for conversions, and situations where optimizations of operations involving "C" pointers do not apply to "UPC" shared pointers. So, we can't just dedicate an address space for UPC pointers-to-shared and be done with it. thanks, - Gary
Index: gcc/explow.c =================================================================== --- gcc/explow.c (.../trunk) (revision 231059) +++ gcc/explow.c (.../branches/gupc) (revision 231080) @@ -794,6 +794,8 @@ promote_mode (const_tree type ATTRIBUTE_ case REFERENCE_TYPE: case POINTER_TYPE: *punsignedp = POINTERS_EXTEND_UNSIGNED; + if (SHARED_TYPE_P (TREE_TYPE (type))) + return TYPE_MODE (upc_pts_type_node); return targetm.addr_space.address_mode (TYPE_ADDR_SPACE (TREE_TYPE (type))); break;