Graphite phone call from 2009_03_03
Attendees: Konrad, David, Kenny, Tobias, Christophe, Sebastian, Albert, Razya
Discussed topics:
- Scop detection discards conditions with EQ_EXPR, as these could produce in the else part non convex domains. Because of this, Graphite cannot handle FORTRAN do loops. Two solutions: a first quick fix would handle only the special case of FORTRAN do loops, a second more general solution would make the iteration domains use a union of convex polyhedra, such that conditions with EQ_EXPR or NE_EXPR could be handled.
- Handle reductions in the code generation
- Kenny's patch for CFG structural analysis splits basic blocks, and that pass does not conserve the loop closed SSA form, causing ICE in verify_ssa.
- limit_scop limitations
- handle array accesses with unknown subscripts, this would avoid splitting SCoPs because a data reference access function was not analyzed.
- scattering domains:
- transforms will be implemented as scattering functions
- read and write scattering functions on disk with -fgraphite-write and -fgraphite-read
- access functions with parameters