On Mon, 2010-10-11 at 14:49 -0700, Ian Romanick wrote: > As has been discussed numerous times, the > assembly-like IRs in Mesa (classic Mesa IR and TGSI) are completely > useless for generating code for GPUs. llvmpipe has shown that they are > also completely useless for generating code for CPUs.
Ian, I don't fully understand the above. I'm not sure if there's actually an disagreement or just different terminology. TGSI is indeed inadequate for optimization (as in TGSI -> TGSI transformations), but it was never intended for that role: there was always the assumption it would be feed into an optimization compiler (ie. TGSI -> xxx -> GPU code). The role of TGSI was to be API agnostic (i.e., support xxx -> TGSI, yyy -> TGSI). I admit this is not advantage the only frontend is Mesa, but TGSI proved to be very useful when used in conjunction with other APIs beyond GL. It also makes hashing easy. FWIW, I'm fine with TGSI being replaced with something else (LLVM IR, mesa/src/glsl's), provided that IR is a superset of TGSI. Also, for the record, I don't think the problem is resemblance to assembly. Take LLVM IR for example: it is an assembly-like IR in SSA form, as every instruction consists of an opcode plus operands, and operands can't be complex expressions. What makes TGSI and Mesa IR inadequate for compilation is not their assembly resemblance, but the fact they lack any other auxiliary structure (a tree form or SSA form would be advisable, but they are merely an one directional list), and therefore are very difficult to walk, analyze, or rewrite. Jose _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
