https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88767
--- Comment #14 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Jiu Fu Guo from comment #13) > Hi Richard， > > As checking the changed code as in comment 9, it seems there is another > opportunity to improve the performance: By improving locality of array A > usage. > > Unroll and jam loop1 into loop4 (or unroll and jam loop1 into loop3 after > loop2/loop4 are unrolled completely), this would reduce memory access by > reusing elements of array A. > > It seems not hard to implement this improvement from the source code aspect > (as the example code shown in comment 9). > While I'm thinking about how to implement this in GCC. > > Some concerns are here. It is not a `perfect nest` for these loops: there > are stmts/instructions that belong to the outer loop (loop1) but outside the > inner loop(loop4). > And even delete loop2 (or distribute loop2 out) and unroll loop4, 'store to > array C: C[(l_n*10)+l_m] +=xx` is moved out of the inner loop (loop3), but > still inside the outer loop(loop1). This is not in favor of 'unroll and > jam'. > > Thanks for any comments! > > BR. > Jiufu Guo I've only quickly tried to understand what you are proposing but I think this is out-of scope of our "separate" distribution / interchange / unroll-and-jam transforms but requires interaction of them. Which means the theory is that the graphite based loop nest optimization should catch this kind of locality transform. Which it for sure doesn't do in it's current state (without checking).
