http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47046
--- Comment #3 from Sebastian Pop <spop at gcc dot gnu.org> 2011-07-26 18:48:57 UTC --- Author: spop Date: Tue Jul 26 18:48:49 2011 New Revision: 176805 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=176805 Log: Fix PR47046: correct evolution_function_is_affine_p "Bug 47046 - gcc.target/i386/sse4_1-movntdqa.c ICEs with -fgraphite-identity" The problem here is that we are left with the following code to be translated in the new representation following the transform that Graphite has chosen: D.2709_14 = j_33 * i_32; D.2710_15 = D.2709_14 * i_32; D.2711_16 = D.2710_15 * sign_34; *D.2708_13 = D.2711_16; In this particular case we have a nonlinear expression "i * i" for which we have to generate code following the new graphite_iv variables. The patch fixes the function that detects whether we are passing non linear stuff to graphite: evolution_function_is_affine_p. It seems like for the moment evolution_function_is_affine_p is testing whether an evolution function is affine only in the innermost loop, without looking recursively at what happens in outer loops. The chrec for this case is: {0, +, {0, +, {1, +, 2}_1}_1}_2 and we are testing whether the evolution is affine only for the loop_2, which is true as we have {0, +, blah}_2 with blah invariant in loop_2. The patch adds the recursive call to evolution_function_is_affine_p. Bootstrapped and tested on amd64-linux. 2011-07-26 Sebastian Pop <sebastian....@amd.com> PR middle-end/47046 * tree-chrec.h (evolution_function_is_affine_p): Recursively call evolution_function_is_affine_p on CHREC_RIGHT. * gcc.dg/graphite/id-pr47046.c: New. Added: trunk/gcc/testsuite/gcc.dg/graphite/id-pr47046.c Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/tree-chrec.h
