On Fri, Sep 10, 2010 at 3:15 PM, Paulo J. Matos <pocma...@gmail.com> wrote:
> Hello,
>
> I am working on the new pass (previously discussed), to optimise switch
> cases.
>
> I am almost finishing it, however, for practical reasons I am
> implementing it first over GCC4.3 and once tested, will port it to svn
> trunk and post it on gcc-patches.
>
> There are something that are still not working. Mainly the issue is that
> GCC doesn't really accept the code I generate during the patch.
>
> Some issues:
>
> 1. To generate a conditional (like if(...){} else {}) I am using a
> COND_EXPR;
> 2. And when I have a complex condition I generate a TRUTH_ANDIF_EXPR and
> then corresponding LE_EXPR, or LT_EXPR, etc.
> 3. To iterate through the tree I am using a tree_stmt_iterator. And
> using tsi_delink to remove stmts from the tree and tsi_link_before/after
> to add stmts to the tree.Which is wrong. You need to use block_stmt_iterator and bsi_remove and ... (I don't remember, 4.3 is so old). > After my pass is run I get: > ,---- > | switch_compact.c: In function 'g': > | switch_compact.c:7: error: bb_for_stmt (stmt) is set to a wrong basic > | block > | switch_compact.c:7: error: invalid conditional operand > | 8 <= n_1(D) && n_1(D) <= 17; > | > | switch_compact.c:7: internal compiler error: verify_stmts failed > `---- > > Could you please comment on the above issues so I can try to track down > the problem? You need to properly update the CFG > PS. Note that the reason why I am not posting the patch is simply > because I assume you only look at patches for gcc-trunk. However, if you > don't mind taking a look at a pass for GCC4.3 at this point, do let me > know and I will post it (either here on in gcc-patches). You will have to transition to a completely different GIMPLE representation when moving to trunk, so I think you are wasting time here. Richard.
