On Wed, Oct 26, 2016 at 6:42 PM, Bin Cheng <bin.ch...@arm.com> wrote:
> Hi,
> For stmt defining reduction, GCC vectorizer assumes that the reduction
> variable is always the last (second) operand. Another fact is that
> vectorizer doesn't swap operands for cond_reduction during analysis stage.
> The problem is GCC middle-end may canonicalize cond_expr into a form that
> reduction variable is not the last one. At the moment, such case cannot be
> vectorized.
> The patch fixes this issue by swapping operands in cond_reduction when it's
> necessary. The patch also swaps it back if vectorization fails. The patch
> resolves failures introduced by previous match.pd patches. In addition,
> couple cases are XPASSed on AArch64 now, which means more loops are
> vectorized. I will send following patch addressing those XPASS tests.
> Bootstrap and test on x86_64 and AArch64 ongoing, is it OK?
@@ -1225,6 +1225,20 @@ destroy_loop_vec_info (loop_vec_info
loop_vinfo, bool clean_stmts)
swap_ssa_operands (stmt,
gimple_assign_rhs1_ptr (stmt),
gimple_assign_rhs2_ptr (stmt));
+ else if (code == COND_EXPR
+ && CONSTANT_CLASS_P (gimple_assign_rhs2 (stmt)))
+ {
+ tree cond_expr = gimple_assign_rhs1 (stmt);
+ enum tree_code cond_code = TREE_CODE (cond_expr);
+
+ gcc_assert (TREE_CODE_CLASS (cond_code) == tcc_comparison);
+ /* HONOR_NANS doesn't matter when inverting it back. */
I think this doesn't hold true for COND_EXPRs that were originally
this way as canonicalization
is also inhibited by this. I suggest to simply not invert back when
cond_code == ERROR_MARK
as we can't possibly have swapped it to the current non-canonical way.
Ok with that change.
Thanks,
Richard.
+ cond_code = invert_tree_comparison (cond_code, false);
+ gcc_assert (cond_code != ERROR_MARK);
+ TREE_SET_CODE (cond_expr, cond_code);
+ swap_ssa_operands (stmt, gimple_assign_rhs2_ptr (stmt),
+ gimple_assign_rhs3_ptr (stmt));
> Thanks,
> bin
>
> 2016-10-25 Bin Cheng <bin.ch...@arm.com>
>
> * tree-vect-loop.c (destroy_loop_vec_info): Handle cond_expr.
> (vect_is_simple_reduction): Swap cond_reduction by inversion.
