On 12 October 2011 17:54, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
Hi,
>
> This patch allows vectorization of some loops that use
> bool (which is especially important now that we use bool more often
> even for stmts that weren't originally using bool in the sources),
> in particular (when bool is cast to an integer type, and the bool rhs
> has def stmts within the loop as either BIT_{AND,IOR,XOR}_EXPR,
> or just SSA_NAME assigns or bool -> another bool casts, or comparisons
> (tested recursively). In that case the pattern recognizer transforms
> the comparisons into COND_EXPRs using suitable integer type (the same width
> as the comparison operands) and other bools to suitable integer types
> with casts added where needed.
>
> The patch doesn't yet handle vectorization of storing into a bool array,
> I'll work on that later.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux. Ok for trunk?
OK with:
> +
> +/* Helper function of adjust_bool_pattern. Add a cast to TYPE to a previous
> + stmt (SSA_NAME_DEF_STMT of VAR), but moving the COND_EXPR from
> RELATED_STMT
by moving?
> + to PATTERN_DEF_STMT and adding a cast as RELATED_STMT. */
> +
> +static tree
> +adjust_bool_pattern_cast (tree type, tree var)
> +{
> + stmt_vec_info stmt_vinfo = vinfo_for_stmt (SSA_NAME_DEF_STMT (var));
> + gimple cast_stmt, pattern_stmt;
> +
> + gcc_assert (!STMT_VINFO_PATTERN_DEF_STMT (stmt_vinfo));
> + pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_vinfo);
> + STMT_VINFO_PATTERN_DEF_STMT (stmt_vinfo) = pattern_stmt;
> + cast_stmt
> + = gimple_build_assign_with_ops (NOP_EXPR,
> + vect_recog_temp_ssa_var (type, NULL),
> + gimple_assign_lhs (pattern_stmt),
> + NULL_TREE);
> + STMT_VINFO_RELATED_STMT (stmt_vinfo) = cast_stmt;
> + return gimple_assign_lhs (cast_stmt);
> +}
> +
> +
> +/* Helper function of vect_recog_bool_pattern. Do the actual
> transformations,
> + recursively. VAR is an SSA_NAME that should be transformed from bool
> + to a wider integer type, OUT_TYPE is the desired final integer type of
> + the whole pattern, TRUEVAL should be NULL unless optimizing
> + BIT_AND_EXPR into a COND_EXPR with one integer from one of the operands
> + in the then_clause, STMTS is where statements with added pattern stmts
> + should be pushed to. */
> +
> +static tree
> +adjust_bool_pattern (tree var, tree out_type, tree trueval,
> + VEC (gimple, heap) **stmts)
> +{
> + gimple stmt = SSA_NAME_DEF_STMT (var);
> + enum tree_code rhs_code, def_rhs_code;
> + tree itype, cond_expr, rhs1, rhs2, irhs1, irhs2;
> + location_t loc;
> + gimple pattern_stmt, def_stmt;
> +
> + rhs1 = gimple_assign_rhs1 (stmt);
> + rhs2 = gimple_assign_rhs2 (stmt);
> + rhs_code = gimple_assign_rhs_code (stmt);
> + loc = gimple_location (stmt);
> + switch (rhs_code)
> + {
> + case SSA_NAME:
> + CASE_CONVERT:
> + irhs1 = adjust_bool_pattern (rhs1, out_type, NULL_TREE, stmts);
> + itype = TREE_TYPE (irhs1);
> + pattern_stmt
> + = gimple_build_assign_with_ops (SSA_NAME,
> + vect_recog_temp_ssa_var (itype, NULL),
> + irhs1, NULL_TREE);
> + break;
> +
> + case BIT_NOT_EXPR:
> + irhs1 = adjust_bool_pattern (rhs1, out_type, NULL_TREE, stmts);
> + itype = TREE_TYPE (irhs1);
> + pattern_stmt
> + = gimple_build_assign_with_ops (BIT_XOR_EXPR,
> + vect_recog_temp_ssa_var (itype, NULL),
> + irhs1, build_int_cst (itype, 1));
> + break;
> +
> + case BIT_AND_EXPR:
> + /* Try to optimize x = y & (a < b ? 1 : 0); into
> + x = (a < b ? y : 0); */
Could you please add some more explanations here? I found it very
difficult to follow. It would be nice to have an example here (similar
to vect_recog_bool_pattern) to illustrate what these statements and
operands are.
> + def_stmt = SSA_NAME_DEF_STMT (rhs2);
> + def_rhs_code = gimple_assign_rhs_code (def_stmt);
> + if (TREE_CODE_CLASS (def_rhs_code) == tcc_comparison)
> + {
> + tree def_rhs1 = gimple_assign_rhs1 (def_stmt);
> + irhs1 = adjust_bool_pattern (rhs1, out_type, NULL_TREE, stmts);
> + if (TYPE_PRECISION (TREE_TYPE (irhs1))
> + == GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (def_rhs1))))
> + {
> + gimple tstmt;
> + stmt_vec_info stmt_def_vinfo = vinfo_for_stmt (def_stmt);
> + irhs2 = adjust_bool_pattern (rhs2, out_type, irhs1, stmts);
> + tstmt = VEC_pop (gimple, *stmts);
> + gcc_assert (tstmt == def_stmt);
> + VEC_quick_push (gimple, *stmts, stmt);
> + STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt))
> + = STMT_VINFO_RELATED_STMT (stmt_def_vinfo);
> + gcc_assert (!STMT_VINFO_PATTERN_DEF_STMT (stmt_def_vinfo));
> + STMT_VINFO_RELATED_STMT (stmt_def_vinfo) = NULL;
> + return irhs2;
> + }
> + else
> + irhs2 = adjust_bool_pattern (rhs2, out_type, NULL_TREE, stmts);
> + goto and_ior_xor;
> + }
> + def_stmt = SSA_NAME_DEF_STMT (rhs1);
> + def_rhs_code = gimple_assign_rhs_code (def_stmt);
> + if (TREE_CODE_CLASS (def_rhs_code) == tcc_comparison)
> + {
> + tree def_rhs1 = gimple_assign_rhs1 (def_stmt);
> + irhs2 = adjust_bool_pattern (rhs2, out_type, NULL_TREE, stmts);
> + if (TYPE_PRECISION (TREE_TYPE (irhs2))
> + == GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (def_rhs1))))
> + {
> + gimple tstmt;
> + stmt_vec_info stmt_def_vinfo = vinfo_for_stmt (def_stmt);
> + irhs1 = adjust_bool_pattern (rhs1, out_type, irhs2, stmts);
> + tstmt = VEC_pop (gimple, *stmts);
> + gcc_assert (tstmt == def_stmt);
> + VEC_quick_push (gimple, *stmts, stmt);
> + STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt))
> + = STMT_VINFO_RELATED_STMT (stmt_def_vinfo);
> + gcc_assert (!STMT_VINFO_PATTERN_DEF_STMT (stmt_def_vinfo));
> + STMT_VINFO_RELATED_STMT (stmt_def_vinfo) = NULL;
> + return irhs1;
> + }
> + else
> + irhs1 = adjust_bool_pattern (rhs1, out_type, NULL_TREE, stmts);
> + goto and_ior_xor;
> + }
> + /* FALLTHRU */
Thanks,
Ira
