* File name: ir/ir/irflag.h
* Purpose: Flags to control optimizations.
* Author: Christian Schaefer, Goetz Lindenmaier
- * Modified by:
+ * Modified by: Michael Beck
* Created:
* CVS-ID: $Id$
* Copyright: (c) 1999-2003 Universität Karlsruhe
/**
* @file irflag.h
*
- * Optimization flags.
+ * Flags to customize the behavior of libfirm.
*
* @author Christian Schaefer
+ *
+ * There are the following groups of flags:
+ * 1. Optimization flags.
+ * a) There is a flag, 'optimize' to turn on/off all optimizations.
+ * b) There are flags for each individual optimization. Some flags turns
+ * transformations in several algorithms on/off.
+ * 2. Normalization flags.
+ * These flags steer transformations of the ir that improve it, as removing
+ * dump Phi nodes (one predecessor, all predecessors are equal ...), Ids, Tuples ...
+ * 3. Verbosity flags.
+ * a) Flags to steer the level of the information.
+ * b) Flags to steer in which phase information should be dumped.
+ * 4. Verification flag
+ * This one controls the behavior of node and type verifications
*/
+#ifndef _FIRM_IR_IRFLAG_H_
+#define _FIRM_IR_IRFLAG_H_
-#ifndef _IRFLAG_H_
-#define _IRFLAG_H_
+#include "firm_types.h"
/**
* A container type to load/restore all optimizations
* Default: optimize == 1.
*/
void set_optimize (int value);
-int get_optimize(void);
+int get_optimize(void);
+
+/** This function enables/disables output of information about phases and
+ * controls the verbosity level.
+ *
+ * 0: no output at all.
+ * 1: very short output
+ * >>1: very verbose output.
+ */
+void set_firm_verbosity (int value);
+int get_firm_verbosity (void);
/** Enables/Disables constant folding optimization.
*
* - constant expression evaluation (2 + 5 ==> 7, 3 < 2 ==> false)
* - algebraic simplification (a * 0 ==> 0, a or a ==> a)
* - simplification of tests ( !(a < b) ==> (a >= b))
- * - refining the memory representation
- * - remove store after load
* Default: opt_constant_folding == 1.
*/
void set_opt_constant_folding (int value);
-/** Enables/Disables constant subexpression elimination.
+/** Enables/Disables loop unrolling.
+ *
+ * If opt_loop_unrolling == 1 perform loop_unrolling.
+ * See loop_unrolling.h.
*
- * If opt_cse == 1 perform constant subexpression elimination.
+ * Default: opt_loop_unrolling = 1;
+ */
+void set_opt_loop_unrolling (int value);
+
+/** Enables/Disables output of information about loop unrolling.
+ */
+void set_opt_loop_unrolling_verbose (int value);
+
+/** Enables/Disables removal of redundant Loads and Stores.
+ *
+ * - Remove Store that overwrites a just stored value (WAW).
+ * - Remove Store if it stores a value just loaded (WAR with the same value).
+ * - Remove Load that loads a value just saved (RAW with the same value).
+ * - remove Load that loads a value already loaded (RAR)
+ * - replace Load of constant values with constants (RC)
+ */
+void set_opt_redundant_loadstore(int value);
+
+/** Enables/Disables common subexpression elimination.
+ *
+ * If opt_cse == 1 perform common subexpression elimination.
* Default: opt_cse == 1.
*/
void set_opt_cse (int value);
+/** Returns constant folding optimization setting. */
+int get_opt_cse(void);
+
/** Enables/Disables global constant subexpression elimination.
*
* If opt_global_cse == 1 and opt_cse == 1 perform intra procedure
* constant subexpression elimination for floating nodes. Intra
* procedure cse gets the graph into state "floating". It is necessary
- * to run pre/code motion to get the graph back into state "pinned".
+ * to run pre/code motion to get the graph back into state "op_pin_state_pinned".
* right after a call to local_optimize with global cse turned on.
* Default: opt_global_cse == 0.
*/
void set_opt_global_cse (int value);
+/** Enables/Disables strength reduction.
+ *
+ * If opt_strength_red == 1 perform strength reduction.
+ * See strenth_red.h.
+ *
+ * Default: opt_strength_red = 1;
+ */
+void set_opt_strength_red (int value);
+
+/** Enables/Disables output of information about strength reduction.
+ */
+void set_opt_strength_red_verbose (int value);
+
/** Enables/Disables unreachable code elimination.
+ *
+ * If set, evaluate conditions of conditional branch and replace the
+ * branch with a Jmp/Bad Tuple.
*
* If opt_unreachable_code == 1 replace nodes (except Block,
* Phi and Tuple) with a Bad predecessor by the Bad node.
/** Enables/Disables strong if and loop simplification (in optimize_cf). */
void set_opt_control_flow_strong_simplification(int value);
-/** Enables/Disables removal of critical control flow edges. */
-void set_opt_critical_edges(int value);
-
/** Enables/Disables reassociation.
*
* If opt_reassociation == 1 reassociation is performed.
/** Enables/Disables dead method elimination.
*
- * If opt_dead_node_elimination == 1 methods never called are
+ * If opt_dead_method_elimination == 1 methods never called are
* removed.
- * Default: opt_dead_node_elimination == 1. */
+ * Default: opt_dead_method_elimination == 1.
+ */
void set_opt_dead_method_elimination (int value);
void set_opt_dead_method_elimination_verbose (int value);
-/** Enable/Disables inlining.
+/** Enable/Disables method inlining.
*
* If opt_inline == 1 the inlining transformation is performed.
*/
void set_opt_inline (int value);
-/** Enable/Disable optimization of dynamic method dispatch
+/** Enable/Disable optimization of dynamic method dispatch.
*
* This flag enables/disables the optimization of dynamic method dispatch.
* If the flag is turned on Sel nodes can be replaced by Const nodes representing
* the address of a function.
*/
void set_opt_dyn_meth_dispatch (int value);
+int get_opt_dyn_meth_dispatch (void);
+
+/** Enable/Disable type optimization of cast nodes.
+ *
+ * Controls the optimizations in tropt.h. Default: on.
+ */
+void set_opt_optimize_class_casts (int value);
+void set_opt_optimize_class_casts_verbose (int value);
+
+/** Restricts the behavior of cast optimization.
+ *
+ * If set, downcast are not optimized if they might be
+ * illegal as in (Super)(Sub) (new Super()). Default:
+ * 0 == not suppressed.
+ */
+void set_opt_suppress_downcast_optimization(int value);
+int get_opt_suppress_downcast_optimization(void);
+
+/** Enable/Disable optimization of tail-recursion calls.
+ *
+ * This flag enables/disables the optimization tail-recursion call.
+ * If the flag is turned on tail-recursion calls are optimized into loops.
+ */
+void set_opt_tail_recursion(int value);
+void set_opt_tail_recursion_verbose(int value);
+
+/** Enable/Disable floating of fragile ops.
+ *
+ * This flags enables/disables the floating of fragile operations.
+ * If this flag is on, fragile operations which are known to NOT raise
+ * an exception can be place to other basic blocks.
+ * Otherwise they remain in the block they were created.
+ */
+void set_opt_fragile_ops(int value);
+
+/**
+ * Enable/Disable if conversion.
+ *
+ * If conversion tries to turn Conds into Mux nodes to eliminate
+ * control flow.
+ */
+void set_opt_if_conversion(int value);
+
+/**
+ * Enable/Disable function call optimization.
+ *
+ * Function call optimization detects const and pure functions and
+ * allows the CSE of Call nodes. A const function is one that
+ * do only evaluate it's parameters and did not read or write memory
+ * to compute its results. Pure functions are allowed to read global memory.
+ */
+void set_opt_function_call(int value);
+
+/**
+ * Enable/Disable Confirm node removal during local optimization.
+ */
+void set_opt_remove_confirm(int value);
+
+/**
+ * Enable/Disable scalar replacement optimization.
+ */
+void set_opt_scalar_replacement(int value);
+void set_opt_scalar_replacement_verbose(int value);
+
+/**
+ * Enable/Disable Null exception in Load and Store nodes only.
+ *
+ * If enabled, only Null pointer exception can occur at Load and
+ * store nodes. If it can be proved that the address input of these
+ * nodes is non-null, the exception edge can safely be removed.
+ * If disabled, other exceptions (like unaligned access, read-only memory,
+ * etc.) can occur.
+ *
+ * This flag is enabled by default.
+ */
+void set_opt_ldst_only_null_ptr_exceptions(int value);
+
+/**
+ * Enable/Disable Selection based Null pointer check elimination.
+ *
+ * In languages, where all addresses are always Sel nodes, Null
+ * pointers can only occur as input to Sel nodes.
+ * If Null pointers are the only source for exceptions in Load and
+ * Store nodes (as typical in high level languages), we can eliminate
+ * exception edges from Load and Store when can prove that the Sel
+ * nodes representing the Load/Store address have non-null inputs.
+ * Enabling this flag enables this elimination.
+ *
+ * Enabling this flag is meaningless if ldst_non_null_exceptions is
+ * enabled.
+ *
+ * This flags should be set for Java style languages.
+ */
+void set_opt_sel_based_null_check_elim(int value);
+
+/**
+ * Enable/Disable Automatic construction of Sync nodes during
+ * Firm construction.
+ *
+ * If this flags is set, sequential non-volatile Loads are automatically
+ * rearranged so that they can be executed in parallel by creating Sync nodes.
+ *
+ * This flags should be set for Java style languages.
+ */
+void set_opt_auto_create_sync(int value);
/** Enable/Disable normalizations of the firm representation.
*
- * This flag guards transformations that normalize the firm representation
+ * This flag guards transformations that normalize the Firm representation
* as removing Ids and Tuples, useless Phis, replacing SymConst(id) by
* Const(entity) and others.
* The transformations guarded by this flag are not guarded by flag
* "optimize".
- * Many algorithms operating on firm can not deal with constructs in
+ * Many algorithms operating on Firm can not deal with constructs in
* the non-normalized representation.
- * default: 1
- * @@@ ATTENTION: not all such transformations are guarded by a flag.
+ * default: ON
+ *
+ * @note ATTENTION: not all such transformations are guarded by a flag.
*/
void set_opt_normalize (int value);
+/** Enable/Disable precise exception context.
+ *
+ * If enabled, all exceptions form a barrier for values, as in the
+ * following example:
+ *
+ * @code
+ * a = 1;
+ * b = 3 / 0;
+ * a = 2;
+ * @endcode
+ *
+ * If precise exception handling is enabled, an exception handler see a == 1,
+ * else it might see a == 2.
+ * Enable this for languages with strict exception order like Java.
+ */
+void set_opt_precise_exc_context(int value);
+
+/** Enable/Disable closed world assumption.
+ *
+ * If enabled, optimizations expect to know the "whole world", i.e. no
+ * external types or callers exist.
+ * This enables some powerful optimizations.
+ */
+void set_opt_closed_world(int value);
+
/**
* Save the current optimization state.
*/
*/
void restore_optimization_state(const optimization_state_t *state);
-#endif
+/**
+ * Switches ALL optimizations off.
+ */
+void all_optimizations_off(void);
+
+/**
+ * Possible verification modes.
+ */
+typedef enum _firm_verification_t {
+ FIRM_VERIFICATION_OFF = 0, /**< do not verify nodes at all */
+ FIRM_VERIFICATION_ON = 1, /**< do node verification and assert on error in debug version */
+ FIRM_VERIFICATION_REPORT = 2, /**< do node verification, but report to stderr only */
+ FIRM_VERIFICATION_ERROR_ONLY = 3 /**< do node verification, but NEVER do assert nor report */
+} firm_verification_t;
+
+/** Select verification of IR nodes and types.
+ *
+ * Per default the verification is in mode NODE_VERIFICATION_ASSERT.
+ * Turn the verification off during development to check partial implementations.
+ */
+void do_node_verification(firm_verification_t mode);
+
+#endif /* _FIRM_IR_IRFLAG_H_ */
projects / libfirm / blobdiff