Next: , Previous: Costs, Up: Target Macros


17.17 Adjusting the Instruction Scheduler

The instruction scheduler may need a fair amount of machine-specific adjustment in order to produce good code. GCC provides several target hooks for this purpose. It is usually enough to define just a few of them: try the first ones in this list first.

— Target Hook: int TARGET_SCHED_ISSUE_RATE (void)

This hook returns the maximum number of instructions that can ever issue at the same time on the target machine. The default is one. Although the insn scheduler can define itself the possibility of issue an insn on the same cycle, the value can serve as an additional constraint to issue insns on the same simulated processor cycle (see hooks ‘TARGET_SCHED_REORDER’ and ‘TARGET_SCHED_REORDER2’). This value must be constant over the entire compilation. If you need it to vary depending on what the instructions are, you must use ‘TARGET_SCHED_VARIABLE_ISSUE’.

— Target Hook: int TARGET_SCHED_VARIABLE_ISSUE (FILE *file, int verbose, rtx_insn *insn, int more)

This hook is executed by the scheduler after it has scheduled an insn from the ready list. It should return the number of insns which can still be issued in the current cycle. The default is ‘more - 1’ for insns other than CLOBBER and USE, which normally are not counted against the issue rate. You should define this hook if some insns take more machine resources than others, so that fewer insns can follow them in the same cycle. file is either a null pointer, or a stdio stream to write any debug output to. verbose is the verbose level provided by -fsched-verbose-n. insn is the instruction that was scheduled.

— Target Hook: int TARGET_SCHED_ADJUST_COST (rtx_insn *insn, rtx link, rtx_insn *dep_insn, int cost)

This function corrects the value of cost based on the relationship between insn and dep_insn through the dependence link. It should return the new value. The default is to make no adjustment to cost. This can be used for example to specify to the scheduler using the traditional pipeline description that an output- or anti-dependence does not incur the same cost as a data-dependence. If the scheduler using the automaton based pipeline description, the cost of anti-dependence is zero and the cost of output-dependence is maximum of one and the difference of latency times of the first and the second insns. If these values are not acceptable, you could use the hook to modify them too. See also see Processor pipeline description.

— Target Hook: int TARGET_SCHED_ADJUST_PRIORITY (rtx_insn *insn, int priority)

This hook adjusts the integer scheduling priority priority of insn. It should return the new priority. Increase the priority to execute insn earlier, reduce the priority to execute insn later. Do not define this hook if you do not need to adjust the scheduling priorities of insns.

— Target Hook: int TARGET_SCHED_REORDER (FILE *file, int verbose, rtx_insn **ready, int *n_readyp, int clock)

This hook is executed by the scheduler after it has scheduled the ready list, to allow the machine description to reorder it (for example to combine two small instructions together on ‘VLIW’ machines). file is either a null pointer, or a stdio stream to write any debug output to. verbose is the verbose level provided by -fsched-verbose-n. ready is a pointer to the ready list of instructions that are ready to be scheduled. n_readyp is a pointer to the number of elements in the ready list. The scheduler reads the ready list in reverse order, starting with ready[*n_readyp − 1] and going to ready[0]. clock is the timer tick of the scheduler. You may modify the ready list and the number of ready insns. The return value is the number of insns that can issue this cycle; normally this is just issue_rate. See also ‘TARGET_SCHED_REORDER2’.

— Target Hook: int TARGET_SCHED_REORDER2 (FILE *file, int verbose, rtx_insn **ready, int *n_readyp, int clock)

Like ‘TARGET_SCHED_REORDER’, but called at a different time. That function is called whenever the scheduler starts a new cycle. This one is called once per iteration over a cycle, immediately after ‘TARGET_SCHED_VARIABLE_ISSUE’; it can reorder the ready list and return the number of insns to be scheduled in the same cycle. Defining this hook can be useful if there are frequent situations where scheduling one insn causes other insns to become ready in the same cycle. These other insns can then be taken into account properly.

— Target Hook: bool TARGET_SCHED_MACRO_FUSION_P (void)

This hook is used to check whether target platform supports macro fusion.

— Target Hook: bool TARGET_SCHED_MACRO_FUSION_PAIR_P (rtx_insn *prev, rtx_insn *curr)

This hook is used to check whether two insns should be macro fused for a target microarchitecture. If this hook returns true for the given insn pair (prev and curr), the scheduler will put them into a sched group, and they will not be scheduled apart. The two insns will be either two SET insns or a compare and a conditional jump and this hook should validate any dependencies needed to fuse the two insns together.

— Target Hook: void TARGET_SCHED_DEPENDENCIES_EVALUATION_HOOK (rtx_insn *head, rtx_insn *tail)

This hook is called after evaluation forward dependencies of insns in chain given by two parameter values (head and tail correspondingly) but before insns scheduling of the insn chain. For example, it can be used for better insn classification if it requires analysis of dependencies. This hook can use backward and forward dependencies of the insn scheduler because they are already calculated.

— Target Hook: void TARGET_SCHED_INIT (FILE *file, int verbose, int max_ready)

This hook is executed by the scheduler at the beginning of each block of instructions that are to be scheduled. file is either a null pointer, or a stdio stream to write any debug output to. verbose is the verbose level provided by -fsched-verbose-n. max_ready is the maximum number of insns in the current scheduling region that can be live at the same time. This can be used to allocate scratch space if it is needed, e.g. by ‘TARGET_SCHED_REORDER’.

— Target Hook: void TARGET_SCHED_FINISH (FILE *file, int verbose)

This hook is executed by the scheduler at the end of each block of instructions that are to be scheduled. It can be used to perform cleanup of any actions done by the other scheduling hooks. file is either a null pointer, or a stdio stream to write any debug output to. verbose is the verbose level provided by -fsched-verbose-n.

— Target Hook: void TARGET_SCHED_INIT_GLOBAL (FILE *file, int verbose, int old_max_uid)

This hook is executed by the scheduler after function level initializations. file is either a null pointer, or a stdio stream to write any debug output to. verbose is the verbose level provided by -fsched-verbose-n. old_max_uid is the maximum insn uid when scheduling begins.

— Target Hook: void TARGET_SCHED_FINISH_GLOBAL (FILE *file, int verbose)

This is the cleanup hook corresponding to TARGET_SCHED_INIT_GLOBAL. file is either a null pointer, or a stdio stream to write any debug output to. verbose is the verbose level provided by -fsched-verbose-n.

— Target Hook: rtx TARGET_SCHED_DFA_PRE_CYCLE_INSN (void)

The hook returns an RTL insn. The automaton state used in the pipeline hazard recognizer is changed as if the insn were scheduled when the new simulated processor cycle starts. Usage of the hook may simplify the automaton pipeline description for some VLIW processors. If the hook is defined, it is used only for the automaton based pipeline description. The default is not to change the state when the new simulated processor cycle starts.

— Target Hook: void TARGET_SCHED_INIT_DFA_PRE_CYCLE_INSN (void)

The hook can be used to initialize data used by the previous hook.

— Target Hook: rtx_insn * TARGET_SCHED_DFA_POST_CYCLE_INSN (void)

The hook is analogous to ‘TARGET_SCHED_DFA_PRE_CYCLE_INSN’ but used to changed the state as if the insn were scheduled when the new simulated processor cycle finishes.

— Target Hook: void TARGET_SCHED_INIT_DFA_POST_CYCLE_INSN (void)

The hook is analogous to ‘TARGET_SCHED_INIT_DFA_PRE_CYCLE_INSN’ but used to initialize data used by the previous hook.

— Target Hook: void TARGET_SCHED_DFA_PRE_ADVANCE_CYCLE (void)

The hook to notify target that the current simulated cycle is about to finish. The hook is analogous to ‘TARGET_SCHED_DFA_PRE_CYCLE_INSN’ but used to change the state in more complicated situations - e.g., when advancing state on a single insn is not enough.

— Target Hook: void TARGET_SCHED_DFA_POST_ADVANCE_CYCLE (void)

The hook to notify target that new simulated cycle has just started. The hook is analogous to ‘TARGET_SCHED_DFA_POST_CYCLE_INSN’ but used to change the state in more complicated situations - e.g., when advancing state on a single insn is not enough.

— Target Hook: int TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD (void)

This hook controls better choosing an insn from the ready insn queue for the DFA-based insn scheduler. Usually the scheduler chooses the first insn from the queue. If the hook returns a positive value, an additional scheduler code tries all permutations of ‘TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD ()’ subsequent ready insns to choose an insn whose issue will result in maximal number of issued insns on the same cycle. For the VLIW processor, the code could actually solve the problem of packing simple insns into the VLIW insn. Of course, if the rules of VLIW packing are described in the automaton.

This code also could be used for superscalar RISC processors. Let us consider a superscalar RISC processor with 3 pipelines. Some insns can be executed in pipelines A or B, some insns can be executed only in pipelines B or C, and one insn can be executed in pipeline B. The processor may issue the 1st insn into A and the 2nd one into B. In this case, the 3rd insn will wait for freeing B until the next cycle. If the scheduler issues the 3rd insn the first, the processor could issue all 3 insns per cycle.

Actually this code demonstrates advantages of the automaton based pipeline hazard recognizer. We try quickly and easy many insn schedules to choose the best one.

The default is no multipass scheduling.

— Target Hook: int TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD (rtx_insn *insn, int ready_index)

This hook controls what insns from the ready insn queue will be considered for the multipass insn scheduling. If the hook returns zero for insn, the insn will be considered in multipass scheduling. Positive return values will remove insn from consideration on the current round of multipass scheduling. Negative return values will remove insn from consideration for given number of cycles. Backends should be careful about returning non-zero for highest priority instruction at position 0 in the ready list. ready_index is passed to allow backends make correct judgements.

The default is that any ready insns can be chosen to be issued.

— Target Hook: void TARGET_SCHED_FIRST_CYCLE_MULTIPASS_BEGIN (void *data, signed char *ready_try, int n_ready, bool first_cycle_insn_p)

This hook prepares the target backend for a new round of multipass scheduling.

— Target Hook: void TARGET_SCHED_FIRST_CYCLE_MULTIPASS_ISSUE (void *data, signed char *ready_try, int n_ready, rtx_insn *insn, const void *prev_data)

This hook is called when multipass scheduling evaluates instruction INSN.

— Target Hook: void TARGET_SCHED_FIRST_CYCLE_MULTIPASS_BACKTRACK (const void *data, signed char *ready_try, int n_ready)

This is called when multipass scheduling backtracks from evaluation of an instruction.

— Target Hook: void TARGET_SCHED_FIRST_CYCLE_MULTIPASS_END (const void *data)

This hook notifies the target about the result of the concluded current round of multipass scheduling.

— Target Hook: void TARGET_SCHED_FIRST_CYCLE_MULTIPASS_INIT (void *data)

This hook initializes target-specific data used in multipass scheduling.

— Target Hook: void TARGET_SCHED_FIRST_CYCLE_MULTIPASS_FINI (void *data)

This hook finalizes target-specific data used in multipass scheduling.

— Target Hook: int TARGET_SCHED_DFA_NEW_CYCLE (FILE *dump, int verbose, rtx_insn *insn, int last_clock, int clock, int *sort_p)

This hook is called by the insn scheduler before issuing insn on cycle clock. If the hook returns nonzero, insn is not issued on this processor cycle. Instead, the processor cycle is advanced. If *sort_p is zero, the insn ready queue is not sorted on the new cycle start as usually. dump and verbose specify the file and verbosity level to use for debugging output. last_clock and clock are, respectively, the processor cycle on which the previous insn has been issued, and the current processor cycle.

— Target Hook: bool TARGET_SCHED_IS_COSTLY_DEPENDENCE (struct _dep *_dep, int cost, int distance)

This hook is used to define which dependences are considered costly by the target, so costly that it is not advisable to schedule the insns that are involved in the dependence too close to one another. The parameters to this hook are as follows: The first parameter _dep is the dependence being evaluated. The second parameter cost is the cost of the dependence as estimated by the scheduler, and the third parameter distance is the distance in cycles between the two insns. The hook returns true if considering the distance between the two insns the dependence between them is considered costly by the target, and false otherwise.

Defining this hook can be useful in multiple-issue out-of-order machines, where (a) it's practically hopeless to predict the actual data/resource delays, however: (b) there's a better chance to predict the actual grouping that will be formed, and (c) correctly emulating the grouping can be very important. In such targets one may want to allow issuing dependent insns closer to one another—i.e., closer than the dependence distance; however, not in cases of “costly dependences”, which this hooks allows to define.

— Target Hook: void TARGET_SCHED_H_I_D_EXTENDED (void)

This hook is called by the insn scheduler after emitting a new instruction to the instruction stream. The hook notifies a target backend to extend its per instruction data structures.

— Target Hook: void * TARGET_SCHED_ALLOC_SCHED_CONTEXT (void)

Return a pointer to a store large enough to hold target scheduling context.

— Target Hook: void TARGET_SCHED_INIT_SCHED_CONTEXT (void *tc, bool clean_p)

Initialize store pointed to by tc to hold target scheduling context. It clean_p is true then initialize tc as if scheduler is at the beginning of the block. Otherwise, copy the current context into tc.

— Target Hook: void TARGET_SCHED_SET_SCHED_CONTEXT (void *tc)

Copy target scheduling context pointed to by tc to the current context.

— Target Hook: void TARGET_SCHED_CLEAR_SCHED_CONTEXT (void *tc)

Deallocate internal data in target scheduling context pointed to by tc.

— Target Hook: void TARGET_SCHED_FREE_SCHED_CONTEXT (void *tc)

Deallocate a store for target scheduling context pointed to by tc.

— Target Hook: int TARGET_SCHED_SPECULATE_INSN (rtx_insn *insn, unsigned int dep_status, rtx *new_pat)

This hook is called by the insn scheduler when insn has only speculative dependencies and therefore can be scheduled speculatively. The hook is used to check if the pattern of insn has a speculative version and, in case of successful check, to generate that speculative pattern. The hook should return 1, if the instruction has a speculative form, or −1, if it doesn't. request describes the type of requested speculation. If the return value equals 1 then new_pat is assigned the generated speculative pattern.

— Target Hook: bool TARGET_SCHED_NEEDS_BLOCK_P (unsigned int dep_status)

This hook is called by the insn scheduler during generation of recovery code for insn. It should return true, if the corresponding check instruction should branch to recovery code, or false otherwise.

— Target Hook: rtx TARGET_SCHED_GEN_SPEC_CHECK (rtx_insn *insn, rtx_insn *label, unsigned int ds)

This hook is called by the insn scheduler to generate a pattern for recovery check instruction. If mutate_p is zero, then insn is a speculative instruction for which the check should be generated. label is either a label of a basic block, where recovery code should be emitted, or a null pointer, when requested check doesn't branch to recovery code (a simple check). If mutate_p is nonzero, then a pattern for a branchy check corresponding to a simple check denoted by insn should be generated. In this case label can't be null.

— Target Hook: void TARGET_SCHED_SET_SCHED_FLAGS (struct spec_info_def *spec_info)

This hook is used by the insn scheduler to find out what features should be enabled/used. The structure *spec_info should be filled in by the target. The structure describes speculation types that can be used in the scheduler.

— Target Hook: int TARGET_SCHED_SMS_RES_MII (struct ddg *g)

This hook is called by the swing modulo scheduler to calculate a resource-based lower bound which is based on the resources available in the machine and the resources required by each instruction. The target backend can use g to calculate such bound. A very simple lower bound will be used in case this hook is not implemented: the total number of instructions divided by the issue rate.

— Target Hook: bool TARGET_SCHED_DISPATCH (rtx_insn *insn, int x)

This hook is called by Haifa Scheduler. It returns true if dispatch scheduling is supported in hardware and the condition specified in the parameter is true.

— Target Hook: void TARGET_SCHED_DISPATCH_DO (rtx_insn *insn, int x)

This hook is called by Haifa Scheduler. It performs the operation specified in its second parameter.

— Target Hook: bool TARGET_SCHED_EXPOSED_PIPELINE

True if the processor has an exposed pipeline, which means that not just the order of instructions is important for correctness when scheduling, but also the latencies of operations.

— Target Hook: int TARGET_SCHED_REASSOCIATION_WIDTH (unsigned int opc, machine_mode mode)

This hook is called by tree reassociator to determine a level of parallelism required in output calculations chain.

— Target Hook: void TARGET_SCHED_FUSION_PRIORITY (rtx_insn *insn, int max_pri, int *fusion_pri, int *pri)

This hook is called by scheduling fusion pass. It calculates fusion priorities for each instruction passed in by parameter. The priorities are returned via pointer parameters.

insn is the instruction whose priorities need to be calculated. max_pri is the maximum priority can be returned in any cases. fusion_pri is the pointer parameter through which insn's fusion priority should be calculated and returned. pri is the pointer parameter through which insn's priority should be calculated and returned.

Same fusion_pri should be returned for instructions which should be scheduled together. Different pri should be returned for instructions with same fusion_pri. fusion_pri is the major sort key, pri is the minor sort key. All instructions will be scheduled according to the two priorities. All priorities calculated should be between 0 (exclusive) and max_pri (inclusive). To avoid false dependencies, fusion_pri of instructions which need to be scheduled together should be smaller than fusion_pri of irrelevant instructions.

Given below example:

ldr r10, [r1, 4] add r4, r4, r10 ldr r15, [r2, 8] sub r5, r5, r15 ldr r11, [r1, 0] add r4, r4, r11 ldr r16, [r2, 12] sub r5, r5, r16

On targets like ARM/AArch64, the two pairs of consecutive loads should be merged. Since peephole2 pass can't help in this case unless consecutive loads are actually next to each other in instruction flow. That's where this scheduling fusion pass works. This hook calculates priority for each instruction based on its fustion type, like:

ldr r10, [r1, 4] ; fusion_pri=99, pri=96 add r4, r4, r10 ; fusion_pri=100, pri=100 ldr r15, [r2, 8] ; fusion_pri=98, pri=92 sub r5, r5, r15 ; fusion_pri=100, pri=100 ldr r11, [r1, 0] ; fusion_pri=99, pri=100 add r4, r4, r11 ; fusion_pri=100, pri=100 ldr r16, [r2, 12] ; fusion_pri=98, pri=88 sub r5, r5, r16 ; fusion_pri=100, pri=100

Scheduling fusion pass then sorts all ready to issue instructions according to the priorities. As a result, instructions of same fusion type will be pushed together in instruction flow, like:

ldr r11, [r1, 0] ldr r10, [r1, 4] ldr r15, [r2, 8] ldr r16, [r2, 12] add r4, r4, r10 sub r5, r5, r15 add r4, r4, r11 sub r5, r5, r16

Now peephole2 pass can simply merge the two pairs of loads.

Since scheduling fusion pass relies on peephole2 to do real fusion work, it is only enabled by default when peephole2 is in effect.

This is firstly introduced on ARM/AArch64 targets, please refer to the hook implementation for how different fusion types are supported.