- Generated by
1.8.1.1
|
GCC Middle and Back End API Reference
|
Data Structures | |
| struct | mem_ref_group |
| struct | mem_ref |
| struct | ar_data |
Functions | |
| static void | dump_mem_details (FILE *file, tree base, tree step, HOST_WIDE_INT delta, bool write_p) |
| static void | dump_mem_ref () |
| static struct mem_ref_group * | find_or_create_group () |
| static void | record_ref (struct mem_ref_group *group, gimple stmt, tree mem, HOST_WIDE_INT delta, bool write_p) |
| static void | release_mem_refs () |
| static bool | idx_analyze_ref () |
| static bool | analyze_ref (struct loop *loop, tree *ref_p, tree *base, tree *step, HOST_WIDE_INT *delta, gimple stmt) |
| static bool | gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs, tree ref, bool write_p, gimple stmt) |
| static struct mem_ref_group * | gather_memory_references () |
| static void | prune_ref_by_self_reuse () |
| static HOST_WIDE_INT | ddown () |
| static bool | is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size, HOST_WIDE_INT step, HOST_WIDE_INT delta, unsigned HOST_WIDE_INT distinct_iters, int align_unit) |
| static void | prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, bool by_is_before) |
| static void | prune_ref_by_reuse () |
| static void | prune_group_by_reuse () |
| static void | prune_by_reuse () |
| static bool | should_issue_prefetch_p () |
| static bool | schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, unsigned ahead) |
| static bool | nothing_to_prefetch_p () |
| static int | estimate_prefetch_count () |
| static void | issue_prefetch_ref () |
| static void | issue_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, unsigned ahead) |
| static bool | nontemporal_store_p () |
| static bool | mark_nontemporal_store () |
| static void | emit_mfence_after_loop () |
| static bool | may_use_storent_in_loop_p () |
| static void | mark_nontemporal_stores () |
| static bool | should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc, unsigned factor) |
| static unsigned | determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs, unsigned ninsns, struct tree_niter_desc *desc, HOST_WIDE_INT est_niter) |
| static unsigned | volume_of_references () |
| static unsigned | volume_of_dist_vector () |
| static void | add_subscript_strides (tree access_fn, unsigned stride, HOST_WIDE_INT *strides, unsigned n, struct loop *loop) |
| static unsigned | self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n, struct loop *loop) |
| static bool | determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs, bool no_other_refs) |
| static bool | trip_count_to_ahead_ratio_too_small_p () |
| static bool | mem_ref_count_reasonable_p () |
| static bool | insn_to_prefetch_ratio_too_small_p (unsigned ninsns, unsigned prefetch_count, unsigned unroll_factor) |
| static bool | loop_prefetch_arrays () |
| unsigned int | tree_ssa_prefetch_arrays () |
| static unsigned int | tree_ssa_loop_prefetch () |
| static bool | gate_tree_ssa_loop_prefetch () |
| gimple_opt_pass * | make_pass_loop_prefetch () |
|
static |
Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE at the position corresponding to the loop of the step. N is the depth of the considered loop nest, and, LOOP is its innermost loop.
References HOST_WIDE_INT, loop::inner, loop_outer(), loop::next, and vNULL.
|
static |
Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and STEP are integer constants and iter is number of iterations of LOOP. The reference occurs in statement STMT. Strips nonaddressable component references from REF_P.
First strip off the component references. Ignore bitfields.
Also strip off the real and imagine parts of a complex, so that
they can have the same base.
|
static |
Divides X by BY, rounding down.
Referenced by is_miss_rate_acceptable().
|
static |
Determines the distance till the first reuse of each reference in REFS in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other memory references in the loop. Return false if the analysis fails.
Find the outermost loop of the loop nest of loop (we require that
there are no sibling loops inside the nest). For each loop, determine the amount of data accessed in each iteration.
We use this to estimate whether the reference is evicted from the
cache before its reuse. Bound the volume by the L2 cache size, since above this bound,
all dependence distances are equivalent. Prepare the references in the form suitable for data dependence
analysis. We ignore unanalyzable data references (the results
are used just as a heuristics to estimate temporality of the
references, hence we do not need to worry about correctness). If the dependence cannot be analyzed, assume that there might be
a reuse. The distance vectors are normalized to be always lexicographically
positive, hence we cannot tell just from them whether DDR_A comes
before DDR_B or vice versa. However, it is not important,
anyway -- if DDR_A is close to DDR_B, then it is either reused in
DDR_B (and it is not nontemporal), or it reuses the value of DDR_B
in cache (and marking it as nontemporal would not affect
anything). If this is a dependence in the innermost loop (i.e., the
distances in all superloops are zero) and it is not
the trivial self-dependence with distance zero, record that
the references are not completely independent. Ignore accesses closer than
L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
so that we use nontemporal prefetches e.g. if single memory
location is accessed several times in a single iteration of
the loop.
|
static |
Determine the coefficient by that unroll LOOP, from the information contained in the list of memory references REFS. Description of umber of iterations of LOOP is stored to DESC. NINSNS is the number of insns of the LOOP. EST_NITER is the estimated number of iterations of the loop, or -1 if no estimate is available.
First check whether the loop is not too large to unroll. We ignore
PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us
from unrolling them enough to make exactly one cache line covered by each
iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent
us from unrolling the loops too many times in cases where we only expect
gains from better scheduling and decreasing loop overhead, which is not
the case here. If we unrolled the loop more times than it iterates, the unrolled version
of the loop would be never entered. Choose the factor so that we may prefetch each cache just once,
but bound the unrolling by UPPER_BOUND.
|
static |
Dumps information about memory reference
References mem_ref_group::base, mem_ref::delta, mem_ref::group, mem_ref_group::step, and mem_ref::write_p.
|
static |
Dumps information about reference REF to FILE.
|
static |
Issue a memory fence instruction after LOOP.
If possible, we prefer not to insert the fence on other paths
in cfg.
|
static |
Estimate the number of prefetches in the given GROUPS. UNROLL_FACTOR is the factor by which LOOP was unrolled.
References mem_ref::issue_prefetch_p, issue_prefetch_ref(), mem_ref_group::next, mem_ref::next, and mem_ref_group::refs.
|
staticread |
Finds a group with BASE and STEP in GROUPS, or creates one if it does not exist.
If step is an integer constant, keep the list of groups sorted
by decreasing step.
|
static |
|
staticread |
Record the suitable memory references in LOOP. NO_OTHER_REFS is set to true if there are no other memory references inside the loop.
Scan the loop body in order, so that the former references precede the
later ones.
References gather_memory_references_ref().
|
static |
Record a memory reference REF to the list REFS. The reference occurs in LOOP in statement STMT and it is write if WRITE_P. Returns true if the reference was recorded, false otherwise.
If analyze_ref fails the default is a NULL_TREE. We can stop here.
Stop if the address of BASE could not be taken.
Limit non-constant step prefetching only to the innermost loops and
only when the step is loop invariant in the entire loop nest. Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
are integer constants.
Referenced by gather_memory_references().
|
static |
Analyzes a single INDEX of a memory reference to obtain information described at analyze_ref. Callback for for_each_index.
|
static |
Determine whether or not the instruction to prefetch ratio in the loop is too small based on the profitablity consideration. NINSNS: estimated number of instructions in the loop, PREFETCH_COUNT: an estimate of the number of prefetches, UNROLL_FACTOR: the factor to unroll the loop if prefetching.
Prefetching most likely causes performance degradation when the instruction
to prefetch ratio is too small. Too many prefetch instructions in a loop
may reduce the I-cache performance.
(unroll_factor * ninsns) is used to estimate the number of instructions in
the unrolled loop. This implementation is a bit simplistic -- the number
of issued prefetch instructions is also affected by unrolling. So,
prefetch_mod and the unroll factor should be taken into account when
determining prefetch_count. Also, the number of insns of the unrolled
loop will usually be significantly smaller than the number of insns of the
original loop * unroll_factor (at least the induction variable increases
and the exit branches will get eliminated), so it might be better to use
tree_estimate_loop_size + estimated_unrolled_size.
References single_dom_exit(), and tree_unroll_loop().
|
static |
Given a CACHE_LINE_SIZE and two inductive memory references with a common STEP greater than CACHE_LINE_SIZE and an address difference DELTA, compute the probability that they will fall in different cache lines. Return true if the computed miss rate is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the number of distinct iterations after which the pattern repeats itself. ALIGN_UNIT is the unit of alignment in bytes.
It always misses if delta is greater than or equal to the cache
line size. Iterate through all possible alignments of the first
memory reference within its cache line. Iterate through all distinct iterations.
References cst_and_fits_in_hwi(), ddown(), mem_ref::delta, mem_ref::group, HOST_WIDE_INT, int_cst_value(), mem_ref::prefetch_before, and mem_ref_group::step.
|
static |
Issue prefetches for the reference REF into loop as decided before. HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the factor by which LOOP was unrolled.
Determine the address to prefetch.
The step size is non-constant but loop-invariant. We use the
heuristic to simply prefetch ahead iterations ahead. Create the prefetch instruction.
Referenced by estimate_prefetch_count().
|
static |
Issue prefetches for the references in GROUPS into loop as decided before. HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the factor by that LOOP was unrolled.
|
static |
Issue prefetch instructions for array references in LOOP. Returns true if the LOOP was unrolled.
FIXME: the time should be weighted by the probabilities of the blocks in
the loop body. Prefetching is not likely to be profitable if the trip count to ahead
ratio is too small. Step 1: gather the memory references.
Give up prefetching if the number of memory references in the
loop is not reasonable based on profitablity and compilation time
considerations. Step 2: estimate the reuse effects.
Step 3: determine unroll factor.
Estimate prefetch count for the unrolled loop.
Prefetching is not likely to be profitable if the instruction to prefetch
ratio is too small. Step 4: what to prefetch?
Step 5: unroll the loop. TODO -- peeling of first and last few
iterations so that we do not issue superfluous prefetches. Step 6: issue the prefetches.
| gimple_opt_pass* make_pass_loop_prefetch | ( | ) |
|
static |
If REF is a nontemporal store, we mark the corresponding modify statement and return true. Otherwise, we return false.
|
static |
Marks nontemporal stores in LOOP. GROUPS contains the description of memory references in the loop.
References mem_ref_group::next, mem_ref::next, mem_ref::prefetch_before, and mem_ref_group::refs.
|
static |
Returns true if we can use storent in loop, false otherwise.
If we must issue a mfence insn after using storent, check that there
is a suitable place for it at each of the loop exits.
References least_common_multiple(), and mem_ref::prefetch_mod.
|
static |
Determine whether or not the number of memory references in the loop is reasonable based on the profitablity and compilation time considerations. NINSNS: estimated number of instructions in the loop, MEM_REF_COUNT: total number of memory references in the loop.
Miss rate computation (is_miss_rate_acceptable) and dependence analysis
(compute_all_dependences) have high costs based on quadratic complexity.
To avoid huge compilation time, we give up prefetching if mem_ref_count
is too large. Prefetching improves performance by overlapping cache missing
memory accesses with CPU operations. If the loop does not have
enough CPU operations to overlap with memory operations, prefetching
won't give a significant benefit. One approximate way of checking
this is to require the ratio of instructions to memory references to
be above a certain limit. This approximation works well in practice.
TODO: Implement a more precise computation by estimating the time
for each CPU or memory op in the loop. Time estimates for memory ops
should account for cache misses.
|
static |
Returns true if REF is a memory write for that a nontemporal store insn can be used.
REF must be a write that is not reused. We require it to be independent
on all other memory references in the loop, as the nontemporal stores may
be reordered with respect to other memory references. Check that we have the storent instruction for the mode.
References edge_def::dest, edge_def::flags, and get_loop_exit_edges().
|
static |
Return TRUE if no prefetch is going to be generated in the given GROUPS.
References force_gimple_operand_gsi(), mem_ref::group, GSI_SAME_STMT, mem_ref_group::step, and unshare_expr().
|
static |
Prune the list of prefetch candidates GROUPS using the reuse analysis.
References mem_ref::prefetch_mod, and should_issue_prefetch_p().
|
static |
Prune the prefetch candidates in GROUP using the reuse analysis.
References dump_file, and dump_flags.
|
static |
Prune the prefetch candidate REF using the reuse with BY. If BY_IS_BEFORE is true, BY is before REF in the loop.
If the step is non constant we cannot calculate prefetch_before.
If the references has the same address, only prefetch the
former. If the reference addresses are invariant and fall into the
same cache line, prefetch just the first one. Only prune the reference that is behind in the array.
Transform the data so that we may assume that the accesses
are forward. Check whether the two references are likely to hit the same cache
line, and how distant the iterations in that it occurs are from
each other. The accesses are sure to meet. Let us check when.
Do not reduce prefetch_before if we meet beyond cache size.
A more complicated case with step > prefetch_block. First reduce
the ratio between the step and the cache line size to its simplest
terms. The resulting denominator will then represent the number of
distinct iterations after which each address will go back to its
initial location within the cache line. This computation assumes
that PREFETCH_BLOCK is a power of two. Do not reduce prefetch_before if we meet beyond cache size.
Try also the following iteration.
The ref probably does not reuse by.
References mem_ref_group::step.
|
static |
Prune the prefetch candidate REF using the reuses with other references in REFS.
|
static |
Prune the prefetch candidate REF using the self-reuse.
If the step size is non constant, we cannot calculate prefetch_mod.
Prefetch references to invariant address just once.
|
static |
Records a memory reference MEM in GROUP with offset DELTA and write status WRITE_P. The reference occurs in statement STMT.
Do not record the same address twice.
It does not have to be possible for write reference to reuse the read
prefetch, or vice versa.
|
static |
Release memory references in GROUPS.
References affine_iv_d::base, cst_and_fits_in_hwi(), HOST_WIDE_INT, int_cst_value(), ar_data::loop, loop_containing_stmt(), simple_iv(), affine_iv_d::step, ar_data::step, and ar_data::stmt.
|
static |
Decide which of the prefetch candidates in GROUPS to prefetch. AHEAD is the number of iterations to prefetch ahead (which corresponds to the number of simultaneous instances of one prefetch running at a time). UNROLL_FACTOR is the factor by that the loop is going to be unrolled. Returns true if there is anything to prefetch.
At most SIMULTANEOUS_PREFETCHES should be running at the same time.
The prefetch will run for AHEAD iterations of the original loop, i.e.,
AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration,
it will need a prefetch slot. For now we just take memory references one by one and issue
prefetches for as many as possible. The groups are sorted
starting with the largest step, since the references with
large step are more likely to cause many cache misses. The loop is far from being sufficiently unrolled for this
prefetch. Do not generate prefetch to avoid many redudant
prefetches. If we need to prefetch the reference each PREFETCH_MOD iterations,
and we unroll the loop UNROLL_FACTOR times, we need to insert
ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each
iteration. If more than half of the prefetches would be lost anyway, do not
issue the prefetch.
|
static |
Returns the volume of memory references accessed between two consecutive self-reuses of the reference DR. We consider the subscripts of DR in N loops, and LOOP_SIZES contains the volumes of accesses in each of the loops. LOOP is the innermost loop of the current loop nest.
In the following example:
for (i = 0; i < N; i++)
for (j = 0; j < N; j++)
use (a[j][i]);
the same cache line is accessed each N steps (except if the change from
i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse,
we cannot rely purely on the results of the data dependence analysis.
Instead, we compute the stride of the reference in each loop, and consider
the innermost loop in that the stride is less than cache size. Keep track of the reference corresponding to the subscript, so that we
know its stride.
|
static |
Returns true if we should issue prefetch for REF.
For now do not issue prefetches for only first few of the
iterations. Do not prefetch nontemporal stores.
Referenced by prune_by_reuse().
|
static |
Determines whether we can profitably unroll LOOP FACTOR times, and if this is the case, fill in DESC by the description of number of iterations.
We only consider loops without control flow for unrolling. This is not
a hard restriction -- tree_unroll_loop works with arbitrary loops
as well; but the unrolling/prefetching is usually more profitable for
loops consisting of a single basic block, and we want to limit the
code growth.
|
static |
Prefetching.
| unsigned int tree_ssa_prefetch_arrays | ( | ) |
Issue prefetch instructions for array references in loops.
It is possible to ask compiler for say -mtune=i486 -march=pentium4.
-mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
of processor costs and i486 does not have prefetch, but
-march=pentium4 causes HAVE_prefetch to be true. Ugh. We assume that size of cache line is a power of two, so verify this
here.
|
static |
Determine whether or not the trip count to ahead ratio is too small based on prefitablility consideration. AHEAD: the iteration ahead distance, EST_NITER: the estimated trip count.
Assume trip count to ahead ratio is big enough if the trip count could not
be estimated at compile time.
References dump_file, and dump_flags.
|
static |
Returns the volume of memory references accessed across VEC iterations of loops, whose sizes are described in the LOOP_SIZES array. N is the number of the loops in the nest (length of VEC and LOOP_SIZES vectors).
We ignore the parts of the distance vector in subloops, since usually
the numbers of iterations are much smaller.
|
static |
Returns the total volume of the memory references REFS, taking into account reuses in the innermost loop and cache line size. TODO -- we should also take into account reuses across the iterations of the loops in the loop nest.
Almost always reuses another value?
If several iterations access the same cache line, use the size of
the line divided by this number. Otherwise, a cache line is
accessed in each iteration. TODO -- in the latter case, we should
take the size of the reference into account, rounding it up on cache
line size multiple.
1.8.1.1