- Generated by
1.8.1.1
|
GCC Middle and Back End API Reference
|
Data Structures | |
| struct | omp_region |
| struct | omp_context |
| struct | omp_for_data_loop |
| struct | omp_for_data |
| struct | omp_taskcopy_context |
Typedefs | |
| typedef struct omp_context | omp_context |
Variables | |
| static splay_tree | all_contexts |
| static int | taskreg_nesting_level |
| static int | target_nesting_level |
| static struct omp_region * | root_omp_region |
| static bitmap | task_shared_vars |
| static unsigned int | tmp_ompfn_id_num |
| static splay_tree | critical_name_mutexes |
| static splay_tree | all_labels |
| typedef struct omp_context omp_context |
Context structure. Used to store information about each parallel directive in the code.
|
static |
Re-gimplification and code generation routines.
Build a call to GOMP_barrier.
|
static |
Scan the CFG and build a tree of OMP regions. Return the root of the OMP region tree.
|
static |
Helper for build_omp_regions. Scan the dominator tree starting at block BB. PARENT is the region that contains BB. If SINGLE_TREE is true, the function ends once a single tree is built (otherwise, whole forest of OMP constructs may be built).
STMT is the return point out of region PARENT. Mark it
as the exit point and make PARENT the immediately
enclosing region. GIMPLE_OMP_ATOMIC_STORE is analoguous to
GIMPLE_OMP_RETURN, but matches with
GIMPLE_OMP_ATOMIC_LOAD. GIMPLE_OMP_SECTIONS_SWITCH is part of
GIMPLE_OMP_SECTIONS, and we do nothing for it. Otherwise, this directive becomes the parent for a new
region.
|
static |
Builds the tree of OMP regions rooted at ROOT, storing it to root_omp_region.
|
static |
Build tree nodes to access VAR in the scope outer to CTX. In the case of a parallel, this is a component reference; for workshare constructs this is some variable.
#pragma omp simd isn't a worksharing construct, and can reference even
private vars in its linear etc. clauses. This can happen with orphaned constructs. If var is reference, it is
possible it is shared and as such valid.
Referenced by omp_copy_decl_2().
|
static |
Build tree nodes to access the field for VAR on the receiver side.
If the receiver record type was remapped in the child function,
remap the field into the new record type.
Referenced by omp_copy_decl_1().
|
static |
Build tree nodes to access the field for VAR on the sender side.
Referenced by lower_reduction_clauses().
|
static |
Callback for walk_stmts. Check if the current statement only contains GIMPLE_OMP_FOR or GIMPLE_OMP_PARALLEL.
|
static |
Check OpenMP nesting restrictions.
FALLTHRU
FALLTHRU
|
static |
Build a decl for the omp child function. It'll not contain a body yet, just the bare decl.
Allocate memory for the function structure. The call to
allocate_struct_function clobbers CFUN, so we need to restore
it afterward.
References GF_OMP_FOR_KIND_FOR, gimple_omp_for_combined_into_p(), gimple_omp_for_kind(), gsi_stmt(), and walk_stmt_info::info.
|
static |
|
static |
Create task copyfn.
Reset DECL_CONTEXT on function arguments.
Populate the function.
Remap src and dst argument types if needed.
First pass: initialize temporaries used in record_type and srecord_type
sizes and field offsets. Second pass: copy shared var pointers and copy construct non-VLA
firstprivate vars. Last pass: handle VLA firstprivates.
| DEBUG_FUNCTION void debug_all_omp_regions | ( | void | ) |
References all_contexts.
| void debug_omp_region | ( | struct omp_region * | ) |
| DEBUG_FUNCTION void debug_omp_region | ( | ) |
|
static |
Destroy a omp_context data structures. Called through the splay tree value delete callback.
We hijacked DECL_ABSTRACT_ORIGIN earlier. We need to clear it before
it produces corrupt debug information.
|
static |
Discover whether REGION is a combined parallel+workshare region.
We only support parallel+for and parallel+sections.
Check for perfect nesting PAR_ENTRY_BB -> WS_ENTRY_BB and
WS_EXIT_BB -> PAR_EXIT_BB. If this is a combined parallel loop, we need to determine
whether or not to use the combined library calls. There
are two cases where we do not apply the transformation:
static loops and any kind of ordered loop. In the first
case, we already open code the loop so there is no need
to do anything else. In the latter case, the combined
parallel loop call would still need extra synchronization
to implement ordered semantics, so there would not be any
gain in using the combined call.
|
static |
|
static |
Check for mismatched contexts and generate an error if needed. Return true if an error is detected.
Try to avoid confusing the user by producing and error message
with correct "exit" or "enter" verbiage. We prefer "exit"
unless we can show that LABEL_CTX is nested within BRANCH_CTX. If it's obvious we have an invalid entry, be specific about the error.
Otherwise, be vague and lazy, but efficient.
|
static |
Pass 1: Create a minimal tree of OpenMP structured blocks, and record where each label is found.
The minimal context here is just the current OMP construct.
gimple_omp_for_{index,initial,final} are all DECLs; no need to
walk them.
|
static |
Pass 2: Check each branch and see if its context differs from that of the destination label's context.
gimple_omp_for_{index,initial,final} are all DECLs; no need to
walk them.
| void dump_omp_region | ( | FILE * | , |
| struct omp_region * | , | ||
| int | |||
| ) |
Return the parallel region associated with STMT.
Debugging dumps for parallel regions.
| void dump_omp_region | ( | ) |
Dump the parallel region tree rooted at REGION.
References free(), free_omp_region_1(), omp_region::inner, and omp_region::next.
|
static |
Main entry point for expanding OMP-GIMPLE into runtime calls.
|
static |
Main entry point.
This pass always runs, to provide PROP_gimple_lomp.
But there is nothing to do unless -fopenmp is given.
|
static |
Expand the parallel region tree rooted at REGION. Expansion proceeds in depth-first order. Innermost regions are expanded first. This way, parallel regions that require a new function to be created (e.g., GIMPLE_OMP_PARALLEL) can be expanded without having any internal dependencies in their body.
First, determine whether this is a combined parallel+workshare
region. Individual omp sections are handled together with their
parent GIMPLE_OMP_SECTIONS region.
|
static |
Expand an GIMPLE_OMP_ATOMIC statement. We try to expand using expand_omp_atomic_fetch_op. If it failed, we try to call expand_omp_atomic_pipeline, and if it fails too, the ultimate fallback is wrapping the operation in a mutex (expand_omp_atomic_mutex). REGION is the atomic region built by build_omp_regions_1().
Make sure the type is one of the supported sizes.
__sync builtins require strict data alignment.
Atomic load.
Atomic store.
When possible, use specialized atomic update functions.
If we don't have specialized __sync builtins, try and implement
as a compare and swap loop. The ultimate fallback is wrapping the operation in a mutex.
|
static |
A subroutine of expand_omp_atomic. Attempt to implement the atomic operation as a __atomic_fetch_op builtin. INDEX is log2 of the size of the data type, and thus usable to find the index of the builtin decl. Returns false if the expression is not of the proper form.
We expect to find the following sequences:
load_bb:
GIMPLE_OMP_ATOMIC_LOAD (tmp, mem)
store_bb:
val = tmp OP something; (or: something OP tmp)
GIMPLE_OMP_STORE (val)
???FIXME: Allow a more flexible sequence.
Perhaps use data flow to pick the statements.Check for one of the supported fetch-op operations.
Make sure the expression is of the proper form.
We could test all of the various optabs involved, but the fact of the
matter is that (with the exception of i486 vs i586 and xadd) all targets
that support any atomic operaton optab also implements compare-and-swap.
Let optabs.c take care of expanding any compare-and-swap loop. OpenMP does not imply any barrier-like semantics on its atomic ops.
It only requires that the operation happen atomically. Thus we can
use the RELAXED memory model.
|
static |
A subroutine of expand_omp_atomic. Attempt to implement the atomic operation as a normal volatile load.
??? If the target does not implement atomic_load_optab[mode], and mode
is smaller than word size, then expand_atomic_load assumes that the load
is atomic. We could avoid the builtin entirely in this case.
References build_pointer_type_for_mode(), builtin_decl_explicit(), can_compare_and_swap_p(), cfun, create_tmp_reg(), create_tmp_var(), force_gimple_operand_gsi(), gimple_in_ssa_p(), gsi_insert_before(), gsi_last_bb(), GSI_SAME_STMT, gsi_stmt(), ptr_mode, si, single_succ(), and type().
|
static |
@verbatim
A subroutine of expand_omp_atomic. Implement the atomic operation as:
GOMP_atomic_start ();
*addr = rhs;
GOMP_atomic_end ();
The result is not globally atomic, but works so long as all parallel references are within #pragma omp atomic directives. According to responses received from omp@openmp.org, appears to be within spec. Which makes sense, since that's how several other compilers handle this situation as well. LOADED_VAL and ADDR are the operands of GIMPLE_OMP_ATOMIC_LOAD we're expanding. STORED_VAL is the operand of the matching GIMPLE_OMP_ATOMIC_STORE.
We replace GIMPLE_OMP_ATOMIC_LOAD (loaded_val, addr) with loaded_val = *addr;
and replace GIMPLE_OMP_ATOMIC_STORE (stored_val) with *addr = stored_val;
|
static |
A subroutine of expand_omp_atomic. Implement the atomic operation as:
oldval = *addr;
repeat:
newval = rhs; // with oldval replacing *addr in rhs
oldval = __sync_val_compare_and_swap (addr, oldval, newval);
if (oldval != newval)
goto repeat;
INDEX is log2 of the size of the data type, and thus usable to find the
index of the builtin decl. ??? We need a non-pointer interface to __atomic_compare_exchange in
order to use the RELAXED memory model effectively. Load the initial value, replacing the GIMPLE_OMP_ATOMIC_LOAD.
For floating-point values, we'll need to view-convert them to integers
so that we can perform the atomic compare and swap. Simplify the
following code by always setting up the "i"ntegral variables. Move the value to the LOADEDI temporary.
Build the compare&swap statement.
Note that we always perform the comparison as an integer, even for
floating point. This allows the atomic operation to properly
succeed even with NaNs and -0.0. Update cfg.
Copy the new value to loadedi (we already did that before the condition
if we are not in SSA). Remove GIMPLE_OMP_ATOMIC_STORE.
|
static |
A subroutine of expand_omp_atomic. Attempt to implement the atomic operation as a normal volatile store.
If the load value is needed, then this isn't a store but an exchange.
??? If the target does not implement atomic_store_optab[mode], and mode
is smaller than word size, then expand_atomic_store assumes that the store
is atomic. We could avoid the builtin entirely in this case. Dig out the type of the function's second argument.
Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above.
|
static |
Prepend TO = FROM assignment before *GSI_P.
|
static |
Expand the OpenMP loop defined by REGION.
If there isn't a continue then this is a degerate case where
the introduction of abnormal edges during lowering will prevent
original loops from being detected. Fix that up.
References builtin_decl_explicit(), edge_def::flags, gimple_build_call(), gimple_call_set_lhs(), gsi_insert_after(), gsi_last_bb(), gsi_remove(), GSI_SAME_STMT, gsi_stmt(), and single_succ_edge().
|
static |
A subroutine of expand_omp_for. Generate code for a parallel
loop with any schedule. Given parameters:
for (V = N1; V cond N2; V += STEP) BODY;
where COND is "<" or ">", we generate pseudocode
more = GOMP_loop_foo_start (N1, N2, STEP, CHUNK, &istart0, &iend0);
if (more) goto L0; else goto L3;
L0:
V = istart0;
iend = iend0;
L1:
BODY;
V += STEP;
if (V cond iend) goto L1; else goto L2;
L2:
if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3;
L3:
If this is a combined omp parallel loop, instead of the call to
GOMP_loop_foo_start, we call GOMP_loop_foo_next.
If this is gimple_omp_for_combined_p loop, then instead of assigning
V and iend in L0 we assign the first two _looptemp_ clause decls of the
inner GIMPLE_OMP_FOR and V += STEP; and
if (V cond iend) goto L1; else goto L2; are removed.
For collapsed loops, given parameters:
collapse(3)
for (V1 = N11; V1 cond1 N12; V1 += STEP1)
for (V2 = N21; V2 cond2 N22; V2 += STEP2)
for (V3 = N31; V3 cond3 N32; V3 += STEP3)
BODY;
we generate pseudocode
if (__builtin_expect (N32 cond3 N31, 0)) goto Z0;
if (cond3 is <)
adj = STEP3 - 1;
else
adj = STEP3 + 1;
count3 = (adj + N32 - N31) / STEP3;
if (__builtin_expect (N22 cond2 N21, 0)) goto Z0;
if (cond2 is <)
adj = STEP2 - 1;
else
adj = STEP2 + 1;
count2 = (adj + N22 - N21) / STEP2;
if (__builtin_expect (N12 cond1 N11, 0)) goto Z0;
if (cond1 is <)
adj = STEP1 - 1;
else
adj = STEP1 + 1;
count1 = (adj + N12 - N11) / STEP1;
count = count1 * count2 * count3;
goto Z1;
Z0:
count = 0;
Z1:
more = GOMP_loop_foo_start (0, count, 1, CHUNK, &istart0, &iend0);
if (more) goto L0; else goto L3;
L0:
V = istart0;
T = V;
V3 = N31 + (T % count3) * STEP3;
T = T / count3;
V2 = N21 + (T % count2) * STEP2;
T = T / count2;
V1 = N11 + T * STEP1;
iend = iend0;
L1:
BODY;
V += 1;
if (V < iend) goto L10; else goto L2;
L10:
V3 += STEP3;
if (V3 cond3 N32) goto L1; else goto L11;
L11:
V3 = N31;
V2 += STEP2;
if (V2 cond2 N22) goto L1; else goto L12;
L12:
V2 = N21;
V1 += STEP1;
goto L1;
L2:
if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3;
L3: See if we need to bias by LLONG_MIN.
Some counts[i] vars might be uninitialized if
some loop has zero iterations. But the body shouldn't
be executed in that case, so just avoid uninit warnings. In a combined parallel loop, emit a call to
GOMP_loop_foo_next. If this is not a combined parallel loop, emit a call to
GOMP_loop_foo_start in ENTRY_BB. Avoid casting pointers to integer of a different size.
The GOMP_loop_ull_*start functions have additional boolean
argument, true for < loops and false for > loops.
In Fortran, the C bool type can be different from
boolean_type_node. Remove the GIMPLE_OMP_FOR statement.
Iteration setup for sequential loop goes in L0_BB.
Code to control the increment and predicate for the sequential
loop goes in the CONT_BB. Remove GIMPLE_OMP_CONTINUE.
Emit code to get the next parallel iteration in L2_BB.
Add the loop cleanup function.
Connect the new blocks.
The loop may have multiple latches.
|
static |
Helper function for expand_omp_{for_*,simd}. If this is the outermost
of the combined collapse > 1 loop constructs, generate code like:
if (__builtin_expect (N32 cond3 N31, 0)) goto ZERO_ITER_BB;
if (cond3 is <)
adj = STEP3 - 1;
else
adj = STEP3 + 1;
count3 = (adj + N32 - N31) / STEP3;
if (__builtin_expect (N22 cond2 N21, 0)) goto ZERO_ITER_BB;
if (cond2 is <)
adj = STEP2 - 1;
else
adj = STEP2 + 1;
count2 = (adj + N22 - N21) / STEP2;
if (__builtin_expect (N12 cond1 N11, 0)) goto ZERO_ITER_BB;
if (cond1 is <)
adj = STEP1 - 1;
else
adj = STEP1 + 1;
count1 = (adj + N12 - N11) / STEP1;
count = count1 * count2 * count3;
Furthermore, if ZERO_ITER_BB is NULL, create a BB which does:
count = 0;
and set ZERO_ITER_BB to that bb. If this isn't the outermost
of the combined loop constructs, just initialize COUNTS array
from the _looptemp_ clauses. NOTE: It *could* be better to moosh all of the BBs together, creating one larger BB with all the computation and the unexpected jump at the end. I.e. bool zero3, zero2, zero1, zero; zero3 = N32 c3 N31; count3 = (N32 - N31) /[cl] STEP3; zero2 = N22 c2 N21; count2 = (N22 - N21) /[cl] STEP2; zero1 = N12 c1 N11; count1 = (N12 - N11) /[cl] STEP1; zero = zero3 || zero2 || zero1; count = count1 * count2 * count3; if (__builtin_expect(zero, false)) goto zero_iter_bb; After all, we expect the zero=false, and thus we expect to have to evaluate all of the comparison expressions, so short-circuiting oughtn't be a win. Since the condition isn't protecting a denominator, we're not concerned about divide-by-zero, so we can fully evaluate count even if a numerator turned out to be wrong. It seems like putting this all together would create much better scheduling opportunities, and less pressure on the chip's branch predictor.
Collapsed loops need work for expansion into SSA form.
First two _looptemp_ clauses are for istart/iend, counts[0]
isn't supposed to be handled, as the inner loop doesn't
use it. ?? We could probably use CEIL_DIV_EXPR instead of
TRUNC_DIV_EXPR and adjusting by hand. Unless we can't
generate the same code in the end because generically we
don't know that the values involved must be negative for
GT??
|
static |
Helper function for expand_omp_{for_*,simd}. Generate code like:
T = V;
V3 = N31 + (T % count3) * STEP3;
T = T / count3;
V2 = N21 + (T % count2) * STEP2;
T = T / count2;
V1 = N11 + T * STEP1;
if this loop doesn't have an inner loop construct combined with it.
If it does have an inner loop construct combined with it and the
iteration count isn't known constant, store values from counts array
into its _looptemp_ temporaries instead. If fd->loop.n2 is constant, then no propagation of the counts
is needed, they are constant. First two _looptemp_ clauses are for istart/iend, counts[0]
isn't supposed to be handled, as the inner loop doesn't
use it.
References omp_for_data_loop::cond_code, omp_region::cont, create_tmp_var(), is_combined_parallel(), omp_for_data::iter_type, omp_for_data::loop, omp_for_data_loop::n1, omp_for_data_loop::n2, omp_for_data_loop::step, type(), and omp_for_data_loop::v.
|
static |
@verbatim
A subroutine of expand_omp_for. Generate code for a parallel loop with static schedule and a specified chunk size. Given parameters:
for (V = N1; V cond N2; V += STEP) BODY;
where COND is "<" or ">", we generate pseudocode
if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2;
if (cond is <)
adj = STEP - 1;
else
adj = STEP + 1;
if ((__typeof (V)) -1 > 0 && cond is >)
n = -(adj + N2 - N1) / -STEP;
else
n = (adj + N2 - N1) / STEP;
trip = 0;
V = threadid * CHUNK * STEP + N1; -- this extra definition of V is
here so that V is defined
if the loop is not entered
L0: s0 = (trip * nthreads + threadid) * CHUNK; e0 = min(s0 + CHUNK, n); if (s0 < n) goto L1; else goto L4; L1: V = s0 * STEP + N1; e = e0 * STEP + N1; L2: BODY; V += STEP; if (V cond e) goto L2; else goto L3; L3: trip += 1; goto L0; L4:
Trip and adjustment setup goes in ENTRY_BB.
Remove the GIMPLE_OMP_FOR.
Iteration space partitioning goes in ITER_PART_BB.
Setup code for sequential iteration goes in SEQ_START_BB.
The code controlling the sequential loop goes in CONT_BB,
replacing the GIMPLE_OMP_CONTINUE. Remove GIMPLE_OMP_CONTINUE.
Trip update code goes into TRIP_UPDATE_BB.
Replace the GIMPLE_OMP_RETURN with a barrier, or nothing.
Connect the new blocks.
When we redirect the edge from trip_update_bb to iter_part_bb, we
remove arguments of the phi nodes in fin_bb. We need to create
appropriate phi nodes in iter_part_bb instead. A special case -- fd->loop.v is not yet computed in
iter_part_bb, we need to use v_extra instead. Make phi node for trip.
|
static |
@verbatim
A subroutine of expand_omp_for. Generate code for a parallel loop with static schedule and no specified chunk size. Given parameters:
for (V = N1; V cond N2; V += STEP) BODY;
where COND is "<" or ">", we generate pseudocode
if ((__typeof (V)) -1 > 0 && N2 cond N1) goto L2; if (cond is <) adj = STEP - 1; else adj = STEP + 1; if ((__typeof (V)) -1 > 0 && cond is >) n = -(adj + N2 - N1) / -STEP; else n = (adj + N2 - N1) / STEP; q = n / nthreads; tt = n % nthreads; if (threadid < tt) goto L3; else goto L4;
L3: tt = 0; q = q + 1; L4: s0 = q * threadid + tt; e0 = s0 + q; V = s0 * STEP + N1; if (s0 >= e0) goto L2; else goto L0; L0: e = e0 * STEP + N1; L1: BODY; V += STEP; if (V cond e) goto L1; L2:
Iteration space partitioning goes in ENTRY_BB.
Remove the GIMPLE_OMP_FOR statement.
Setup code for sequential iteration goes in SEQ_START_BB.
The code controlling the sequential loop replaces the
GIMPLE_OMP_CONTINUE. Remove the GIMPLE_OMP_CONTINUE statement.
Replace the GIMPLE_OMP_RETURN with a barrier, or nothing.
Connect all the blocks.
|
static |
Callback for expand_omp_build_assign. Return non-NULL if *tp needs to be regimplified.
Any variable with DECL_VALUE_EXPR needs to be regimplified.
|
static |
Expand code for an OpenMP sections directive. In pseudo code, we generate
v = GOMP_sections_start (n);
L0:
switch (v)
{
case 0:
goto L2;
case 1:
section 1;
goto L1;
case 2:
...
case n:
...
default:
abort ();
}
L1:
v = GOMP_sections_next ();
goto L0;
L2:
reduction;
If this is a combined parallel sections, replace the call to
GOMP_sections_start with call to GOMP_sections_next. This can happen if there are reductions.
We will build a switch() with enough cases for all the
GIMPLE_OMP_SECTION regions, a '0' case to handle the end of more work
and a default case to abort if something goes wrong. Use vec::quick_push on label_vec throughout, since we know the size
in advance. The call to GOMP_sections_start goes in ENTRY_BB, replacing the
GIMPLE_OMP_SECTIONS statement. If we are not inside a combined parallel+sections region,
call GOMP_sections_start. Otherwise, call GOMP_sections_next.
The switch() statement replacing GIMPLE_OMP_SECTIONS_SWITCH goes in
L0_BB. Convert each GIMPLE_OMP_SECTION into a CASE_LABEL_EXPR.
Skip optional reduction region.
Error handling code goes in DEFAULT_BB.
Code to get the next section goes in L1_BB.
Cleanup function replaces GIMPLE_OMP_RETURN in EXIT_BB.
References build_call_expr_loc(), build_int_cst(), builtin_decl_explicit(), cfun, force_gimple_operand_gsi(), gimple_in_ssa_p(), gimple_location(), gimple_omp_atomic_seq_cst_p(), gsi_last_bb(), gsi_remove(), GSI_SAME_STMT, gsi_stmt(), MEMMODEL_RELAXED, MEMMODEL_SEQ_CST, single_succ(), type(), update_ssa(), and useless_type_conversion_p().
|
static |
A subroutine of expand_omp_for. Generate code for a simd non-worksharing
loop. Given parameters:
for (V = N1; V cond N2; V += STEP) BODY;
where COND is "<" or ">", we generate pseudocode
V = N1;
goto L1;
L0:
BODY;
V += STEP;
L1:
if (V cond N2) goto L0; else goto L2;
L2:
For collapsed loops, given parameters:
collapse(3)
for (V1 = N11; V1 cond1 N12; V1 += STEP1)
for (V2 = N21; V2 cond2 N22; V2 += STEP2)
for (V3 = N31; V3 cond3 N32; V3 += STEP3)
BODY;
we generate pseudocode
if (cond3 is <)
adj = STEP3 - 1;
else
adj = STEP3 + 1;
count3 = (adj + N32 - N31) / STEP3;
if (cond2 is <)
adj = STEP2 - 1;
else
adj = STEP2 + 1;
count2 = (adj + N22 - N21) / STEP2;
if (cond1 is <)
adj = STEP1 - 1;
else
adj = STEP1 + 1;
count1 = (adj + N12 - N11) / STEP1;
count = count1 * count2 * count3;
V = 0;
V1 = N11;
V2 = N21;
V3 = N31;
goto L1;
L0:
BODY;
V += 1;
V3 += STEP3;
V2 += (V3 cond3 N32) ? 0 : STEP2;
V3 = (V3 cond3 N32) ? V3 : N31;
V1 += (V2 cond2 N22) ? 0 : STEP1;
V2 = (V2 cond2 N22) ? V2 : N21;
L1:
if (V < count) goto L0; else goto L2;
L2: Not needed in SSA form right now.
Remove the GIMPLE_OMP_FOR statement.
Code to control the increment goes in the CONT_BB.
Remove GIMPLE_OMP_CONTINUE.
Emit the condition in L1_BB.
Remove GIMPLE_OMP_RETURN.
Connect the new blocks.
If not -fno-tree-loop-vectorize, hint that we want to vectorize
the loop.
|
static |
Expand code for an OpenMP single directive. We've already expanded much of the code, here we simply place the GOMP_barrier call.
|
static |
Generic expansion for OpenMP synchronization directives: master, ordered and critical. All we need to do here is remove the entry and exit markers for REGION.
|
static |
Expand the OpenMP target{, data, update} directive starting at REGION. If the target region needs data sent from the parent
function, then the very first statement (except possible
tree profile counter updates) of the parallel body
is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since
&.OMP_DATA_O is passed as an argument to the child function,
we need to replace it with the argument as seen by the child
function.
In most cases, this will end up being the identity assignment
.OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had
a function call that has been inlined, the original PARM_DECL
.OMP_DATA_I may have been converted into a different local
variable. In which case, we need to keep the assignment. We're ignoring the subcode because we're
effectively doing a STRIP_NOPS. Declare local variables needed in CHILD_CFUN.
The gimplifier could record temporaries in target block
rather than in containing function's local_decls chain,
which would mean cgraph missed finalizing them. Do it now. We'll create a CFG for child_fn, so no gimple body is needed.
Reset DECL_CONTEXT on function arguments.
Split ENTRY_BB at GIMPLE_OMP_TARGET,
so that it can be moved to the child function. Convert GIMPLE_OMP_RETURN into a RETURN_EXPR.
Move the target region into CHILD_CFUN.
When the OMP expansion process cannot guarantee an up-to-date
loop tree arrange for the child function to fixup loops. Remove non-local VAR_DECLs from child_cfun->local_decls list.
Inform the callgraph about the new function.
Fix the callgraph edges for child_cfun. Those for cfun will be
fixed in a following pass. Some EH regions might become dead, see PR34608. If
pass_cleanup_cfg isn't the first pass to happen with the
new child, these dead EH edges might cause problems.
Clean them up now. Emit a library call to launch the target region, or do data
transfers. By default, the value of DEVICE is -1 (let runtime library choose)
and there is no conditional. Ensure 'device' is of the correct type.
If we found the clause 'if (cond)', build
(cond ? device : -2). FIXME: This will be address of
extern char __OPENMP_TARGET__[] __attribute__((visibility ("hidden")))
symbol, as soon as the linker plugin is able to create it for us.
References omp_region::exit, gsi_last_bb(), gsi_remove(), and gsi_stmt().
|
static |
Expand the OpenMP parallel or task directive starting at REGION.
Due to inlining, it may happen that we have already outlined
the region, in which case all we need to do is make the
sub-graph unreachable and emit the parallel call. If the parallel region needs data sent from the parent
function, then the very first statement (except possible
tree profile counter updates) of the parallel body
is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since
&.OMP_DATA_O is passed as an argument to the child function,
we need to replace it with the argument as seen by the child
function.
In most cases, this will end up being the identity assignment
.OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had
a function call that has been inlined, the original PARM_DECL
.OMP_DATA_I may have been converted into a different local
variable. In which case, we need to keep the assignment. We're ignore the subcode because we're
effectively doing a STRIP_NOPS. ?? Is setting the subcode really necessary ??
If we are in ssa form, we must load the value from the default
definition of the argument. That should not be defined now,
since the argument is not used uninitialized. ?? Is setting the subcode really necessary ??
Declare local variables needed in CHILD_CFUN.
The gimplifier could record temporaries in parallel/task block
rather than in containing function's local_decls chain,
which would mean cgraph missed finalizing them. Do it now. We'll create a CFG for child_fn, so no gimple body is needed.
Reset DECL_CONTEXT on function arguments.
Split ENTRY_BB at GIMPLE_OMP_PARALLEL or GIMPLE_OMP_TASK,
so that it can be moved to the child function. Convert GIMPLE_OMP_RETURN into a RETURN_EXPR.
Move the parallel region into CHILD_CFUN.
When the OMP expansion process cannot guarantee an up-to-date
loop tree arrange for the child function to fixup loops. Remove non-local VAR_DECLs from child_cfun->local_decls list.
Inform the callgraph about the new function.
Fix the callgraph edges for child_cfun. Those for cfun will be
fixed in a following pass. Some EH regions might become dead, see PR34608. If
pass_cleanup_cfg isn't the first pass to happen with the
new child, these dead EH edges might cause problems.
Clean them up now. Emit a library call to launch the children threads.
|
static |
Build the function calls to GOMP_parallel_start etc to actually generate the parallel operation. REGION is the parallel region being expanded. BB is the block where to insert the code. WS_ARGS will be set if this is a call to a combined parallel+workshare construct, it contains the list of additional arguments needed by the workshare construct.
Determine what flavor of GOMP_parallel we will be
emitting. By default, the value of NUM_THREADS is zero (selected at run time)
and there is no conditional. Ensure 'val' is of the correct type.
If we found the clause 'if (cond)', build either
(cond != 0) or (cond ? val : 1u).
|
static |
Build the function call to GOMP_task to actually generate the task operation. BB is the block where to insert the code.
References gimple_block().
|
static |
Extract the header elements of parallel loop FOR_STMT and store them into *FD.
FIXME: for now map schedule(auto) to schedule(static).
There should be analysis to determine whether all iterations
are approximately the same amount of work (then schedule(static)
is best) or if it varies (then schedule(dynamic,N) is better). We only need to compute a default chunk size for ordered
static loops and dynamic loops.
Referenced by workshare_safe_to_combine_p().
|
static |
Helper function for expand_omp_for_*. Generate code like:
L10:
V3 += STEP3;
if (V3 cond3 N32) goto BODY_BB; else goto L11;
L11:
V3 = N31;
V2 += STEP2;
if (V2 cond2 N22) goto BODY_BB; else goto L12;
L12:
V2 = N21;
V1 += STEP1;
goto BODY_BB;
References omp_for_data::loop, omp_for_data_loop::n1, omp_for_data_loop::n2, and omp_for_data_loop::step.
|
static |
Finalize task copyfn.
Inform the callgraph about the new function.
References omp_context::cb, copy_node(), copy_tree_body_r(), omp_context::field_map, layout_type(), lang_hooks_for_types::make_type, nreverse(), omp_context::receiver_decl, omp_context::record_type, remap_type(), type(), gdbhooks::TYPE_DECL, and lang_hooks::types.
|
static |
Callback for walk_gimple_seq. Check if combined parallel contains gimple_omp_for_combined_into_p OMP_FOR.
References fixup_child_record_type(), layout_type(), and omp_context::record_type.
| tree find_omp_clause | ( | ) |
Find an OpenMP clause of type KIND within CLAUSES.
References omp_context::stmt.
Referenced by lower_send_clauses(), scan_omp_teams(), and workshare_safe_to_combine_p().
|
static |
Fix up RECEIVER_DECL with a type that has been remapped to the child context.
??? It isn't sufficient to just call remap_type here, because
variably_modified_type_p doesn't work the way we expect for
record types. Testing each field for whether it needs remapping
and creating a new record by hand works, however. Arrange to be able to look up the receiver field
given the sender field.
Referenced by find_combined_for().
|
static |
Adjust the replacement for DECL in CTX for the new context. This means copying the DECL_VALUE_EXPR, and fixing up the type.
|
static |
Release the memory associated with the region tree rooted at REGION.
Referenced by dump_omp_region().
| void free_omp_regions | ( | void | ) |
Release the memory for the entire omp region tree.
|
static |
|
static |
OMP expansion -- the default pass, run before creation of SSA form.
Collect additional arguments needed to emit a combined parallel+workshare call. WS_STMT is the workshare directive being expanded.
Number of sections is equal to the number of edges from the
GIMPLE_OMP_SECTIONS_SWITCH statement, except for the one to
the exit of the sections region.
References omp_for_data::chunk_size, and fold_convert_loc().
|
static |
A convenience function to build an empty GIMPLE_COND with just the condition.
|
static |
Add a new field for VAR inside the structure CTX->SENDER_DECL.
Remember what variable this field was created for. This does have a
side effect of making dwarf2out ignore this member, so for helpful
debugging we clear it later in delete_omp_context.
|
static |
|
inlinestatic |
Return true if REGION is a combined parallel+workshare region.
References OMP_CLAUSE_NOWAIT.
Referenced by expand_omp_for_init_vars(), and lower_send_clauses().
|
inlinestatic |
Return true if CTX is for an omp parallel.
References omp_region::is_combined_parallel.
|
inlinestatic |
Return true if DECL is a reference type.
References omp_context::field_map.
Referenced by lower_reduction_clauses().
|
inlinestatic |
Return true if CTX is for an omp task.
References count, GF_OMP_FOR_KIND_SIMD, gimple_location(), and gimple_omp_for_kind().
Referenced by lower_reduction_clauses().
|
inlinestatic |
Return true if CTX is for an omp parallel or omp task.
Referenced by maybe_lookup_field().
|
inlinestatic |
Return true if EXPR is variable sized.
Referenced by lower_reduction_clauses(), and omp_copy_decl_2().
|
inlinestatic |
Lookup variables in the decl or field splay trees. The "maybe" form allows for the variable form to not have been entered, otherwise we assert that the variable must have been entered.
|
static |
Referenced by lower_reduction_clauses().
|
static |
Find the mapping for DECL in CTX or the immediately enclosing
context that has a mapping for DECL.
If CTX is a nested parallel directive, we may have to use the decl
mappings created in CTX's parent context. Suppose that we have the
following parallel nesting (variable UIDs showed for clarity):
iD.1562 = 0;
#omp parallel shared(iD.1562) -> outer parallel
iD.1562 = iD.1562 + 1;
#omp parallel shared (iD.1562) -> inner parallel
iD.1562 = iD.1562 - 1;
Each parallel structure will create a distinct .omp_data_s structure
for copying iD.1562 in/out of the directive:
outer parallel .omp_data_s.1.i -> iD.1562
inner parallel .omp_data_s.2.i -> iD.1562
A shared variable mapping will produce a copy-out operation before
the parallel directive and a copy-in operation after it. So, in
this case we would have:
iD.1562 = 0;
.omp_data_o.1.i = iD.1562;
#omp parallel shared(iD.1562) -> outer parallel
.omp_data_i.1 = &.omp_data_o.1
.omp_data_i.1->i = .omp_data_i.1->i + 1;
.omp_data_o.2.i = iD.1562; -> **
#omp parallel shared(iD.1562) -> inner parallel
.omp_data_i.2 = &.omp_data_o.2
.omp_data_i.2->i = .omp_data_i.2->i - 1;
** This is a problem. The symbol iD.1562 cannot be referenced
inside the body of the outer parallel region. But since we are
emitting this copy operation while expanding the inner parallel
directive, we need to access the CTX structure of the outer
parallel directive to get the correct mapping:
.omp_data_o.2.i = .omp_data_i.1->i
Since there may be other workshare or parallel directives enclosing
the parallel directive, it may be necessary to walk up the context
parent chain. This is not a problem in general because nested
parallelism happens only rarely.
|
inlinestatic |
|
inlinestatic |
|
static |
Generate code to implement the COPYPRIVATE clauses.
|
static |
|
static |
Generate code to implement the LASTPRIVATE clauses. This is used for both parallel and workshare constructs. PREDICATE may be NULL if it's always true.
Early exit if there are no lastprivate or linear clauses.
If this was a workshare clause, see if it had been combined
with its parallel. In that case, look for the clauses on the
parallel statement itself. If this was a workshare clause, see if it had been combined
with its parallel. In that case, continue looking for the
clauses also on the parallel statement itself.
|
static |
Referenced by lower_omp_sections(), and maybe_add_implicit_barrier_cancel().
|
static |
Inside target region we haven't called fold_stmt during gimplification,
because it can break code by adding decl references that weren't in the
source. Call fold_stmt now.
|
static |
If we have issued syntax errors, avoid doing any heavy lifting.
Just replace the OpenMP directives with a NOP to avoid
confusing RTL expansion. FALLTHRU
FALLTHRU
|
static |
|
static |
Lower code for an OpenMP loop directive.
Replace at gsi right away, so that 'stmt' is no member
of a sequence anymore as we're going to add to to a different
one below. Move declaration of temporaries in the loop body before we make
it go away. We need two temporaries with fd.loop.v type (istart/iend)
and then (fd.collapse - 1) temporaries with the same
type for count2 ... countN-1 vars if not constant. The pre-body and input clauses go before the lowered GIMPLE_OMP_FOR.
Lower the header expressions. At this point, we can assume that
the header is of the form:
#pragma omp for (V = VAL1; V {<|>|<=|>=} VAL2; V = V [+-] VAL3)
We just need to make sure that VAL1, VAL2 and VAL3 are lowered
using the .omp_data_s mapping, if needed. Once lowered, extract the bounds and clauses.
After the loop, add exit clauses.
Region exit marker goes at the end of the loop body.
|
static |
A subroutine of lower_omp_for. Generate code to emit the predicate for a lastprivate clause. Given a loop control predicate of (V cond N2), we gate the clause on (!(V cond N2)). The lowered form is appended to *DLIST, iterator initialization is appended to *BODY_P.
When possible, use a strict equality expression. This can let VRP
type optimizations deduce the value and remove a copy. Optimize: v = 0; is usually cheaper than v = some_other_constant.
Initialize the iterator variable, so that threads that don't execute
any iterations don't execute the lastprivate clauses by accident.
|
static |
Expand code for an OpenMP master directive.
|
static |
Expand code for an OpenMP ordered directive.
Callback for lower_omp_1. Return non-NULL if *tp needs to be regimplified. If DATA is non-NULL, lower_omp_1 is outside of OpenMP context, but with task_shared_vars set.
Any variable with DECL_VALUE_EXPR needs to be regimplified.
If a global variable has been privatized, TREE_CONSTANT on
ADDR_EXPR might be wrong.
|
static |
Lower the OpenMP sections directive in the current statement in GSI_P. CTX is the enclosing OMP context for the current statement.
References omp_context::block_vars, builtin_decl_explicit(), gimple_bind_add_seq(), gimple_bind_add_stmt(), gimple_bind_append_vars(), gimple_bind_vars(), gimple_build_bind(), gimple_build_call(), gimple_build_omp_return(), gimple_omp_body(), gimple_omp_body_ptr(), gimple_omp_set_body(), gsi_replace(), gsi_stmt(), lower_omp(), maybe_catch_exception(), pop_gimplify_context(), and push_gimplify_context().
|
static |
Expand code for an OpenMP single directive.
|
static |
A subroutine of lower_omp_single. Expand the simple form of
a GIMPLE_OMP_SINGLE, with a copyprivate clause:
#pragma omp single copyprivate (a, b, c)
Create a new structure to hold copies of 'a', 'b' and 'c' and emit:
{
if ((copyout_p = GOMP_single_copy_start ()) == NULL)
{
BODY;
copyout.a = a;
copyout.b = b;
copyout.c = c;
GOMP_single_copy_end (©out);
}
else
{
a = copyout_p->a;
b = copyout_p->b;
c = copyout_p->c;
}
GOMP_barrier ();
}
FIXME. It may be better to delay expanding the logic of this until
pass_expand_omp. The expanded logic may make the job more difficult
to a synchronization analysis pass.
|
static |
A subroutine of lower_omp_single. Expand the simple form of
a GIMPLE_OMP_SINGLE, without a copyprivate clause:
if (GOMP_single_start ())
BODY;
[ GOMP_barrier (); ] -> unless 'nowait' is present.
FIXME. It may be better to delay expanding the logic of this until
pass_expand_omp. The expanded logic may make the job more difficult
to a synchronization analysis pass.
|
static |
Lower the OpenMP target directive in the current statement in GSI_P. CTX holds context information for the directive.
Declare all the variables created by mapping and the variables
declared in the scope of the target body. Once all the expansions are done, sequence all the different
fragments inside gimple_omp_body. fixup_child_record_type might have changed receiver_decl's type.
References bitmap_bit_p(), and recompute_tree_invariant_for_addr_expr().
|
static |
Expand code for an OpenMP taskgroup directive.
|
static |
Lower the OpenMP parallel or task directive in the current statement in GSI_P. CTX holds context information for the directive.
Declare all the variables created by mapping and the variables
declared in the scope of the parallel body. Once all the expansions are done, sequence all the different
fragments inside gimple_omp_body. fixup_child_record_type might have changed receiver_decl's type.
|
static |
Expand code for an OpenMP teams directive.
|
static |
Generate code to implement the input clauses, FIRSTPRIVATE and COPYIN, from the receiver (aka child) side and initializers for REFERENCE_TYPE private variables. Initialization statements go in ILIST, while calls to destructors go in DLIST.
Set max_vf=1 (which will later enforce safelen=1) in simd loops
with data sharing clauses referencing variable sized vars. That
is unnecessarily hard to support and very unlikely to result in
vectorized code anyway. Do all the fixed sized types in the first pass, and the variable sized
types in the second pass. This makes sure that the scalar arguments to
the variable sized types are processed before we use them in the
variable sized operations. Ignore shared directives in teams construct.
Handle _looptemp_ clauses only on parallel.
For variable sized types, we need to allocate the
actual storage here. Call alloca and store the
result in the pointer decl that we created elsewhere. void *tmp = __builtin_alloca
For references that are being privatized for Fortran,
allocate new backing storage for the new pointer
variable. This allows us to avoid changing all the
code that expects a pointer to something that expects
a direct variable. Ignore shared directives in teams construct.
Shared global vars are just accessed directly.
Set up the DECL_VALUE_EXPR for shared variables now. This
needs to be delayed until after fixup_child_record_type so
that we get the correct type during the dereference. ??? If VAR is not passed by reference, and the variable
hasn't been initialized yet, then we'll get a warning for
the store into the omp_data_s structure. Ideally, we'd be
able to notice this and not store anything at all, but
we're generating code too early. Suppress the warning. FALLTHRU
FALLTHRU
reduction(-:var) sums up the partial results, so it
acts identically to reduction(+:var). Don't want uninit warnings on simduid, it is always uninitialized,
but we use it not for the value, but for the DECL_UID only. The copyin sequence is not to be executed by the main thread, since
that would result in self-copies. Perhaps not visible to scalars,
but it certainly is to C++ operator=. If any copyin variable is passed by reference, we must ensure the
master thread doesn't modify it before it is copied over in all
threads. Similarly for variables in both firstprivate and
lastprivate clauses we need to ensure the lastprivate copying
happens after firstprivate copying in all threads. And similarly
for UDRs if initializer expression refers to omp_orig. Don't add any barrier for #pragma omp simd or
#pragma omp distribute. If max_vf is non-zero, then we can use only a vectorization factor
up to the max_vf we chose. So stick it into the safelen clause.
|
static |
Helper function of lower_rec_input_clauses, used for #pragma omp simd privatization.
|
static |
Generate code to implement the REDUCTION clauses.
SIMD reductions are handled in lower_rec_input_clauses.
First see if there is exactly one reduction clause. Use OMP_ATOMIC
update in that case, otherwise use a lock. Never use OMP_ATOMIC for array reductions or UDRs.
reduction(-:var) sums up the partial results, so it acts
identically to reduction(+:var).
References build_fold_addr_expr_loc(), build_sender_ref(), lang_hooks::decls, gimplify_assign(), is_global_var(), is_reference(), is_task_ctx(), is_variable_sized(), lookup_decl_in_outer_ctx(), OMP_CLAUSE__LOOPTEMP_, OMP_CLAUSE_COPYIN, OMP_CLAUSE_FIRSTPRIVATE, OMP_CLAUSE_LASTPRIVATE, OMP_CLAUSE_PRIVATE, OMP_CLAUSE_REDUCTION, lang_hooks_for_decls::omp_private_outer_ref, and use_pointer_for_field().
|
static |
Generate code to implement the clauses, FIRSTPRIVATE, COPYIN, LASTPRIVATE, and REDUCTION from the sender (aka parent) side.
References build_int_cst(), cfun, create_tmp_var(), find_omp_clause(), fold_convert_loc(), gimple_boolify(), gimple_in_ssa_p(), gimple_location(), gimple_omp_parallel_clauses(), walk_stmt_info::gsi, omp_region::inner, integer_zerop(), is_combined_parallel(), make_ssa_name(), OMP_CLAUSE_IF, OMP_CLAUSE_NUM_THREADS, OMP_CLAUSE_PROC_BIND, OMP_CLAUSE_SCHEDULE_AUTO, OMP_CLAUSE_SCHEDULE_RUNTIME, omp_region::sched_kind, and omp_region::type.
|
static |
Generate code to implement SHARED from the sender (aka parent) side. This is trickier, since GIMPLE_OMP_PARALLEL_CLAUSES doesn't list things that got automatically shared.
If CTX is a nested parallel directive. Find the immediately
enclosing parallel or workshare construct that contains a
mapping for OVAR. We don't need to receive a new reference to a result
or parm decl. In fact we may not store to it as we will
invalidate any pending RSO and generate wrong gimple
during inlining.
| bool make_gimple_omp_edges | ( | ) |
Called from tree-cfg.c::make_edges to create cfg edges for all GIMPLE_OMP codes.
In the case of a GIMPLE_OMP_SECTION, the edge will go
somewhere other than the next block. This will be
created later. Mark all GIMPLE_OMP_FOR and GIMPLE_OMP_CONTINUE
succs edges as abnormal to prevent splitting
them. Make the loopback edge.
Create an edge from GIMPLE_OMP_FOR to exit, which
corresponds to the case that the body of the loop
is not executed at all. Wire up the edges into and out of the nested sections.
Make the loopback edge to the block with
GIMPLE_OMP_SECTIONS_SWITCH. Make the edge from the switch to exit.
| gimple_opt_pass* make_pass_diagnose_omp_blocks | ( | ) |
| gimple_opt_pass* make_pass_expand_omp | ( | ) |
| gimple_opt_pass* make_pass_lower_omp | ( | ) |
|
static |
Routines to lower OpenMP directives into OMP-GIMPLE.
If ctx is a worksharing context inside of a cancellable parallel region and it isn't nowait, add lhs to its GIMPLE_OMP_RETURN and conditional branch to parallel's cancel_label to handle cancellation in the implicit barrier.
References omp_context::block_vars, builtin_decl_explicit(), gimple_bind_add_seq(), gimple_bind_add_stmt(), gimple_bind_append_vars(), gimple_build_bind(), gimple_build_call(), gimple_build_omp_return(), gimple_omp_body(), gimple_omp_body_ptr(), gimple_omp_set_body(), gsi_replace(), gsi_stmt(), and lower_omp().
|
static |
Referenced by lower_omp_sections().
|
static |
If exceptions are enabled, wrap the statements in BODY in a MUST_NOT_THROW catch handler and return it. This prevents programs from violating the structured block semantics with throws.
References builtin_decl_explicit(), gimple_call_fndecl(), gsi_stmt(), and is_gimple_call().
|
static |
If a context was created for STMT when it was scanned, return it.
References real_arithmetic(), and real_inf().
|
inlinestatic |
Referenced by maybe_lookup_field(), and omp_max_vf().
|
static |
Referenced by omp_copy_decl_2().
|
static |
Similar to lookup_decl_in_outer_ctx, but return DECL if not found in outer contexts.
References targetm.
|
inlinestatic |
References is_taskreg_ctx(), maybe_lookup_decl(), and omp_context::outer.
|
static |
Create a new context, with OUTER_CTX being the surrounding context.
Referenced by scan_omp_for().
|
staticread |
Create a new parallel region starting at STMT inside region PARENT.
This is a nested region. Add it to the list of inner
regions in PARENT. This is a toplevel region. Add it to the list of toplevel
regions in ROOT_OMP_REGION.
|
static |
Build COMPONENT_REF and set TREE_THIS_VOLATILE and TREE_READONLY on it as appropriate.
|
static |
Return alignment to be assumed for var in CLAUSE, which should be OMP_CLAUSE_ALIGNED.
Otherwise return implementation defined alignment.
|
static |
The callback for remap_decl. Search all containing contexts for a mapping of the variable; this avoids having to duplicate the splay tree ahead of time. We know a mapping doesn't already exist in the given context. Create new mappings to implement default semantics.
|
static |
References build_receiver_ref(), and use_pointer_for_field().
|
static |
Construct a new automatic decl similar to VAR.
References build_outer_var_ref(), is_global_var(), is_variable_sized(), and maybe_lookup_decl_in_outer_ctx().
| void omp_expand_local | ( | ) |
Expands omp construct (and its subconstructs) starting in HEAD.
|
static |
Return maximum possible vectorization factor for the target.
References is_global_var(), maybe_lookup_decl(), OMP_CLAUSE__LOOPTEMP_, OMP_CLAUSE_COPYIN, OMP_CLAUSE_FIRSTPRIVATE, OMP_CLAUSE_LINEAR, OMP_CLAUSE_PRIVATE, OMP_CLAUSE_REDUCTION, OMP_CLAUSE_SHARED, and omp_context::stmt.
| tree omp_reduction_init | ( | ) |
Construct the initialization value for reduction CLAUSE.
|
static |
Optimize omp_get_thread_num () and omp_get_num_threads ()
calls. These can't be declared as const functions, but
within one parallel body they are constant, so they can be
transformed there into __builtin_omp_get_{thread_num,num_threads} ()
which are declared const. Similarly for task body, except
that in untied task omp_get_thread_num () can change at any task
scheduling point. In #pragma omp task untied omp_get_thread_num () can change
during the execution of the task region.
References gimple_assign_rhs1(), and gimple_omp_taskreg_data_arg().
|
static |
Remove barriers in REGION->EXIT's block. Note that this is only valid for GIMPLE_OMP_PARALLEL regions. Since the end of a parallel region is an implicit barrier, any workshare inside the GIMPLE_OMP_PARALLEL that left a barrier at the end of the GIMPLE_OMP_PARALLEL region can now be removed.
If the parallel region doesn't return, we don't have REGION->EXIT
block at all. The last insn in the block will be the parallel's GIMPLE_OMP_RETURN. The
workshare's GIMPLE_OMP_RETURN will be in a preceding block. The kinds of
statements that can appear in between are extremely limited -- no
memory operations at all. Here, we allow nothing at all, so the
only thing we allow to precede this GIMPLE_OMP_RETURN is a label. OpenMP 3.0 tasks unfortunately prevent this optimization
in many cases. If there could be tasks queued, the barrier
might be needed to let the tasks run before some local
variable of the parallel that the task uses as shared
runs out of scope. The task can be spawned either
from within current function (this would be easy to check)
or from some function it calls and gets passed an address
of such a variable.
References recompute_tree_invariant_for_addr_expr().
|
static |
References gsi_end_p(), gsi_next(), gsi_start_bb(), gsi_stmt(), and single_succ().
|
static |
Referenced by scan_omp_for().
|
static |
Scan all the statements starting at the current statement. CTX contains context information about the OpenMP directives and clauses found during the scan.
|
static |
Helper function scan_omp. Callback for walk_tree or operators in walk_gimple_stmt used to scan for OpenMP directives in TP.
|
static |
Helper function for scan_omp. Callback for walk_gimple_stmt used to scan for OpenMP directives in the current statement in GSI.
Check the OpenMP nesting restrictions.
|
static |
Scan an OpenMP loop directive.
References gimple_omp_body_ptr(), gimple_omp_teams_clauses(), new_omp_context(), scan_omp(), and scan_sharing_clauses().
|
inlinestatic |
Convenience function for calling scan_omp_1_op on tree operands.
References omp_context::stmt.
|
static |
Scan an OpenMP parallel directive.
Ignore parallel directives with empty bodies, unless there
are copyin clauses. We need two temporaries with fd.loop.v type (istart/iend)
and then (fd.collapse - 1) temporaries with the same
type for count2 ... countN-1 vars if not constant.
|
static |
Scan an OpenMP sections directive.
References error_at(), GF_OMP_FOR_KIND_DISTRIBUTE, gimple_location(), and gimple_omp_for_kind().
|
static |
Scan an OpenMP single directive.
References error_at(), GF_OMP_FOR_KIND_DISTRIBUTE, GF_OMP_FOR_KIND_SIMD, gimple_call_fndecl(), gimple_location(), gimple_omp_for_kind(), is_gimple_call(), and omp_context::stmt.
|
static |
Scan an OpenMP target{, data, update} directive.
References error_at(), gimple_call_fndecl(), and gimple_location().
|
static |
Scan an OpenMP task directive.
Ignore task directives with empty bodies.
Move VLA fields to the end.
|
static |
Scan an OpenMP teams directive.
References omp_context::cancellable, find_omp_clause(), gimple_location(), gimple_omp_sections_clauses(), OMP_CLAUSE_NOWAIT, omp_context::stmt, and warning_at().
|
static |
Instantiate decls as necessary in CTX to satisfy the data sharing specified by CLAUSES.
Ignore shared directives in teams construct.
Global variables don't need to be copied,
the receiver side will use them directly. We don't need to copy const scalar vars back.
Let the corresponding firstprivate clause create
the variable. FALLTHRU
Global variables with "omp declare target" attribute
don't need to be copied, the receiver side will use them
directly. Ignore OMP_CLAUSE_MAP_POINTER kind for arrays in
#pragma omp target data, there is nothing to map for
those. Let the corresponding firstprivate clause create
the variable. FALLTHRU
Ignore shared directives in teams construct.
Referenced by scan_omp_for().
|
static |
|
static |
|
static |
Return true if DECL should be copied by pointer. SHARED_CTX is the parallel context if DECL is to be shared.
We can only use copy-in/copy-out semantics for shared variables
when we know the value is not accessible from an outer scope. ??? Trivially accessible from anywhere. But why would we even
be passing an address in this case? Should we simply assert
this to be false, or should we have a cleanup pass that removes
these from the list of mappings? For variables with DECL_HAS_VALUE_EXPR_P set, we cannot tell
without analyzing the expression whether or not its location
is accessible to anyone else. In the case of nested parallel
regions it certainly may be. Do not use copy-in/copy-out for variables that have their
address taken. lower_send_shared_vars only uses copy-in, but not copy-out
for these. Disallow copy-in/out in nested parallel if
decl is shared in outer parallel, otherwise
each thread could store the shared variable
in its own copy-in location, making the
variable no longer really shared. For tasks avoid using copy-in/out. As tasks can be
deferred or executed in different thread, when GOMP_task
returns, the task hasn't necessarily terminated. Taking address of OUTER in lower_send_shared_vars
might need regimplification of everything that uses the
variable.
References gimple_omp_taskreg_clauses(), OMP_CLAUSE_SHARED, and omp_context::stmt.
Referenced by lower_reduction_clauses(), and omp_copy_decl_1().
|
static |
Chain all the DECLs in LIST by their TREE_CHAIN fields.
|
static |
Given two blocks PAR_ENTRY_BB and WS_ENTRY_BB such that WS_ENTRY_BB
is the immediate dominator of PAR_ENTRY_BB, return true if there
are no data dependencies that would prevent expanding the parallel
directive at PAR_ENTRY_BB as a combined parallel+workshare region.
When expanding a combined parallel+workshare region, the call to
the child function may need additional arguments in the case of
GIMPLE_OMP_FOR regions. In some cases, these arguments are
computed out of variables passed in from the parent to the child
via 'struct .omp_data_s'. For instance:
#pragma omp parallel for schedule (guided, i * 4)
for (j ...)
Is lowered into:
# BLOCK 2 (PAR_ENTRY_BB)
.omp_data_o.i = i;
#pragma omp parallel [child fn: bar.omp_fn.0 ( ..., D.1598)
# BLOCK 3 (WS_ENTRY_BB)
.omp_data_i = &.omp_data_o;
D.1667 = .omp_data_i->i;
D.1598 = D.1667 * 4;
#pragma omp for schedule (guided, D.1598)
When we outline the parallel region, the call to the child function
'bar.omp_fn.0' will need the value D.1598 in its argument list, but
that value is computed *after* the call site. So, in principle we
cannot do the transformation.
To see whether the code in WS_ENTRY_BB blocks the combined
parallel+workshare call, we collect all the variables used in the
GIMPLE_OMP_FOR header check whether they appear on the LHS of any
statement in WS_ENTRY_BB. If so, then we cannot emit the combined
call.
FIXME. If we had the SSA form built at this point, we could merely
hoist the code in block 3 into block 2 and be done with it. But at
this point we don't have dataflow information and though we could
hack something up here, it is really not worth the aggravation. FIXME. We give up too easily here. If any of these arguments
are not constants, they will likely involve variables that have
been mapped into fields of .omp_data_s for sharing with the child
function. With appropriate data flow, it would be possible to
see through this.
References omp_for_data::chunk_size, extract_omp_for_data(), find_omp_clause(), fold_convert_loc(), gimple_location(), gimple_omp_for_combined_into_p(), gimple_omp_parallel_clauses(), omp_for_data::loop, omp_for_data_loop::n1, omp_for_data_loop::n2, OMP_CLAUSE__LOOPTEMP_, omp_for_data_loop::step, and vec_alloc().
|
static |
Referenced by debug_all_omp_regions().
|
static |
The following is a utility to diagnose OpenMP structured block violations. It is not part of the "omplower" pass, as that's invoked too late. It should be invoked by the respective front ends after gimplification.
|
static |
Gimplify a GIMPLE_OMP_CRITICAL statement. This is a relatively simple substitution of a couple of function calls. But in the NAMED case, requires that languages coordinate a symbol name. It is therefore best put here in common code.
|
static |
|
static |
|
static |
|
static |
|
static |
Create a new name for omp child function. Returns an identifier.
1.8.1.1