Skip to content

Commit 62596cd

Browse files
committed
[flang][OpenMP] Map basic local specifiers to private clauses
Starts the effort to map `do concurrent` locality specifiers to OpenMP clauses. This PR adds support for basic specifiers (no `init` or `copy` regions yet).
1 parent fd943c2 commit 62596cd

File tree

2 files changed

+101
-2
lines changed

2 files changed

+101
-2
lines changed

flang/lib/Optimizer/OpenMP/DoConcurrentConversion.cpp

Lines changed: 53 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -7,9 +7,11 @@
77
//===----------------------------------------------------------------------===//
88

99
#include "flang/Optimizer/Builder/FIRBuilder.h"
10+
#include "flang/Optimizer/Builder/Todo.h"
1011
#include "flang/Optimizer/Dialect/FIROps.h"
1112
#include "flang/Optimizer/OpenMP/Passes.h"
1213
#include "flang/Optimizer/OpenMP/Utils.h"
14+
#include "flang/Support/OpenMP-utils.h"
1315
#include "mlir/Analysis/SliceAnalysis.h"
1416
#include "mlir/Dialect/OpenMP/OpenMPDialect.h"
1517
#include "mlir/IR/IRMapping.h"
@@ -308,10 +310,47 @@ class DoConcurrentConversion
308310
fir::DoConcurrentLoopOp loop, mlir::IRMapping &mapper,
309311
const mlir::omp::LoopNestOperands &clauseOps,
310312
bool isComposite) const {
313+
mlir::omp::WsloopOperands wsloopClauseOps;
314+
315+
// For `local` (and `local_init`) opernads, emit corresponding `private`
316+
// clauses and attach these clauses to the workshare loop.
317+
if (!loop.getLocalOperands().empty())
318+
for (auto [op, sym, arg] : llvm::zip_equal(
319+
loop.getLocalOperands(),
320+
loop.getLocalSymsAttr().getAsRange<mlir::SymbolRefAttr>(),
321+
loop.getRegionLocalArgs())) {
322+
auto localizer = mlir::SymbolTable::lookupNearestSymbolFrom<
323+
fir::LocalitySpecifierOp>(loop, sym);
324+
if (localizer.getLocalitySpecifierType() ==
325+
fir::LocalitySpecifierType::LocalInit)
326+
TODO(localizer.getLoc(),
327+
"local_init conversion is not supported yet");
328+
329+
if (!localizer.getInitRegion().empty())
330+
TODO(localizer.getLoc(),
331+
"non-empty `init` regions are not supported yet");
332+
333+
auto oldIP = rewriter.saveInsertionPoint();
334+
rewriter.setInsertionPointAfter(localizer);
335+
auto privatizer = rewriter.create<mlir::omp::PrivateClauseOp>(
336+
localizer.getLoc(), sym.getLeafReference().str() + ".omp",
337+
localizer.getTypeAttr().getValue(),
338+
mlir::omp::DataSharingClauseType::Private);
339+
rewriter.restoreInsertionPoint(oldIP);
340+
341+
wsloopClauseOps.privateVars.push_back(op);
342+
wsloopClauseOps.privateSyms.push_back(
343+
mlir::SymbolRefAttr::get(privatizer));
344+
}
311345

312-
auto wsloopOp = rewriter.create<mlir::omp::WsloopOp>(loop.getLoc());
346+
auto wsloopOp =
347+
rewriter.create<mlir::omp::WsloopOp>(loop.getLoc(), wsloopClauseOps);
313348
wsloopOp.setComposite(isComposite);
314-
rewriter.createBlock(&wsloopOp.getRegion());
349+
350+
Fortran::common::openmp::EntryBlockArgs wsloopArgs;
351+
wsloopArgs.priv.vars = wsloopClauseOps.privateVars;
352+
Fortran::common::openmp::genEntryBlock(rewriter, wsloopArgs,
353+
wsloopOp.getRegion());
315354

316355
auto loopNestOp =
317356
rewriter.create<mlir::omp::LoopNestOp>(loop.getLoc(), clauseOps);
@@ -324,6 +363,18 @@ class DoConcurrentConversion
324363
rewriter.setInsertionPointToEnd(&loopNestOp.getRegion().back());
325364
rewriter.create<mlir::omp::YieldOp>(loop->getLoc());
326365

366+
// `local` region arguments are transferred/cloned from the `do concurrent`
367+
// loop to the loopnest op when the region is cloned above. Instead, these
368+
// region arguments should be on the workshare loop's region.
369+
for (auto [wsloopArg, loopNestArg] :
370+
llvm::zip_equal(wsloopOp.getRegion().getArguments(),
371+
loopNestOp.getRegion().getArguments().drop_front(
372+
clauseOps.loopLowerBounds.size())))
373+
rewriter.replaceAllUsesWith(loopNestArg, wsloopArg);
374+
375+
for (unsigned i = 0; i < loop.getLocalVars().size(); ++i)
376+
loopNestOp.getRegion().eraseArgument(clauseOps.loopLowerBounds.size());
377+
327378
return loopNestOp;
328379
}
329380

Lines changed: 48 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,48 @@
1+
// Tests mapping `local` locality specifier to `private` clauses for a simple
2+
// case (not `init` or `copy` regions).
3+
4+
// RUN: fir-opt --omp-do-concurrent-conversion="map-to=host" %s | FileCheck %s
5+
6+
fir.local {type = local} @_QFlocal_spec_translationElocal_var_private_f32 : f32
7+
8+
func.func @_QPlocal_spec_translation() {
9+
%3 = fir.alloca f32 {bindc_name = "local_var", uniq_name = "_QFlocal_spec_translationElocal_var"}
10+
%4:2 = hlfir.declare %3 {uniq_name = "_QFlocal_spec_translationElocal_var"} : (!fir.ref<f32>) -> (!fir.ref<f32>, !fir.ref<f32>)
11+
12+
%c4_i32 = arith.constant 4 : index
13+
%c11_i32 = arith.constant 11 : index
14+
%c1 = arith.constant 1 : index
15+
16+
fir.do_concurrent {
17+
%7 = fir.alloca i32 {bindc_name = "i"}
18+
%8:2 = hlfir.declare %7 {uniq_name = "_QFlocal_spec_translationEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
19+
20+
fir.do_concurrent.loop (%arg0) = (%c4_i32) to (%c11_i32) step (%c1)
21+
local(@_QFlocal_spec_translationElocal_var_private_f32 %4#0 -> %arg1 : !fir.ref<f32>) {
22+
%9 = fir.convert %arg0 : (index) -> i32
23+
fir.store %9 to %8#0 : !fir.ref<i32>
24+
25+
%10:2 = hlfir.declare %arg1 {uniq_name = "_QFlocal_spec_translationElocal_var"} : (!fir.ref<f32>) -> (!fir.ref<f32>, !fir.ref<f32>)
26+
%cst = arith.constant 4.200000e+01 : f32
27+
hlfir.assign %cst to %10#0 : f32, !fir.ref<f32>
28+
}
29+
}
30+
return
31+
}
32+
33+
// CHECK: omp.private {type = private} @[[PRIVATIZER:.*local_spec_translationElocal_var.*.omp]] : f32
34+
35+
// CHECK: func.func @_QPlocal_spec_translation
36+
// CHECK: %[[LOCAL_VAR:.*]] = fir.alloca f32 {bindc_name = "local_var", {{.*}}}
37+
// CHECK: %[[LOCAL_VAR_DECL:.*]]:2 = hlfir.declare %[[LOCAL_VAR]]
38+
// CHECK: omp.parallel {
39+
// CHECK: omp.wsloop private(@[[PRIVATIZER]] %[[LOCAL_VAR_DECL]]#0 -> %[[LOCAL_ARG:.*]] : !fir.ref<f32>) {
40+
// CHECK: omp.loop_nest {{.*}} {
41+
// CHECK: %[[PRIV_DECL:.*]]:2 = hlfir.declare %[[LOCAL_ARG]]
42+
// CHECK: %[[C42:.*]] = arith.constant
43+
// CHECK: hlfir.assign %[[C42]] to %[[PRIV_DECL]]#0
44+
// CHECK: omp.yield
45+
// CHECK: }
46+
// CHECK: }
47+
// CHECK: omp.terminator
48+
// CHECK: }

0 commit comments

Comments
 (0)