Skip to content

[flang][OpenMP] Add alias analysis for omp private #113566

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
DominikAdamski merged 5 commits into from
Nov 5, 2024
Merged

[flang][OpenMP] Add alias analysis for omp private #113566

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8a6bcdc
[flang][OpenMP] Add alias analysis for omp private
DominikAdamski Oct 21, 2024
362cf24
refactor
DominikAdamski Oct 25, 2024
14808d9
Extend alias analysis for other OpenMP operations
DominikAdamski Oct 25, 2024
a027f1d
applied remarks
DominikAdamski Oct 28, 2024
a0f2424
Merge branch 'main' into aa_omp_private
DominikAdamski Nov 5, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 54 additions & 14 deletions flang/lib/Optimizer/Analysis/AliasAnalysis.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -372,6 +372,29 @@ getAttrsFromVariable(fir::FortranVariableOpInterface var) {
return attrs;
}

template <typename OMPTypeOp, typename DeclTypeOp>
static Value getPrivateArg(omp::BlockArgOpenMPOpInterface &argIface,
OMPTypeOp &op, DeclTypeOp &declOp) {
Value privateArg;
if (!op.getPrivateSyms().has_value())
return privateArg;
for (auto [opSym, blockArg] :
llvm::zip_equal(*op.getPrivateSyms(), argIface.getPrivateBlockArgs())) {
if (blockArg == declOp.getMemref()) {
omp::PrivateClauseOp privateOp =
SymbolTable::lookupNearestSymbolFrom<omp::PrivateClauseOp>(
op, cast<SymbolRefAttr>(opSym));
privateOp.walk([&](omp::YieldOp yieldOp) {
llvm::TypeSwitch<Operation *>(yieldOp.getResults()[0].getDefiningOp())
.template Case<fir::DeclareOp, hlfir::DeclareOp>(
[&](auto declOp) { privateArg = declOp.getMemref(); });
});
return privateArg;
}
}
return privateArg;
}

AliasAnalysis::Source AliasAnalysis::getSource(mlir::Value v,
bool getInstantiationPoint) {
auto *defOp = v.getDefiningOp();
Expand Down Expand Up @@ -470,20 +493,37 @@ AliasAnalysis::Source AliasAnalysis::getSource(mlir::Value v,
breakFromLoop = true;
})
.Case<hlfir::DeclareOp, fir::DeclareOp>([&](auto op) {
// If declare operation is inside omp target region,
// continue alias analysis outside the target region
if (auto targetOp =
llvm::dyn_cast<omp::TargetOp>(op->getParentOp())) {
auto argIface = cast<omp::BlockArgOpenMPOpInterface>(*targetOp);
for (auto [opArg, blockArg] : llvm::zip_equal(
targetOp.getMapVars(), argIface.getMapBlockArgs())) {
if (blockArg == op.getMemref()) {
omp::MapInfoOp mapInfo =
llvm::cast<omp::MapInfoOp>(opArg.getDefiningOp());
v = mapInfo.getVarPtr();
defOp = v.getDefiningOp();
return;
}
if (omp::BlockArgOpenMPOpInterface argIface =
dyn_cast<omp::BlockArgOpenMPOpInterface>(op->getParentOp())) {
Value ompValArg;
llvm::TypeSwitch<Operation *>(op->getParentOp())
.template Case<omp::TargetOp>([&](auto targetOp) {
// If declare operation is inside omp target region,
// continue alias analysis outside the target region
for (auto [opArg, blockArg] : llvm::zip_equal(
targetOp.getMapVars(), argIface.getMapBlockArgs())) {
if (blockArg == op.getMemref()) {
omp::MapInfoOp mapInfo =
llvm::cast<omp::MapInfoOp>(opArg.getDefiningOp());
ompValArg = mapInfo.getVarPtr();
break;
}
}
// If given operation does not reflect mapping item,
// check private clause
if (!ompValArg)
ompValArg = getPrivateArg(argIface, targetOp, op);
})
.template Case<omp::DistributeOp, omp::ParallelOp,
omp::SectionsOp, omp::SimdOp, omp::SingleOp,
omp::TaskloopOp, omp::TaskOp, omp::WsloopOp>(
[&](auto privateOp) {
ompValArg = getPrivateArg(argIface, privateOp, op);
});
if (ompValArg) {
v = ompValArg;
defOp = ompValArg.getDefiningOp();
return;
}
}
auto varIf = llvm::cast<fir::FortranVariableOpInterface>(defOp);
Expand Down
102 changes: 102 additions & 0 deletions flang/test/Analysis/AliasAnalysis/alias-analysis-omp-teams-distribute-private-ptr.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,102 @@
// Use --mlir-disable-threading so that the AA queries are serialized
// as well as its diagnostic output.
// RUN: fir-opt %s -pass-pipeline='builtin.module(func.func(test-fir-alias-analysis))' -split-input-file --mlir-disable-threading 2>&1 | FileCheck %s

// Fortran code:
// program main
// integer, target :: arrayA(10)
// integer, pointer, dimension(:) :: ptrA
// integer :: i
// ptrA => arrayA
// !$omp teams distribute parallel do firstprivate(ptrA)
// do i = 1, 10
// arrayA(i) = arrayA(i) + ptrA(i);
// end do
// end program main

// CHECK-LABEL: Testing : "_QQmain"
// CHECK-DAG: ptrA#0 <-> ArrayA#0: MayAlias

omp.private {type = private} @_QFEi_private_ref_i32 : !fir.ref<i32> alloc {
^bb0(%arg0: !fir.ref<i32>):
%0 = fir.alloca i32 {bindc_name = "i", pinned, uniq_name = "_QFEi"}
%1:2 = hlfir.declare %0 {uniq_name = "_QFEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
omp.yield(%1#0 : !fir.ref<i32>)
}
omp.private {type = firstprivate} @_QFEptra_firstprivate_ref_box_ptr_Uxi32 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>> alloc {
^bb0(%arg0: !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>):
%0 = fir.alloca !fir.box<!fir.ptr<!fir.array<?xi32>>> {bindc_name = "ptra", pinned, uniq_name = "_QFEptra"}
%1:2 = hlfir.declare %0 {fortran_attrs = #fir.var_attrs<pointer>, uniq_name = "_QFEptra"} : (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>) -> (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>)
omp.yield(%1#0 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>)
} copy {
^bb0(%arg0: !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>, %arg1: !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>):
%0 = fir.load %arg0 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
fir.store %0 to %arg1 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
omp.yield(%arg1 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>)
}
func.func @_QQmain() attributes {fir.bindc_name = "main"} {
%0 = fir.address_of(@_QFEarraya) : !fir.ref<!fir.array<10xi32>>
%c10 = arith.constant 10 : index
%1 = fir.shape %c10 : (index) -> !fir.shape<1>
%2:2 = hlfir.declare %0(%1) {fortran_attrs = #fir.var_attrs<target>, uniq_name = "_QFEarraya"} : (!fir.ref<!fir.array<10xi32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<10xi32>>, !fir.ref<!fir.array<10xi32>>)
%3 = fir.address_of(@_QFEarrayb) : !fir.ref<!fir.array<10xi32>>
%c10_0 = arith.constant 10 : index
%4 = fir.shape %c10_0 : (index) -> !fir.shape<1>
%5:2 = hlfir.declare %3(%4) {uniq_name = "_QFEarrayb"} : (!fir.ref<!fir.array<10xi32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<10xi32>>, !fir.ref<!fir.array<10xi32>>)
%6 = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFEi"}
%7:2 = hlfir.declare %6 {uniq_name = "_QFEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
%8 = fir.address_of(@_QFEptra) : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
%9:2 = hlfir.declare %8 {fortran_attrs = #fir.var_attrs<pointer>, uniq_name = "_QFEptra"} : (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>) -> (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>)
%10 = fir.shape %c10 : (index) -> !fir.shape<1>
%11 = fir.embox %2#1(%10) : (!fir.ref<!fir.array<10xi32>>, !fir.shape<1>) -> !fir.box<!fir.ptr<!fir.array<?xi32>>>
fir.store %11 to %9#1 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
omp.teams {
omp.parallel private(@_QFEptra_firstprivate_ref_box_ptr_Uxi32 %9#0 -> %arg0, @_QFEi_private_ref_i32 %7#0 -> %arg1 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>, !fir.ref<i32>) {
%12:2 = hlfir.declare %arg0 {fortran_attrs = #fir.var_attrs<pointer>, uniq_name = "_QFEptra"} : (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>) -> (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>)
%13:2 = hlfir.declare %arg1 {uniq_name = "_QFEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
%c1_i32 = arith.constant 1 : i32
%c10_i32 = arith.constant 10 : i32
%c1_i32_1 = arith.constant 1 : i32
omp.distribute {
omp.wsloop {
omp.loop_nest (%arg2) : i32 = (%c1_i32) to (%c10_i32) inclusive step (%c1_i32_1) {
fir.store %arg2 to %13#1 : !fir.ref<i32>
%14 = fir.load %13#0 : !fir.ref<i32>
%15 = fir.convert %14 : (i32) -> i64
%16 = hlfir.designate %2#0 (%15) : (!fir.ref<!fir.array<10xi32>>, i64) -> !fir.ref<i32>
%17 = fir.load %16 : !fir.ref<i32>
%18 = fir.load %12#0 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
%19 = fir.load %13#0 : !fir.ref<i32>
%20 = fir.convert %19 : (i32) -> i64
%21 = hlfir.designate %18 (%20) {test.ptr = "ptrA" } : (!fir.box<!fir.ptr<!fir.array<?xi32>>>, i64) -> !fir.ref<i32>
%22 = fir.load %21 : !fir.ref<i32>
%23 = arith.addi %17, %22 : i32
%24 = fir.load %13#0 : !fir.ref<i32>
%25 = fir.convert %24 : (i32) -> i64
%26 = hlfir.designate %2#0 (%25) {test.ptr = "ArrayA"} : (!fir.ref<!fir.array<10xi32>>, i64) -> !fir.ref<i32>
hlfir.assign %23 to %26 : i32, !fir.ref<i32>
omp.yield
}
} {omp.composite}
} {omp.composite}
omp.terminator
} {omp.composite}
omp.terminator
}
return
}
fir.global internal @_QFEarraya target : !fir.array<10xi32> {
%0 = fir.zero_bits !fir.array<10xi32>
fir.has_value %0 : !fir.array<10xi32>
}
fir.global internal @_QFEarrayb : !fir.array<10xi32> {
%0 = fir.zero_bits !fir.array<10xi32>
fir.has_value %0 : !fir.array<10xi32>
}
fir.global internal @_QFEptra : !fir.box<!fir.ptr<!fir.array<?xi32>>> {
%0 = fir.zero_bits !fir.ptr<!fir.array<?xi32>>
%c0 = arith.constant 0 : index
%1 = fir.shape %c0 : (index) -> !fir.shape<1>
%2 = fir.embox %0(%1) : (!fir.ptr<!fir.array<?xi32>>, !fir.shape<1>) -> !fir.box<!fir.ptr<!fir.array<?xi32>>>
fir.has_value %2 : !fir.box<!fir.ptr<!fir.array<?xi32>>>
}
121 changes: 121 additions & 0 deletions flang/test/Analysis/AliasAnalysis/alias-analysis-omp-teams-distribute-private.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,121 @@
// Use --mlir-disable-threading so that the AA queries are serialized
// as well as its diagnostic output.
// RUN: fir-opt %s -pass-pipeline='builtin.module(func.func(test-fir-alias-analysis))' -split-input-file --mlir-disable-threading 2>&1 | FileCheck %s

// Fortran code:
//
// program main
// integer :: arrayA(10,10)
// integer :: tmp(2)
// integer :: i,j
// !$omp teams distribute parallel do private(tmp)
// do j = 1, 10
// do i = 1,10
// tmp = [i,j]
// arrayA = tmp(1)
// end do
// end do
// end program main

// CHECK-LABEL: Testing : "_QQmain"
// CHECK-DAG: tmp_private_array#0 <-> unnamed_array#0: NoAlias
// CHECK-DAG: tmp_private_array#1 <-> unnamed_array#0: NoAlias
Copy link
Contributor

@jeanPerier jeanPerier Oct 25, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you add a positive case? I understand NoAlias is what matter from an optimization point of view, but MayAlias is the most important to prove from a correctness point of view, so I think it is always good to add both negative and positive test cases here.

Copy link
Contributor Author

@DominikAdamski DominikAdamski Nov 5, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done


omp.private {type = private} @_QFEi_private_ref_i32 : !fir.ref<i32> alloc {
^bb0(%arg0: !fir.ref<i32>):
%0 = fir.alloca i32 {bindc_name = "i", pinned, uniq_name = "_QFEi"}
%1:2 = hlfir.declare %0 {uniq_name = "_QFEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
omp.yield(%1#0 : !fir.ref<i32>)
}
omp.private {type = private} @_QFEj_private_ref_i32 : !fir.ref<i32> alloc {
^bb0(%arg0: !fir.ref<i32>):
%0 = fir.alloca i32 {bindc_name = "j", pinned, uniq_name = "_QFEj"}
%1:2 = hlfir.declare %0 {uniq_name = "_QFEj"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
omp.yield(%1#0 : !fir.ref<i32>)
}
omp.private {type = private} @_QFEtmp_private_ref_2xi32 : !fir.ref<!fir.array<2xi32>> alloc {
^bb0(%arg0: !fir.ref<!fir.array<2xi32>>):
%c2 = arith.constant 2 : index
%0 = fir.alloca !fir.array<2xi32> {bindc_name = "tmp", pinned, uniq_name = "_QFEtmp"}
%1 = fir.shape %c2 : (index) -> !fir.shape<1>
%2:2 = hlfir.declare %0(%1) {uniq_name = "_QFEtmp"} : (!fir.ref<!fir.array<2xi32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<2xi32>>, !fir.ref<!fir.array<2xi32>>)
omp.yield(%2#0 : !fir.ref<!fir.array<2xi32>>)
}
func.func @_QQmain() attributes {fir.bindc_name = "main"} {
%0 = fir.address_of(@_QFEarraya) : !fir.ref<!fir.array<10x10xi32>>
%c10 = arith.constant 10 : index
%c10_0 = arith.constant 10 : index
%1 = fir.shape %c10, %c10_0 : (index, index) -> !fir.shape<2>
%2:2 = hlfir.declare %0(%1) {uniq_name = "_QFEarraya"} : (!fir.ref<!fir.array<10x10xi32>>, !fir.shape<2>) -> (!fir.ref<!fir.array<10x10xi32>>, !fir.ref<!fir.array<10x10xi32>>)
%3 = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFEi"}
%4:2 = hlfir.declare %3 {uniq_name = "_QFEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
%5 = fir.alloca i32 {bindc_name = "j", uniq_name = "_QFEj"}
%6:2 = hlfir.declare %5 {uniq_name = "_QFEj"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
%c2 = arith.constant 2 : index
%7 = fir.alloca !fir.array<2xi32> {bindc_name = "tmp", uniq_name = "_QFEtmp"}
%8 = fir.shape %c2 : (index) -> !fir.shape<1>
%9:2 = hlfir.declare %7(%8) {uniq_name = "_QFEtmp"} : (!fir.ref<!fir.array<2xi32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<2xi32>>, !fir.ref<!fir.array<2xi32>>)
omp.teams {
omp.parallel private(@_QFEtmp_private_ref_2xi32 %9#0 -> %arg0, @_QFEj_private_ref_i32 %6#0 -> %arg1, @_QFEi_private_ref_i32 %4#0 -> %arg2 : !fir.ref<!fir.array<2xi32>>, !fir.ref<i32>, !fir.ref<i32>) {
%c2_1 = arith.constant 2 : index
%10 = fir.shape %c2_1 : (index) -> !fir.shape<1>
%11:2 = hlfir.declare %arg0(%10) {uniq_name = "_QFEtmp", test.ptr = "tmp_private_array"} : (!fir.ref<!fir.array<2xi32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<2xi32>>, !fir.ref<!fir.array<2xi32>>)
%12:2 = hlfir.declare %arg1 {uniq_name = "_QFEj"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
%13:2 = hlfir.declare %arg2 {uniq_name = "_QFEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
%c1_i32 = arith.constant 1 : i32
%c10_i32 = arith.constant 10 : i32
%c1_i32_2 = arith.constant 1 : i32
omp.distribute {
omp.wsloop {
omp.loop_nest (%arg3) : i32 = (%c1_i32) to (%c10_i32) inclusive step (%c1_i32_2) {
fir.store %arg3 to %12#1 : !fir.ref<i32>
%c1_i32_3 = arith.constant 1 : i32
%14 = fir.convert %c1_i32_3 : (i32) -> index
%c10_i32_4 = arith.constant 10 : i32
%15 = fir.convert %c10_i32_4 : (i32) -> index
%c1 = arith.constant 1 : index
%16 = fir.convert %14 : (index) -> i32
%17:2 = fir.do_loop %arg4 = %14 to %15 step %c1 iter_args(%arg5 = %16) -> (index, i32) {
fir.store %arg5 to %13#1 : !fir.ref<i32>
%c2_5 = arith.constant 2 : index
%c1_6 = arith.constant 1 : index
%c1_7 = arith.constant 1 : index
%18 = fir.allocmem !fir.array<2xi32> {bindc_name = ".tmp.arrayctor", uniq_name = ""}
%19 = fir.shape %c2_5 : (index) -> !fir.shape<1>
%20:2 = hlfir.declare %18(%19) {uniq_name = ".tmp.arrayctor"} : (!fir.heap<!fir.array<2xi32>>, !fir.shape<1>) -> (!fir.heap<!fir.array<2xi32>>, !fir.heap<!fir.array<2xi32>>)
%21 = fir.load %13#0 : !fir.ref<i32>
%22 = arith.addi %c1_6, %c1_7 : index
%23 = hlfir.designate %20#0 (%c1_6) : (!fir.heap<!fir.array<2xi32>>, index) -> !fir.ref<i32>
hlfir.assign %21 to %23 : i32, !fir.ref<i32>
%24 = fir.load %12#0 : !fir.ref<i32>
%25 = hlfir.designate %20#0 (%22) : (!fir.heap<!fir.array<2xi32>>, index) -> !fir.ref<i32>
hlfir.assign %24 to %25 : i32, !fir.ref<i32>
%true = arith.constant true
%26 = hlfir.as_expr %20#0 move %true {test.ptr = "unnamed_array"} : (!fir.heap<!fir.array<2xi32>>, i1) -> !hlfir.expr<2xi32>
hlfir.assign %26 to %11#0 : !hlfir.expr<2xi32>, !fir.ref<!fir.array<2xi32>>
hlfir.destroy %26 : !hlfir.expr<2xi32>
%c1_8 = arith.constant 1 : index
%27 = hlfir.designate %11#0 (%c1_8) : (!fir.ref<!fir.array<2xi32>>, index) -> !fir.ref<i32>
%28 = fir.load %27 : !fir.ref<i32>
hlfir.assign %28 to %2#0 : i32, !fir.ref<!fir.array<10x10xi32>>
%29 = arith.addi %arg4, %c1 : index
%30 = fir.convert %c1 : (index) -> i32
%31 = fir.load %13#1 : !fir.ref<i32>
%32 = arith.addi %31, %30 : i32
fir.result %29, %32 : index, i32
}
fir.store %17#1 to %13#1 : !fir.ref<i32>
omp.yield
}
} {omp.composite}
} {omp.composite}
omp.terminator
} {omp.composite}
omp.terminator
}
return
}
fir.global internal @_QFEarraya : !fir.array<10x10xi32> {
%0 = fir.zero_bits !fir.array<10x10xi32>
fir.has_value %0 : !fir.array<10x10xi32>
}
Loading