[flang][OpenMP] Generalize fixing alloca IP pre-condition for private ops
#122866
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…ate` ops This PR generalizes a fix that we implemented previously for `omp.wsloop`s. The fix makes sure the pre-condtion that the `alloca` block has a single successor whenever we inline delayed privatizers is respected. I simply moved the fix to `allocatePrivateVars` so that it kicks in for any op not just `omp.wsloop`. This handles a bug uncovered by [a test](https://github.com/OpenMP-Validation-and-Verification/OpenMP_VV/blob/master/tests/4.5/target_simd/test_target_simd_safelen.F90) in the OpenMP_VV test suite.
|
@llvm/pr-subscribers-mlir-openmp @llvm/pr-subscribers-flang-openmp Author: Kareem Ergawy (ergawy) ChangesThis PR generalizes a fix that we implemented previously for This handles a bug uncovered by a test in the OpenMP_VV test suite. Full diff: https://github.com/llvm/llvm-project/pull/122866.diff 2 Files Affected:
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index d6112fa9af1182..6ffe169038c54e 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -1340,13 +1340,23 @@ allocatePrivateVars(llvm::IRBuilderBase &builder,
llvm::SmallVectorImpl<llvm::Value *> &llvmPrivateVars,
const llvm::OpenMPIRBuilder::InsertPointTy &allocaIP,
llvm::DenseMap<Value, Value> *mappedPrivateVars = nullptr) {
- llvm::IRBuilderBase::InsertPointGuard guard(builder);
// Allocate private vars
llvm::BranchInst *allocaTerminator =
llvm::cast<llvm::BranchInst>(allocaIP.getBlock()->getTerminator());
+ if (allocaTerminator->getNumSuccessors() != 1) {
+ splitBB(llvm::OpenMPIRBuilder::InsertPointTy(
+ allocaIP.getBlock(), allocaTerminator->getIterator()),
+ true, "omp.region.after_alloca");
+ }
+
+ llvm::IRBuilderBase::InsertPointGuard guard(builder);
+ // Update the allocaTerminator in case the alloca block was split above.
+ allocaTerminator =
+ llvm::cast<llvm::BranchInst>(allocaIP.getBlock()->getTerminator());
builder.SetInsertPoint(allocaTerminator);
assert(allocaTerminator->getNumSuccessors() == 1 &&
"This is an unconditional branch created by OpenMPIRBuilder");
+
llvm::BasicBlock *afterAllocas = allocaTerminator->getSuccessor(0);
// FIXME: Some of the allocation regions do more than just allocating.
@@ -1876,11 +1886,6 @@ convertOmpWsloop(Operation &opInst, llvm::IRBuilderBase &builder,
SmallVector<llvm::Value *> privateReductionVariables(
wsloopOp.getNumReductionVars());
- splitBB(llvm::OpenMPIRBuilder::InsertPointTy(
- allocaIP.getBlock(),
- allocaIP.getBlock()->getTerminator()->getIterator()),
- true, "omp.region.after_alloca");
-
llvm::Expected<llvm::BasicBlock *> afterAllocas = allocatePrivateVars(
builder, moduleTranslation, privateBlockArgs, privateDecls,
mlirPrivateVars, llvmPrivateVars, allocaIP);
diff --git a/mlir/test/Target/LLVMIR/openmp-target-simd-on_device.mlir b/mlir/test/Target/LLVMIR/openmp-target-simd-on_device.mlir
new file mode 100644
index 00000000000000..0ce90578ea9d62
--- /dev/null
+++ b/mlir/test/Target/LLVMIR/openmp-target-simd-on_device.mlir
@@ -0,0 +1,34 @@
+// RUN: mlir-translate -mlir-to-llvmir %s | FileCheck %s
+
+module attributes {omp.is_target_device = true} {
+ omp.private {type = private} @simd_privatizer : !llvm.ptr alloc {
+ ^bb0(%arg0: !llvm.ptr):
+ omp.yield(%arg0 : !llvm.ptr)
+ }
+
+ llvm.func @test_target_simd() {
+ omp.target {
+ %5 = llvm.mlir.constant(1 : i32) : i32
+ %x = llvm.alloca %5 x i32 {bindc_name = "x"} : (i32) -> !llvm.ptr
+ omp.simd private(@simd_privatizer %x -> %arg1 : !llvm.ptr) {
+ omp.loop_nest (%arg2) : i32 = (%5) to (%5) inclusive step (%5) {
+ omp.yield
+ }
+ }
+ omp.terminator
+ }
+ llvm.return
+ }
+
+}
+
+// CHECK-LABEL: define {{.*}} @__omp_offloading_{{.*}}_test_target_simd_{{.*}}
+
+// CHECK: %[[INT:.*]] = alloca i32, align 4
+// CHECK: br label %[[LATE_ALLOC_BB:.*]]
+
+// CHECK: [[LATE_ALLOC_BB]]:
+// CHECK: br label %[[AFTER_ALLOC_BB:.*]]
+
+// CHECK: [[AFTER_ALLOC_BB]]:
+// CHECK: br i1 %{{.*}}, label %{{.*}}, label %{{.*}}
|
|
@llvm/pr-subscribers-mlir Author: Kareem Ergawy (ergawy) ChangesThis PR generalizes a fix that we implemented previously for This handles a bug uncovered by a test in the OpenMP_VV test suite. Full diff: https://github.com/llvm/llvm-project/pull/122866.diff 2 Files Affected:
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index d6112fa9af1182..6ffe169038c54e 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -1340,13 +1340,23 @@ allocatePrivateVars(llvm::IRBuilderBase &builder,
llvm::SmallVectorImpl<llvm::Value *> &llvmPrivateVars,
const llvm::OpenMPIRBuilder::InsertPointTy &allocaIP,
llvm::DenseMap<Value, Value> *mappedPrivateVars = nullptr) {
- llvm::IRBuilderBase::InsertPointGuard guard(builder);
// Allocate private vars
llvm::BranchInst *allocaTerminator =
llvm::cast<llvm::BranchInst>(allocaIP.getBlock()->getTerminator());
+ if (allocaTerminator->getNumSuccessors() != 1) {
+ splitBB(llvm::OpenMPIRBuilder::InsertPointTy(
+ allocaIP.getBlock(), allocaTerminator->getIterator()),
+ true, "omp.region.after_alloca");
+ }
+
+ llvm::IRBuilderBase::InsertPointGuard guard(builder);
+ // Update the allocaTerminator in case the alloca block was split above.
+ allocaTerminator =
+ llvm::cast<llvm::BranchInst>(allocaIP.getBlock()->getTerminator());
builder.SetInsertPoint(allocaTerminator);
assert(allocaTerminator->getNumSuccessors() == 1 &&
"This is an unconditional branch created by OpenMPIRBuilder");
+
llvm::BasicBlock *afterAllocas = allocaTerminator->getSuccessor(0);
// FIXME: Some of the allocation regions do more than just allocating.
@@ -1876,11 +1886,6 @@ convertOmpWsloop(Operation &opInst, llvm::IRBuilderBase &builder,
SmallVector<llvm::Value *> privateReductionVariables(
wsloopOp.getNumReductionVars());
- splitBB(llvm::OpenMPIRBuilder::InsertPointTy(
- allocaIP.getBlock(),
- allocaIP.getBlock()->getTerminator()->getIterator()),
- true, "omp.region.after_alloca");
-
llvm::Expected<llvm::BasicBlock *> afterAllocas = allocatePrivateVars(
builder, moduleTranslation, privateBlockArgs, privateDecls,
mlirPrivateVars, llvmPrivateVars, allocaIP);
diff --git a/mlir/test/Target/LLVMIR/openmp-target-simd-on_device.mlir b/mlir/test/Target/LLVMIR/openmp-target-simd-on_device.mlir
new file mode 100644
index 00000000000000..0ce90578ea9d62
--- /dev/null
+++ b/mlir/test/Target/LLVMIR/openmp-target-simd-on_device.mlir
@@ -0,0 +1,34 @@
+// RUN: mlir-translate -mlir-to-llvmir %s | FileCheck %s
+
+module attributes {omp.is_target_device = true} {
+ omp.private {type = private} @simd_privatizer : !llvm.ptr alloc {
+ ^bb0(%arg0: !llvm.ptr):
+ omp.yield(%arg0 : !llvm.ptr)
+ }
+
+ llvm.func @test_target_simd() {
+ omp.target {
+ %5 = llvm.mlir.constant(1 : i32) : i32
+ %x = llvm.alloca %5 x i32 {bindc_name = "x"} : (i32) -> !llvm.ptr
+ omp.simd private(@simd_privatizer %x -> %arg1 : !llvm.ptr) {
+ omp.loop_nest (%arg2) : i32 = (%5) to (%5) inclusive step (%5) {
+ omp.yield
+ }
+ }
+ omp.terminator
+ }
+ llvm.return
+ }
+
+}
+
+// CHECK-LABEL: define {{.*}} @__omp_offloading_{{.*}}_test_target_simd_{{.*}}
+
+// CHECK: %[[INT:.*]] = alloca i32, align 4
+// CHECK: br label %[[LATE_ALLOC_BB:.*]]
+
+// CHECK: [[LATE_ALLOC_BB]]:
+// CHECK: br label %[[AFTER_ALLOC_BB:.*]]
+
+// CHECK: [[AFTER_ALLOC_BB]]:
+// CHECK: br i1 %{{.*}}, label %{{.*}}, label %{{.*}}
|
|
Seems like the failed simd tests now pass according to @ajarmusch. Can you guys please ✅ if you do not have any objections to the PR? |
tblah
approved these changes
Jan 15, 2025
|
Linaro reported a regression due to this PR: https://linaro.atlassian.net/browse/LLVM-1520. I am taking a look into this ... (cc @luporl) |
ergawy
added a commit
to ergawy/llvm-project
that referenced
this pull request
Jan 16, 2025
…PrivateVars` While llvm#122866 fixed some issues, it introduced a regression in worksharing loops. The new bug comes from the fact that we now conditionally created the `after_alloca` block based on the number of predecessors of the alloca insertion point. This is unneccessary, we can just alway create the block. If we do this, we respect the post condtions expected after calling `allocatePrivateVars` (i.e. that the `afterAlloca` block has a single predecessor.
|
A possible fix for the Fujistu regression: #123168. |
ergawy
added a commit
that referenced
this pull request
Jan 16, 2025
…atePrivateVars` (#123168) While #122866 fixed some issues, it introduced a regression in worksharing loops. The new bug comes from the fact that we now conditionally created the `after_alloca` block based on the number of sucessors of the alloca insertion point. This is unneccessary, we can just alway create the block. If we do this, we respect the post condtions expected after calling `allocatePrivateVars` (i.e. that the `afterAlloca` block has a single predecessor.
github-actions bot
pushed a commit
to arm/arm-toolchain
that referenced
this pull request
Jan 16, 2025
…k in `allocatePrivateVars` (#123168) While llvm/llvm-project#122866 fixed some issues, it introduced a regression in worksharing loops. The new bug comes from the fact that we now conditionally created the `after_alloca` block based on the number of sucessors of the alloca insertion point. This is unneccessary, we can just alway create the block. If we do this, we respect the post condtions expected after calling `allocatePrivateVars` (i.e. that the `afterAlloca` block has a single predecessor.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR generalizes a fix that we implemented previously for
omp.wsloops. The fix makes sure the pre-condtion that theallocablock has a single successor whenever we inline delayed privatizers is respected. I simply moved the fix toallocatePrivateVarsso that it kicks in for any op not justomp.wsloop.This handles a bug uncovered by a test in the OpenMP_VV test suite.