[Mlir-commits] [mlir] [OpenMP][mlir] Add translation for `if` in `omp.teams` (PR #69404)

[llvmlistbot at llvm.org]

https://github.com/shraiysh updated https://github.com/llvm/llvm-project/pull/69404

>From cf9e3e16a8b6f56e76c725b6244154b825cc8cc0 Mon Sep 17 00:00:00 2001
From: Shraiysh Vaishay <shraiysh.vaishay at amd.com>
Date: Tue, 17 Oct 2023 19:59:11 -0500
Subject: [PATCH 1/2] [OpenMP][mlir] Add translation for `if` in `omp.teams`

This patch adds translation for `if` clause on `teams` construct in
OpenMP Dialect.
---
 .../OpenMP/OpenMPToLLVMIRTranslation.cpp      |  8 +++-
 mlir/test/Target/LLVMIR/openmp-teams.mlir     | 48 +++++++++++++++++++
 2 files changed, 54 insertions(+), 2 deletions(-)

diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index e3dc68a1b8b7d72..025f102f0fbf809 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -666,7 +666,7 @@ convertOmpTeams(omp::TeamsOp op, llvm::IRBuilderBase &builder,
                 LLVM::ModuleTranslation &moduleTranslation) {
   using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy;
   LogicalResult bodyGenStatus = success();
-  if (op.getIfExpr() || !op.getAllocatorsVars().empty() || op.getReductions())
+  if (!op.getAllocatorsVars().empty() || op.getReductions())
     return op.emitError("unhandled clauses for translation to LLVM IR");
 
   auto bodyCB = [&](InsertPointTy allocaIP, InsertPointTy codegenIP) {
@@ -689,9 +689,13 @@ convertOmpTeams(omp::TeamsOp op, llvm::IRBuilderBase &builder,
   if (Value threadLimitVar = op.getThreadLimit())
     threadLimit = moduleTranslation.lookupValue(threadLimitVar);
 
+  llvm::Value *ifExpr = nullptr;
+  if (Value ifExprVar = op.getIfExpr())
+    ifExpr = moduleTranslation.lookupValue(ifExprVar);
+
   llvm::OpenMPIRBuilder::LocationDescription ompLoc(builder);
   builder.restoreIP(moduleTranslation.getOpenMPBuilder()->createTeams(
-      ompLoc, bodyCB, numTeamsLower, numTeamsUpper, threadLimit));
+      ompLoc, bodyCB, numTeamsLower, numTeamsUpper, threadLimit, ifExpr));
   return bodyGenStatus;
 }
 
diff --git a/mlir/test/Target/LLVMIR/openmp-teams.mlir b/mlir/test/Target/LLVMIR/openmp-teams.mlir
index 87ef90223ed704a..cc12dff6df1e376 100644
--- a/mlir/test/Target/LLVMIR/openmp-teams.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-teams.mlir
@@ -235,3 +235,51 @@ llvm.func @omp_teams_num_teams_and_thread_limit(%numTeamsLower: i32, %numTeamsUp
 // CHECK: define internal void [[OUTLINED_FN]](ptr {{.+}}, ptr {{.+}})
 // CHECK: call void @duringTeams()
 // CHECK: ret void
+
+// -----
+
+llvm.func @beforeTeams()
+llvm.func @duringTeams()
+llvm.func @afterTeams()
+
+// CHECK-LABEL: @teams_if
+// CHECK-SAME: (i1 [[ARG:.+]])
+llvm.func @teams_if(%arg : i1) {
+    // CHECK-NEXT: call void @beforeTeams()
+    llvm.call @beforeTeams() : () -> ()
+    // CHECK: [[NUM_TEAMS_UPPER:%.+]] = select i1 [[ARG]], i32 0, i32 1
+    // CHECK: [[NUM_TEAMS_LOWER:%.+]] = select i1 [[ARG]], i32 0, i32 1
+    // CHECK: call void @__kmpc_push_num_teams_51(ptr {{.+}}, i32 {{.+}}, i32 [[NUM_TEAMS_LOWER]], i32 [[NUM_TEAMS_UPPER]], i32 0)
+    // CHECK: call void {{.+}} @__kmpc_fork_teams({{.+}})
+    omp.teams if(%arg) {
+        llvm.call @duringTeams() : () -> ()
+        omp.terminator
+    }
+    // CHECK: call void @afterTeams()
+    llvm.call @afterTeams() : () -> ()
+    llvm.return
+}
+
+// -----
+
+llvm.func @beforeTeams()
+llvm.func @duringTeams()
+llvm.func @afterTeams()
+
+// CHECK-LABEL: @teams_if_with_num_teams
+// CHECK-SAME: (i1 [[CONDITION:.+]], i32 [[NUM_TEAMS_LOWER:.+]], i32 [[NUM_TEAMS_UPPER:.+]], i32 [[THREAD_LIMIT:.+]])
+llvm.func @teams_if_with_num_teams(%condition: i1, %numTeamsLower: i32, %numTeamsUpper: i32, %threadLimit: i32) {
+    // CHECK: call void @beforeTeams()
+    llvm.call @beforeTeams() : () -> ()
+    // CHECK: [[NUM_TEAMS_UPPER_NEW:%.+]] = select i1 [[CONDITION]], i32 [[NUM_TEAMS_UPPER]], i32 1
+    // CHECK: [[NUM_TEAMS_LOWER_NEW:%.+]] = select i1 [[CONDITION]], i32 [[NUM_TEAMS_LOWER]], i32 1
+    // CHECK: call void @__kmpc_push_num_teams_51(ptr {{.+}}, i32 {{.+}}, i32 [[NUM_TEAMS_LOWER_NEW]], i32 [[NUM_TEAMS_UPPER_NEW]], i32 [[THREAD_LIMIT]])
+    // CHECK: call void {{.+}} @__kmpc_fork_teams({{.+}})
+    omp.teams if(%condition) num_teams(%numTeamsLower: i32 to %numTeamsUpper: i32) thread_limit(%threadLimit: i32) {
+        llvm.call @duringTeams() : () -> ()
+        omp.terminator
+    }
+    // CHECK: call void @afterTeams()
+    llvm.call @afterTeams() : () -> ()
+    llvm.return
+}

>From f1487f630506bc116597eacaeba4d05bb02d3b35 Mon Sep 17 00:00:00 2001
From: Shraiysh Vaishay <shraiysh.vaishay at amd.com>
Date: Fri, 20 Oct 2023 16:54:51 -0500
Subject: [PATCH 2/2] Added comment about default values.

---
 mlir/test/Target/LLVMIR/openmp-teams.mlir | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/mlir/test/Target/LLVMIR/openmp-teams.mlir b/mlir/test/Target/LLVMIR/openmp-teams.mlir
index cc12dff6df1e376..973d2ed395eb0d0 100644
--- a/mlir/test/Target/LLVMIR/openmp-teams.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-teams.mlir
@@ -247,6 +247,9 @@ llvm.func @afterTeams()
 llvm.func @teams_if(%arg : i1) {
     // CHECK-NEXT: call void @beforeTeams()
     llvm.call @beforeTeams() : () -> ()
+    // If the condition is true, then the value of bounds is zero - which basically means "implementation-defined". 
+    // The runtime sees zero and sets a default value of number of teams. This behavior is according to the standard.
+    // The same is true for `thread_limit`.
     // CHECK: [[NUM_TEAMS_UPPER:%.+]] = select i1 [[ARG]], i32 0, i32 1
     // CHECK: [[NUM_TEAMS_LOWER:%.+]] = select i1 [[ARG]], i32 0, i32 1
     // CHECK: call void @__kmpc_push_num_teams_51(ptr {{.+}}, i32 {{.+}}, i32 [[NUM_TEAMS_LOWER]], i32 [[NUM_TEAMS_UPPER]], i32 0)