[OpenMP] Do not SPMDize generic regions with no parallel

This patch changes SPMDization to not trigger for regions with no parallelism. Otherwise, this will introduce unnecessary barriers that will slow the single-threaded region down. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D109438
llvm · Sep 8, 2021 · 6b9a3ec · 6b9a3ec
1 parent e567356
commit 6b9a3ec
Show file tree

Hide file tree

Showing 4 changed files with 108 additions and 12 deletions.
diff --git a/llvm/lib/Transforms/IPO/OpenMPOpt.cpp b/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
@@ -599,6 +599,12 @@ struct KernelInfoState : AbstractState {
     return true;
   }
 
+  /// Returns true if this kernel contains any OpenMP parallel regions.
+  bool mayContainParallelRegion() {
+    return !ReachedKnownParallelRegions.empty() ||
+           !ReachedUnknownParallelRegions.empty();
+  }
+
   /// Return empty set as the best state of potential values.
   static KernelInfoState getBestState() { return KernelInfoState(true); }
 
@@ -3003,7 +3009,7 @@ struct AAKernelInfoFunction : AAKernelInfo {
 
     // If we can we change the execution mode to SPMD-mode otherwise we build a
     // custom state machine.
-    if (!changeToSPMDMode(A))
+    if (!mayContainParallelRegion() || !changeToSPMDMode(A))
       buildCustomStateMachine(A);
 
     return ChangeStatus::CHANGED;
@@ -3308,8 +3314,7 @@ struct AAKernelInfoFunction : AAKernelInfo {
     // happen if there simply are no parallel regions. In the resulting kernel
     // all worker threads will simply exit right away, leaving the main thread
     // to do the work alone.
-    if (ReachedKnownParallelRegions.empty() &&
-        ReachedUnknownParallelRegions.empty()) {
+    if (!mayContainParallelRegion()) {
       ++NumOpenMPTargetRegionKernelsWithoutStateMachine;
 
       auto Remark = [&](OptimizationRemark OR) {

diff --git a/llvm/test/Transforms/OpenMP/always_inline_device.ll b/llvm/test/Transforms/OpenMP/always_inline_device.ll
@@ -12,11 +12,11 @@ define weak void @__omp_offloading_fd02_c0934fc2_foo_l4() #0 {
 ; CHECK: Function Attrs: convergent norecurse nounwind
 ; CHECK-LABEL: @__omp_offloading_fd02_c0934fc2_foo_l4(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* @[[GLOB1:[0-9]+]], i1 true, i1 false, i1 false)
+; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* @[[GLOB1:[0-9]+]], i1 false, i1 false, i1 true)
 ; CHECK-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1
 ; CHECK-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]
 ; CHECK:       user_code.entry:
-; CHECK-NEXT:    call void @__kmpc_target_deinit(%struct.ident_t* @[[GLOB1]], i1 true, i1 false)
+; CHECK-NEXT:    call void @__kmpc_target_deinit(%struct.ident_t* @[[GLOB1]], i1 false, i1 true)
 ; CHECK-NEXT:    ret void
 ; CHECK:       worker.exit:
 ; CHECK-NEXT:    ret void

diff --git a/llvm/test/Transforms/OpenMP/get_hardware_num_threads_in_block_fold.ll b/llvm/test/Transforms/OpenMP/get_hardware_num_threads_in_block_fold.ll
@@ -54,18 +54,40 @@ define weak void @kernel1() #0 {
 define weak void @kernel2() #0 {
 ; CHECK-LABEL: define {{[^@]+}}@kernel2
 ; CHECK-SAME: () #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [0 x i8*], align 8
 ; CHECK-NEXT:    [[I:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* null, i1 true, i1 false, i1 false)
+; CHECK-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[I]], -1
+; CHECK-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[COMMON_RET:%.*]]
+; CHECK:       common.ret:
+; CHECK-NEXT:    ret void
+; CHECK:       user_code.entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* null) #[[ATTR1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast [0 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8**
 ; CHECK-NEXT:    call void @helper0() #[[ATTR1]]
 ; CHECK-NEXT:    call void @helper1() #[[ATTR1]]
 ; CHECK-NEXT:    call void @helper2() #[[ATTR1]]
+; CHECK-NEXT:    call void @__kmpc_parallel_51(%struct.ident_t* null, i32 [[TMP0]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*)* @__omp_outlined__ to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined___wrapper to i8*), i8** [[TMP1]], i64 0)
 ; CHECK-NEXT:    call void @__kmpc_target_deinit(%struct.ident_t* null, i1 true, i1 false)
 ; CHECK-NEXT:    ret void
 ;
-  %i = call i32 @__kmpc_target_init(%struct.ident_t* null, i1 false, i1 false, i1 false)
+entry:
+  %captured_vars_addrs = alloca [0 x i8*], align 8
+  %i = call i32 @__kmpc_target_init(%struct.ident_t* null, i1 false, i1 true, i1 true)
+  %exec_user_code = icmp eq i32 %i, -1
+  br i1 %exec_user_code, label %user_code.entry, label %common.ret
+
+common.ret:
+  ret void
+
+user_code.entry:
+  %0 = call i32 @__kmpc_global_thread_num(%struct.ident_t* null)
+  %1 = bitcast [0 x i8*]* %captured_vars_addrs to i8**
   call void @helper0()
   call void @helper1()
   call void @helper2()
-  call void @__kmpc_target_deinit(%struct.ident_t* null, i1 false, i1 false)
+  call void @__kmpc_parallel_51(%struct.ident_t* null, i32 %0, i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*)* @__omp_outlined__ to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined___wrapper to i8*), i8** %1, i64 0)
+  call void @__kmpc_target_deinit(%struct.ident_t* null, i1 false, i1 true)
   ret void
 }
 
@@ -136,9 +158,31 @@ define internal void @helper2() {
   ret void
 }
 
+define internal void @__omp_outlined__(i32* noalias %.global_tid., i32* noalias %.bound_tid.) {
+; CHECK-LABEL: define {{[^@]+}}@__omp_outlined__
+; CHECK-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    ret void
+;
+entry:
+  ret void
+}
+
+define internal void @__omp_outlined___wrapper(i16 zeroext %0, i32 %1) {
+; CHECK-LABEL: define {{[^@]+}}@__omp_outlined___wrapper
+; CHECK-SAME: (i16 zeroext [[TMP0:%.*]], i32 [[TMP1:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    ret void
+;
+entry:
+  ret void
+}
+
 declare i32 @__kmpc_get_hardware_num_threads_in_block()
 declare i32 @__kmpc_target_init(%struct.ident_t*, i1 zeroext, i1 zeroext, i1 zeroext) #1
 declare void @__kmpc_target_deinit(%struct.ident_t* nocapture readnone, i1 zeroext, i1 zeroext) #1
+declare void @__kmpc_parallel_51(%struct.ident_t*, i32, i32, i32, i32, i8*, i8*, i8**, i64)
+declare i32 @__kmpc_global_thread_num(%struct.ident_t*)
 
 
 !llvm.module.flags = !{!0, !1}
@@ -155,7 +199,8 @@ attributes #0 = { "omp_target_thread_limit"="666" "omp_target_num_teams"="777"}
 ;.
 ; CHECK: attributes #[[ATTR0]] = { "omp_target_num_teams"="777" "omp_target_thread_limit"="666" }
 ; CHECK: attributes #[[ATTR1]] = { nounwind }
-; CHECK: attributes #[[ATTR2:[0-9]+]] = { convergent nounwind }
+; CHECK: attributes #[[ATTR2:[0-9]+]] = { alwaysinline }
+; CHECK: attributes #[[ATTR3:[0-9]+]] = { convergent nounwind }
 ;.
 ; CHECK: [[META0:![0-9]+]] = !{i32 7, !"openmp", i32 50}
 ; CHECK: [[META1:![0-9]+]] = !{i32 7, !"openmp-device", i32 50}

diff --git a/llvm/test/Transforms/OpenMP/is_spmd_exec_mode_fold.ll b/llvm/test/Transforms/OpenMP/is_spmd_exec_mode_fold.ll
@@ -38,14 +38,36 @@ define weak void @is_spmd() {
 
 define weak void @will_be_spmd() {
 ; CHECK-LABEL: define {{[^@]+}}@will_be_spmd() {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [0 x i8*], align 8
 ; CHECK-NEXT:    [[I:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* null, i1 true, i1 false, i1 false)
+; CHECK-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[I]], -1
+; CHECK-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[COMMON_RET:%.*]]
+; CHECK:       common.ret:
+; CHECK-NEXT:    ret void
+; CHECK:       user_code.entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* null) #[[ATTR2:[0-9]+]]
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast [0 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8**
 ; CHECK-NEXT:    call void @is_spmd_helper2()
+; CHECK-NEXT:    call void @__kmpc_parallel_51(%struct.ident_t* null, i32 [[TMP0]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*)* @__omp_outlined__ to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined___wrapper to i8*), i8** [[TMP1]], i64 0)
 ; CHECK-NEXT:    call void @__kmpc_target_deinit(%struct.ident_t* null, i1 true, i1 false)
 ; CHECK-NEXT:    ret void
 ;
-  %i = call i32 @__kmpc_target_init(%struct.ident_t* null, i1 false, i1 false, i1 false)
+entry:
+  %captured_vars_addrs = alloca [0 x i8*], align 8
+  %i = call i32 @__kmpc_target_init(%struct.ident_t* null, i1 false, i1 true, i1 true)
+  %exec_user_code = icmp eq i32 %i, -1
+  br i1 %exec_user_code, label %user_code.entry, label %common.ret
+
+common.ret:
+  ret void
+
+user_code.entry:
+  %0 = call i32 @__kmpc_global_thread_num(%struct.ident_t* null)
+  %1 = bitcast [0 x i8*]* %captured_vars_addrs to i8**
   call void @is_spmd_helper2()
-  call void @__kmpc_target_deinit(%struct.ident_t* null, i1 false, i1 false)
+  call void @__kmpc_parallel_51(%struct.ident_t* null, i32 %0, i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*)* @__omp_outlined__ to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined___wrapper to i8*), i8** %1, i64 0)
+  call void @__kmpc_target_deinit(%struct.ident_t* null, i1 false, i1 true)
   ret void
 }
 
@@ -153,10 +175,32 @@ define internal void @is_mixed_helper() {
   ret void
 }
 
+define internal void @__omp_outlined__(i32* noalias %.global_tid., i32* noalias %.bound_tid.) {
+; CHECK-LABEL: define {{[^@]+}}@__omp_outlined__
+; CHECK-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    ret void
+;
+entry:
+  ret void
+}
+
+define internal void @__omp_outlined___wrapper(i16 zeroext %0, i32 %1) {
+; CHECK-LABEL: define {{[^@]+}}@__omp_outlined___wrapper
+; CHECK-SAME: (i16 zeroext [[TMP0:%.*]], i32 [[TMP1:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    ret void
+;
+entry:
+  ret void
+}
+
 declare void @spmd_compatible() "llvm.assume"="ompx_spmd_amenable"
 declare i8 @__kmpc_is_spmd_exec_mode()
-declare i32 @__kmpc_target_init(%struct.ident_t*, i1 zeroext, i1 zeroext, i1 zeroext) #1
-declare void @__kmpc_target_deinit(%struct.ident_t* nocapture readnone, i1 zeroext, i1 zeroext) #1
+declare i32 @__kmpc_target_init(%struct.ident_t*, i1 zeroext, i1 zeroext, i1 zeroext)
+declare void @__kmpc_target_deinit(%struct.ident_t* nocapture readnone, i1 zeroext, i1 zeroext)
+declare void @__kmpc_parallel_51(%struct.ident_t*, i32, i32, i32, i32, i8*, i8*, i8**, i64)
+declare i32 @__kmpc_global_thread_num(%struct.ident_t*)
 declare void @foo()
 declare void @bar()
 
@@ -171,6 +215,8 @@ declare void @bar()
 !5 = !{void ()* @will_not_be_spmd, !"kernel", i32 1}
 ;.
 ; CHECK: attributes #[[ATTR0:[0-9]+]] = { "llvm.assume"="ompx_spmd_amenable" }
+; CHECK: attributes #[[ATTR1:[0-9]+]] = { alwaysinline }
+; CHECK: attributes #[[ATTR2]] = { nounwind }
 ;.
 ; CHECK: [[META0:![0-9]+]] = !{i32 7, !"openmp", i32 50}
 ; CHECK: [[META1:![0-9]+]] = !{i32 7, !"openmp-device", i32 50}