[llvm] f97de4c - [OpenMPOpt] Move dedup runtime calls after init for target regions

[Giorgis Georgakoudis via llvm-commits]

Author: Giorgis Georgakoudis
Date: 2021-07-23T05:54:01-07:00
New Revision: f97de4cb0be43730bd91b800fa7f5418364ba3fd

URL: https://github.com/llvm/llvm-project/commit/f97de4cb0be43730bd91b800fa7f5418364ba3fd
DIFF: https://github.com/llvm/llvm-project/commit/f97de4cb0be43730bd91b800fa7f5418364ba3fd.diff

LOG: [OpenMPOpt] Move dedup runtime calls after init for target regions

Deduplication in OpenMPOpt finds redundant OpenMP runtime calls and replaces them with a single call placed in the earliest safe location in the IR. When deduplication happens in a target region this patch makes sure replacement calls are put after target_init.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D106556

Added: 
    llvm/test/Transforms/OpenMP/deduplication_target.ll

Modified: 
    llvm/lib/Transforms/IPO/OpenMPOpt.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/IPO/OpenMPOpt.cpp b/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
index 6c4f410b3871..4532ab791eec 100644
--- a/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
+++ b/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
@@ -1573,7 +1573,27 @@ struct OpenMPOpt {
           if (!CanBeMoved(*CI))
             continue;
 
-          CI->moveBefore(&*F.getEntryBlock().getFirstInsertionPt());
+          // If the function is a kernel, dedup will move
+          // the runtime call right after the kernel init callsite. Otherwise,
+          // it will move it to the beginning of the caller function.
+          if (isKernel(F)) {
+            auto &KernelInitRFI = OMPInfoCache.RFIs[OMPRTL___kmpc_target_init];
+            auto *KernelInitUV = KernelInitRFI.getUseVector(F);
+
+            if (KernelInitUV->empty())
+              continue;
+
+            assert(KernelInitUV->size() == 1 &&
+                   "Expected a single __kmpc_target_init in kernel\n");
+
+            CallInst *KernelInitCI =
+                getCallIfRegularCall(*KernelInitUV->front(), &KernelInitRFI);
+            assert(KernelInitCI &&
+                   "Expected a call to __kmpc_target_init in kernel\n");
+
+            CI->moveAfter(KernelInitCI);
+          } else
+            CI->moveBefore(&*F.getEntryBlock().getFirstInsertionPt());
           ReplVal = CI;
           break;
         }

diff  --git a/llvm/test/Transforms/OpenMP/deduplication_target.ll b/llvm/test/Transforms/OpenMP/deduplication_target.ll
new file mode 100644
index 000000000000..4b9b86559ed6
--- /dev/null
+++ b/llvm/test/Transforms/OpenMP/deduplication_target.ll
@@ -0,0 +1,66 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --scrub-attributes
+; RUN: opt -openmp-opt-cgscc -S < %s | FileCheck %s
+; RUN: opt -passes=openmp-opt-cgscc -S < %s | FileCheck %s
+
+target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64"
+
+%struct.ident_t = type { i32, i32, i32, i32, i8* }
+
+ at 0 = private unnamed_addr constant [23 x i8] c";unknown;unknown;0;0;;\00", align 1
+ at 1 = private unnamed_addr constant %struct.ident_t { i32 0, i32 2, i32 0, i32 0, i8* getelementptr inbounds ([23 x i8], [23 x i8]* @0, i32 0, i32 0) }, align 8
+ at 2 = private unnamed_addr constant %struct.ident_t { i32 0, i32 2, i32 1, i32 0, i8* getelementptr inbounds ([23 x i8], [23 x i8]* @0, i32 0, i32 0) }, align 8
+ at __omp_offloading_50_a3e09bf8_foo_l2_exec_mode = weak constant i8 0
+ at llvm.compiler.used = appending global [1 x i8*] [i8* @__omp_offloading_50_a3e09bf8_foo_l2_exec_mode], section "llvm.metadata"
+
+declare void @use(i32)
+
+define weak void @__omp_offloading_50_a3e09bf8_foo_l2() #0 {
+; CHECK-LABEL: define {{[^@]+}}@__omp_offloading_50_a3e09bf8_foo_l2
+; CHECK-SAME: () #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [0 x i8*], align 8
+; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* @[[GLOB1:[0-9]+]], i1 true, i1 false, i1 true)
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB2:[0-9]+]])
+; 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 true)
+; CHECK-NEXT:    ret void
+; CHECK:       worker.exit:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %captured_vars_addrs = alloca [0 x i8*], align 8
+  %0 = call i32 @__kmpc_target_init(%struct.ident_t* @1, i1 true, i1 false, i1 true)
+  %exec_user_code = icmp eq i32 %0, -1
+  br i1 %exec_user_code, label %user_code.entry, label %worker.exit
+
+user_code.entry:                                  ; preds = %entry
+  %1 = call i32 @__kmpc_global_thread_num(%struct.ident_t* @2)
+  %2 = call i32 @__kmpc_global_thread_num(%struct.ident_t* @2)
+  call void @__kmpc_target_deinit(%struct.ident_t* @1, i1 true, i1 true)
+  ret void
+
+worker.exit:                                      ; preds = %entry
+  ret void
+}
+
+declare i32 @__kmpc_target_init(%struct.ident_t*, i1, i1, i1)
+
+declare i32 @__kmpc_global_thread_num(%struct.ident_t*) #1
+
+declare void @__kmpc_target_deinit(%struct.ident_t*, i1, i1)
+
+attributes #0 = { convergent noinline norecurse nounwind "frame-pointer"="none" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-features"="+ptx32,+sm_20" }
+attributes #1 = { nounwind }
+
+!omp_offload.info = !{!0}
+!nvvm.annotations = !{!1}
+!llvm.module.flags = !{!2, !3, !4}
+
+!0 = !{i32 0, i32 80, i32 -1545561096, !"foo", i32 2, i32 0}
+!1 = !{void ()* @__omp_offloading_50_a3e09bf8_foo_l2, !"kernel", i32 1}
+!2 = !{i32 1, !"wchar_size", i32 4}
+!3 = !{i32 7, !"openmp", i32 50}
+!4 = !{i32 7, !"openmp-device", i32 50}