[SYCL][NVPTX] Use ptx_kernel cc instead of metadata

llvm/llvm-project@4583f6d3443c switch to use
ptx_kernel cc instead of metadata to mark kernels.

Update TargetHelpers /SYCLCreateNVVMAnnotations and tests.

Author: jsji

llvm/lib/SYCLLowerIR/SYCLCreateNVVMAnnotations.cpp

@@ -17,17 +17,18 @@
 
 using namespace llvm;
 
-static void addNVVMMetadata(GlobalValue &GV, StringRef Name, int Operand) {
-  Module *M = GV.getParent();
+static void addNVVMMetadata(Function &F, StringRef Name, int Operand) {
+  Module *M = F.getParent();
   LLVMContext &Ctx = M->getContext();
 
   // Get "nvvm.annotations" metadata node
   NamedMDNode *MD = M->getOrInsertNamedMetadata("nvvm.annotations");
 
-  Metadata *MDVals[] = {ConstantAsMetadata::get(&GV), MDString::get(Ctx, Name),
+  Metadata *MDVals[] = {ConstantAsMetadata::get(&F), MDString::get(Ctx, Name),
                         ConstantAsMetadata::get(
                             ConstantInt::get(Type::getInt32Ty(Ctx), Operand))};
   // Append metadata to nvvm.annotations
+  F.setCallingConv(CallingConv::PTX_Kernel);
   MD->addOperand(MDNode::get(Ctx, MDVals));
 }
 

llvm/lib/SYCLLowerIR/TargetHelpers.cpp

@@ -43,14 +43,7 @@ void KernelCache::handleNewCloneOf(Function &OldF, Function &NewF,
                                    bool KernelOnly) {
   assert(KernelData.contains(&OldF) && "Unknown kernel");
   if (auto &KP = KernelData[&OldF]; KP.hasAnnotations()) {
-    if (KernelOnly) {
-      // We know this is a kernel, so add a single "kernel" annotation.
-      auto &Ctx = OldF.getContext();
-      Metadata *NewKernelMD[] = {
-          ConstantAsMetadata::get(&NewF), MDString::get(Ctx, "kernel"),
-          ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(Ctx), 1))};
-      KP.ModuleAnnotationsMD->addOperand(MDNode::get(Ctx, NewKernelMD));
-    } else {
+    if (!KernelOnly) {
       // Otherwise we'll need to clone all metadata, possibly dropping ones
       // which we can't assume are safe to clone.
       llvm_unreachable("Unimplemented cloning logic");
@@ -73,12 +66,16 @@ void KernelCache::populateKernels(Module &M) {
     return;
   }
 
-  // NVPTX kernels are identified by the global annotations metadata.
+  // NVPTX kernels are identified by their calling convention, and may have
+  // annotations.
   if (T.isNVPTX()) {
-    // Access `nvvm.annotations` to determine which functions are kernel
-    // entry points.
     auto *AnnotationMetadata = M.getNamedMetadata("nvvm.annotations");
-    // No kernels in the module, early exit.
+    for (auto &F : M) {
+      if (F.getCallingConv() == CallingConv::PTX_Kernel) {
+        Kernels.push_back(&F);
+	KernelData[&F] = KernelPayload{AnnotationMetadata};
+      }
+    }
     if (!AnnotationMetadata)
       return;
 
@@ -93,36 +90,17 @@ void KernelCache::populateKernels(Module &M) {
 
       Visited.insert(MDN);
 
-      // Kernel entry points are identified using metadata nodes of the form:
-      //   !X = !{<function>[, !"kind", i32 X]+}
-      // Where "kind" == "kernel" and X == 1.
-      bool IsKernel = false;
-      for (size_t I = 1, E = MDN->getNumOperands() - 1; I < E && !IsKernel;
-           I += 2) {
-        if (auto *Type = dyn_cast<MDString>(MDN->getOperand(I)))
-          if (Type->getString() == "kernel") {
-            if (auto *Val =
-                    mdconst::dyn_extract<ConstantInt>(MDN->getOperand(I + 1)))
-              IsKernel = Val->getZExtValue() == 1;
-          }
-      }
-
       // Get a pointer to the entry point function from the metadata.
       const MDOperand &FuncOperand = MDN->getOperand(0);
       if (!FuncOperand)
         continue;
 
-      if (auto *FuncConstant = dyn_cast<ConstantAsMetadata>(FuncOperand)) {
-        if (auto *Func = dyn_cast<Function>(FuncConstant->getValue())) {
-          if (IsKernel && !KernelData.contains(Func)) {
-            Kernels.push_back(Func);
-            KernelData[Func] = KernelPayload{AnnotationMetadata};
-          }
-          DependentMDNodes[Func].push_back(MDN);
-        }
-      }
+      if (auto *FuncConstant = dyn_cast<ConstantAsMetadata>(FuncOperand))
+        if (auto *Func = dyn_cast<Function>(FuncConstant->getValue()))
+          if (Func->getCallingConv() == CallingConv::PTX_Kernel)
+             DependentMDNodes[Func].push_back(MDN);
     }
-
+      
     // We need to match non-kernel metadata nodes using the kernel name to the
     // kernel nodes. To avoid checking matched nodes multiple times keep track
     // of handled entries.

llvm/test/CodeGen/NVPTX/global-offset-annotations.ll

@@ -3,10 +3,7 @@
 target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"
 target triple = "nvptx64-nvidia-cuda"
 
-; This test checks that the transformation is applied to kernels found using
-; less common annotation formats, and that annotations are correctly updated.
-; We don't currently know it's safe to clone all metadata, so only add a
-; "kernel" annotation and leave others in place.
+; This test checks that the transformation is applied to kernels found 
 
 declare ptr @llvm.nvvm.implicit.offset()
 ; CHECK-NOT: llvm.nvvm.implicit.offset
@@ -25,14 +22,13 @@ entry:
   ret void
 }
 
-; CHECK: !nvvm.annotations = !{![[OLDMD0:[0-9]+]], ![[OLDMD1:[0-9]+]], ![[OLDMD1]], ![[OLDMD0]], ![[NEWKERNELMD:[0-9]+]]}
+; CHECK: !nvvm.annotations = !{![[OLDMD0:[0-9]+]], ![[OLDMD1:[0-9]+]], ![[OLDMD1]], ![[OLDMD0]]}
 
 !llvm.module.flags = !{!0}
 !nvvm.annotations = !{!1, !2, !2, !1}
 
 ; CHECK: ![[OLDMD0]] = distinct !{ptr @_ZTS14example_kernel, !"maxnreg", i32 256, !"kernel", i32 1}
 ; CHECK: ![[OLDMD1]] = !{ptr @_ZTS14example_kernel, !"maxntidx", i32 8, !"maxntidy", i32 16, !"maxntidz", i32 32}
-; CHECK: ![[NEWKERNELMD]] = !{ptr @_ZTS14example_kernel_with_offset, !"kernel", i32 1}
 
 !0 = !{i32 1, !"sycl-device", i32 1}
 !1 = distinct !{ptr @_ZTS14example_kernel, !"maxnreg", i32 256, !"kernel", i32 1}

llvm/test/CodeGen/NVPTX/global-offset-dbg.ll

@@ -8,41 +8,40 @@ target triple = "nvptx64-nvidia-cuda"
 
 declare ptr @llvm.nvvm.implicit.offset()
 
-define i64 @_ZTS14other_function() !dbg !11 {
+define i64 @_ZTS14other_function() !dbg !10 {
   %1 = tail call ptr @llvm.nvvm.implicit.offset()
   %2 = getelementptr inbounds i32, ptr %1, i64 2
   %3 = load i32, ptr %2, align 4
   %4 = zext i32 %3 to i64
   ret i64 %4
 }
 
-define void @_ZTS14example_kernel() !dbg !14 {
+define ptx_kernel void @_ZTS14example_kernel() !dbg !13 {
 entry:
-  %0 = call i64 @_ZTS14other_function(), !dbg !15
+  %0 = call i64 @_ZTS14other_function(), !dbg !14
   ret void
 }
 
 !llvm.dbg.cu = !{!0}
-!llvm.module.flags = !{!3, !4, !10}
-!nvvm.annotations = !{!5, !6, !7, !6, !8, !8, !8, !8, !9, !9, !8}
-!nvvmir.version = !{!6}
+!llvm.module.flags = !{!3, !4, !9}
+!nvvm.annotations = !{!5, !6, !5, !7, !7, !7, !7, !8, !8, !7}
+!nvvmir.version = !{!5}
 
 !0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !1, producer: "clang version 0.0.0", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !2, nameTableKind: None)
 !1 = !DIFile(filename: "global-offset-debug.cpp", directory: "/tmp")
 !2 = !{}
 !3 = !{i32 2, !"Dwarf Version", i32 4}
 !4 = !{i32 2, !"Debug Info Version", i32 3}
-!5 = distinct !{ptr @_ZTS14example_kernel, !"kernel", i32 1}
-!6 = !{i32 1, i32 4}
-!7 = !{null, !"align", i32 8, !"align", i32 65544, !"align", i32 131080}
-!8 = !{null, !"align", i32 16}
-!9 = !{null, !"align", i32 16, !"align", i32 65552, !"align", i32 131088}
-!10 = !{i32 1, !"sycl-device", i32 1}
-!11 = distinct !DISubprogram(name: "other_function", scope: !1, file: !1, line: 3, type: !12, scopeLine: 3, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition, unit: !0, retainedNodes: !2)
-!12 = !DISubroutineType(types: !13)
-!13 = !{null}
-!14 = distinct !DISubprogram(name: "example_kernel", scope: !1, file: !1, line: 10, type: !12, scopeLine: 10, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition, unit: !0, retainedNodes: !2)
-!15 = !DILocation(line: 1, column: 2, scope: !14)
+!5 = !{i32 1, i32 4}
+!6 = !{null, !"align", i32 8, !"align", i32 65544, !"align", i32 131080}
+!7 = !{null, !"align", i32 16}
+!8 = !{null, !"align", i32 16, !"align", i32 65552, !"align", i32 131088}
+!9 = !{i32 1, !"sycl-device", i32 1}
+!10 = distinct !DISubprogram(name: "other_function", scope: !1, file: !1, line: 3, type: !11, scopeLine: 3, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition, unit: !0, retainedNodes: !2)
+!11 = !DISubroutineType(types: !12)
+!12 = !{null}
+!13 = distinct !DISubprogram(name: "example_kernel", scope: !1, file: !1, line: 10, type: !11, scopeLine: 10, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition, unit: !0, retainedNodes: !2)
+!14 = !DILocation(line: 1, column: 2, scope: !13)
 ; CHECK-LABEL: define i64 @_ZTS14other_function(
 ; CHECK-SAME: ) !dbg [[DBG12:![0-9]+]] {
 ; CHECK-NEXT:    [[TMP1:%.*]] = zext i32 0 to i64
@@ -57,14 +56,14 @@ entry:
 ; CHECK-NEXT:    ret i64 [[TMP4]]
 ;
 ;
-; CHECK-LABEL: define void @_ZTS14example_kernel(
+; CHECK-LABEL: define ptx_kernel void @_ZTS14example_kernel(
 ; CHECK-SAME: ) !dbg [[DBG16:![0-9]+]] {
 ; CHECK-NEXT:  [[ENTRY:.*:]]
 ; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @_ZTS14other_function(), !dbg [[DBG17:![0-9]+]]
 ; CHECK-NEXT:    ret void
 ;
 ;
-; CHECK-LABEL: define void @_ZTS14example_kernel_with_offset(
+; CHECK-LABEL: define ptx_kernel void @_ZTS14example_kernel_with_offset(
 ; CHECK-SAME: ptr byval([3 x i32]) [[TMP0:%.*]]) !dbg [[DBG18:![0-9]+]] {
 ; CHECK-NEXT:  [[ENTRY:.*:]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = call i64 @_ZTS14other_function_with_offset(ptr [[TMP0]]), !dbg [[DBG19:![0-9]+]]
@@ -77,12 +76,10 @@ entry:
 ; CHECK: [[META3:![0-9]+]] = !{i32 2, !"Dwarf Version", i32 4}
 ; CHECK: [[META4:![0-9]+]] = !{i32 2, !"Debug Info Version", i32 3}
 ; CHECK: [[META5:![0-9]+]] = !{i32 1, !"sycl-device", i32 1}
-; CHECK: [[META6:![0-9]+]] = distinct !{ptr @_ZTS14example_kernel, !"kernel", i32 1}
 ; CHECK: [[META7:![0-9]+]] = !{i32 1, i32 4}
 ; CHECK: [[META8:![0-9]+]] = !{null, !"align", i32 8, !"align", i32 65544, !"align", i32 131080}
 ; CHECK: [[META9:![0-9]+]] = !{null, !"align", i32 16}
 ; CHECK: [[META10:![0-9]+]] = !{null, !"align", i32 16, !"align", i32 65552, !"align", i32 131088}
-; CHECK: [[META11:![0-9]+]] = !{ptr @_ZTS14example_kernel_with_offset, !"kernel", i32 1}
 ; CHECK: [[DBG12]] = distinct !DISubprogram(name: "other_function", scope: [[META1]], file: [[META1]], line: 3, type: [[META13:![0-9]+]], scopeLine: 3, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition, unit: [[META0]], retainedNodes: [[META2]])
 ; CHECK: [[META13]] = !DISubroutineType(types: [[META14:![0-9]+]])
 ; CHECK: [[META14]] = !{null}

llvm/test/CodeGen/NVPTX/global-offset-intrinsic-function-mix.ll

@@ -47,28 +47,27 @@ define i64 @_ZTS14mixed_function() {
 ; CHECK: %5 = zext i32 %4 to i64
 ; CHECK: ret i64 %2
 
-define void @_ZTS12some_kernel() {
+define ptx_kernel void @_ZTS12some_kernel() {
 entry:
   %0 = call i64 @_ZTS14mixed_function()
 ; CHECK: %0 = call i64 @_ZTS14mixed_function()
   ret void
 }
 
-; CHECK: define void @_ZTS12some_kernel_with_offset(ptr byval([3 x i32]) %0) {
+; Check cc in our new kernel
+; CHECK: define ptx_kernel void @_ZTS12some_kernel_with_offset(ptr byval([3 x i32]) %0) {
 ; CHECK: entry:
 ; CHECK:   %1 = call i64 @_ZTS14mixed_function_with_offset(ptr %0)
 ; CHECK:   ret void
 ; CHECK: }
 
-; Check the last annotation is our new kernel
-; CHECK: !nvvm.annotations = {{.*}}, ![[NEWKERNELMD:[0-9]+]]}
+; CHECK: !nvvm.annotations = {{.*}}
 
 !llvm.module.flags = !{!6}
 !nvvm.annotations = !{!0, !1, !2, !1, !3, !3, !3, !3, !4, !4, !3, !5}
 !nvvmir.version = !{!5}
 
-; CHECK: ![[NEWKERNELMD]] = !{ptr @_ZTS12some_kernel_with_offset, !"kernel", i32 1}
-!0 = distinct !{ptr @_ZTS12some_kernel, !"kernel", i32 1}
+!0 = distinct !{ptr @_ZTS12some_kernel, !"dummy", i32 1}
 !1 = !{null, !"align", i32 8}
 !2 = !{null, !"align", i32 8, !"align", i32 65544, !"align", i32 131080}
 !3 = !{null, !"align", i32 16}

llvm/test/CodeGen/NVPTX/global-offset-multiple-calls-from-one-function.ll

@@ -43,7 +43,7 @@ define i64 @_ZTS14other_function() {
 ; CHECK: ret i64 %4
 ; CHECK: }
 
-define void @_ZTS14example_kernel() {
+define ptx_kernel void @_ZTS14example_kernel() {
 entry:
   %0 = call i64 @_ZTS14other_function()
 ; CHECK: %0 = call i64 @_ZTS14other_function()
@@ -52,22 +52,20 @@ entry:
   ret void
 }
 
-; CHECK: define void @_ZTS14example_kernel_with_offset(ptr byval([3 x i32]) %0) {
+; CHECK: define ptx_kernel void @_ZTS14example_kernel_with_offset(ptr byval([3 x i32]) %0) {
 ; CHECK: entry:
 ; CHECK:   %1 = call i64 @_ZTS14other_function_with_offset(ptr %0)
 ; CHECK:   %2 = call i64 @_ZTS14other_function_with_offset(ptr %0)
 ; CHECK:   ret void
 ; CHECK: }
 
-; Check the last annotation is our new kernel
-; CHECK: !nvvm.annotations = {{.*}}, ![[NEWKERNELMD:[0-9]+]]}
+; CHECK: !nvvm.annotations = {{.*}}
 
 !llvm.module.flags = !{!7}
 !nvvm.annotations = !{!0, !1, !2, !1, !3, !3, !3, !3, !4, !4, !3}
 !nvvmir.version = !{!6}
 
-; CHECK: ![[NEWKERNELMD]] = !{ptr @_ZTS14example_kernel_with_offset, !"kernel", i32 1}
-!0 = distinct !{ptr @_ZTS14example_kernel, !"kernel", i32 1}
+!0 = distinct !{ptr @_ZTS14example_kernel, !"dummy", i32 1}
 !1 = !{null, !"align", i32 8}
 !2 = !{null, !"align", i32 8, !"align", i32 65544, !"align", i32 131080}
 !3 = !{null, !"align", i32 16}

llvm/test/CodeGen/NVPTX/global-offset-multiple-entry-points.ll

@@ -11,7 +11,7 @@ declare ptr @llvm.nvvm.implicit.offset()
 
 ; This function is a kernel entry point that does not use global offset. It will
 ; not get a clone with a global offset parameter.
-define dso_local void @_ZTS12third_kernel() {
+define dso_local ptx_kernel void @_ZTS12third_kernel() {
 entry:
   ret void
 }
@@ -45,14 +45,14 @@ define i64 @_ZTS14first_function() {
 ; CHECK: %2 = call i64 @_ZTS15common_function_with_offset(ptr %0)
 ; CHECK: ret i64 %2
 
-define void @_ZTS12first_kernel() {
+define ptx_kernel void @_ZTS12first_kernel() {
 entry:
   %0 = call i64 @_ZTS14first_function()
 ; CHECK: %0 = call i64 @_ZTS14first_function()
   ret void
 }
 
-; CHECK: define void @_ZTS12first_kernel_with_offset(ptr byval([3 x i32]) %0) {
+; CHECK: define ptx_kernel void @_ZTS12first_kernel_with_offset(ptr byval([3 x i32]) %0) {
 ; CHECK: entry:
 ; CHECK:   %1 = call i64 @_ZTS14first_function_with_offset(ptr %0)
 ; CHECK:   ret void
@@ -69,14 +69,14 @@ define i64 @_ZTS15second_function() {
 ; CHECK: %2 = call i64 @_ZTS15common_function_with_offset(ptr %0)
 ; CHECK: ret i64 %2
 
-define void @_ZTS13second_kernel() {
+define ptx_kernel void @_ZTS13second_kernel() {
 entry:
   %0 = call i64 @_ZTS15second_function()
 ; CHECK: %0 = call i64 @_ZTS15second_function()
   ret void
 }
 
-; CHECK: define void @_ZTS13second_kernel_with_offset(ptr byval([3 x i32]) %0) {
+; CHECK: define ptx_kernel void @_ZTS13second_kernel_with_offset(ptr byval([3 x i32]) %0) {
 ; CHECK: entry:
 ; CHECK:   %1 = call i64 @_ZTS15second_function_with_offset(ptr %0)
 ; CHECK:   ret void
@@ -102,21 +102,18 @@ define i64 @_ZTS15no_entry_point() {
 ; CHECK: }
 
 ; Check the last two annotations are our new kernels
-; CHECK: !nvvm.annotations = {{.*}}, ![[NEWKERNEL0MD:[0-9]+]], ![[NEWKERNEL1MD:[0-9]+]]}
+; CHECK: !nvvm.annotations = {{.*}}
 
 !llvm.module.flags = !{!10}
 !nvvm.annotations = !{!0, !1, !2, !1, !3, !3, !3, !3, !4, !4, !3, !5, !6}
 !nvvmir.version = !{!9}
 
-; CHECK: ![[NEWKERNEL0MD]] = !{ptr @_ZTS13second_kernel_with_offset, !"kernel", i32 1}
-; CHECK: ![[NEWKERNEL1MD]] = !{ptr @_ZTS12first_kernel_with_offset, !"kernel", i32 1}
-
-!0 = distinct !{ptr @_ZTS12first_kernel, !"kernel", i32 1}
+!0 = distinct !{ptr @_ZTS12first_kernel, !"dummy", i32 1}
 !1 = !{null, !"align", i32 8}
 !2 = !{null, !"align", i32 8, !"align", i32 65544, !"align", i32 131080}
 !3 = !{null, !"align", i32 16}
 !4 = !{null, !"align", i32 16, !"align", i32 65552, !"align", i32 131088}
-!5 = distinct !{ptr @_ZTS13second_kernel, !"kernel", i32 1}
-!6 = distinct !{ptr @_ZTS12third_kernel, !"kernel", i32 1}
+!5 = distinct !{ptr @_ZTS13second_kernel, !"dummy", i32 1}
+!6 = distinct !{ptr @_ZTS12third_kernel, !"dummy", i32 1}
 !9 = !{i32 1, i32 4}
 !10 = !{i32 1, !"sycl-device", i32 1}

llvm/test/CodeGen/NVPTX/global-offset-simple.ll

@@ -26,28 +26,26 @@ define i64 @_ZTS14other_function() {
 ; CHECK: ret i64 %4
 ; CHECK: }
 
-define void @_ZTS14example_kernel() {
+define ptx_kernel void @_ZTS14example_kernel() {
 entry:
   %0 = call i64 @_ZTS14other_function()
 ; CHECK: %0 = call i64 @_ZTS14other_function()
   ret void
 }
 
-; CHECK: define void @_ZTS14example_kernel_with_offset(ptr byval([3 x i32]) %0) {
+; CHECK: define ptx_kernel void @_ZTS14example_kernel_with_offset(ptr byval([3 x i32]) %0) {
 ; CHECK: entry:
 ; CHECK:   %1 = call i64 @_ZTS14other_function_with_offset(ptr %0)
 ; CHECK:   ret void
 ; CHECK: }
 
-; Check the last annotation is our new kernel
-; CHECK: !nvvm.annotations = {{.*}}, ![[NEWKERNELMD:[0-9]+]]}
+; CHECK: !nvvm.annotations = {{.*}}
 
 !llvm.module.flags = !{!7}
 !nvvm.annotations = !{!0, !1, !2, !1, !3, !3, !3, !3, !4, !4, !3}
 !nvvmir.version = !{!6}
 
-; CHECK: ![[NEWKERNELMD]] = !{ptr @_ZTS14example_kernel_with_offset, !"kernel", i32 1}
-!0 = distinct !{ptr @_ZTS14example_kernel, !"kernel", i32 1}
+!0 = distinct !{ptr @_ZTS14example_kernel, !"dummy", i32 1}
 !1 = !{null, !"align", i32 8}
 !2 = !{null, !"align", i32 8, !"align", i32 65544, !"align", i32 131080}
 !3 = !{null, !"align", i32 16}