[AMDGPU] fix amdgpu_max_num_work_groups in templates (#141633)

Clang does not instantiate amdgpu_max_num_work_groups attribute with one
template argument, causing
assertion codegen.

Fixes: #139570

Author: yxsamliu

clang/lib/Sema/SemaTemplateInstantiateDecl.cpp

@@ -648,21 +648,30 @@ static void instantiateDependentAMDGPUMaxNumWorkGroupsAttr(
   EnterExpressionEvaluationContext Unevaluated(
       S, Sema::ExpressionEvaluationContext::ConstantEvaluated);
 
-  ExprResult ResultX = S.SubstExpr(Attr.getMaxNumWorkGroupsX(), TemplateArgs);
-  if (!ResultX.isUsable())
-    return;
-  ExprResult ResultY = S.SubstExpr(Attr.getMaxNumWorkGroupsY(), TemplateArgs);
-  if (!ResultY.isUsable())
-    return;
-  ExprResult ResultZ = S.SubstExpr(Attr.getMaxNumWorkGroupsZ(), TemplateArgs);
-  if (!ResultZ.isUsable())
-    return;
+  Expr *XExpr = nullptr;
+  Expr *YExpr = nullptr;
+  Expr *ZExpr = nullptr;
+
+  if (Attr.getMaxNumWorkGroupsX()) {
+    ExprResult ResultX = S.SubstExpr(Attr.getMaxNumWorkGroupsX(), TemplateArgs);
+    if (ResultX.isUsable())
+      XExpr = ResultX.getAs<Expr>();
+  }
+
+  if (Attr.getMaxNumWorkGroupsY()) {
+    ExprResult ResultY = S.SubstExpr(Attr.getMaxNumWorkGroupsY(), TemplateArgs);
+    if (ResultY.isUsable())
+      YExpr = ResultY.getAs<Expr>();
+  }
 
-  Expr *XExpr = ResultX.getAs<Expr>();
-  Expr *YExpr = ResultY.getAs<Expr>();
-  Expr *ZExpr = ResultZ.getAs<Expr>();
+  if (Attr.getMaxNumWorkGroupsZ()) {
+    ExprResult ResultZ = S.SubstExpr(Attr.getMaxNumWorkGroupsZ(), TemplateArgs);
+    if (ResultZ.isUsable())
+      ZExpr = ResultZ.getAs<Expr>();
+  }
 
-  S.AMDGPU().addAMDGPUMaxNumWorkGroupsAttr(New, Attr, XExpr, YExpr, ZExpr);
+  if (XExpr)
+    S.AMDGPU().addAMDGPUMaxNumWorkGroupsAttr(New, Attr, XExpr, YExpr, ZExpr);
 }
 
 // This doesn't take any template parameters, but we have a custom action that

clang/test/CodeGenCUDA/amdgpu-kernel-attrs.cu

@@ -78,6 +78,24 @@ __global__ void template_32_4_a_max_num_work_groups() {}
 template __global__ void template_32_4_a_max_num_work_groups<2>();
 // CHECK: define{{.*}} amdgpu_kernel void @_Z35template_32_4_a_max_num_work_groupsILj2EEvv() [[MAX_NUM_WORK_GROUPS_32_4_2:#[0-9]+]]
 
+template<unsigned a>
+__attribute__((amdgpu_max_num_work_groups(a)))
+__global__ void template_a_max_num_work_groups() {}
+template __global__ void template_a_max_num_work_groups<32>();
+// CHECK: define{{.*}} amdgpu_kernel void @_Z30template_a_max_num_work_groupsILj32EEvv() [[MAX_NUM_WORK_GROUPS_32_1_1]]
+
+template<unsigned a, unsigned b>
+__attribute__((amdgpu_max_num_work_groups(a, b)))
+__global__ void template_a_b_max_num_work_groups() {}
+template __global__ void template_a_b_max_num_work_groups<32, 1>();
+// CHECK: define{{.*}} amdgpu_kernel void @_Z32template_a_b_max_num_work_groupsILj32ELj1EEvv() [[MAX_NUM_WORK_GROUPS_32_1_1]]
+
+template<unsigned a, unsigned b, unsigned c>
+__attribute__((amdgpu_max_num_work_groups(a, b, c)))
+__global__ void template_a_b_c_max_num_work_groups() {}
+template __global__ void template_a_b_c_max_num_work_groups<32, 4, 2>();
+// CHECK: define{{.*}} amdgpu_kernel void @_Z34template_a_b_c_max_num_work_groupsILj32ELj4ELj2EEvv() [[MAX_NUM_WORK_GROUPS_32_4_2]]
+
 // Make sure this is silently accepted on other targets.
 // NAMD-NOT: "amdgpu-flat-work-group-size"
 // NAMD-NOT: "amdgpu-waves-per-eu"