[SPIR-V] Fix out-of-range value for NumWorkgroups builtin

[svenvh]

The OpenCL C specification states that for out-of-range dimension indices, get_num_groups must return 1 instead of 0.

[llvmbot]

@llvm/pr-subscribers-backend-spir-v

Author: Sven van Haastregt (svenvh)

Changes

The OpenCL C specification states that for out-of-range dimension indices, get_num_groups must return 1 instead of 0.

---

Full diff: https://github.com/llvm/llvm-project/pull/127198.diff

2 Files Affected:

• (modified) llvm/lib/Target/SPIRV/SPIRVBuiltins.cpp (+1)
• (added) llvm/test/CodeGen/SPIRV/opencl/get_num_groups.ll (+55)

diff --git a/llvm/lib/Target/SPIRV/SPIRVBuiltins.cpp b/llvm/lib/Target/SPIRV/SPIRVBuiltins.cpp
index a7a34e0439ab1..aec2aaccb33a8 100644
--- a/llvm/lib/Target/SPIRV/SPIRVBuiltins.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVBuiltins.cpp
@@ -1779,6 +1779,7 @@ static bool generateGetQueryInst(const SPIRV::IncomingCall *Call,
   SPIRV::BuiltIn::BuiltIn Value =
       SPIRV::lookupGetBuiltin(Call->Builtin->Name, Call->Builtin->Set)->Value;
   uint64_t IsDefault = (Value == SPIRV::BuiltIn::GlobalSize ||
+                        Value == SPIRV::BuiltIn::NumWorkgroups ||
                         Value == SPIRV::BuiltIn::WorkgroupSize ||
                         Value == SPIRV::BuiltIn::EnqueuedWorkgroupSize);
   return genWorkgroupQuery(Call, MIRBuilder, GR, Value, IsDefault ? 1 : 0);
diff --git a/llvm/test/CodeGen/SPIRV/opencl/get_num_groups.ll b/llvm/test/CodeGen/SPIRV/opencl/get_num_groups.ll
new file mode 100644
index 0000000000000..3f1d1dc248fc4
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/opencl/get_num_groups.ll
@@ -0,0 +1,55 @@
+; RUN: llc -O0 -mtriple=spirv64-unknown-unknown %s -o - | FileCheck %s
+
+;; The set of valid inputs for get_num_groups depends on the runtime NDRange,
+;; but inputs outside of [0, 2] always return 1.
+;; Here we assume Itanium mangling for function name.
+declare i64 @_Z14get_num_groupsj(i32)
+
+define i64 @foo(i32 %dim) {
+  %x = call i64 @_Z14get_num_groupsj(i32 0)
+  %y = call i64 @_Z14get_num_groupsj(i32 5)
+  %acc = add i64 %x, %y
+  %unknown = call i64 @_Z14get_num_groupsj(i32 %dim)
+  %ret = add i64 %acc, %unknown
+  ret i64 %ret
+}
+
+;; Capabilities:
+; CHECK-DAG: OpCapability Kernel
+; CHECK-DAG: OpCapability Int64
+
+; CHECK-NOT: DAG-FENCE
+
+;; Decorations:
+; CHECK-DAG: OpDecorate %[[#GET_NUM_GROUPS:]] BuiltIn NumWorkgroups
+; CHECK-DAG: OpDecorate %[[#GET_NUM_GROUPS]] Constant
+
+; CHECK-NOT: DAG-FENCE
+
+;; Types, Constants and Variables:
+; CHECK-DAG: %[[#BOOL:]] = OpTypeBool
+; CHECK-DAG: %[[#I32:]] = OpTypeInt 32 0
+; CHECK-DAG: %[[#I64:]] = OpTypeInt 64 0
+; CHECK-DAG: %[[#VEC:]] = OpTypeVector %[[#I64]] 3
+; CHECK-DAG: %[[#PTR:]] = OpTypePointer Input %[[#VEC]]
+; CHECK-DAG: %[[#FN:]] = OpTypeFunction %[[#I64]] %[[#I32]]
+; CHECK-DAG: %[[#GET_NUM_GROUPS]] = OpVariable %[[#PTR]] Input
+; CHECK-DAG: %[[#ONE:]] = OpConstant %[[#I64]] 1
+; CHECK-DAG: %[[#THREE:]] = OpConstant %[[#I32]] 3
+
+;; Functions:
+; CHECK:     OpFunction %[[#I64]] None %[[#FN]]
+; CHECK:     %[[#DIM:]] = OpFunctionParameter %[[#I32]]
+
+;; get_num_groups(0): OpLoad + OpCompositeExtract.
+; CHECK:     %[[#TMP1:]] = OpLoad %[[#VEC]] %[[#GET_NUM_GROUPS]]
+; CHECK:     %[[#X:]] = OpCompositeExtract %[[#I64]] %[[#TMP1]] 0
+
+;; get_num_groups(5): OpConstant of one.
+; CHECK:     OpIAdd %[[#I64]] %[[#X]] %[[#ONE]]
+
+;; get_num_groups(dim): Implementation using OpSelect.
+; CHECK-DAG: %[[#TMP2:]] = OpLoad %[[#VEC]] %[[#GET_NUM_GROUPS]]
+; CHECK-DAG: %[[#TMP3:]] = OpVectorExtractDynamic %[[#I64]] %[[#TMP2]] %[[#DIM]]
+; CHECK-DAG: %[[#COND:]] = OpULessThan %[[#BOOL]] %[[#DIM]] %[[#THREE]]
+; CHECK:     %[[#UNKNOWN:]] = OpSelect %[[#I64]] %[[#COND]] %[[#TMP3]] %[[#ONE]]