Move SPIRV global variables into a context variable (#4741)

9 files changed, 292 insertions, 4 deletions

diff --git a/tests/bugs/spirv-opt-SROA-of-globals.slang b/tests/bugs/spirv-opt-SROA-of-globals.slang
new file mode 100644
index 000000000..140559b4f
--- /dev/null
+++ b/tests/bugs/spirv-opt-SROA-of-globals.slang
@@ -0,0 +1,37 @@
+//TEST:SIMPLE(filecheck=CHECK): -target spirv -fvk-use-entrypoint-name -enable-experimental-passes
+
+// This test checks that spirv-opt is running SROA (scalar replacement of aggregates) to 
+// hoist out all member variables of a struct. If this sucsessfully runs, we should not see
+// any `OpCompositeConstruct` of our `%Data` struct.
+// Note: SROA will only run for 100 elements if spirv-opt does not manually set the spirv-opt `scalar-replacement` option
+
+//CHECK-NOT: OpCompositeConstruct %Data
+struct Data
+{
+    uint data0;
+    uint data1;
+    uint data2;
+};
+
+static Data globalVar;
+ByteAddressBuffer bab;
+RWStructuredBuffer<uint> outputBuffer;
+
+struct Payload_t
+{
+    uint dataOut;
+};
+
+[shader("anyhit")]
+void main1(inout Payload_t payload)
+{
+    globalVar = bab.Load<Data>(0);
+    payload.dataOut = globalVar.data0;
+}
+
+[shader("anyhit")]
+void main2(inout Payload_t payload)
+{
+    globalVar = bab.Load<Data>(0);
+    payload.dataOut = globalVar.data0;
+}
\ No newline at end of file
diff --git a/tests/spirv/address-space-specialize.slang b/tests/spirv/address-space-specialize.slang
index e2b48489a..f8201838b 100644
--- a/tests/spirv/address-space-specialize.slang
+++ b/tests/spirv/address-space-specialize.slang
@@ -1,3 +1,4 @@
+//TEST:SIMPLE(filecheck=CHECK_EXPERIMENTAL):-target spirv -entry main -stage compute -emit-spirv-directly -O0 -enable-experimental-passes
 //TEST:SIMPLE(filecheck=CHECK):-target spirv -entry main -stage compute -emit-spirv-directly -O0
 
 // Test that we can pass arguments in different address space to an `inout` parameter, and have
@@ -7,6 +8,12 @@
 static int gArray0[2];
 groupshared int gArray1[2];
 
+// Note: static globals are inside a context variable
+// CHECK_EXPERIMENTAL: OpTypeArray %int %int_2
+// CHECK_EXPERIMENTAL: OpVariable %_ptr_Function__arr_int_int_2 Function
+// CHECK_EXPERIMENTAL: %array_0 = OpFunctionParameter %_ptr_Workgroup__arr_int_int_2
+
+
 // CHECK: %array = OpFunctionParameter %_ptr_Private__arr_int_int_2
 // CHECK: %array_0 = OpFunctionParameter %_ptr_Workgroup__arr_int_int_2
 
diff --git a/tests/spirv/explicit-context-validation-builtin-compute-glsl.slang b/tests/spirv/explicit-context-validation-builtin-compute-glsl.slang
new file mode 100644
index 000000000..9fb0947ae
--- /dev/null
+++ b/tests/spirv/explicit-context-validation-builtin-compute-glsl.slang
@@ -0,0 +1,19 @@
+//TEST:SIMPLE(filecheck=CHECK): -allow-glsl -target spirv -emit-spirv-directly -stage compute -entry computeMain -enable-experimental-passes
+
+// Check to ensure builtin is not moved into a kernelContext (part of entry-point). Ensure builtin is referenced directly.
+
+// CHECK: OpEntryPoint {{.*}} %gl_NumWorkGroups
+// CHECK: OpDecorate %gl_NumWorkGroups BuiltIn NumWorkgroups
+// CHECK: %gl_NumWorkGroups = OpVariable {{.*}} Input
+// CHECK: %[[NUM_WORK_GROUP_LOAD:[A-Za-z0-9_]+]] = OpLoad %v3uint %gl_NumWorkGroups
+// CHECK: OpCompositeExtract %uint %[[NUM_WORK_GROUP_LOAD]] 0
+
+RWStructuredBuffer<uint> outputBuffer;
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+void computeMain()
+{
+    outputBuffer[0] = gl_NumWorkGroups.x;
+    outputBuffer[1] = gl_NumWorkGroups.y;
+    outputBuffer[2] = gl_NumWorkGroups.z;
+}
\ No newline at end of file
diff --git a/tests/spirv/explicit-context-validation-builtin-compute-hlsl.slang b/tests/spirv/explicit-context-validation-builtin-compute-hlsl.slang
new file mode 100644
index 000000000..24184c9f4
--- /dev/null
+++ b/tests/spirv/explicit-context-validation-builtin-compute-hlsl.slang
@@ -0,0 +1,16 @@
+//TEST:SIMPLE(filecheck=CHECK): -target spirv -emit-spirv-directly -stage compute -entry computeMain -enable-experimental-passes
+
+// Check to ensure builtin is not moved into a kernelContext (part of entry-point). Ensure builtin is referenced directly.
+
+// CHECK: OpEntryPoint {{.*}} %gl_GlobalInvocationID
+// CHECK: OpDecorate %gl_GlobalInvocationID BuiltIn GlobalInvocationId
+// CHECK: %gl_GlobalInvocationID = OpVariable {{.*}} Input
+// CHECK: %[[NUM_WORK_GROUP_LOAD:[A-Za-z0-9_]+]] = OpLoad %v3uint %gl_GlobalInvocationID
+// CHECK: OpCompositeExtract %uint %[[NUM_WORK_GROUP_LOAD]] 0
+
+RWStructuredBuffer<uint> outputBuffer;
+[numthreads(1, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    outputBuffer[dispatchThreadID.x] = 1;
+}
\ No newline at end of file
diff --git a/tests/spirv/explicit-context-validation-builtin-vertex-glsl.slang b/tests/spirv/explicit-context-validation-builtin-vertex-glsl.slang
new file mode 100644
index 000000000..8e37d4814
--- /dev/null
+++ b/tests/spirv/explicit-context-validation-builtin-vertex-glsl.slang
@@ -0,0 +1,35 @@
+//TEST:SIMPLE(filecheck=CHECK): -allow-glsl -target spirv -emit-spirv-directly -stage vertex -entry vertexMain -enable-experimental-passes
+
+// Check to ensure varying output/input and builtin is not moved into a kernelContext (part of entry-point). Ensure varying output/input and builtin is referenced directly.
+
+// CHECK: OpEntryPoint{{((.*)(%.*data1In|%.*data2In|%.*data1Out)(.*))|((.*)(%.*data1In|%.*data2In|%.*data1Out)(.*))|((.*)(%.*data1In|%.*data2In|%.*data1Out)(.*))}}
+
+//CHECK-DAG: OpDecorate %[[DATA_IN_1:.*data1In]] Location 1
+//CHECK-DAG: OpDecorate %[[DATA_IN_2:.*data2In]] Location 2
+//CHECK-DAG: OpDecorate %[[DATA_OUT_1:.*data1Out]] Location 0
+//CHECK-DAG: OpDecorate %gl_Position BuiltIn Position
+
+//CHECK-DAG: %[[DATA_IN_1]] = OpVariable{{.*}} Input
+//CHECK-DAG: %[[DATA_IN_2]] = OpVariable{{.*}} Input
+//CHECK-DAG: %[[DATA_OUT_1]] = OpVariable{{.*}} Output
+
+//CHECK-DAG: OpLoad{{.*}} %[[DATA_IN_1]]
+//CHECK-DAG: OpLoad{{.*}} %[[DATA_IN_2]]
+
+//CHECK-DAG: OpStore %gl_Position
+//CHECK-DAG: OpStore %[[DATA_OUT_1]]
+
+layout(location = 1) in vec4 data1In;
+layout(location = 2) in int data2In;
+layout(location = 0) out int data1Out;
+
+void nestedCall()
+{
+    gl_Position = data1In;
+    data1Out = data2In;
+}
+
+void vertexMain()
+{
+    nestedCall();
+}
diff --git a/tests/spirv/explicit-context-validation-groupshared.slang b/tests/spirv/explicit-context-validation-groupshared.slang
new file mode 100644
index 000000000..7fb27377d
--- /dev/null
+++ b/tests/spirv/explicit-context-validation-groupshared.slang
@@ -0,0 +1,30 @@
+//TEST:SIMPLE(filecheck=CHECK): -target spirv -emit-spirv-directly -stage compute -entry computeMain -enable-experimental-passes
+
+// Check to ensure builtin is not moved into a kernelContext (part of entry-point). Ensure builtin is referenced directly.
+
+// CHECK: OpEntryPoint{{.*}} %val
+// CHECK: %[[GROUP_SHARED_VAR:(.*)]] = OpVariable{{.*}} Workgroup
+
+// CHECK: OpControlBarrier
+// CHECK: OpControlBarrier
+// CHECK-DAG: %[[GROUP_SHARED_VAL:(.*)]] = OpLoad{{.*}} %[[GROUP_SHARED_VAR]]
+// CHECK-DAG: %[[OUTPUT_BUFFER_LOC:(.*)]] = OpAccessChain{{.*}} %outputBuffer
+// CHECK: OpStore %[[OUTPUT_BUFFER_LOC]] %[[GROUP_SHARED_VAL]]
+
+groupshared uint val;
+RWStructuredBuffer<uint> outputBuffer;
+
+void nestedCall(uint index)
+{
+    val += 1;
+    GroupMemoryBarrierWithGroupSync();
+    outputBuffer[index] = val;
+}
+
+[numthreads(4, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    val = 0;
+    GroupMemoryBarrierWithGroupSync();
+    nestedCall(dispatchThreadID.x);
+}
\ No newline at end of file
diff --git a/tests/spirv/explicit-context-validation-raytracing-glsl.slang b/tests/spirv/explicit-context-validation-raytracing-glsl.slang
new file mode 100644
index 000000000..1e7b56a9f
--- /dev/null
+++ b/tests/spirv/explicit-context-validation-raytracing-glsl.slang
@@ -0,0 +1,78 @@
+//TEST:SIMPLE(filecheck=CHECK): -allow-glsl -target spirv -emit-spirv-directly -stage raygeneration -entry main -enable-experimental-passes
+
+// Check to ensure we make global ray-tracing objects. Ensure we store into these variables directly and not through a context-pointer.
+
+// CHECK-DAG: %[[RAYTRACING_AS_TYPE:[A-Za-z0-9_]+]] = OpTypeAccelerationStructureKHR
+// CHECK-DAG: %[[RAYTRACING_AS_PTR_TYPE:[A-Za-z0-9_]+]] = OpTypePointer UniformConstant %[[RAYTRACING_AS_TYPE]]
+// CHECK-DAG: %[[RAYTRACING_AS:[A-Za-z0-9_]+]] = OpVariable %[[RAYTRACING_AS_PTR_TYPE]] UniformConstant
+
+// CHECK-DAG: %[[RAY_PAYLOAD:[A-Za-z0-9_]+]] = OpVariable %{{.*}} RayPayloadKHR
+// CHECK-DAG: %[[HIT_ATTR:[A-Za-z0-9_]+]] = OpVariable %{{.*}} HitObjectAttributeNV
+// CHECK-DAG: %[[CALL_DATA:[A-Za-z0-9_]+]] = OpVariable %{{.*}} CallableDataKHR
+
+
+// CHECK: OpLoad {{.*}} %[[RAYTRACING_AS]]
+// CHECK: OpStore %[[RAY_PAYLOAD]]
+// CHECK: OpStore %[[HIT_ATTR]]
+// CHECK: OpStore %[[CALL_DATA]]
+
+layout(binding = 0) uniform accelerationStructureEXT as;
+
+buffer MyBlockName
+{
+    uint data[];
+} outputBuffer;
+layout(location = 2) rayPayloadEXT vec4 payload;
+layout(location = 2) hitObjectAttributeNV vec4 attrMain;
+layout(location = 0) callableDataEXT vec4 outcall;
+
+bool testHitObjectTraceRay() {
+    hitObjectNV hit;
+
+
+    hitObjectTraceRayNV(hit, as, gl_RayFlagsNoneEXT, 0xff, 0, 0, 0, vec3(0.1, 0.1, 0.0), 0.01f, vec3(0, 0, 1), 1e4f, 2);
+    return true
+        && hitObjectIsHitNV(hit) == true
+        ;
+}
+
+
+bool testPayloadReadWrite() {
+    payload = vec4(2);
+    vec4 read = payload;
+    return true
+        && read != vec4(0)
+        ;
+}
+
+bool testAttributeReadWrite() {
+    attrMain = vec4(2);
+    vec4 read = attrMain;
+    return true
+        && read != vec4(0)
+        ;
+}
+
+bool testCallableReadWrite() {
+    outcall = vec4(2);
+    vec4 read = outcall;
+    return true
+        && read != vec4(0)
+        ;
+}
+
+bool testReadWriteOfObjects(){
+    return true
+        && testPayloadReadWrite()
+        && testAttributeReadWrite()
+        && testCallableReadWrite();
+        ;       
+}
+
+void main()
+{
+    outputBuffer.data[0] = true
+        && testHitObjectTraceRay()
+        && testReadWriteOfObjects()
+        ;
+}
diff --git a/tests/spirv/explicit-context-validation-raytracing-hlsl.slang b/tests/spirv/explicit-context-validation-raytracing-hlsl.slang
new file mode 100644
index 000000000..28420eeec
--- /dev/null
+++ b/tests/spirv/explicit-context-validation-raytracing-hlsl.slang
@@ -0,0 +1,53 @@
+//TEST:SIMPLE(filecheck=CHECK): -target spirv -emit-spirv-directly -stage raygeneration -entry main -enable-experimental-passes
+
+// Check to ensure we make global ray-tracing objects. Ensure we store into these variables directly and not through a context-pointer.
+
+// CHECK-DAG: %[[RAYTRACING_AS_TYPE:[A-Za-z0-9_]+]] = OpTypeAccelerationStructureKHR
+// CHECK-DAG: %[[RAYTRACING_AS_PTR_TYPE:[A-Za-z0-9_]+]] = OpTypePointer UniformConstant %[[RAYTRACING_AS_TYPE]]
+// CHECK-DAG: %[[RAYTRACING_AS:[A-Za-z0-9_]+]] = OpVariable %[[RAYTRACING_AS_PTR_TYPE]] UniformConstant
+
+// CHECK-DAG: %[[RAY_PAYLOAD:[A-Za-z0-9_]+]] = OpVariable %{{.*}} RayPayloadKHR
+
+// CHECK: OpLoad {{.*}} %[[RAYTRACING_AS]]
+// CHECK: OpTraceRayKHR{{.*}} %[[RAY_PAYLOAD]]
+
+RaytracingAccelerationStructure as;
+
+struct ShadowRay
+{
+    float hitDistance;
+};
+
+void nestedNestedCall()
+{
+    RayDesc ray = {};
+
+    ShadowRay shadowRay;
+    shadowRay.hitDistance = 0;
+
+    TraceRay(as,
+            // ray flags
+            1,
+            // cull mask
+            0xff,
+            // sbt record offset
+            0,
+            // sbt record stride
+            0,
+            // missIndex
+            2,
+            // ray
+            ray,
+            // payload
+            shadowRay);
+}
+
+void nestedCall()
+{
+    nestedNestedCall();
+}
+
+void main()
+{
+    nestedCall();
+}
\ No newline at end of file
diff --git a/tests/spirv/subgroup-size-2.slang b/tests/spirv/subgroup-size-2.slang
index bd5ae5eec..500bd63c8 100644
--- a/tests/spirv/subgroup-size-2.slang
+++ b/tests/spirv/subgroup-size-2.slang
@@ -1,8 +1,11 @@
 // Test that using workgroup size from more than one entrypoint result in
 // correct lowering into global variable.
 
+//TEST:SIMPLE(filecheck=CHECK_EXPERIMENTAL): -target spirv -emit-spirv-directly -fvk-use-entrypoint-name -O0 -enable-experimental-passes
 //TEST:SIMPLE(filecheck=CHECK): -target spirv -emit-spirv-directly -fvk-use-entrypoint-name -O0
 
+
+
 RWStructuredBuffer<int> outputBuffer;
 
 uint3 f() { return WorkgroupSize(); }
@@ -11,16 +14,26 @@ uint3 f() { return WorkgroupSize(); }
 [numthreads(1u, 2u, 3)]
 void compute1()
 {
+    // CHECK_EXPERIMENTAL-DAG: %[[VAR:[A-Za-z0-9_]+]] = OpTypePointer Function %v3int
+    // CHECK_EXPERIMENTAL: OpAccessChain %[[VAR]]
+    //
+    // CHECK_EXPERIMENTAL-DAG: %[[CALL_RS:[A-Za-z0-9_]+]] = OpFunctionCall %v3uint %f
+    // CHECK_EXPERIMENTAL: OpCompositeExtract %uint %[[CALL_RS]] 0
+    //
+    // CHECK_EXPERIMENTAL-DAG: %[[PTR:[A-Za-z0-9_]+]] = OpAccessChain %_ptr_StorageBuffer_int %outputBuffer %int_0 %int_1
+    // CHECK_EXPERIMENTAL: OpStore %[[PTR]] %int_2
+
     // CHECK-DAG: %[[VAR:[A-Za-z0-9_]+]] = OpVariable %_ptr_Private_v3int Private
     // CHECK: OpStore %[[VAR]]
-
+    //
     // CHECK-DAG: %[[CALL_RS:[A-Za-z0-9_]+]] = OpFunctionCall %v3uint %f
     // CHECK: OpCompositeExtract %uint %[[CALL_RS]] 0
-    const int x = f().x;
-    outputBuffer[0] = x;
-
+    //
     // CHECK-DAG: %[[PTR:[A-Za-z0-9_]+]] = OpAccessChain %_ptr_StorageBuffer_int %outputBuffer %int_0 %int_1
     // CHECK: OpStore %[[PTR]] %int_2
+
+    const int x = f().x;
+    outputBuffer[0] = x;
     outputBuffer[1] = WorkgroupSize().y;
 }