Don't claim alignment above 4 for raw buffer due to base (#2905)

lib/HLSL/HLOperationLower.cpp

@@ -3575,7 +3575,12 @@ void TranslateLoad(ResLoadHelper &helper, HLResource::Kind RK,
   Type *i64Ty = Builder.getInt64Ty();
   Type *doubleTy = Builder.getDoubleTy();
   Type *EltTy = Ty->getScalarType();
-  Constant *Alignment = OP->GetI32Const(OP->GetAllocSizeForType(EltTy));
+  // If RawBuffer load of 64-bit value, don't set alignment to 8,
+  // since buffer alignment isn't known to be anything over 4.
+  unsigned alignValue = OP->GetAllocSizeForType(EltTy);
+  if (RK == HLResource::Kind::RawBuffer && alignValue > 4)
+    alignValue = 4;
+  Constant *Alignment = OP->GetI32Const(alignValue);
   unsigned numComponents = 1;
   if (Ty->isVectorTy()) {
     numComponents = Ty->getVectorNumElements();
@@ -3804,7 +3809,12 @@ void TranslateStore(DxilResource::Kind RK, Value *handle, Value *val,
     val = Builder.CreateZExt(val, Ty);
   }
 
-  Constant *Alignment = OP->GetI32Const(OP->GetAllocSizeForType(EltTy));
+  // If RawBuffer store of 64-bit value, don't set alignment to 8,
+  // since buffer alignment isn't known to be anything over 4.
+  unsigned alignValue = OP->GetAllocSizeForType(EltTy);
+  if (RK == HLResource::Kind::RawBuffer && alignValue > 4)
+    alignValue = 4;
+  Constant *Alignment = OP->GetI32Const(alignValue);
   bool is64 = EltTy == i64Ty || EltTy == doubleTy;
   if (is64 && isTyped) {
     EltTy = i32Ty;

tools/clang/test/HLSLFileCheck/hlsl/objects/ByteAddressBuffer/raw_buf4.hlsl

@@ -12,12 +12,12 @@
 
 // CHECK-NOT: call %dx.types.ResRet.f16
 
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf1_texture_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 3, i32 8)
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf1_texture_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 15, i32 8)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf1_texture_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 3, i32 4)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf1_texture_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 15, i32 4)
 // CHECK: call double @dx.op.makeDouble.f64
 
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf2_UAV_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 3, i32 8)
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf2_UAV_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 15, i32 8)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf2_UAV_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 3, i32 4)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf2_UAV_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 15, i32 4)
 // CHECK: call double @dx.op.makeDouble.f64
 
 // CHECK: call void @dx.op.rawBufferStore.f32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, float %{{.*}}, float undef, float undef, float undef, i8 1, i32 4)
@@ -28,8 +28,8 @@
 // CHECK: call void @dx.op.rawBufferStore.f32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, float %{{.*}}, float %{{.*}}, float undef, float undef, i8 3, i32 4)
 // CHECK: call void @dx.op.rawBufferStore.f32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, float %{{.*}}, float %{{.*}}, float %{{.*}}, float undef, i8 7, i32 4)
 // CHECK: call void @dx.op.rawBufferStore.f32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, float %{{.*}}, float %{{.*}}, float %{{.*}}, float %{{.*}}, i8 15, i32 4)
-// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, i32 %{{.*}}, i32 %{{.*}}, i32 undef, i32 undef, i8 3, i32 8)
-// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, i32 %{{.*}}, i32 %{{.*}}, i32 %{{.*}}, i32 %{{.*}}, i8 15, i32 8)
+// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, i32 %{{.*}}, i32 %{{.*}}, i32 undef, i32 undef, i8 3, i32 4)
+// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, i32 %{{.*}}, i32 %{{.*}}, i32 %{{.*}}, i32 %{{.*}}, i8 15, i32 4)
 
 ByteAddressBuffer buf1;
 RWByteAddressBuffer buf2;

tools/clang/test/HLSLFileCheck/hlsl/objects/ByteAddressBuffer/raw_buf5.hlsl

@@ -20,12 +20,12 @@
 // CHECK: call %dx.types.ResRet.f16 @dx.op.rawBufferLoad.f16(i32 139, %dx.types.Handle %buf2_UAV_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 7, i32 2)
 // CHECK: call %dx.types.ResRet.f16 @dx.op.rawBufferLoad.f16(i32 139, %dx.types.Handle %buf2_UAV_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 15, i32 2)
 
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf1_texture_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 3, i32 8)
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf1_texture_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 15, i32 8)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf1_texture_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 3, i32 4)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf1_texture_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 15, i32 4)
 // CHECK: call double @dx.op.makeDouble.f64
 
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf2_UAV_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 3, i32 8)
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf2_UAV_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 15, i32 8)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf2_UAV_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 3, i32 4)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %buf2_UAV_rawbuf, i32 %{{[0-9]+}}, i32 undef, i8 15, i32 4)
 // CHECK: call double @dx.op.makeDouble.f64
 
 // Store by default will store what's passed in
@@ -41,8 +41,8 @@
 // CHECK: call void @dx.op.rawBufferStore.f32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, float %{{.*}}, float %{{.*}}, float undef, float undef, i8 3, i32 4)
 // CHECK: call void @dx.op.rawBufferStore.f32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, float %{{.*}}, float %{{.*}}, float %{{.*}}, float undef, i8 7, i32 4)
 // CHECK: call void @dx.op.rawBufferStore.f32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, float %{{.*}}, float %{{.*}}, float %{{.*}}, float %{{.*}}, i8 15, i32 4)
-// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, i32 %{{.*}}, i32 %{{.*}}, i32 undef, i32 undef, i8 3, i32 8)
-// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, i32 %{{.*}}, i32 %{{.*}}, i32 %{{.*}}, i32 %{{.*}}, i8 15, i32 8)
+// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, i32 %{{.*}}, i32 %{{.*}}, i32 undef, i32 undef, i8 3, i32 4)
+// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %buf2_UAV_rawbuf, i32 1, i32 undef, i32 %{{.*}}, i32 %{{.*}}, i32 %{{.*}}, i32 %{{.*}}, i8 15, i32 4)
 
 ByteAddressBuffer buf1;
 RWByteAddressBuffer buf2;

tools/clang/test/HLSLFileCheck/hlsl/objects/ByteAddressBuffer/store_type_shapes.hlsl

@@ -16,7 +16,7 @@ void main() {
   buf.Store(104, (int)42);
   // CHECK: call void @dx.op.rawBufferStore.i16(i32 140, {{.*}}, i32 108, i32 undef, i16 42, i16 undef, i16 undef, i16 undef, i8 1, i32 2)
   buf.Store(108, (int16_t)42);
-  // CHECK: call void @dx.op.rawBufferStore.i32(i32 140, {{.*}}, i32 112, i32 undef, i32 42, i32 0, i32 undef, i32 undef, i8 3, i32 8)
+  // CHECK: call void @dx.op.rawBufferStore.i32(i32 140, {{.*}}, i32 112, i32 undef, i32 42, i32 0, i32 undef, i32 undef, i8 3, i32 4)
   buf.Store(112, (int64_t)42);
 
   // CHECK: call void @dx.op.rawBufferStore.i32(i32 140, {{.*}}, i32 200, i32 undef, i32 42, i32 42, i32 undef, i32 undef, i8 3, i32 4)

tools/clang/test/HLSLFileCheck/hlsl/objects/StructuredBuffer/rawbufferloadstore_64bit_6_2.hlsl

@@ -15,12 +15,12 @@ RWStructuredBuffer<TestData> uav1 : register(u1);
 void main(uint GI : SV_GroupIndex) {
 
   int64_t3 vec3 = srv0.Load<int64_t3>(0);
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 0, i32 undef, i8 15, i32 8)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 0, i32 undef, i8 15, i32 4)
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 0
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 1
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 2
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 3
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 16, i32 undef, i8 3, i32 8)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 16, i32 undef, i8 3, i32 4)
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 0
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 1
 // CHECK: zext i32 %{{[0-9]+}} to i64
@@ -32,16 +32,16 @@ void main(uint GI : SV_GroupIndex) {
 // CHECK: trunc i64 %{{[0-9]+}} to i32
 // CHECK: lshr i64 %{{[0-9]+}}, 32
 // CHECK: trunc i64 %{{[0-9]+}} to i32
-// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 0, i32 undef, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i8 15, i32 8) 
-// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 16, i32 undef, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 undef, i32 undef, i8 3, i32 8) 
+// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 0, i32 undef, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i8 15, i32 4)
+// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 16, i32 undef, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 undef, i32 undef, i8 3, i32 4)
 
   int64_t4 vec4 = srv0.Load<int64_t4>(0);
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 0, i32 undef, i8 15, i32 8)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 0, i32 undef, i8 15, i32 4)
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 0
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 1
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 2
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 3
-// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 16, i32 undef, i8 15, i32 8)
+// CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 16, i32 undef, i8 15, i32 4)
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 0
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 1
 // CHECK: extractvalue %dx.types.ResRet.i32 %{{[a-zA-Z0-9]+}}, 2
@@ -55,8 +55,8 @@ void main(uint GI : SV_GroupIndex) {
 // CHECK: trunc i64 %{{[0-9]+}} to i32
 // CHECK: lshr i64 %{{[0-9]+}}, 32
 // CHECK: trunc i64 %{{[0-9]+}} to i32
-// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 0, i32 undef, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i8 15, i32 8) 
-// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 16, i32 undef, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i8 15, i32 8) 
+// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 0, i32 undef, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i8 15, i32 4)
+// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 16, i32 undef, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i32 %{{[0-9]+}}, i8 15, i32 4)
 
   int64_t3 svec3 = srv1[0].v3;
 // CHECK: call %dx.types.ResRet.i32 @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %srv1_texture_structbuf, i32 0, i32 0, i8 15, i32 8)

tools/clang/test/HLSLFileCheck/hlsl/objects/StructuredBuffer/rawbufferloadstore_64bit_6_3.hlsl

@@ -15,23 +15,23 @@ RWStructuredBuffer<TestData> uav1 : register(u1);
 void main(uint GI : SV_GroupIndex) {
 
   int64_t3 vec3 = srv0.Load<int64_t3>(0);
-// CHECK: call %dx.types.ResRet.i64 @dx.op.rawBufferLoad.i64(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 0, i32 undef, i8 7, i32 8)
+// CHECK: call %dx.types.ResRet.i64 @dx.op.rawBufferLoad.i64(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 0, i32 undef, i8 7, i32 4)
 // CHECK: extractvalue %dx.types.ResRet.i64 %{{[0-9a-zA-Z]+}}, 0
 // CHECK: extractvalue %dx.types.ResRet.i64 %{{[0-9a-zA-Z]+}}, 1
 // CHECK: extractvalue %dx.types.ResRet.i64 %{{[0-9a-zA-Z]+}}, 2
 
   uav0.Store(0, vec3);
-// CHECK: call void @dx.op.rawBufferStore.i64(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 0, i32 undef, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}, i64 undef, i8 7, i32 8)
+// CHECK: call void @dx.op.rawBufferStore.i64(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 0, i32 undef, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}, i64 undef, i8 7, i32 4)
 
   int64_t4 vec4 = srv0.Load<int64_t4>(0);
-// CHECK: call %dx.types.ResRet.i64 @dx.op.rawBufferLoad.i64(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 0, i32 undef, i8 15, i32 8)
+// CHECK: call %dx.types.ResRet.i64 @dx.op.rawBufferLoad.i64(i32 139, %dx.types.Handle %srv0_texture_rawbuf, i32 0, i32 undef, i8 15, i32 4)
 // CHECK: extractvalue %dx.types.ResRet.i64 %{{[0-9a-zA-Z]+}}, 0
 // CHECK: extractvalue %dx.types.ResRet.i64 %{{[0-9a-zA-Z]+}}, 1
 // CHECK: extractvalue %dx.types.ResRet.i64 %{{[0-9a-zA-Z]+}}, 2
 // CHECK: extractvalue %dx.types.ResRet.i64 %{{[0-9a-zA-Z]+}}, 3
 
   uav0.Store(0, vec4);
-// CHECK: call void @dx.op.rawBufferStore.i64(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 0, i32 undef, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}, i8 15, i32 8)
+// CHECK: call void @dx.op.rawBufferStore.i64(i32 140, %dx.types.Handle %uav0_UAV_rawbuf, i32 0, i32 undef, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}, i8 15, i32 4)
 
   int64_t3 svec3 = srv1[0].v3;
 // CHECK: call %dx.types.ResRet.i64 @dx.op.rawBufferLoad.i64(i32 139, %dx.types.Handle %srv1_texture_structbuf, i32 0, i32 0, i8 7, i32 8)