diff --git a/libs/vkd3d-shader/hlsl_codegen.c b/libs/vkd3d-shader/hlsl_codegen.c
index ce431ee6..6fab8349 100644
--- a/libs/vkd3d-shader/hlsl_codegen.c
+++ b/libs/vkd3d-shader/hlsl_codegen.c
@@ -6579,6 +6579,42 @@ static void generate_vsir_signature(struct hlsl_ctx *ctx,
     }
 }
 
+static enum vkd3d_data_type vsir_data_type_from_hlsl_type(struct hlsl_ctx *ctx, const struct hlsl_type *type)
+{
+    if (hlsl_version_lt(ctx, 4, 0))
+        return VKD3D_DATA_FLOAT;
+
+    if (type->class == HLSL_CLASS_ARRAY)
+        return vsir_data_type_from_hlsl_type(ctx, type->e.array.type);
+    if (type->class == HLSL_CLASS_STRUCT)
+        return VKD3D_DATA_MIXED;
+    if (type->class <= HLSL_CLASS_LAST_NUMERIC)
+    {
+        switch (type->e.numeric.type)
+        {
+            case HLSL_TYPE_DOUBLE:
+                return VKD3D_DATA_DOUBLE;
+            case HLSL_TYPE_FLOAT:
+                return VKD3D_DATA_FLOAT;
+            case HLSL_TYPE_HALF:
+                return VKD3D_DATA_HALF;
+            case HLSL_TYPE_INT:
+                return VKD3D_DATA_INT;
+            case HLSL_TYPE_UINT:
+            case HLSL_TYPE_BOOL:
+                return VKD3D_DATA_UINT;
+        }
+    }
+
+    vkd3d_unreachable();
+}
+
+static enum vkd3d_data_type vsir_data_type_from_hlsl_instruction(struct hlsl_ctx *ctx,
+        const struct hlsl_ir_node *instr)
+{
+    return vsir_data_type_from_hlsl_type(ctx, instr->data_type);
+}
+
 static uint32_t sm1_generate_vsir_get_src_swizzle(uint32_t src_writemask, uint32_t dst_writemask)
 {
     uint32_t swizzle;
@@ -7327,8 +7363,8 @@ static void sm1_generate_vsir_instr_resource_load(struct hlsl_ctx *ctx,
     }
 }
 
-static void sm1_generate_vsir_instr_swizzle(struct hlsl_ctx *ctx, struct vsir_program *program,
-        struct hlsl_ir_swizzle *swizzle_instr)
+static void generate_vsir_instr_swizzle(struct hlsl_ctx *ctx,
+        struct vsir_program *program, struct hlsl_ir_swizzle *swizzle_instr)
 {
     struct hlsl_ir_node *instr = &swizzle_instr->node, *val = swizzle_instr->val.node;
     struct vkd3d_shader_dst_param *dst_param;
@@ -7342,8 +7378,9 @@ static void sm1_generate_vsir_instr_swizzle(struct hlsl_ctx *ctx, struct vsir_pr
         return;
 
     dst_param = &ins->dst[0];
-    vsir_register_init(&dst_param->reg, VKD3DSPR_TEMP, VKD3D_DATA_FLOAT, 1);
+    vsir_register_init(&dst_param->reg, VKD3DSPR_TEMP, vsir_data_type_from_hlsl_instruction(ctx, instr), 1);
     dst_param->reg.idx[0].offset = instr->reg.id;
+    dst_param->reg.dimension = VSIR_DIMENSION_VEC4;
     dst_param->write_mask = instr->reg.writemask;
 
     swizzle = hlsl_swizzle_from_writemask(val->reg.writemask);
@@ -7352,8 +7389,9 @@ static void sm1_generate_vsir_instr_swizzle(struct hlsl_ctx *ctx, struct vsir_pr
     swizzle = vsir_swizzle_from_hlsl(swizzle);
 
     src_param = &ins->src[0];
-    vsir_register_init(&src_param->reg, VKD3DSPR_TEMP, VKD3D_DATA_FLOAT, 1);
+    vsir_register_init(&src_param->reg, VKD3DSPR_TEMP, vsir_data_type_from_hlsl_instruction(ctx, val), 1);
     src_param->reg.idx[0].offset = val->reg.id;
+    src_param->reg.dimension = VSIR_DIMENSION_VEC4;
     src_param->swizzle = swizzle;
 }
 
@@ -7496,7 +7534,7 @@ static void sm1_generate_vsir_block(struct hlsl_ctx *ctx, struct hlsl_block *blo
                 break;
 
             case HLSL_IR_SWIZZLE:
-                sm1_generate_vsir_instr_swizzle(ctx, program, hlsl_ir_swizzle(instr));
+                generate_vsir_instr_swizzle(ctx, program, hlsl_ir_swizzle(instr));
                 break;
 
             default:
@@ -7557,6 +7595,25 @@ static void add_last_vsir_instr_to_block(struct hlsl_ctx *ctx, struct vsir_progr
     hlsl_block_add_instr(block, vsir_instr);
 }
 
+static void replace_instr_with_last_vsir_instr(struct hlsl_ctx *ctx,
+        struct vsir_program *program, struct hlsl_ir_node *instr)
+{
+    struct vkd3d_shader_location *loc;
+    struct hlsl_ir_node *vsir_instr;
+
+    loc = &program->instructions.elements[program->instructions.count - 1].location;
+
+    if (!(vsir_instr = hlsl_new_vsir_instruction_ref(ctx,
+            program->instructions.count - 1, instr->data_type, &instr->reg, loc)))
+    {
+        ctx->result = VKD3D_ERROR_OUT_OF_MEMORY;
+        return;
+    }
+
+    list_add_before(&instr->entry, &vsir_instr->entry);
+    hlsl_replace_node(instr, vsir_instr);
+}
+
 static void sm4_generate_vsir_instr_dcl_temps(struct hlsl_ctx *ctx, struct vsir_program *program,
         uint32_t temp_count, struct hlsl_block *block, const struct vkd3d_shader_location *loc)
 {
@@ -7589,6 +7646,41 @@ static void sm4_generate_vsir_instr_dcl_indexable_temp(struct hlsl_ctx *ctx,
     add_last_vsir_instr_to_block(ctx, program, block);
 }
 
+static void sm4_generate_vsir_block(struct hlsl_ctx *ctx, struct hlsl_block *block, struct vsir_program *program)
+{
+    struct hlsl_ir_node *instr, *next;
+
+    LIST_FOR_EACH_ENTRY_SAFE(instr, next, &block->instrs, struct hlsl_ir_node, entry)
+    {
+        if (instr->data_type)
+        {
+            if (instr->data_type->class != HLSL_CLASS_SCALAR && instr->data_type->class != HLSL_CLASS_VECTOR)
+            {
+                hlsl_fixme(ctx, &instr->loc, "Class %#x should have been lowered or removed.", instr->data_type->class);
+                break;
+            }
+        }
+
+        switch (instr->type)
+        {
+            case HLSL_IR_CALL:
+                vkd3d_unreachable();
+
+            case HLSL_IR_CONSTANT:
+                /* In SM4 all constants are inlined. */
+                break;
+
+            case HLSL_IR_SWIZZLE:
+                generate_vsir_instr_swizzle(ctx, program, hlsl_ir_swizzle(instr));
+                replace_instr_with_last_vsir_instr(ctx, program, instr);
+                break;
+
+            default:
+                break;
+        }
+    }
+}
+
 static void sm4_generate_vsir_add_function(struct hlsl_ctx *ctx,
         struct hlsl_ir_function_decl *func, uint64_t config_flags, struct vsir_program *program)
 {
@@ -7631,6 +7723,8 @@ static void sm4_generate_vsir_add_function(struct hlsl_ctx *ctx,
     list_move_head(&func->body.instrs, &block.instrs);
 
     hlsl_block_cleanup(&block);
+
+    sm4_generate_vsir_block(ctx, &func->body, program);
 }
 
 /* OBJECTIVE: Translate all the information from ctx and entry_func to the
diff --git a/libs/vkd3d-shader/tpf.c b/libs/vkd3d-shader/tpf.c
index e128d26a..8158f3ab 100644
--- a/libs/vkd3d-shader/tpf.c
+++ b/libs/vkd3d-shader/tpf.c
@@ -674,6 +674,7 @@ struct sm4_index_range_array
 struct vkd3d_sm4_lookup_tables
 {
     const struct vkd3d_sm4_opcode_info *opcode_info_from_sm4[VKD3D_SM4_OP_COUNT];
+    const struct vkd3d_sm4_opcode_info *opcode_info_from_vsir[VKD3DSIH_COUNT];
     const struct vkd3d_sm4_register_type_info *register_type_info_from_sm4[VKD3D_SM4_REGISTER_TYPE_COUNT];
     const struct vkd3d_sm4_register_type_info *register_type_info_from_vkd3d[VKD3DSPR_COUNT];
     const struct vkd3d_sm4_stat_field_info *stat_field_from_sm4[VKD3D_SM4_OP_COUNT];
@@ -1412,6 +1413,8 @@ struct tpf_compiler
     struct vkd3d_sm4_lookup_tables lookup;
     struct sm4_stat *stat;
 
+    int result;
+
     struct vkd3d_bytecode_buffer *buffer;
     struct dxbc_writer dxbc;
 };
@@ -1903,6 +1906,7 @@ static void init_sm4_lookup_tables(struct vkd3d_sm4_lookup_tables *lookup)
         const struct vkd3d_sm4_opcode_info *info = &opcode_table[i];
 
         lookup->opcode_info_from_sm4[info->opcode] = info;
+        lookup->opcode_info_from_vsir[info->handler_idx] = info;
     }
 
     for (i = 0; i < ARRAY_SIZE(register_type_table); ++i)
@@ -1929,6 +1933,24 @@ static const struct vkd3d_sm4_opcode_info *get_info_from_sm4_opcode(
     return lookup->opcode_info_from_sm4[sm4_opcode];
 }
 
+static const struct vkd3d_sm4_opcode_info *get_info_from_vsir_opcode(
+        const struct vkd3d_sm4_lookup_tables *lookup, enum vkd3d_shader_opcode vsir_opcode)
+{
+    if (vsir_opcode >= VKD3DSIH_COUNT)
+        return NULL;
+    return lookup->opcode_info_from_vsir[vsir_opcode];
+}
+
+static unsigned int opcode_info_get_dst_count(const struct vkd3d_sm4_opcode_info *info)
+{
+    return strnlen(info->dst_info, SM4_MAX_DST_COUNT);
+}
+
+static unsigned int opcode_info_get_src_count(const struct vkd3d_sm4_opcode_info *info)
+{
+    return strnlen(info->src_info, SM4_MAX_SRC_COUNT);
+}
+
 static const struct vkd3d_sm4_register_type_info *get_info_from_sm4_register_type(
         const struct vkd3d_sm4_lookup_tables *lookup, enum vkd3d_sm4_register_type sm4_type)
 {
@@ -2651,8 +2673,8 @@ static void shader_sm4_read_instruction(struct vkd3d_shader_sm4_parser *sm4, str
     ins->raw = false;
     ins->structured = false;
     ins->predicate = NULL;
-    ins->dst_count = strnlen(opcode_info->dst_info, SM4_MAX_DST_COUNT);
-    ins->src_count = strnlen(opcode_info->src_info, SM4_MAX_SRC_COUNT);
+    ins->dst_count = opcode_info_get_dst_count(opcode_info);
+    ins->src_count = opcode_info_get_src_count(opcode_info);
     ins->src = src_params = vsir_program_get_src_params(program, ins->src_count);
     if (!src_params && ins->src_count)
     {
@@ -2971,7 +2993,7 @@ int tpf_parse(const struct vkd3d_shader_compile_info *compile_info, uint64_t con
     return VKD3D_OK;
 }
 
-static void write_sm4_block(const struct tpf_compiler *tpf, const struct hlsl_block *block);
+static void write_sm4_block(struct tpf_compiler *tpf, const struct hlsl_block *block);
 
 static bool type_is_integer(const struct hlsl_type *type)
 {
@@ -6094,7 +6116,7 @@ static void write_sm4_expr(const struct tpf_compiler *tpf, const struct hlsl_ir_
     hlsl_release_string_buffer(tpf->ctx, dst_type_string);
 }
 
-static void write_sm4_if(const struct tpf_compiler *tpf, const struct hlsl_ir_if *iff)
+static void write_sm4_if(struct tpf_compiler *tpf, const struct hlsl_ir_if *iff)
 {
     struct sm4_instruction instr =
     {
@@ -6211,7 +6233,7 @@ static void write_sm4_load(const struct tpf_compiler *tpf, const struct hlsl_ir_
     write_sm4_instruction(tpf, &instr);
 }
 
-static void write_sm4_loop(const struct tpf_compiler *tpf, const struct hlsl_ir_loop *loop)
+static void write_sm4_loop(struct tpf_compiler *tpf, const struct hlsl_ir_loop *loop)
 {
     struct sm4_instruction instr =
     {
@@ -6395,7 +6417,7 @@ static void write_sm4_store(const struct tpf_compiler *tpf, const struct hlsl_ir
     write_sm4_instruction(tpf, &instr);
 }
 
-static void write_sm4_switch(const struct tpf_compiler *tpf, const struct hlsl_ir_switch *s)
+static void write_sm4_switch(struct tpf_compiler *tpf, const struct hlsl_ir_switch *s)
 {
     const struct hlsl_ir_node *selector = s->selector.node;
     struct hlsl_ir_switch_case *c;
@@ -6456,7 +6478,46 @@ static void write_sm4_swizzle(const struct tpf_compiler *tpf, const struct hlsl_
     write_sm4_instruction(tpf, &instr);
 }
 
-static void tpf_handle_instruction(const struct tpf_compiler *tpf, const struct vkd3d_shader_instruction *ins)
+static void tpf_simple_instruction(struct tpf_compiler *tpf, const struct vkd3d_shader_instruction *ins)
+{
+    const struct vkd3d_sm4_opcode_info *info;
+    struct sm4_instruction instr = {0};
+    unsigned int dst_count, src_count;
+
+    info = get_info_from_vsir_opcode(&tpf->lookup, ins->opcode);
+    VKD3D_ASSERT(info);
+
+    dst_count = opcode_info_get_dst_count(info);
+    src_count = opcode_info_get_src_count(info);
+
+    if (ins->dst_count != dst_count)
+    {
+        ERR("Invalid destination count %u for vsir instruction %#x (expected %u).\n",
+                ins->dst_count, ins->opcode, dst_count);
+        tpf->result = VKD3D_ERROR_INVALID_SHADER;
+        return;
+    }
+    if (ins->src_count != src_count)
+    {
+        ERR("Invalid source count %u for vsir instruction %#x (expected %u).\n",
+                ins->src_count, ins->opcode, src_count);
+        tpf->result = VKD3D_ERROR_INVALID_SHADER;
+        return;
+    }
+
+    instr.opcode = info->opcode;
+    instr.dst_count = ins->dst_count;
+    instr.src_count = ins->src_count;
+
+    for (unsigned int i = 0; i < ins->dst_count; ++i)
+        instr.dsts[i] = ins->dst[i];
+    for (unsigned int i = 0; i < ins->src_count; ++i)
+        instr.srcs[i] = ins->src[i];
+
+    write_sm4_instruction(tpf, &instr);
+}
+
+static void tpf_handle_instruction(struct tpf_compiler *tpf, const struct vkd3d_shader_instruction *ins)
 {
     switch (ins->opcode)
     {
@@ -6468,13 +6529,17 @@ static void tpf_handle_instruction(const struct tpf_compiler *tpf, const struct
             tpf_dcl_indexable_temp(tpf, &ins->declaration.indexable_temp);
             break;
 
+        case VKD3DSIH_MOV:
+            tpf_simple_instruction(tpf, ins);
+            break;
+
         default:
             vkd3d_unreachable();
             break;
     }
 }
 
-static void write_sm4_block(const struct tpf_compiler *tpf, const struct hlsl_block *block)
+static void write_sm4_block(struct tpf_compiler *tpf, const struct hlsl_block *block)
 {
     const struct hlsl_ir_node *instr;
     unsigned int vsir_instr_idx;
@@ -6766,7 +6831,13 @@ int tpf_compile(struct vsir_program *program, uint64_t config_flags,
     tpf_write_sfi0(&tpf);
     tpf_write_stat(&tpf);
 
-    if (!(ret = ctx->result))
+    ret = VKD3D_OK;
+    if (ctx->result)
+        ret = ctx->result;
+    if (tpf.result)
+        ret = tpf.result;
+
+    if (!ret)
         ret = dxbc_writer_write(&tpf.dxbc, out);
     for (i = 0; i < tpf.dxbc.section_count; ++i)
         vkd3d_shader_free_shader_code(&tpf.dxbc.sections[i].data);
diff --git a/libs/vkd3d-shader/vkd3d_shader_private.h b/libs/vkd3d-shader/vkd3d_shader_private.h
index a2afa0fe..07590b66 100644
--- a/libs/vkd3d-shader/vkd3d_shader_private.h
+++ b/libs/vkd3d-shader/vkd3d_shader_private.h
@@ -585,6 +585,8 @@ enum vkd3d_shader_opcode
     VKD3DSIH_XOR,
 
     VKD3DSIH_INVALID,
+
+    VKD3DSIH_COUNT,
 };
 
 enum vkd3d_shader_register_type