UVM phase机制源码探微

忿忿的泥巴坨 · 2024-9-19 16:19:21

1. uvm_phase

UVM的phase机制承袭在差别时间做差别事情的计划哲学，重要提供了function和task两种类型的phase，在function phase中重要做一些实例化，各组件的连接以及仿真结束后的统计、report等工作，这些事情是不必要泯灭仿真时间的；重要的仿真过程是在task phase即run_phase中运行的，与run_phase同步运行的还有12个run-time task phase, 这些细分的phase可以实现对仿真行为更加精致的控制，比如在reset_phase对DUT做reset，configure_phase对DUT做一些设置, 在main_phase中灌输激励等等，各个function phase是严格按照下图中从上到下的顺序来执行的（图中蓝色为function phase，绿色为task phase）：

图1.1 uvm_phase graph

2. uvm_phase的运行和调度

我们知道，UVM通过phase的调度来进行差别阶段的仿真，那当我们跑一个testcase的时候，验证平台是怎么开始仿真的呢？我们一般会在平台的顶层module TB里边来调用run_test()开始仿真：

//uvm_root.svh
task uvm_root::run_test(string test_name="");
...
process phase_runner_proc;
...
if (test_name != "") begin
uvm_coreservice_t cs = uvm_coreservice_t::get();
uvm_factory factory=cs.get_factory();
if(m_children.exists("uvm_test_top")) begin
uvm_report_fatal("TTINST",
"An uvm_test_top already exists via a previous call to run_test", UVM_NONE);
#0; // forces shutdown because $finish is forked
end
$cast(uvm_test_top, factory.create_component_by_name(test_name,
"", "uvm_test_top", null));
...
fork begin
// spawn the phase runner task
phase_runner_proc = process::self();
uvm_phase::m_run_phases();
end
join_none
#0; // let the phase runner start
wait (m_phase_all_done == 1);
// clean up after ourselves
phase_runner_proc.kill();
...
if (finish_on_completion)
$finish;
endtask

复制代码

在这个函数中UVM会通过传入的下令行参数+UVM_TESTNAME=xxx来获取testcase的名字，然后创建一个uvm_test_top的component节点，这个uvm_test_top句柄指向当前仿真testcase class的实例，随后会在此中例化env及整个平台hirearchy。别的，SV中有process进程的用法，uvm会用fork…join_none来启动一个进程phase_runner_proc，在此中调用m_run_phases()，在这个task中完成进行对phase的部分调度（之所以说是部分调度，在于其实还有一个background process在同步运行，下文会提到），同时uvm不停等待全部phase全部完成的信号m_phase_all_done，一旦该信号被拉高则杀掉phase_runner_proc进程，调用$finish结束本次仿真。

//uvm_phase.svh
class uvm_phase extends uvm_object;
local static mailbox #(uvm_phase) m_phase_hopper = new();
extern static task m_run_phases();
...
endclass
task uvm_phase::m_run_phases();
uvm_root top;
uvm_coreservice_t cs;
cs = uvm_coreservice_t::get();
top = cs.get_root();
// initiate by starting first phase in common domain
begin
uvm_phase ph = uvm_domain::get_common_domain();
void'(m_phase_hopper.try_put(ph));
end
forever begin
uvm_phase phase;
m_phase_hopper.get(phase);
fork
begin
phase.execute_phase();
end
join_none
#0; // let the process start running
end
endtask

复制代码

m_run_phases()是界说在uvm_phase类的静态task，其调用get_common_domain()来获得一个uvm_phase类型的句柄，然后将其放入邮箱m_phase_hopper中，这个邮箱是static类型的，存放的对象是uvm_phase类型，UVM平台中全部的phase实例对象会共用这一个邮箱。随后在forever循环中该邮箱会不绝地被查抄，一旦有新的phase放入，则立即取出来用fork…join_none启动新线程调用该phase的execute_phase()。
2.1. get_common_domain()

这个函数做的事情就是把全部的function phase和task phase按照先后顺序排排队，安排好先后顺序，按照图1.1中所示的那样编成一个domain，取个名字叫做"common"，作为uvm default的domain。这个函数中先查抄m_common_domain是否为null，我们可以大致猜到这是一个common_domain的实例句柄，很明显这里是采用了单例模式，若系统中已经存在这个实例，则直接返回该实例句柄。若是第一次调用该函数，m_common_domain为null，则调用new()函数创建一个句柄为domain，名字为"common"的实例，然后调用add()函数将8个function phase和run_phase按顺序加入，add()函数做的其实就是给这些phase排排队加入名为"common_domain"的phase运行图，我们后面会详细来看add()函数的实现。之后把该domain句柄分别存入以"common"为索引的数组m_domains和赋给m_common_domain。get_uvm_domain()实际上就是获取12个run-time task phase的运行图，末了会再次调用add()将其加入m_common_domain的运行图中，with_phase参数传入的是之前run_phase的句柄，表明12个run-time phase是以跟run_phase并行运行的方式加入到m_common_domain中。

static function uvm_domain get_common_domain();
uvm_domain domain;
uvm_phase schedule;
if (m_common_domain != null)
return m_common_domain;
domain = new("common");
domain.add(uvm_build_phase::get());
domain.add(uvm_connect_phase::get());
domain.add(uvm_end_of_elaboration_phase::get());
domain.add(uvm_start_of_simulation_phase::get());
domain.add(uvm_run_phase::get());
domain.add(uvm_extract_phase::get());
domain.add(uvm_check_phase::get());
domain.add(uvm_report_phase::get());
domain.add(uvm_final_phase::get());
m_domains["common"] = domain;
...
m_common_domain = domain;
domain = get_uvm_domain();
m_common_domain.add(domain,
.with_phase(m_common_domain.find(uvm_run_phase::get())));
return m_common_domain;
endfunction

复制代码

2.1.1. phase_type & phase_state

为了更方便的对phase进行调度，UVM界说了差别的phase_type，如UVM_PHASE_NODE代表全部的（无论function/task phase）phase的一个实例(节点)；UVM_PHASE_DOMAIN代表在此中界说了差别phase线性/并行运行关系的一组phase的运行图（雷同于从上海发往北京的高铁调度时刻表，phase就像是运行在其间的一列列班车，哪个时间段该发行哪班车，都会清清楚楚纪录在时刻表里，固然高铁不可能每一个时刻都会发出一班车，不像phase之间是无缝衔接的），此中可以包含许多的UVM_PHASE_NODE；而UVM_PHASE_SCHEDULE表示一个微缩版的UVM_PHASE_DOMAIN，它此中只界说了12个run-time task phase的运行图，等等。

//uvm_object_globals.svh
typedef enum { UVM_PHASE_IMP,
UVM_PHASE_NODE,
UVM_PHASE_TERMINAL,
UVM_PHASE_SCHEDULE,
UVM_PHASE_DOMAIN,
UVM_PHASE_GLOBAL
} uvm_phase_type;

复制代码

别的UVM还界说了phase_state，用来表示某phase当前的差别运行状态，如UVM_PHASE_SYNCING表示差别domain的run-time task phase正在进行同步，UVM_PHASE_EXECUTING表示当前phase正在执行，等等。

//uvm_object_globals.svh
typedef enum { UVM_PHASE_UNINITIALIZED = 0,
UVM_PHASE_DORMANT = 1,
UVM_PHASE_SCHEDULED = 2,
UVM_PHASE_SYNCING = 4,
UVM_PHASE_STARTED = 8,
UVM_PHASE_EXECUTING = 16,
UVM_PHASE_READY_TO_END = 32,
UVM_PHASE_ENDED = 64,
UVM_PHASE_CLEANUP = 128,
UVM_PHASE_DONE = 256,
UVM_PHASE_JUMPING = 512
} uvm_phase_state;

复制代码

2.1.2. uvm_phase class

2.1.2.1. new()

uvm_phase类继承自uvm_object，在此中界说了一些变量如m_phase_type和m_state, 分别是uvm_phase_type和uvm_phase_state类型的，uvm_phase类型的句柄m_parent指向的是该phase的上一个层级，此中还界说了uvm_phase类型的m_end_node，这个变量是用来表示一个uvm_domain和uvm_schedule的结束节点。别的，两个bit类型的联合数组以uvm_phase类型为索引，用来纪录各个phase间的顺序关系，此中m_predecessors[]用来纪录当前phase的全部前置phase,比如若当前phase是一个connect_phase，则build_phase在其之前运行，是它的一个前置phase，则connect_phase的实例中m_predecessors[build_phase] = 1；同理，m_successors[]用于纪录当前phase的全部后继phase，对于一个build_phase实例来说，其m_successors[connect_phase] = 1；

//uvm_phase.svh
class uvm_phase extends uvm_object;
protected uvm_phase_type m_phase_type;
protected uvm_phase m_parent; // our 'schedule' node [or points 'up' one level]
uvm_phase m_imp; // phase imp to call when we execute this node
local uvm_phase_state m_state;
...
protected bit m_predecessors[uvm_phase];
protected bit m_successors[uvm_phase];
protected uvm_phase m_end_node;
...
endclass
function uvm_phase::new(string name="uvm_phase",
uvm_phase_type phase_type=UVM_PHASE_SCHEDULE,
uvm_phase parent=null);
super.new(name);
m_phase_type = phase_type;
// The common domain is the only thing that initializes m_state. All
// other states are initialized by being 'added' to a schedule.
if ((name == "common") &&
(phase_type == UVM_PHASE_DOMAIN))
m_state = UVM_PHASE_DORMANT;
m_parent = parent;
...
if (parent == null && (phase_type == UVM_PHASE_SCHEDULE ||
phase_type == UVM_PHASE_DOMAIN )) begin
//m_parent = this;
m_end_node = new({name,"_end"}, UVM_PHASE_TERMINAL, this);
this.m_successors[m_end_node] = 1;
m_end_node.m_predecessors[this] = 1;
end
endfunction

复制代码

来看uvm_phase类的new()函数，它有三个参数，在创建一个uvm_phase类型的实例时分别传入名字，phase_type和和parent，它分别把传入的phase_type和parent赋给m_phase_type和m_parent，若传入的phase_type是UVM_PHASE_SCHEDULE或UVM_PHASE_DOMAIN，表明当前正在创建一个多phase的运行图（domain/schedule），则调用new()创建一个m_end_node的实例对象，这个m_end_node的phase_type是UVM_PHASE_TERMINAL类型，parent就是这个运行图，然后把m_end_node作为这个运行图的后继phase纪录下来，这段代码的意图就是一旦我们想创建一个新的domain/schedule对象，就必要先给它安排一个UVM_PHASE_TERMINAL类型的节点，无论运行图内部怎么调度这些phase，最终运行完毕总是要结束的。
2.1.2.1. uvm_common_phases

UVM的build_phase类和final_phase类继承自uvm_topdown_phase类，之所以叫"topdown"是因为在UVM平台hireachy各个component节点实例化的时候是自上而下执行的，与之相反其余的function phase类型都继承自uvm_bottomup_phase类，比如connect_phase的执行在UVM平台hireachy各个component节点间是自下而上执行。别的run_phase和12个run-time phase类都继承自uvm_task_phase类。这几种类型new()函数中phase_type参数都会传入UVM_PHASE_IMP来创建一个phase实例。此中的关键是traverse()函数，此中会根据当前phase执行的状态来调用当前component的差别函数，此中UVM预先界说了许多callback如phase_started()/phase_ended()，用户可以在component中来扩展这些callback，这些callback会在全部的phase（包括全部function和task phase）执行前后进行调用。我们留意到，在当前component的traverse()执行到末了的时候，会查抄它全部的子节点并调用traverse()来以此执行此phase的内容，所以我们说继承自uvm_bottomup_phase的phase如build_phase的执行是自上而下的。与之相反，uvm_bottomup_phase类的traverse()函数是先在其全部子节点中调用traverse()执行，末了执行当前component，因而继承自uvm_bottomup_phase的phase如connect_phase的执行是自下而上的。uvm_task_phase的执行顺序与uvm_bottomup_phase雷同，也是自下而上发起，但却是同时执行的，并不会等待子节点phase执行完毕才会执行上一级节点。别的，function phase在UVM_PHASE_EXECUTING状态，会调用execute()函数，这个函数会启动一个新进程并调用exec_func()。

//uvm_topdown_phase.svh
virtual class uvm_topdown_phase extends uvm_phase;
// Function: new
//
// Create a new instance of a top-down phase
//
function new(string name);
super.new(name,UVM_PHASE_IMP);
endfunction
virtual function void traverse(uvm_component comp,
uvm_phase phase,
uvm_phase_state state);
string name;
uvm_domain phase_domain = phase.get_domain();
uvm_domain comp_domain = comp.get_domain();
if (m_phase_trace)
`uvm_info("PH_TRACE",$sformatf("topdown-phase phase=%s state=%s comp=%s comp.domain=%s phase.domain=%s",
phase.get_name(), state.name(), comp.get_full_name(),comp_domain.get_name(),phase_domain.get_name()),
UVM_DEBUG)
if (phase_domain == uvm_domain::get_common_domain() ||
phase_domain == comp_domain) begin
case (state)
UVM_PHASE_STARTED: begin
comp.m_current_phase = phase;
comp.m_apply_verbosity_settings(phase);
comp.phase_started(phase);
end
UVM_PHASE_EXECUTING: begin
if (!(phase.get_name() == "build" && comp.m_build_done)) begin
uvm_phase ph = this;
comp.m_phasing_active++;
if (comp.m_phase_imps.exists(this))
ph = comp.m_phase_imps[this];
ph.execute(comp, phase);
comp.m_phasing_active--;
end
end
UVM_PHASE_READY_TO_END: begin
comp.phase_ready_to_end(phase);
end
UVM_PHASE_ENDED: begin
comp.phase_ended(phase);
comp.m_current_phase = null;
end
default:
`uvm_fatal("PH_BADEXEC","topdown phase traverse internal error")
endcase
end
if(comp.get_first_child(name))
do
traverse(comp.get_child(name), phase, state);
while(comp.get_next_child(name));
endfunction
// Function: execute
//
// Executes the top-down phase ~phase~ for the component ~comp~.
//
virtual function void execute(uvm_component comp,
uvm_phase phase);
// reseed this process for random stability
process proc = process::self();
proc.srandom(uvm_create_random_seed(phase.get_type_name(), comp.get_full_name()));
comp.m_current_phase = phase;
exec_func(comp,phase);
endfunction
endclass
//uvm_common_phases.svh
class uvm_build_phase extends uvm_topdown_phase;
virtual function void exec_func(uvm_component comp, uvm_phase phase);
comp.build_phase(phase);
endfunction
local static uvm_build_phase m_inst;
static const string type_name = "uvm_build_phase";
// Function: get
// Returns the singleton phase handle
//
static function uvm_build_phase get();
if(m_inst == null)
m_inst = new();
return m_inst;
endfunction
protected function new(string name="build");
super.new(name);
endfunction
virtual function string get_type_name();
return type_name;
endfunction
endclass

复制代码

以uvm_build_phase为例，其实exec_func()就是调用了我们界说在某component的build_phase来执行。此中界说了uvm_build_phase类型的静态变量m_inst，代表这个phase的一个实例，当调用静态函数get()时会返回该唯一实例对象句柄。而task phase在UVM_PHASE_EXECUTING状态，也会调用execute()函数，这个函数会用fork…join_none启动一个新进程并调用exec_task()，在exec_task()返回前用m_num_procs_not_yet_returned来纪录当前有多少个component在运行这个phase，以判断何时可以结束该phase。与function phase雷同，exec_task()也是界说在task phase中调用当前component的同名task phase执行，get()函数也会返回该phase唯一实例对象句柄m_inst。

//uvm_task_phase.svh
virtual class uvm_task_phase extends uvm_phase
...
virtual function void execute(uvm_component comp,
uvm_phase phase);
fork
begin
process proc;
// reseed this process for random stability
proc = process::self();
proc.srandom(uvm_create_random_seed(phase.get_type_name(), comp.get_full_name()));
phase.m_num_procs_not_yet_returned++;
exec_task(comp,phase);
phase.m_num_procs_not_yet_returned--;
end
join_none
endfunction
endclass

复制代码

2.1.3. uvm_domain class

uvm_domain类继承自uvm_phase，此中界说了以字符串为索引的内容为uvm_domain类型的静态联合数组m_domains[]，在uvm_domain的new()函数中首先调用super即uvm_phase类型的new()函数创建一个phase_type为UVM_PHASE_DOMAIN的实例，然后把传入的名字作为索引，将该实例对象存入m_domains[]。

//uvm_domain.svh
class uvm_domain extends uvm_phase;
static local uvm_domain m_common_domain;
static local uvm_domain m_domains[string];
...
function new(string name);
super.new(name,UVM_PHASE_DOMAIN);
if (m_domains.exists(name))
`uvm_error("UNIQDOMNAM", $sformatf("Domain created with non-unique name '%s'", name))
m_domains[name] = this;
endfunction
...
endclass

复制代码

2.1.4. get_uvm_domain()

这个函数就是将12个run-time phases的运行顺序排好然后返回一个uvm_domain类型的运行图句柄m_uvm_domain。

//uvm_domain.svh
class uvm_domain extends uvm_phase;
static local uvm_domain m_uvm_domain; // run-time
static local uvm_phase m_uvm_schedule;
...
static function void add_uvm_phases(uvm_phase schedule);
schedule.add(uvm_pre_reset_phase::get());
schedule.add(uvm_reset_phase::get());
schedule.add(uvm_post_reset_phase::get());
schedule.add(uvm_pre_configure_phase::get());
schedule.add(uvm_configure_phase::get());
schedule.add(uvm_post_configure_phase::get());
schedule.add(uvm_pre_main_phase::get());
schedule.add(uvm_main_phase::get());
schedule.add(uvm_post_main_phase::get());
schedule.add(uvm_pre_shutdown_phase::get());
schedule.add(uvm_shutdown_phase::get());
schedule.add(uvm_post_shutdown_phase::get());
endfunction
// Function: get_uvm_domain
//
// Get a handle to the singleton ~uvm~ domain
//
static function uvm_domain get_uvm_domain();
if (m_uvm_domain == null) begin
m_uvm_domain = new("uvm");
m_uvm_schedule = new("uvm_sched", UVM_PHASE_SCHEDULE);
add_uvm_phases(m_uvm_schedule);
m_uvm_domain.add(m_uvm_schedule);
end
return m_uvm_domain;
endfunction
endclass

复制代码

静态函数get_uvm_domain()首先分别创建了名为"uvm"和"uvm_sched"的对象并分别赋值给uvm_domain和uvm_phase类型的句柄m_uvm_domain和m_uvm_schedule，二者phase_type分别为UVM_PHASE_COMMON和UVM_PHASE_SCHEDULE，然后调用add_uvm_phases()函数，将12个run-time phases通过add()函数按先后顺序加入m_uvm_schedule中，此时m_uvm_schedule就是一个运行图，末了再将这个运行图schedule加入m_uvm_domain这个domain中，末了回到get_common_domain()中，把m_uvm_domain加入到m_common_domain中，同时设置with_phase为run_phase，就是把m_uvm_domain中的这12个run-time phase和run_phase同步运行。
2.2. add()

我们在之前调用add()函数时无外乎是四种情况，分别为
1， domain.add(uvm_build_phase::get()); //向uvm_common_domain中添加funtion phase和run_phase
2， schedule.add(uvm_pre_reset_phase::get()); //向m_uvm_schedule中添加12个run-time phases
3， m_uvm_domain.add(m_uvm_schedule); //将m_uvm_schedule加入m_uvm_domain
4， m_common_domain.add(domain,
.with_phase(m_common_domain.find(uvm_run_phase::get()))); //将包含12个run-time phases 的m_uvm_domain加入uvm_common_domain并使它们跟run_phase同步运行
来看add()函数，我们隐去了此中关于debug的部分：

function void uvm_phase::add(uvm_phase phase,
uvm_phase with_phase=null,
uvm_phase after_phase=null,
uvm_phase before_phase=null);
uvm_phase new_node, begin_node, end_node, tmp_node;
uvm_phase_state_change state_chg;
if (phase == null)
`uvm_fatal("PH/NULL", "add: phase argument is null")
if (with_phase != null && with_phase.get_phase_type() == UVM_PHASE_IMP) begin
string nm = with_phase.get_name();
with_phase = find(with_phase);
if (with_phase == null)
`uvm_fatal("PH_BAD_ADD",
{"cannot find with_phase '",nm,"' within node '",get_name(),"'"})
end
if (before_phase != null && before_phase.get_phase_type() == UVM_PHASE_IMP) begin
string nm = before_phase.get_name();
before_phase = find(before_phase);
if (before_phase == null)
`uvm_fatal("PH_BAD_ADD",
{"cannot find before_phase '",nm,"' within node '",get_name(),"'"})
end
if (after_phase != null && after_phase.get_phase_type() == UVM_PHASE_IMP) begin
string nm = after_phase.get_name();
after_phase = find(after_phase);
if (after_phase == null)
`uvm_fatal("PH_BAD_ADD",
{"cannot find after_phase '",nm,"' within node '",get_name(),"'"})
end
if (with_phase != null && (after_phase != null || before_phase != null))
`uvm_fatal("PH_BAD_ADD",
"cannot specify both 'with' and 'before/after' phase relationships")
if (before_phase == this || after_phase == m_end_node || with_phase == m_end_node)
`uvm_fatal("PH_BAD_ADD",
"cannot add before begin node, after end node, or with end nodes")
// If we are inserting a new "leaf node"
if (phase.get_phase_type() == UVM_PHASE_IMP) begin
uvm_task_phase tp;
new_node = new(phase.get_name(),UVM_PHASE_NODE,this);
new_node.m_imp = phase;
begin_node = new_node;
end_node = new_node;
// The phase_done objection is only required
// for task-based nodes
if ($cast(tp, phase)) begin
if (new_node.get_name() == "run") begin
new_node.phase_done = uvm_test_done_objection::get();
end
else begin
new_node.phase_done = uvm_objection::type_id::create({phase.get_name(), "_objection"});
end
end
end
// We are inserting an existing schedule
else begin
begin_node = phase;
end_node = phase.m_end_node;
phase.m_parent = this;
end
// If 'with_phase' is us, then insert node in parallel
/*
if (with_phase == this) begin
after_phase = this;
before_phase = m_end_node;
end
*/
// If no before/after/with specified, insert at end of this schedule
if (with_phase == null && after_phase == null && before_phase == null) begin
before_phase = m_end_node;
end
// INSERT IN PARALLEL WITH 'WITH' PHASE
if (with_phase != null) begin
begin_node.m_predecessors = with_phase.m_predecessors;
end_node.m_successors = with_phase.m_successors;
foreach (with_phase.m_predecessors[pred])
pred.m_successors[begin_node] = 1;
foreach (with_phase.m_successors[succ])
succ.m_predecessors[end_node] = 1;
end
// INSERT BEFORE PHASE
else if (before_phase != null && after_phase == null) begin
begin_node.m_predecessors = before_phase.m_predecessors;
end_node.m_successors[before_phase] = 1;
foreach (before_phase.m_predecessors[pred]) begin
pred.m_successors.delete(before_phase);
pred.m_successors[begin_node] = 1;
end
before_phase.m_predecessors.delete();
before_phase.m_predecessors[end_node] = 1;
end
// INSERT AFTER PHASE
else if (before_phase == null && after_phase != null) begin
end_node.m_successors = after_phase.m_successors;
begin_node.m_predecessors[after_phase] = 1;
foreach (after_phase.m_successors[succ]) begin
succ.m_predecessors.delete(after_phase);
succ.m_predecessors[end_node] = 1;
end
after_phase.m_successors.delete();
after_phase.m_successors[begin_node] = 1;
end
// IN BETWEEN 'BEFORE' and 'AFTER' PHASES
else if (before_phase != null && after_phase != null) begin
if (!after_phase.is_before(before_phase)) begin
`uvm_fatal("PH_ADD_PHASE",{"Phase '",before_phase.get_name(),
"' is not before phase '",after_phase.get_name(),"'"})
end
// before and after? add 1 pred and 1 succ
begin_node.m_predecessors[after_phase] = 1;
end_node.m_successors[before_phase] = 1;
after_phase.m_successors[begin_node] = 1;
before_phase.m_predecessors[end_node] = 1;
if (after_phase.m_successors.exists(before_phase)) begin
after_phase.m_successors.delete(before_phase);
before_phase.m_predecessors.delete(after_phase);
end
end // if (before_phase != null && after_phase != null)
// Transition nodes to DORMANT state
if (new_node == null)
tmp_node = phase;
else
tmp_node = new_node;
state_chg = uvm_phase_state_change::type_id::create(tmp_node.get_name());
state_chg.m_phase = tmp_node;
state_chg.m_jump_to = null;
state_chg.m_prev_state = tmp_node.m_state;
tmp_node.m_state = UVM_PHASE_DORMANT;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(tmp_node, state_chg))
endfunction

复制代码

对于第一种调用，函数首先会判断参数中phase是否为null，然后依次判断各个参数with_phase/after_phase/before_phase是否为null，我们可以略过这些语句。接下来判断传入的phase是否为UVM_PHASE_IMP类型，很明显uvm_build_phase::get()返回值符合这个条件，于是创建一个UVM_PHASE_NODE为phase_type的uvm_phase类型的新实例对象new_node，这个new_node的唯一对象句柄m_imp就指向当前传入的phase_type为UVM_PHASE_IMP的对象，这个phase_type为UVM_PHASE_NODE的实例句柄会更换参数传入的phase_type为UVM_PHASE_IMP的对象存入phase的运行图。接下来把new_node分别赋值给begin_node和end_node。接下来判断当前phase是否是一个task phase类型，若是，则在此中实例化uvm_objection的对象，这里临时按下不表。接下来判断若with_phase/after_phase/before_phase均为null，则before_phase指向m_end_node，这个m_end_node就是当前调用add()函数的phase_type为UVM_PHASE_DOMAIN大概UVM_PHASE_SCHEDULE的结束节点。这里before_phase指向某个运行图（domain/schedule）的结束节点，意图很明显，就是要把当前phase加入到这个运行图的m_end_node节点之前，作为运行图的末了一个phase。后面的代码判断若before_phase不为null，则
1，把当前运行图的m_end_node.m_predecessors[]作为新加入phase的m_predecessors[]；
2，把m_end_node作为新加入phase的m_successors，即phase.m_successors[m_end_node]=1；
3，对于m_end_node的全部m_predecessors，删除m_end_node作为他们的m_successors的纪录，并把新加入的phase作为它们的m_successors；
4，把m_end_node的全部m_predecessors纪录删除，并把新加入的phase作为其唯一的m_predecessors。
对于第一种情况，若调用domain.add(uvm_build_phase::get());之后，则运行图如下：

若依次调用add()添加全部function phase和run_phase到uvm_common_domain之后，所形成的运行图如如下：

第二种调用情况雷同，是把12个run-time task phases加入到m_uvm_schedule中形成顺序运行图，如下：

接下来看第三种调用add()的方式：m_uvm_domain.add(m_uvm_schedule);将m_uvm_schedule加入到m_uvm_domain的运行图中，由于传入的phase参数不是UVM_PHASE_IMP类型，所以函数会直接将before_phase指向m_uvm_domain的m_end_node，将m_uvm_schedule放在m_uvm_domain::m_end_node之前：

第四种调用方式给参数with_phase传入了find()返回值，这个find()函数是界说在uvm_phase类中的函数，它会调用m_find_predecessor()和m_find_successor()函数遍历当前domain中的全部前置和后置phase，返回要查找phase的唯一实例句柄m_imp。当with_phase传入run_phase的实例句柄时，add()函数做了以下事情：
1，把run_phase的全部m_predecessors作为m_uvm_domain的m_predecessors；
2，把run_phase的全部m_successors作为m_uvm_domain::m_end_node的m_successors；
3，对于run_phase全部的m_predecessors，把m_uvm_domain作为它们的m_successors；
4，对于run_phase全部的m_successors，把m_uvm_domain::m_end_node作为它们的m_predecessors。
最终调用get_common_domain()所形成的m_common_domain的phase运行图如下：

2.3. execute_phase()

在task m_run_phases()中调用uvm_domain::get_common_domain()拿到UVM default domain m_common_domain的句柄ph后，会把ph放入邮箱m_phase_hopper，forever循环从邮箱中拿到ph后会启动fork…join_none线程调用execute_phase()。该task代码如下：

//uvm_phase.svh
task uvm_phase::execute_phase();
uvm_task_phase task_phase;
uvm_root top;
uvm_phase_state_change state_chg;
uvm_coreservice_t cs;
cs = uvm_coreservice_t::get();
top = cs.get_root();
// If we got here by jumping forward, we must wait for
// all its predecessor nodes to be marked DONE.
// (the next conditional speeds this up)
// Also, this helps us fast-forward through terminal (end) nodes
foreach (m_predecessors[pred])
wait (pred.m_state == UVM_PHASE_DONE);
// If DONE (by, say, a forward jump), return immed
if (m_state == UVM_PHASE_DONE)
return;
...
// If we're a schedule or domain, then "fake" execution
if (m_phase_type != UVM_PHASE_NODE) begin
state_chg.m_prev_state = m_state;
m_state = UVM_PHASE_STARTED;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(this, state_chg))
#0;
state_chg.m_prev_state = m_state;
m_state = UVM_PHASE_EXECUTING;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(this, state_chg))
#0;
end
else begin // PHASE NODE
//---------
// STARTED:
//---------
state_chg.m_prev_state = m_state;
m_state = UVM_PHASE_STARTED;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(this, state_chg))
m_imp.traverse(top,this,UVM_PHASE_STARTED);
m_ready_to_end_count = 0 ; // reset the ready_to_end count when phase starts
#0; // LET ANY WAITERS WAKE UP
//if (m_imp.get_phase_type() != UVM_PHASE_TASK) begin
if (!$cast(task_phase,m_imp)) begin
//-----------
// EXECUTING: (function phases)
//-----------
state_chg.m_prev_state = m_state;
m_state = UVM_PHASE_EXECUTING;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(this, state_chg))
#0; // LET ANY WAITERS WAKE UP
m_imp.traverse(top,this,UVM_PHASE_EXECUTING);
end
else begin
m_executing_phases[this] = 1;
state_chg.m_prev_state = m_state;
m_state = UVM_PHASE_EXECUTING;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(this, state_chg))
fork : master_phase_process
begin
m_phase_proc = process::self();
//-----------
// EXECUTING: (task phases)
//-----------
task_phase.traverse(top,this,UVM_PHASE_EXECUTING);
wait(0); // stay alive for later kill
end
join_none
uvm_wait_for_nba_region(); //Give sequences, etc. a chance to object
// Now wait for one of three criterion for end-of-phase.
fork
begin // guard
fork
// JUMP
begin
wait (m_premature_end);
`UVM_PH_TRACE("PH/TRC/EXE/JUMP","PHASE EXIT ON JUMP REQUEST",this,UVM_DEBUG)
end
// WAIT_FOR_ALL_DROPPED
begin
bit do_ready_to_end ; // bit used for ready_to_end iterations
// OVM semantic: don't end until objection raised or stop request
if (phase_done.get_objection_total(top) ||
m_use_ovm_run_semantic && m_imp.get_name() == "run") begin
if (!phase_done.m_top_all_dropped)
phase_done.wait_for(UVM_ALL_DROPPED, top);
`UVM_PH_TRACE("PH/TRC/EXE/ALLDROP","PHASE EXIT ALL_DROPPED",this,UVM_DEBUG)
end
else begin
if (m_phase_trace) `UVM_PH_TRACE("PH/TRC/SKIP","No objections raised, skipping phase",this,UVM_LOW)
end
wait_for_self_and_siblings_to_drop() ;
do_ready_to_end = 1;
//--------------
// READY_TO_END:
//--------------
while (do_ready_to_end) begin
uvm_wait_for_nba_region(); // Let all siblings see no objections before traverse might raise another
`UVM_PH_TRACE("PH_READY_TO_END","PHASE READY TO END",this,UVM_DEBUG)
m_ready_to_end_count++;
if (m_phase_trace)
`UVM_PH_TRACE("PH_READY_TO_END_CB","CALLING READY_TO_END CB",this,UVM_HIGH)
state_chg.m_prev_state = m_state;
m_state = UVM_PHASE_READY_TO_END;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(this, state_chg))
if (m_imp != null)
m_imp.traverse(top,this,UVM_PHASE_READY_TO_END);
uvm_wait_for_nba_region(); // Give traverse targets a chance to object
wait_for_self_and_siblings_to_drop();
do_ready_to_end = (m_state == UVM_PHASE_EXECUTING) && (m_ready_to_end_count < max_ready_to_end_iter) ; //when we don't wait in task above, we drop out of while loop
end
end
// TIMEOUT
begin
if (this.get_name() == "run") begin
if (top.phase_timeout == 0)
wait(top.phase_timeout != 0);
if (m_phase_trace)
`UVM_PH_TRACE("PH/TRC/TO_WAIT", $sformatf("STARTING PHASE TIMEOUT WATCHDOG (timeout == %t)", top.phase_timeout), this, UVM_HIGH)
`uvm_delay(top.phase_timeout)
if ($time == `UVM_DEFAULT_TIMEOUT) begin
if (m_phase_trace)
`UVM_PH_TRACE("PH/TRC/TIMEOUT", "PHASE TIMEOUT WATCHDOG EXPIRED", this, UVM_LOW)
foreach (m_executing_phases[p]) begin
if ((p.phase_done != null) && (p.phase_done.get_objection_total() > 0)) begin
if (m_phase_trace)
`UVM_PH_TRACE("PH/TRC/TIMEOUT/OBJCTN",
$sformatf("Phase '%s' has outstanding objections:\n%s", p.get_full_name(), p.phase_done.convert2string()),
this,
UVM_LOW)
end
end
`uvm_fatal("PH_TIMEOUT",
$sformatf("Default timeout of %0t hit, indicating a probable testbench issue",
`UVM_DEFAULT_TIMEOUT))
end
else begin
if (m_phase_trace)
`UVM_PH_TRACE("PH/TRC/TIMEOUT", "PHASE TIMEOUT WATCHDOG EXPIRED", this, UVM_LOW)
foreach (m_executing_phases[p]) begin
if ((p.phase_done != null) && (p.phase_done.get_objection_total() > 0)) begin
if (m_phase_trace)
`UVM_PH_TRACE("PH/TRC/TIMEOUT/OBJCTN",
$sformatf("Phase '%s' has outstanding objections:\n%s", p.get_full_name(), p.phase_done.convert2string()),
this,
UVM_LOW)
end
end
`uvm_fatal("PH_TIMEOUT",
$sformatf("Explicit timeout of %0t hit, indicating a probable testbench issue",
top.phase_timeout))
end
if (m_phase_trace)
`UVM_PH_TRACE("PH/TRC/EXE/3","PHASE EXIT TIMEOUT",this,UVM_DEBUG)
end // if (this.get_name() == "run")
else begin
wait (0); // never unblock for non-run phase
end
end // if (m_phase_trace)
join_any
disable fork;
end
join // guard
end
end
m_executing_phases.delete(this);
//---------
// JUMPING:
//---------
// If jump_to() was called then we need to kill all the successor
// phases which may still be running and then initiate the new
// phase. The return is necessary so we don't start new successor
// phases. If we are doing a forward jump then we want to set the
// state of this phase's successors to UVM_PHASE_DONE. This
// will let us pretend that all the phases between here and there
// were executed and completed. Thus any dependencies will be
// satisfied preventing deadlocks.
// GSA TBD insert new jump support
if (m_phase_type == UVM_PHASE_NODE) begin
if(m_premature_end) begin
if(m_jump_phase != null) begin
state_chg.m_jump_to = m_jump_phase;
`uvm_info("PH_JUMP",
$sformatf("phase %s (schedule %s, domain %s) is jumping to phase %s",
get_name(), get_schedule_name(), get_domain_name(), m_jump_phase.get_name()),
UVM_MEDIUM);
end
else begin
`uvm_info("PH_JUMP",
$sformatf("phase %s (schedule %s, domain %s) is ending prematurely",
get_name(), get_schedule_name(), get_domain_name()),
UVM_MEDIUM);
end
#0; // LET ANY WAITERS ON READY_TO_END TO WAKE UP
if (m_phase_trace)
`UVM_PH_TRACE("PH_END","ENDING PHASE PREMATURELY",this,UVM_HIGH)
end
else begin
// WAIT FOR PREDECESSORS: // WAIT FOR PREDECESSORS:
// function phases only
if (task_phase == null)
m_wait_for_pred();
end
//-------
// ENDED:
//-------
// execute 'phase_ended' callbacks
if (m_phase_trace)
`UVM_PH_TRACE("PH_END","ENDING PHASE",this,UVM_HIGH)
state_chg.m_prev_state = m_state;
m_state = UVM_PHASE_ENDED;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(this, state_chg))
if (m_imp != null)
m_imp.traverse(top,this,UVM_PHASE_ENDED);
#0; // LET ANY WAITERS WAKE UP
//---------
// CLEANUP:
//---------
// kill this phase's threads
state_chg.m_prev_state = m_state;
if(m_premature_end) m_state = UVM_PHASE_JUMPING;
else m_state = UVM_PHASE_CLEANUP ;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(this, state_chg))
if (m_phase_proc != null) begin
m_phase_proc.kill();
m_phase_proc = null;
end
#0; // LET ANY WAITERS WAKE UP
if (phase_done != null)
phase_done.clear();
end
//------
// DONE:
//------
m_premature_end = 0 ;
if(m_jump_fwd || m_jump_bkwd) begin
if(m_jump_fwd) begin
clear_successors(UVM_PHASE_DONE,m_jump_phase);
end
m_jump_phase.clear_successors();
uvmkit_gettimeofday(phase_finish_time);
void'(uvmkit_timeval_subtract(phase_diff_time, phase_finish_time, phase_start_time));
`uvm_info("TIMED_PHASE",
$sformatf("'%s' phase took a total of %f seconds",
this.get_full_name(),
uvmkit_timeval_to_real(phase_diff_time)),
UVM_LOW)
end
else begin
if (m_phase_trace)
`UVM_PH_TRACE("PH/TRC/DONE","Completed phase",this,UVM_LOW)
state_chg.m_prev_state = m_state;
m_state = UVM_PHASE_DONE;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(this, state_chg))
m_phase_proc = null;
uvmkit_gettimeofday(phase_finish_time);
void'(uvmkit_timeval_subtract(phase_diff_time, phase_finish_time, phase_start_time));
`uvm_info("TIMED_PHASE",
$sformatf("'%s' phase took a total of %f seconds",
this.get_full_name(),
uvmkit_timeval_to_real(phase_diff_time)),
UVM_LOW)
#0; // LET ANY WAITERS WAKE UP
end
#0; // LET ANY WAITERS WAKE UP
if (phase_done != null)
phase_done.clear();
//-----------
// SCHEDULED:
//-----------
if(m_jump_fwd || m_jump_bkwd) begin
void'(m_phase_hopper.try_put(m_jump_phase));
m_jump_phase = null;
m_jump_fwd = 0;
m_jump_bkwd = 0;
uvmkit_gettimeofday(phase_finish_time);
void'(uvmkit_timeval_subtract(phase_diff_time, phase_finish_time, phase_start_time));
`uvm_info("TIMED_PHASE",
$sformatf("'%s' phase took a total of %f seconds",
this.get_full_name(),
uvmkit_timeval_to_real(phase_diff_time)),
UVM_LOW)
end
// If more successors, schedule them to run now
else if (m_successors.size() == 0) begin
top.m_phase_all_done=1;
end
else begin
// execute all the successors
foreach (m_successors[succ]) begin
if(succ.m_state < UVM_PHASE_SCHEDULED) begin
state_chg.m_prev_state = succ.m_state;
state_chg.m_phase = succ;
succ.m_state = UVM_PHASE_SCHEDULED;
`uvm_do_callbacks(uvm_phase, uvm_phase_cb, phase_state_change(succ, state_chg))
#0; // LET ANY WAITERS WAKE UP
void'(m_phase_hopper.try_put(succ));
if (m_phase_trace)
`UVM_PH_TRACE("PH/TRC/SCHEDULED",{"Scheduled from phase ",get_full_name()},succ,UVM_LOW)
end
end
end
endtask

复制代码

第一次调用这个task是句柄m_common_domain，首先我们必要等待调用这个task的phase的m_predecessors[]执行完毕，即它们的m_state 为UVM_PHASE_DONE，这里m_common_domain是没有任何前置phase的，所以略过。若当前phase的m_state是UVM_PHASE_DONE，表明已经执行过了，task会直接return。若当前phase_type不是UVM_PHASE_NODE，即对于domain/schedule这两张phase运行图来说，会进行"fake" execution，task只会将其m_state分别转换为UVM_PHASE_STARTED和UVM_PHASE_EXECUTING，而并不会真的做什么事情。在task的末了判断m_jump_fwd和m_jump_bkwd是否为1，这两个变量是跟domain的jump操纵有关的，这里它们的值都是0，则对于当前phase的全部后续phase m_successors，判断其m_state是否已经走到了UVM_PHASE_SCHEDULED这一步，若没有，则将其m_state置为UVM_PHASE_SCHEDULED并分别将后续phase的句柄放入m_phase_hopper邮箱。对于m_common_domain来说，它的后续phase只有一个即build_phase，于是这里就会将m_common_domain运行图中的build_phase唯一实例句柄放入m_phase_hopper。
在task m_run_phases()中forever循环从邮箱中拿到build_phase的句柄，会立即用fork…join_none启动新线程调用build_phase的execute_phase()执行。由于这次调用execute_phase()的句柄类型的phase_type是UVM_PHASE_NODE，而且是function phase，所以会依次将其m_state进入UVM_PHASE_STARTED、UVM_PHASE_EXECUTING和UVM_PHASE_ENDED状态，并将这些状态作为参数调用traverse()。traverse()我们之前已经研究过，这里传入的第一个参数是top，即uvm_root的唯一实例，则执行这些phase的component是从uvm_root作为入口，对于build_phase是自上而下执行的。在后面的状态中执行一些清理的工作，这里不再赘述。依然是task末了，会查抄build_phase的后续phase m_successors[]并将其放入m_phase_hopper邮箱来执行，这样UVM中的各个function phase就这样按固定顺序依次执行。直到执行到末了m_common_domain的m_end_node不再有m_successors，则把变量m_phase_all_done置为1，UVM会调用$finish结束仿真。
在执行到start_of_simulation_phse的时候，该phase有两个后继phase，分别为run_phase和m_uvm_domain，他们会分别被放入m_phase_hopper邮箱并被取走调用fork…join_none分别启动并行的新线程来调用各自的execute_phase()，在m_uvm_domain执行execute_phase()时，并不会真的执行，所以也不会有任何仿真延时，它也会在task末了把后继phase句柄即m_uvm_schedule放入邮箱并被取走开启新线程调用execute_phase()，m_uvm_schedule同样也不会真的被执行，而是继承把后继phase即pre_reset_phase实例句柄放入邮箱，当新的并行线程被启动调用execute_phase()时，这个phase_type为UVM_PHASE_NODE的phase会被执行。全部这统统都是并行发生的，所以我们说run_phase和12个run-time phase是同时被并行启动运行的。
在运行task phase时，execute_phase()中会在调用traverse()执行当前task phase的同时启动数个并行进程，同时监测当前phase的objection和TIMEOUT，一旦达到退出条件，则立即杀掉当前进程进入下一个task phase。因此，在运行task phase时设置合理的objection和timeout以保证phase正常运行非常重要。

3. objection机制

UVM的objection机制是共同task phase来用的，在task phase如run_phase中若整个平台中没有raise任何objection，则UVM运行到这里会立即跳转到下一个phase，不会执行run_phase里的任何内容。
3.1. uvm_objection class

uvm_objection类扩展自uvm_report_object类，此中界说了一个存储内容为uvm_objection类型的静态队列，在new()函数中会把当前调用该函数的component/object句柄存入队列m_objections。别的，在uvm_objection类中还界说了一个uvm_objection_events类型的索引为uvm_object的联合数组m_events[]，uvm_objection_events类中界说了数个event变量。联合数组m_source_count[]和m_total_count[]均是以uvm_object类型为索引，纪录的是raise objection的sequence大概component节点的objection数量和总的objection数量，联合数组m_drain_time[]以uvm_object类型为索引，纪录的是当前sequence大概component节点的objection的drain_time。

//uvm_objection.svh
class uvm_objection_events;
int waiters;
event raised;
event dropped;
event all_dropped;
endclass
class uvm_objection extends uvm_report_object;
`uvm_register_cb(uvm_objection, uvm_objection_callback)
protected int m_source_count[uvm_object];
protected int m_total_count [uvm_object];
protected time m_drain_time [uvm_object];
protected uvm_objection_events m_events [uvm_object];
/*protected*/ bit m_top_all_dropped;
protected uvm_root m_top;
static uvm_objection m_objections[$];
...
function new(string name="");
super.new(name);
...
m_objections.push_back(this);
endfunction
...
endclass

复制代码

3.2. raise_objection

一般我们在某个sequence大概test的run_phase中调用raise_objection()，如phase.raise_objection()，当前phase会调用phase_done的raise_objection()函数，phase_done是界说在uvm_phase中的uvm_objection类型的句柄，在我们调用add()函数添加各个task phase时，就会例化出各个task phase的uvm_objection类型句柄并赋值给phase_done。

//uvm_phase.svh
class uvm_phase extends uvm_object;
uvm_objection phase_done;
...
endclass
function void uvm_phase::add(uvm_phase phase,
uvm_phase with_phase=null,
uvm_phase after_phase=null,
uvm_phase before_phase=null);
if (phase.get_phase_type() == UVM_PHASE_IMP) begin
uvm_task_phase tp;
...
// The phase_done objection is only required
// for task-based nodes
if ($cast(tp, phase)) begin
if (new_node.get_name() == "run") begin
new_node.phase_done = uvm_test_done_objection::get();
end
else begin
new_node.phase_done = uvm_objection::type_id::create({phase.get_name(), "_objection"});
end
end
end
...
endfunction
function void uvm_phase::raise_objection (uvm_object obj,
string description="",
int count=1);
if (phase_done != null)
phase_done.raise_objection(obj,description,count);
else
m_report_null_objection(obj, description, count, "raise");
endfunction

复制代码

界说在uvm_objection类中的raise_objection()函数有三个参数，第一个参数uvm_object类型obj，我们可以在一个sequence大概某component节点中使用this指针传入当前的sequence大概component，第二个参数可以输入一些字符串类型的描述，第三个参数是要raise的objection数量，默认是1。若参数obj为null，则obj为uvm_root的唯一实例m_top，然后把obj作为前两个参数调用m_raise()函数。

//uvm_objection.svh
class uvm_objection extends uvm_report_object;
virtual function void raise_objection (uvm_object obj=null,
string description="",
int count=1);
if(obj == null)
obj = m_top;
m_cleared = 0;
m_top_all_dropped = 0;
m_raise (obj, obj, description, count);
endfunction
function void m_raise (uvm_object obj,
uvm_object source_obj,
string description="",
int count=1);
int idx;
uvm_objection_context_object ctxt;
// Ignore raise if count is 0
if (count == 0)
return;
if (m_total_count.exists(obj))
m_total_count[obj] += count;
else
m_total_count[obj] = count;
if (source_obj==obj) begin
if (m_source_count.exists(obj))
m_source_count[obj] += count;
else
m_source_count[obj] = count;
end
if (m_trace_mode)
m_report(obj,source_obj,description,count,"raised");
raised(obj, source_obj, description, count);
...
if (ctxt == null) begin
// If there were no drains, just propagate as usual
if (!m_prop_mode && obj != m_top)
m_raise(m_top,source_obj,description,count);
else if (obj != m_top)
m_propagate(obj, source_obj, description, count, 1, 0);
end
else begin
// Otherwise we need to determine what exactly happened
int diff_count;
// Determine the diff count, if it's positive, then we're
// looking at a 'raise' total, if it's negative, then
// we're looking at a 'drop', but not down to 0. If it's
// a 0, that means that there is no change in the total.
diff_count = count - ctxt.count;
if (diff_count != 0) begin
// Something changed
if (diff_count > 0) begin
// we're looking at an increase in the total
if (!m_prop_mode && obj != m_top)
m_raise(m_top, source_obj, description, diff_count);
else if (obj != m_top)
m_propagate(obj, source_obj, description, diff_count, 1, 0);
end
else begin
// we're looking at a decrease in the total
// The count field is always positive...
diff_count = -diff_count;
if (!m_prop_mode && obj != m_top)
m_drop(m_top, source_obj, description, diff_count);
else if (obj != m_top)
m_propagate(obj, source_obj, description, diff_count, 0, 0);
end
end
// Cleanup
ctxt.clear();
m_context_pool.push_back(ctxt);
end
endfunction
endclass

复制代码

在此中界说了一个uvm_objection_context_object类型的变量ctxt，在uvm_objection_context_object类中界说了uvm_object类型的变量obj和source_obj，字符串description以及数量count，其实就是纪录下我们每次raise_objection大概drop_objection的参数信息。

//uvm_objection.svh
class uvm_objection_context_object;
uvm_object obj;
uvm_object source_obj;
string description;
int count;
uvm_objection objection;
// Clears the values stored within the object,
// preventing memory leaks from reused objects
function void clear();
obj = null;
source_obj = null;
description = "";
count = 0;
objection = null;
endfunction : clear
endclass

复制代码

回到m_raise()函数，若传入count为0，函数直接返回。若是第一次在my_seq中raise objection，则m_total_count[my_seq]=count，若不是，则以该obj为索引的m_total_count数量加上此次要raise的objection数量。由于传入的obj和source_obj是同一个参数m_top，所以m_source_count[my_seq]=count，若不是，则略过，所以m_source_count的索引只会纪录在此中raise objection的sequence大概component节点。接下来函数调用raised():

//uvm_objection.svh
class uvm_objection extends uvm_report_object;
virtual function void raised (uvm_object obj,
uvm_object source_obj,
string description,
int count);
uvm_component comp;
if ($cast(comp,obj))
comp.raised(this, source_obj, description, count);
`uvm_do_callbacks(uvm_objection,uvm_objection_callback,raised(this,obj,source_obj,description,count))
if (m_events.exists(obj))
->m_events[obj].raised;
endfunction
endclass

复制代码

若传入的obj参数是component类型，会调用component的callback raised()，我们可以在某component中来重载这个callback。若m_events中有当前obj的索引内容，则触发此中的raised变乱。
回到m_raise()中，此时ctxt为null，且m_prop_mode为1（default 1），则调用m_propagate()：

//uvm_objection.svh
class uvm_objection extends uvm_report_object;
unction void m_propagate (uvm_object obj,
uvm_object source_obj,
string description,
int count,
bit raise,
int in_top_thread);
if (obj != null && obj != m_top) begin
obj = m_get_parent(obj);
if(raise)
m_raise(obj, source_obj, description, count);
else
m_drop(obj, source_obj, description, count, in_top_thread);
end
endfunction
function uvm_object m_get_parent(uvm_object obj);
uvm_component comp;
uvm_sequence_base seq;
if ($cast(comp, obj)) begin
obj = comp.get_parent();
end
else if ($cast(seq, obj)) begin
obj = seq.get_sequencer();
end
else
obj = m_top;
if (obj == null)
obj = m_top;
return obj;
endfunction
endclass

复制代码

若当前传入obj不是m_top，则调用m_get_parent()取得其父节点的句柄。这个obj颠末层层参数转达，实际上还是我们在一开始调用raise_objection()时的sequence大概component，若是obj是一个sequence，则m_get_parent()返回该sequence的sequencer，若是一个component节点，则直接返回其父节点。然后在m_propagate()中再次调用m_raise()函数，不外这次obj参数即为其父节点句柄。颠末层层递归调用m_raise()，直到uvm_root节点，此时
m_source_count[my_seq] = 1;
m_total_count[env.agt.sqr] = 1;
m_total_count[env.agt] = 1;
m_total_count[env] = 1;
m_total_count[m_top] = 1;
当我们在某个seq大概component中raise objection数为count时，UVM会将count纪录在以自身为索引的m_source_count[]中。当m_prop_mode默认为1时其会将count以当前节点的全部上级节点为索引纪录在m_total_count[]中，m_prop_mode为0时只会将自身和顶层uvm_root为索引纪录在m_total_count[]中，不会纪录中央节点。
3.3. drop_objection

同样在某sequence大概component中drop_objection()时，也会在此中调用drop_done的drop_objection()函数：

//uvm_phase.svh
function void uvm_phase::drop_objection (uvm_object obj,
string description="",
int count=1);
if (phase_done != null)
phase_done.drop_objection(obj,description,count);
else
m_report_null_objection(obj, description, count, "drop");
endfunction
//uvm_objection.svh
class uvm_objection extends uvm_report_object;
virtual function void drop_objection (uvm_object obj=null,
string description="",
int count=1);
if(obj == null)
obj = m_top;
m_drop (obj, obj, description, count, 0);
endfunction
// Function- m_drop
function void m_drop (uvm_object obj,
uvm_object source_obj,
string description="",
int count=1,
int in_top_thread=0);
// Ignore drops if the count is 0
if (count == 0)
return;
if (!m_total_count.exists(obj) || (count > m_total_count[obj])) begin
if(m_cleared)
return;
uvm_report_fatal("OBJTN_ZERO", {"Object "", obj.get_full_name(),
"" attempted to drop objection '",this.get_name(),"' count below zero"});
return;
end
if (obj == source_obj) begin
if (!m_source_count.exists(obj) || (count > m_source_count[obj])) begin
if(m_cleared)
return;
uvm_report_fatal("OBJTN_ZERO", {"Object "", obj.get_full_name(),
"" attempted to drop objection '",this.get_name(),"' count below zero"});
return;
end
m_source_count[obj] -= count;
end
m_total_count[obj] -= count;
if (m_trace_mode)
m_report(obj,source_obj,description,count,"dropped");
dropped(obj, source_obj, description, count);
// if count != 0, no reason to fork
if (m_total_count[obj] != 0) begin
if (!m_prop_mode && obj != m_top)
m_drop(m_top,source_obj,description, count, in_top_thread);
else if (obj != m_top) begin
this.m_propagate(obj, source_obj, description, count, 0, in_top_thread);
end
end
else begin
uvm_objection_context_object ctxt;
if (m_context_pool.size())
ctxt = m_context_pool.pop_front();
else
ctxt = new;
ctxt.obj = obj;
ctxt.source_obj = source_obj;
ctxt.description = description;
ctxt.count = count;
ctxt.objection = this;
// Need to be thread-safe, let the background
// process handle it.
// Why don't we look at in_top_thread here? Because
// a re-raise will kill the drain at object that it's
// currently occuring at, and we need the leaf-level kills
// to not cause accidental kills at branch-levels in
// the propagation.
// Using the background process just allows us to
// separate the links of the chain.
m_scheduled_list.push_back(ctxt);
end // else: !if(m_total_count[obj] != 0)
endfunction
endclass

复制代码

若传入参数obj不是m_top，uvm_objection的drop_objection()会把obj作为前两个参数继承调用m_drop()，若传入的参数count为0，函数直接返回。若m_total_count[]中不存在obj为索引的objection数量纪录，大概当前传入的count数量超出了m_total_count[obj]，函数直接返回。bit m_cleared用来纪录当前objection是否已经全部被drop掉，我们会在每次调用raise_objectiion时候把这个bit置为0，若这个bit为1，则函数也直接返回。当前m_drop()的前两个参数obj和source_obj相等，将m_source_count[obj]减去count，同时将m_total_count[obj]也减去count。接下来调用dropped()函数：

//uvm_objection.svh
class uvm_objection extends uvm_report_object;
virtual function void dropped (uvm_object obj,
uvm_object source_obj,
string description,
int count);
uvm_component comp;
if($cast(comp,obj))
comp.dropped(this, source_obj, description, count);
`uvm_do_callbacks(uvm_objection,uvm_objection_callback,dropped(this,obj,source_obj,description,count))
if (m_events.exists(obj))
->m_events[obj].dropped;
endfunction
endclass

复制代码

若obj是一个component类型，则调用component的callback dropped()，我们可以在component中扩展这个callback。别的会触发m_events[obj]的dropped变乱。
回到m_drop()中来，若减去这次drop的objection数量之后m_total_count[obj]不为0且m_prop_mode为1，则调用m_propagate()并把第五个参数传入0，表示这是一次drop操纵，m_propagate()在之前我们已经见过，这个函数就是从当前的obj开始遍历全部父节点直到uvm_root，根据第五个参数对这些父节点进行raise大概drop objection操纵。
3.4. drain_time

若当前调用m_drop()减去这次drop的objection数量之后m_total_count[obj]为0，阐明对obj来说，全部的objection都已经被dropped了，当前phase是否可以直接结束进入下一个phase中呢，答案是否定的，UVM会等待执行当前phase的其他节点中的objection被dropped，那若全部节点的全部objection都已经被dropped了，是否可以结束当前phase呢？答案依然是否定的，若某节点中当前phase设置drain_time，则在objection都被dropped之后，该phase会等待drain_time时间之后才会结束。这个drain_time考虑到了验证平台激励和DUT输出可能会存在延时，比如我们在driver的main_phase中输入向DUT灌输激励之后drop_objection，DUT处理这些激励必要一些延时，假如此时直接结束当前phase，monitor就会漏掉末了的一些激励处理之后的结果，所以phase会在全部的objection被dropped之后再延伸drain_time时间才会结束。

//uvm_objection.svh
class uvm_objection extends uvm_report_object;
function void set_drain_time (uvm_object obj=null, time drain);
if (obj==null)
obj = m_top;
m_drain_time[obj] = drain;
endfunction
endclass

复制代码

通过调用set_drain_time()函数可以设置当前obj的drain_time，将其存入以obj为索引的联合数组中。若传入参数obj为null，则设置m_top的drain_time。
回到m_drop()中，若当前m_total_count[obj]为0，我们还必要等待当前phase的drain_time时间才能真正结束phase，此时会查抄m_context_pool的内容，若有存储直接弹出最前面的一个，若没有，则创建一个uvm_objection_context_object类型的对象ctxt，并把obj/source_obj/description/count和this指针分别赋值给其成员变量，末了将其放入队列m_scheduled_list[ ] 。其中 m c o n t e x t p o o l [ ]。此中m_context_pool[ ]。此中mcontextpool[]中存放的是用过的uvm_objection_context_object指针，从这里拿跟new()一个没有太大区别。

//uvm_objection.svh
class uvm_objection extends uvm_report_object;
local static uvm_objection_context_object m_context_pool[$];
`ifndef UVM_USE_PROCESS_CONTAINER
local process m_drain_proc[uvm_object];
`else
local process_container_c m_drain_proc[uvm_object];
`endif
local static uvm_objection_context_object m_scheduled_list[$];
local uvm_objection_context_object m_scheduled_contexts[uvm_object];
local uvm_objection_context_object m_forked_list[$];
local uvm_objection_context_object m_forked_contexts[uvm_object];
...
endclass

复制代码

队列m_scheduled_list[$]存放的是当前在等待drain_time的objection信息，联合数组m_scheduled_contexts[]以uvm_object为索引，存放的是关于uvm_object的在等待drain_time的objection信息，队列m_forked_list[$]存放的是当前已经启动fork线程等待drain_time的objection信息，联合数组m_forked_contexts[]以uvm_object为索引，存放的是关于uvm_object的已经启动fork线程在等待drain_time的objection信息。别的联合数组m_drain_proc[]存放的是以uvm_object为索引的等待drain_time的进程句柄。
等待phase的drain_time的进程其实不停作为backgroud进程从仿真一开始调用uvm_root::run_test()就启动了，乃至比调用m_run_phases()还要早，在run_test()中调用静态函数m_init_objections()启动该进程，该函数会用fork…join_none调用m_execute_scheduled_forks()。

//uvm_root.svh
task uvm_root::run_test(string test_name="");
uvm_objection::m_init_objections();
endtask
//uvm_objection.svh
class uvm_objection extends uvm_report_object;
static function void m_init_objections();
fork
uvm_objection::m_execute_scheduled_forks();
join_none
endfunction
// background process; when non
static task m_execute_scheduled_forks();
while(1) begin
wait(m_scheduled_list.size() != 0);
if(m_scheduled_list.size() != 0) begin
uvm_objection_context_object c;
uvm_objection o;
// Save off the context before the fork
c = m_scheduled_list.pop_front();
// A re-raise can use this to figure out props (if any)
c.objection.m_scheduled_contexts[c.obj] = c;
// The fork below pulls out from the forked list
c.objection.m_forked_list.push_back(c);
// The fork will guard the m_forked_drain call, but
// a re-raise can kill m_forked_list contexts in the delta
// before the fork executes.
fork : guard
automatic uvm_objection objection = c.objection;
begin
// Check to maike sure re-raise didn't empty the fifo
if (objection.m_forked_list.size() > 0) begin
uvm_objection_context_object ctxt;
ctxt = objection.m_forked_list.pop_front();
// Clear it out of scheduled
objection.m_scheduled_contexts.delete(ctxt.obj);
// Move it in to forked (so re-raise can figure out props)
objection.m_forked_contexts[ctxt.obj] = ctxt;
// Save off our process handle, so a re-raise can kill it...
`ifndef UVM_USE_PROCESS_CONTAINER
objection.m_drain_proc[ctxt.obj] = process::self();
`else
begin
process_container_c c = new(process::self());
objection.m_drain_proc[ctxt.obj]=c;
end
`endif
// Execute the forked drain
objection.m_forked_drain(ctxt.obj, ctxt.source_obj, ctxt.description, ctxt.count, 1);
// Cleanup if we survived (no re-raises)
objection.m_drain_proc.delete(ctxt.obj);
objection.m_forked_contexts.delete(ctxt.obj);
// Clear out the context object (prevent memory leaks)
ctxt.clear();
// Save the context in the pool for later reuse
m_context_pool.push_back(ctxt);
end
end
join_none : guard
end
end
endtask
endclass

复制代码

静态task m_execute_scheduled_forks()中使用while无穷循环，一旦队列m_scheduled_list[$]不为空，则从中取出uvm_objection_context_object类型句柄放入m_scheduled_contexts[obj]和队列m_forked_list[$]，这些句柄纪录的是当前obj在drop_objection时必要等待drain_time的信息。随纵然用fork…join_none并行线程取出m_forked_list[$]的句柄并清除m_forked_list[$]和m_scheduled_contexts[obj]相关内容，同时启动m_drain_proc[obj]调用m_forked_drain()来执行等待任务，等待该task返回以后清除m_forked_drain[obj]和m_forked_contexts[obj]。

//uvm_objection.svh
class uvm_objection extends uvm_report_object;
task m_forked_drain (uvm_object obj,
uvm_object source_obj,
string description="",
int count=1,
int in_top_thread=0);
int diff_count;
if (m_drain_time.exists(obj))
`uvm_delay(m_drain_time[obj])
if (m_trace_mode)
m_report(obj,source_obj,description,count,"all_dropped");
all_dropped(obj,source_obj,description, count);
// wait for all_dropped cbs to complete
wait fork;
/* NOT NEEDED - Any raise would have killed us!
if(!m_total_count.exists(obj))
diff_count = -count;
else
diff_count = m_total_count[obj] - count;
*/
// we are ready to delete the 0-count entries for the current
// object before propagating up the hierarchy.
if (m_source_count.exists(obj) && m_source_count[obj] == 0)
m_source_count.delete(obj);
if (m_total_count.exists(obj) && m_total_count[obj] == 0)
m_total_count.delete(obj);
if (!m_prop_mode && obj != m_top)
m_drop(m_top,source_obj,description, count, 1);
else if (obj != m_top)
m_propagate(obj, source_obj, description, count, 0, 1);
endtask
virtual task all_dropped (uvm_object obj,
uvm_object source_obj,
string description,
int count);
uvm_component comp;
if($cast(comp,obj))
comp.all_dropped(this, source_obj, description, count);
`uvm_do_callbacks(uvm_objection,uvm_objection_callback,all_dropped(this,obj,source_obj,description,count))
if (m_events.exists(obj))
->m_events[obj].all_dropped;
if (obj == m_top)
m_top_all_dropped = 1;
endtask
endclass

复制代码

在task m_forked_drain()中，若当前obj已经设置了drain_time，会首先等待m_drain_time[obj]时间，然后调用all_dropped()函数，若obj是component类型，该函数中会首先调用all_dropped() callback，我们可以在某component中重载这个callback，其次会触发m_events[obj]的all_dropped变乱，若当前obj是m_top，则把m_top_all_dropped置为1。回到task m_forked_drain()中，接着把m_source_count[]和m_total_count[]中关于obj的部分清除，若m_prop_mode为1则调用m_propagate()继承drop全部父类节点的objection，直到m_top的全部objection被dropped。
若某obj在等待drain_time期间，又有objection被raised，该如那边理？回到之前的m_raise()中：

//uvm_objection.svh
class uvm_objection extends uvm_report_object;
function void m_raise (uvm_object obj,
uvm_object source_obj,
string description="",
int count=1);
int idx;
uvm_objection_context_object ctxt;
// Handle any outstanding drains...
// First go through the scheduled list
idx = 0;
while (idx < m_scheduled_list.size()) begin
if ((m_scheduled_list[idx].obj == obj) &&
(m_scheduled_list[idx].objection == this)) begin
// Caught it before the drain was forked
ctxt = m_scheduled_list[idx];
m_scheduled_list.delete(idx);
break;
end
idx++;
end
// If it's not there, go through the forked list
if (ctxt == null) begin
idx = 0;
while (idx < m_forked_list.size()) begin
if (m_forked_list[idx].obj == obj) begin
// Caught it after the drain was forked,
// but before the fork started
ctxt = m_forked_list[idx];
m_forked_list.delete(idx);
m_scheduled_contexts.delete(ctxt.obj);
break;
end
idx++;
end
end
// If it's not there, go through the forked contexts
if (ctxt == null) begin
if (m_forked_contexts.exists(obj)) begin
// Caught it with the forked drain running
ctxt = m_forked_contexts[obj];
m_forked_contexts.delete(obj);
// Kill the drain
`ifndef UVM_USE_PROCESS_CONTAINER
m_drain_proc[obj].kill();
m_drain_proc.delete(obj);
`else
m_drain_proc[obj].p.kill();
m_drain_proc.delete(obj);
`endif
end
end
if (ctxt == null) begin
// If there were no drains, just propagate as usual
if (!m_prop_mode && obj != m_top)
m_raise(m_top,source_obj,description,count);
else if (obj != m_top)
m_propagate(obj, source_obj, description, count, 1, 0);
end
else begin
// Otherwise we need to determine what exactly happened
int diff_count;
// Determine the diff count, if it's positive, then we're
// looking at a 'raise' total, if it's negative, then
// we're looking at a 'drop', but not down to 0. If it's
// a 0, that means that there is no change in the total.
diff_count = count - ctxt.count;
if (diff_count != 0) begin
// Something changed
if (diff_count > 0) begin
// we're looking at an increase in the total
if (!m_prop_mode && obj != m_top)
m_raise(m_top, source_obj, description, diff_count);
else if (obj != m_top)
m_propagate(obj, source_obj, description, diff_count, 1, 0);
end
else begin
// we're looking at a decrease in the total
// The count field is always positive...
diff_count = -diff_count;
if (!m_prop_mode && obj != m_top)
m_drop(m_top, source_obj, description, diff_count);
else if (obj != m_top)
m_propagate(obj, source_obj, description, diff_count, 0, 0);
end
end
// Cleanup
ctxt.clear();
m_context_pool.push_back(ctxt);
end
endfunction
endclass

复制代码

首先查抄队列m_scheduled_list[$]中是否有当前obj的纪录信息，若有（阐明此时已经drop全部objection并在等待drain_time）则取出赋值给uvm_objection_context_object类型的ctxt句柄并删除队列中相关纪录，若没有接着查抄队列m_forked_list[$]，若有则同样取出赋值给ctxt句柄并删除队列中相关纪录，若没有则查抄数组m_forked_contexts[]是否有相关信息，若有(阐明已经进入等待fork线程)则同样取出赋值给ctxt句柄并删除队列中相关纪录，同时杀掉m_drain_proc[obj]制止等待obj的drain_time进程。当ctxt得到纪录信息，比力当前raise_objection的count和ctxt中的count大小，若raise_objection的count较大，则直接将差值count作为参数调用m_propagate()来raise objection，若raise_objection的count较小，则直接将差值count作为参数调用m_propagate()来drop objection，若相等，则略过。
转头来看uvm_phase::execute_phase()中在m_state为UVM_PHASE_EXECUTING时，若当前运行的是一个task phase时，一方面会用fork…join_none启动并行线程，在此中启动一个process来执行traverse()，另一方面也会再启动两个并行线程，一个用来检测当前phase的total_objection，另一个会检测timeout相关设置，以判断是否可以退出fork线程进入phase的下一个执行阶段。值得一提的是，这个timeout是界说在uvm_root中的一个time类型变量phase_timeout，目的是为了防止仿真不停block在某个phase中而设置的退出机制。这个default timeout是9200s，我们可以通过调用set_timeout()来重置这个phase_timeout时间:

//uvm_root.svh
function void uvm_root::set_timeout(time timeout, bit overridable=1);
static bit m_uvm_timeout_overridable = 1;
if (m_uvm_timeout_overridable == 0) begin
uvm_report_info("NOTIMOUTOVR",
$sformatf("The global timeout setting of %0d is not overridable to %0d due to a previous setting.",
phase_timeout, timeout), UVM_NONE);
return;
end
m_uvm_timeout_overridable = overridable;
phase_timeout = timeout;
endfunction

复制代码

4. phase的高级应用

4.1. jump()

UVM提供了在12个run-time phases间进行phase跳转运行的函数jump()，如我们在main_phase中可以调用jump()跳转到reset_phase： phase.jump(uvm_reset_phase.get());

//uvm_phase.svh
function void uvm_phase::end_prematurely() ;
m_premature_end = 1 ;
endfunction
function void uvm_phase::jump(uvm_phase phase);
set_jump_phase(phase) ;
end_prematurely() ;
endfunction

复制代码

jump()函数会先后调用set_jump_phase()和end_prematurely()，后者很简单就是把变量m_premature_end置为1，这个bit会控制exectute_phase()中当前task phase的运行进入jumping状态。来看set_jump_phase()：

//uvm_phase.svh
function void uvm_phase::set_jump_phase(uvm_phase phase) ;
uvm_phase d;
if ((m_state < UVM_PHASE_STARTED) ||
(m_state > UVM_PHASE_ENDED) )
begin
`uvm_error("JMPPHIDL", { "Attempting to jump from phase "",
get_name(), "" which is not currently active (current state is ",
m_state.name(), "). The jump will not happen until the phase becomes ",
"active."})
end
// A jump can be either forward or backwards in the phase graph.
// If the specified phase (name) is found in the set of predecessors
// then we are jumping backwards. If, on the other hand, the phase is in the set
// of successors then we are jumping forwards. If neither, then we
// have an error.
//
// If the phase is non-existant and thus we don't know where to jump
// we have a situation where the only thing to do is to uvm_report_fatal
// and terminate_phase. By calling this function the intent was to
// jump to some other phase. So, continuing in the current phase doesn't
// make any sense. And we don't have a valid phase to jump to. So we're done.
d = m_find_predecessor(phase,0);
if (d == null) begin
d = m_find_successor(phase,0);
if (d == null) begin
string msg;
$sformat(msg,{"phase %s is neither a predecessor or successor of ",
"phase %s or is non-existant, so we cannot jump to it. ",
"Phase control flow is now undefined so the simulation ",
"must terminate"}, phase.get_name(), get_name());
`uvm_fatal("PH_BADJUMP", msg);
end
else begin
m_jump_fwd = 1;
`uvm_info("PH_JUMPF",$sformatf("jumping forward to phase %s", phase.get_name()),
UVM_DEBUG);
end
end
else begin
m_jump_bkwd = 1;
`uvm_info("PH_JUMPB",$sformatf("jumping backward to phase %s", phase.get_name()),
UVM_DEBUG);
end
m_jump_phase = d;
endfunction

复制代码

若当前phase的m_state还没有走到UVM_PHASE_STARTED大概已经处于UVM_PHASE_ENDED状态，则无法进行phase的跳转，函数直接返回。接着调用m_find_predecessor()在当前phase的全部m_predecessors[]中寻找要跳转的目的phase，若找到则把m_jump_bkwd置为1并把该phase句柄赋值给m_jump_phase；若没有找到，则调用m_find_successor()在该phase的全部m_successors[]中寻找目的phase，若找到则把m_jump_fwd置为1并把该phase句柄赋值给m_jump_phase，若没有找到，则报错。
回到uvm_phase::execute_phase()中，当前phase的m_state为UVM_PHASE_ENDED之后，即当前task phase已经执行完毕。若m_premature_end为1，则把m_state设置为UVM_PHASE_JUMPING，否则设置为UVM_PHASE_CLEANUP，然后做一些当前phase进程的清理工作。若m_jump_fwd为1，则phase会向前跳转，则把调用clear_successors()把当前phase到目的phase之间的phase从m_successors[]中删去。接下来，若m_jump_fwd大概m_jump_bkwd为1，则将m_jump_phase放入m_phase_hopper邮箱中等待运行，之后吧m_jump_phase赋值为null，把m_jump_fwd和m_jump_bkwd赋值为0。
4.2. 创建新的domain

4.2.1. set_domain()

UVM有一个default domain叫做m_common_domain，代表8个function phase和run_phase的运行图，另外还包括一个与run_phase并行运行的m_uvm_domain，这个m_uvm_domain代表12个run-time phases的运行图。在m_uvm_domain之外，我们还可以创建新的domain，可以使新domain的12个run-time phases独立运行，而不必与m_uvm_domain中的phase同步。通常我们可以在某component中调用new()函数来创建一个新的domain:
uvm_domain my_domain = new(“my_domain”);
set_domain(my_domain);
set_domain()是界说在uvm_component类中的函数，在此中会调用define_domain()将新domain的句柄传入。若参数hier为default 1，

//uvm_component.svh
function void uvm_component::set_domain(uvm_domain domain, int hier=1);
// build and store the custom domain
m_domain = domain;
define_domain(domain);
if (hier)
foreach (m_children[c])
m_children[c].set_domain(domain);
endfunction
function void uvm_component::define_domain(uvm_domain domain);
uvm_phase schedule;
//schedule = domain.find(uvm_domain::get_uvm_schedule());
schedule = domain.find_by_name("uvm_sched");
if (schedule == null) begin
uvm_domain common;
schedule = new("uvm_sched", UVM_PHASE_SCHEDULE);
uvm_domain::add_uvm_phases(schedule);
domain.add(schedule);
common = uvm_domain::get_common_domain();
if (common.find(domain,0) == null)
common.add(domain,.with_phase(uvm_run_phase::get()));
end
endfunction

复制代码

此中会首先调用find_by_name()在新domain中寻找是否有名为"uvm_sched"的phase，很明显我们的新domain还没有添加任何phase，所以为null。接下来会创建一个phase_type为UVM_PHASE_SCHEDULE的phase赋值给句柄schedule，并调用add_uvm_phases()将12个run-time phases加入schedule，再将schedule添加进新建的domain中。末了，将新的my_domain加入UVM的default domain m_common_domain中。其实这个过程跟m_uvm_domain的创建过程一样，加入新的my_domain之后，会形成如下图中的运行图（两个domain的run-time phases并不会同步运行）：

4.2.2. sync()

我们可以通过调用sync()来对差别doumain的runtime-phase进行同步，uvm_phase类中有一个队列m_sync[$]，此中存储的是必要与当前phase同步的phases。

//uvm_phase.svh
class uvm_phase extends uvm_object;
local uvm_phase m_sync[$];
...
endclass
function void uvm_phase::sync(uvm_domain target,
uvm_phase phase=null,
uvm_phase with_phase=null);
if (!this.is_domain()) begin
`uvm_fatal("PH_BADSYNC","sync() called from a non-domain phase schedule node");
end
else if (target == null) begin
`uvm_fatal("PH_BADSYNC","sync() called with a null target domain");
end
else if (!target.is_domain()) begin
`uvm_fatal("PH_BADSYNC","sync() called with a non-domain phase schedule node as target");
end
else if (phase == null && with_phase != null) begin
`uvm_fatal("PH_BADSYNC","sync() called with null phase and non-null with phase");
end
else if (phase == null) begin
// whole domain sync - traverse this domain schedule from begin to end node and sync each node
int visited[uvm_phase];
uvm_phase queue[$];
queue.push_back(this);
visited[this] = 1;
while (queue.size()) begin
uvm_phase node;
node = queue.pop_front();
if (node.m_imp != null) begin
sync(target, node.m_imp);
end
foreach (node.m_successors[succ]) begin
if (!visited.exists(succ)) begin
queue.push_back(succ);
visited[succ] = 1;
end
end
end
end else begin
// single phase sync
// this is a 2-way ('with') sync and we check first in case it is already there
uvm_phase from_node, to_node;
int found_to[$], found_from[$];
if(with_phase == null) with_phase = phase;
from_node = find(phase);
to_node = target.find(with_phase);
if(from_node == null || to_node == null) return;
found_to = from_node.m_sync.find_index(node) with (node == to_node);
found_from = to_node.m_sync.find_index(node) with (node == from_node);
if (found_to.size() == 0) from_node.m_sync.push_back(to_node);
if (found_from.size() == 0) to_node.m_sync.push_back(from_node);
end
endfunction

复制代码

sync()函数有三个参数，第一个参数传入想要被同步的新domain句柄，第二个参数传入要被同步的phase，第三个参数传入的phase句柄表示要与新domain中的哪个phase进行同步。若我们必要同步m_common_domain中的pre_reset_phase和my_domain中的main_phase，则可以这么调用：
uvm_domain common_domain = uvm_domain::get_common_domain();
common_domain.sync(my_domain，uvm_pre_reset_phase::get()，uvm_main_phase::get());
sync()函数会分别在这两个要被同步的phase的队列m_sync[$]中检索是否已经有了对方的信息，很明显这里没有，那么就把对方的phase句柄分别存入自己的m_sync[$]中。
若调用时第三个参数with_phase为null，如
common_domain.sync(my_domain，uvm_pre_reset_phase::get());
则with_phase为pre_reset_phase，两个domain的pre_reset_phase的m_sync[$]中会分别放入一条对方的纪录。
若调用时只提供第一个参数，如
common_domain.sync(my_domain);
则会将两个domain的全部12个run-time phase进行sync。
回到uvm_phase::execute_phase()中，若当前phase的m_state运行到UVM_PHASE_SYNCING，若m_sync中纪录有必要sync的信息，则会不停等待必要sync的这些phase的m_state运行到UVM_PHASE_SYNCING才会继承往下一个状态运行，UVM通过这种方式实现phase间的同步。

免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！更多信息从访问主页：qidao123.com:ToB企服之家，中国第一个企服评测及商务社交产业平台。

		自动登录	找回密码
密码			立即注册

UVM phase机制源码探微

本帖子中包含更多资源

0 个回复

快速回复

楼主热帖

标签云