<html>
<head>
<meta content="text/html; charset=utf-8" http-equiv="Content-Type">
</head>
<body bgcolor="#FFFFFF" text="#000000">
Your timing is good. I'm working on docs today and should get to
this by end of day. :)<br>
<br>
Philip<br>
<br>
<div class="moz-cite-prefix">On 02/24/2015 02:37 PM, Sean Silva
wrote:<br>
</div>
<blockquote
cite="mid:CAHnXoanO6ad2SPvrCGaErDsrY3gK2yo25=TfDMeqoPNOGeQqTQ@mail.gmail.com"
type="cite">
<div dir="ltr">Necro-nit (wasn't sure where to post this feedback;
I realize that this has been slightly updated in ToT): please
update the prototypes here to match their current definitions
(e.g. `llvm.experimental.` prefix).
<div><br>
</div>
<div>(sorry for the delay in getting to this)</div>
<div><br>
</div>
<div>-- Sean Silva</div>
</div>
<div class="gmail_extra"><br>
<div class="gmail_quote">On Tue, Dec 2, 2014 at 11:37 AM, Philip
Reames <span dir="ltr"><<a moz-do-not-send="true"
href="mailto:listmail@philipreames.com" target="_blank">listmail@philipreames.com</a>></span>
wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0
.8ex;border-left:1px #ccc solid;padding-left:1ex">Author:
reames<br>
Date: Tue Dec 2 13:37:00 2014<br>
New Revision: 223143<br>
<br>
URL: <a moz-do-not-send="true"
href="http://llvm.org/viewvc/llvm-project?rev=223143&view=rev"
target="_blank">http://llvm.org/viewvc/llvm-project?rev=223143&view=rev</a><br>
Log:<br>
[Statepoints 4/4] Statepoint infrastructure for garbage
collection: Documentation<br>
<br>
This is the fourth and final patch in the statepoint
series. It contains the documentation for the statepoint
intrinsics and their usage.<br>
<br>
There's definitely still room to improve the documentation
here, but I wanted to get this landed so it was available
for others. There will likely be a series of small cleanup
changes over the next few weeks as we work to clarify and
revise the documentation. If you have comments or
questions, please feel free to discuss them either in this
commit thread, the original review thread, or on llvmdev.
Comments are more than welcome.<br>
<br>
Reviewed by: atrick, ributzka<br>
Differential Revision: <a moz-do-not-send="true"
href="http://reviews.llvm.org/D5683" target="_blank">http://reviews.llvm.org/D5683</a><br>
<br>
<br>
<br>
Added:<br>
llvm/trunk/docs/Statepoints.rst<br>
<br>
Added: llvm/trunk/docs/Statepoints.rst<br>
URL: <a moz-do-not-send="true"
href="http://llvm.org/viewvc/llvm-project/llvm/trunk/docs/Statepoints.rst?rev=223143&view=auto"
target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/docs/Statepoints.rst?rev=223143&view=auto</a><br>
==============================================================================<br>
--- llvm/trunk/docs/Statepoints.rst (added)<br>
+++ llvm/trunk/docs/Statepoints.rst Tue Dec 2 13:37:00 2014<br>
@@ -0,0 +1,209 @@<br>
+=====================================<br>
+Garbage Collection Safepoints in LLVM<br>
+=====================================<br>
+<br>
+.. contents::<br>
+ :local:<br>
+ :depth: 2<br>
+<br>
+Status<br>
+=======<br>
+<br>
+This document describes a set of experimental extensions to
LLVM. Use with caution. Because the intrinsics have
experimental status, compatibility across LLVM releases is
not guaranteed.<br>
+<br>
+LLVM currently supports an alternate mechanism for
conservative garbage collection support using the gc_root
intrinsic. The mechanism described here shares little in
common with the alternate implementation and it is hoped
that this mechanism will eventually replace the gc_root
mechanism.<br>
+<br>
+Overview<br>
+========<br>
+<br>
+To collect dead objects, garbage collectors must be able to
identify any references to objects contained within
executing code, and, depending on the collector, potentially
update them. The collector does not need this information
at all points in code - that would make the problem much
harder - but only at well defined points in the execution
known as 'safepoints' For a most collectors, it is
sufficient to track at least one copy of each unique pointer
value. However, for a collector which wishes to relocate
objects directly reachable from running code, a higher
standard is required.<br>
+<br>
+One additional challenge is that the compiler may compute
intermediate results ("derived pointers") which point
outside of the allocation or even into the middle of another
allocation. The eventual use of this intermediate value
must yield an address within the bounds of the allocation,
but such "exterior derived pointers" may be visible to the
collector. Given this, a garbage collector can not safely
rely on the runtime value of an address to indicate the
object it is associated with. If the garbage collector
wishes to move any object, the compiler must provide a
mapping for each pointer to an indication of its allocation.<br>
+<br>
+To simplify the interaction between a collector and the
compiled code, most garbage collectors are organized in
terms of two three abstractions: load barriers, store
barriers, and safepoints.<br>
+<br>
+#. A load barrier is a bit of code executed immediately
after the machine load instruction, but before any use of
the value loaded. Depending on the collector, such a
barrier may be needed for all loads, merely loads of a
particular type (in the original source language), or none
at all.<br>
+#. Analogously, a store barrier is a code fragement that
runs immediately before the machine store instruction, but
after the computation of the value stored. The most common
use of a store barrier is to update a 'card table' in a
generational garbage collector.<br>
+<br>
+#. A safepoint is a location at which pointers visible to
the compiled code (i.e. currently in registers or on the
stack) are allowed to change. After the safepoint
completes, the actual pointer value may differ, but the
'object' (as seen by the source language) pointed to will
not.<br>
+<br>
+ Note that the term 'safepoint' is somewhat overloaded.
It refers to both the location at which the machine state is
parsable and the coordination protocol involved in bring
application threads to a point at which the collector can
safely use that information. The term "statepoint" as used
in this document refers exclusively to the former.<br>
+<br>
+This document focuses on the last item - compiler support
for safepoints in generated code. We will assume that an
outside mechanism has decided where to place safepoints.
From our perspective, all safepoints will be function
calls. To support relocation of objects directly reachable
from values in compiled code, the collector must be able to:<br>
+<br>
+#. identify every copy of a pointer (including copies
introduced by the compiler itself) at the safepoint,<br>
+#. identify which object each pointer relates to, and<br>
+#. potentially update each of those copies.<br>
+<br>
+This document describes the mechanism by which an LLVM
based compiler can provide this information to a language
runtime/collector and ensure that all pointers can be read
and updated if desired. The heart of the approach is to
construct (or rewrite) the IR in a manner where the possible
updates performed by the garbage collector are explicitly
visible in the IR. Doing so requires that we:<br>
+<br>
+#. create a new SSA value for each potentially relocated
pointer, and ensure that no uses of the original (non
relocated) value is reachable after the safepoint,<br>
+#. specify the relocation in a way which is opaque to the
compiler to ensure that the optimizer can not introduce new
uses of an unrelocated value after a statepoint. This
prevents the optimizer from performing unsound
optimizations.<br>
+#. recording a mapping of live pointers (and the allocation
they're associated with) for each statepoint.<br>
+<br>
+At the most abstract level, inserting a safepoint can be
thought of as replacing a call instruction with a call to a
multiple return value function which both calls the original
target of the call, returns it's result, and returns updated
values for any live pointers to garbage collected objects.<br>
+<br>
+ Note that the task of identifying all live pointers to
garbage collected values, transforming the IR to expose a
pointer giving the base object for every such live pointer,
and inserting all the intrinsics correctly is explicitly out
of scope for this document. The recommended approach is
described in the section of Late Safepoint Placement below.<br>
+<br>
+This abstract function call is concretely represented by a
sequence of intrinsic calls known as a 'statepoint
sequence'.<br>
+<br>
+<br>
+Let's consider a simple call in LLVM IR:<br>
+ todo<br>
+<br>
+Depending on our language we may need to allow a safepoint
during the execution of the function called from this site.
If so, we need to let the collector update local values in
the current frame.<br>
+<br>
+Let's say we need to relocate SSA values 'a', 'b', and 'c'
at this safepoint. To represent this, we would generate the
statepoint sequence::<br>
+ put an example sequence here<br>
+<br>
+Ideally, this sequence would have been represented as a M
argument, N return value function (where M is the number of
values being relocated + the original call arguments and N
is the original return value + each relocated value), but
LLVM does not easily support such a representation.<br>
+<br>
+Instead, the statepoint intrinsic marks the actual site of
the safepoint or statepoint. The statepoint returns a token
value (which exists only at compile time). To get back the
original return value of the call, we use the 'gc_result'
intrinsic. To get the relocation of each pointer in turn,
we use the 'gc_relocate' intrinsic with the appropriate
index. Note that both the gc_relocate and gc_result are
tied to the statepoint. The combination forms a "statepoint
sequence" and represents the entitety of a parseable call or
'statepoint'.<br>
+<br>
+When lowered, this example would generate the following x86
assembly::<br>
+ put assembly here<br>
+<br>
+Each of the potentially relocated values has been spilled
to the stack, and a record of that location has been
recorded to the StackMap section. If the garbage collector
needs to update any of these pointers during the call, it
knows exactly what to change.<br>
+<br>
+Intrinsics<br>
+===========<br>
+<br>
+'''gc_statepoint''' Intrinsic<br>
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^<br>
+<br>
+Syntax:<br>
+"""""""<br>
+<br>
+::<br>
+<br>
+ declare i32<br>
+ @gc_statepoint(func_type <target>, i64
<#call args>.<br>
+ i64 <unused>, ... (call
parameters),<br>
+ i64 <# deopt args>, ... (deopt
parameters),<br>
+ ... (gc parameters))<br>
+<br>
+Overview:<br>
+"""""""""<br>
+<br>
+The statepoint intrinsic represents a call which is
parse-able by the runtime.<br>
+<br>
+Operands:<br>
+"""""""""<br>
+<br>
+The 'target' operand is the function actually being
called. The target can be specified as either a symbolic
LLVM funciton, or as an arbitrary Value of appropriate
function type. Note that the function type must match the
signature of the callee and the types of the 'call
parameters' arguments.<br>
+<br>
+The '#call args' operand is the number of arguments to the
actual call. It must exactly match the number of arguments
passed in the 'call parameters' variable length section.<br>
+<br>
+The 'unused' operand is unused and likely to be removed.
Please do not use.<br>
+<br>
+The 'call parameters' arguments are simply the arguments
which need to be passed to the call target. They will be
lowered according to the specified calling convention and
otherwise handled like a normal call instruction. The
number of arguments must exactly match what is specified in
'# call args'. The types must match the signature of
'target'.<br>
+<br>
+The 'deopt parameters' arguments contain an arbitrary list
of Values which is meaningful to the runtime. The runtime
may read any of these values, but is assumed not to modify
them. If the garbage collector might need to modify one of
these values, it must also be listed in the 'gc pointer'
argument list. The '# deopt args' field indicates how many
operands are to be interpreted as 'deopt parameters'.<br>
+<br>
+The 'gc parameters' arguments contain every pointer to a
garbage collector object which potentially needs to be
updated by the garbage collector. Note that the argument
list must explicitly contain a base pointer for every
derived pointer listed. The order of arguments is
unimportant. Unlike the other variable length parameter
sets, this list is not length prefixed.<br>
+<br>
+Semantics:<br>
+""""""""""<br>
+<br>
+A statepoint is assumed to read and write all memory. As a
result, memory operations can not be reordered past a
statepoint. It is illegal to mark a statepoint as being
either 'readonly' or 'readnone'.<br>
+<br>
+Note that legal IR can not perform any memory operation on
a 'gc pointer' argument of the statepoint in a location
statically reachable from the statepoint. Instead, the
explicitly relocated value (from a ''gc_relocate'') must be
used.<br>
+<br>
+'''gc_result''' Intrinsic<br>
+^^^^^^^^^^^^^^^^^^^^^^^^^^<br>
+<br>
+Syntax:<br>
+"""""""<br>
+<br>
+::<br>
+<br>
+ declare type*<br>
+ @gc_result_ptr(i32 %statepoint_token)<br>
+<br>
+ declare fX<br>
+ @gc_result_float(i32 %statepoint_token)<br>
+<br>
+ declare iX<br>
+ @gc_result_int(i32 %statepoint_token)<br>
+<br>
+Overview:<br>
+"""""""""<br>
+<br>
+'''gc_result''' extracts the result of the original call
instruction which was replaced by the '''gc_statepoint'''.
The '''gc_result''' intrinsic is actually a family of three
intrinsics due to an implementation limitation. Other than
the type of the return value, the semantics are the same.<br>
+<br>
+Operands:<br>
+"""""""""<br>
+<br>
+The first and only argument is the '''gc.statepoint'''
which starts the safepoint sequence of which this
'''gc_result'' is a part. Despite the typing of this as a
generic i32, *only* the value defined by a
'''gc.statepoint''' is legal here.<br>
+<br>
+Semantics:<br>
+""""""""""<br>
+<br>
+The ''gc_result'' represents the return value of the call
target of the ''statepoint''. The type of the ''gc_result''
must exactly match the type of the target. If the call
target returns void, there will be no ''gc_result''.<br>
+<br>
+A ''gc_result'' is modeled as a 'readnone' pure function.
It has no side effects since it is just a projection of the
return value of the previous call represented by the
''gc_statepoint''.<br>
+<br>
+'''gc_relocate''' Intrinsic<br>
+^^^^^^^^^^^^^^^^^^^^^^^^^^^<br>
+<br>
+Syntax:<br>
+"""""""<br>
+<br>
+::<br>
+<br>
+ declare <type> addrspace(1)*<br>
+ @gc_relocate(i32 %token, i32 %base_offset, i32
%pointer_offset)<br>
+<br>
+Overview:<br>
+"""""""""<br>
+<br>
+A ''gc_relocate'' returns the potentially relocated value
of a pointer at the safepoint.<br>
+<br>
+Operands:<br>
+"""""""""<br>
+<br>
+The first argument is the '''gc.statepoint''' which starts
the safepoint sequence of which this '''gc_relocation'' is a
part. Despite the typing of this as a generic i32, *only*
the value defined by a '''gc.statepoint''' is legal here.<br>
+<br>
+The second argument is an index into the statepoints list
of arguments which specifies the base pointer for the
pointer being relocated. This index must land within the
'gc parameter' section of the statepoint's argument list.<br>
+<br>
+The third argument is an index into the statepoint's list
of arguments which specify the (potentially) derived pointer
being relocated. It is legal for this index to be the same
as the second argument if-and-only-if a base pointer is
being relocated. This index must land within the 'gc
parameter' section of the statepoint's argument list.<br>
+<br>
+Semantics:<br>
+""""""""""<br>
+The return value of ''gc_relocate'' is the potentially
relocated value of the pointer specified by it's arguments.
It is unspecified how the value of the returned pointer
relates to the argument to the ''gc_statepoint'' other than
that a) it points to the same source language object with
the same offset, and b) the 'based-on' relationship of the
newly relocated pointers is a projection of the unrelocated
pointers. In particular, the integer value of the pointer
returned is unspecified.<br>
+<br>
+A ''gc_relocate'' is modeled as a 'readnone' pure
function. It has no side effects since it is just a way to
extract information about work done during the actual call
modeled by the ''gc_statepoint''.<br>
+<br>
+<br>
+StackMap Format<br>
+================<br>
+<br>
+Locations for each pointer value which may need read and/or
updated by the runtime or collector are provided via the
StackMap format specified in the PatchPoint documentation.<br>
+<br>
+.. TODO: link<br>
+<br>
+Each statepoint generates the following Locations:<br>
+<br>
+* Constant which describes number of following deopt
*Locations* (not operands)<br>
+* Variable number of Locations, one for each deopt
parameter listed in the IR statepoint (same number as
described by previous Constant)<br>
+* Variable number of Locations pairs, one pair for each
unique pointer which needs relocated. The first Location in
each pair describes the base pointer for the object. The
second is the derived pointer actually being relocated. It
is guaranteed that the base pointer must also appear
explicitly as a relocation pair if used after the
statepoint. There may be fewer pairs then gc parameters in
the IR statepoint. Each *unique* pair will occur at least
once; duplicates are possible.<br>
+<br>
+Note that the Locations used in each section may describe
the same physical location. e.g. A stack slot may appear as
a deopt location, a gc base pointer, and a gc derived
pointer.<br>
+<br>
+The ID field of the 'StkMapRecord' for a statepoint is
meaningless and it's value is explicitly unspecified.<br>
+<br>
+The LiveOut section of the StkMapRecord will be empty for a
statepoint record.<br>
+<br>
+Safepoint Semantics & Verification<br>
+==================================<br>
+<br>
+The fundamental correctness property for the compiled
code's correctness w.r.t. the garbage collector is a dynamic
one. It must be the case that there is no dynamic trace
such that a operation involving a potentially relocated
pointer is observably-after a safepoint which could relocate
it. 'observably-after' is this usage means that an outside
observer could observe this sequence of events in a way
which precludes the operation being performed before the
safepoint.<br>
+<br>
+To understand why this 'observable-after' property is
required, consider a null comparison performed on the
original copy of a relocated pointer. Assuming that control
flow follows the safepoint, there is no way to observe
externally whether the null comparison is performed before
or after the safepoint. (Remember, the original Value is
unmodified by the safepoint.) The compiler is free to make
either scheduling choice.<br>
+<br>
+The actual correctness property implemented is slightly
stronger than this. We require that there be no *static
path* on which a potentially relocated pointer is
'observably-after' it may have been relocated. This is
slightly stronger than is strictly necessary (and thus may
disallow some otherwise valid programs), but greatly
simplifies reasoning about correctness of the compiled code.<br>
+<br>
+By construction, this property will be upheld by the
optimizer if correctly established in the source IR. This
is a key invariant of the design.<br>
+<br>
+The existing IR Verifier pass has been extended to check
most of the local restrictions on the intrinsics mentioned
in their respective documentation. The current
implementation in LLVM does not check the key relocation
invariant, but this is ongoing work on developing such a
verifier. Please ask on llvmdev if you're interested in
experimenting with the current version.<br>
+<br>
<br>
<br>
_______________________________________________<br>
llvm-commits mailing list<br>
<a moz-do-not-send="true"
href="mailto:llvm-commits@cs.uiuc.edu">llvm-commits@cs.uiuc.edu</a><br>
<a moz-do-not-send="true"
href="http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits"
target="_blank">http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits</a><br>
</blockquote>
</div>
<br>
</div>
</blockquote>
<br>
</body>
</html>