1 ===================================
2 Stack maps and patch points in LLVM
3 ===================================
12 In this document we refer to the "runtime" collectively as all
13 components that serve as the LLVM client, including the LLVM IR
14 generator, object code consumer, and code patcher.
16 A stack map records the location of ``live values`` at a particular
17 instruction address. These ``live values`` do not refer to all the
18 LLVM values live across the stack map. Instead, they are only the
19 values that the runtime requires to be live at this point. For
20 example, they may be the values the runtime will need to resume
21 program execution at that point independent of the compiled function
22 containing the stack map.
24 LLVM emits stack map data into the object code within a designated
25 :ref:`stackmap-section`. This stack map data contains a record for
26 each stack map. The record stores the stack map's instruction address
27 and contains a entry for each mapped value. Each entry encodes a
28 value's location as a register, stack offset, or constant.
30 A patch point is an instruction address at which space is reserved for
31 patching a new instruction sequence at run time. Patch points look
32 much like calls to LLVM. They take arguments that follow a calling
33 convention and may return a value. They also imply stack map
34 generation, which allows the runtime to locate the patchpoint and
35 find the location of ``live values`` at that point.
40 This functionality is currently experimental but is potentially useful
41 in a variety of settings, the most obvious being a runtime (JIT)
42 compiler. Example applications of the patchpoint intrinsics are
43 implementing an inline call cache for polymorphic method dispatch or
44 optimizing the retrieval of properties in dynamically typed languages
47 The intrinsics documented here are currently used by the JavaScript
48 compiler within the open source WebKit project, see the `FTL JIT
49 <https://trac.webkit.org/wiki/FTLJIT>`_, but they are designed to be
50 used whenever stack maps or code patching are needed. Because the
51 intrinsics have experimental status, compatibility across LLVM
52 releases is not guaranteed.
54 The stack map functionality described in this document is separate
55 from the functionality described in
56 :ref:`stack-map`. `GCFunctionMetadata` provides the location of
57 pointers into a collected heap captured by the `GCRoot` intrinsic,
58 which can also be considered a "stack map". Unlike the stack maps
59 defined above, the `GCFunctionMetadata` stack map interface does not
60 provide a way to associate live register values of arbitrary type with
61 an instruction address, nor does it specify a format for the resulting
62 stack map. The stack maps described here could potentially provide
63 richer information to a garbage collecting runtime, but that usage
64 will not be discussed in this document.
69 The following two kinds of intrinsics can be used to implement stack
70 maps and patch points: ``llvm.experimental.stackmap`` and
71 ``llvm.experimental.patchpoint``. Both kinds of intrinsics generate a
72 stack map record, and they both allow some form of code patching. They
73 can be used independently (i.e. ``llvm.experimental.patchpoint``
74 implicitly generates a stack map without the need for an additional
75 call to ``llvm.experimental.stackmap``). The choice of which to use
76 depends on whether it is necessary to reserve space for code patching
77 and whether any of the intrinsic arguments should be lowered according
78 to calling conventions. ``llvm.experimental.stackmap`` does not
79 reserve any space, nor does it expect any call arguments. If the
80 runtime patches code at the stack map's address, it will destructively
81 overwrite the program text. This is unlike
82 ``llvm.experimental.patchpoint``, which reserves space for in-place
83 patching without overwriting surrounding code. The
84 ``llvm.experimental.patchpoint`` intrinsic also lowers a specified
85 number of arguments according to its calling convention. This allows
86 patched code to make in-place function calls without marshaling.
88 Each instance of one of these intrinsics generates a stack map record
89 in the :ref:`stackmap-section`. The record includes an ID, allowing
90 the runtime to uniquely identify the stack map, and the offset within
91 the code from the beginning of the enclosing function.
93 '``llvm.experimental.stackmap``' Intrinsic
94 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
102 @llvm.experimental.stackmap(i64 <id>, i32 <numShadowBytes>, ...)
107 The '``llvm.experimental.stackmap``' intrinsic records the location of
108 specified values in the stack map without generating any code.
113 The first operand is an ID to be encoded within the stack map. The
114 second operand is the number of shadow bytes following the
115 intrinsic. The variable number of operands that follow are the ``live
116 values`` for which locations will be recorded in the stack map.
118 To use this intrinsic as a bare-bones stack map, with no code patching
119 support, the number of shadow bytes can be set to zero.
124 The stack map intrinsic generates no code in place, unless nops are
125 needed to cover its shadow (see below). However, its offset from
126 function entry is stored in the stack map. This is the relative
127 instruction address immediately following the instructions that
128 precede the stack map.
130 The stack map ID allows a runtime to locate the desired stack map
131 record. LLVM passes this ID through directly to the stack map
132 record without checking uniqueness.
134 LLVM guarantees a shadow of instructions following the stack map's
135 instruction offset during which neither the end of the basic block nor
136 another call to ``llvm.experimental.stackmap`` or
137 ``llvm.experimental.patchpoint`` may occur. This allows the runtime to
138 patch the code at this point in response to an event triggered from
139 outside the code. The code for instructions following the stack map
140 may be emitted in the stack map's shadow, and these instructions may
141 be overwritten by destructive patching. Without shadow bytes, this
142 destructive patching could overwrite program text or data outside the
143 current function. We disallow overlapping stack map shadows so that
144 the runtime does not need to consider this corner case.
146 For example, a stack map with 8 byte shadow:
151 call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 77, i32 8,
153 %val = load i64* %ptr
154 %add = add i64 %val, 3
157 May require one byte of nop-padding:
162 0x05 nop <--- stack map address
163 0x06 movq (%rdi), %rax
166 0x0b ret <---- end of 8-byte shadow
168 Now, if the runtime needs to invalidate the compiled code, it may
169 patch 8 bytes of code at the stack map's address at follows:
174 0x05 movl $0xffff, %rax <--- patched code at stack map address
175 0x0a callq *%rax <---- end of 8-byte shadow
177 This way, after the normal call to the runtime returns, the code will
178 execute a patched call to a special entry point that can rebuild a
179 stack frame from the values located by the stack map.
181 '``llvm.experimental.patchpoint.*``' Intrinsic
182 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
190 @llvm.experimental.patchpoint.void(i64 <id>, i32 <numBytes>,
191 i8* <target>, i32 <numArgs>, ...)
193 @llvm.experimental.patchpoint.i64(i64 <id>, i32 <numBytes>,
194 i8* <target>, i32 <numArgs>, ...)
199 The '``llvm.experimental.patchpoint.*``' intrinsics creates a function
200 call to the specified ``<target>`` and records the location of specified
201 values in the stack map.
206 The first operand is an ID, the second operand is the number of bytes
207 reserved for the patchable region, the third operand is the target
208 address of a function (optionally null), and the fourth operand
209 specifies how many of the following variable operands are considered
210 function call arguments. The remaining variable number of operands are
211 the ``live values`` for which locations will be recorded in the stack
217 The patch point intrinsic generates a stack map. It also emits a
218 function call to the address specified by ``<target>`` if the address
219 is not a constant null. The function call and its arguments are
220 lowered according to the calling convention specified at the
221 intrinsic's callsite. Variants of the intrinsic with non-void return
222 type also return a value according to calling convention.
224 On PowerPC, note that ``<target>`` must be the ABI function pointer for the
225 intended target of the indirect call. Specifically, when compiling for the
226 ELF V1 ABI, ``<target>`` is the function-descriptor address normally used as
227 the C/C++ function-pointer representation.
229 Requesting zero patch point arguments is valid. In this case, all
230 variable operands are handled just like
231 ``llvm.experimental.stackmap.*``. The difference is that space will
232 still be reserved for patching, a call will be emitted, and a return
235 The location of the arguments are not normally recorded in the stack
236 map because they are already fixed by the calling convention. The
237 remaining ``live values`` will have their location recorded, which
238 could be a register, stack location, or constant. A special calling
239 convention has been introduced for use with stack maps, anyregcc,
240 which forces the arguments to be loaded into registers but allows
241 those register to be dynamically allocated. These argument registers
242 will have their register locations recorded in the stack map in
243 addition to the remaining ``live values``.
245 The patch point also emits nops to cover at least ``<numBytes>`` of
246 instruction encoding space. Hence, the client must ensure that
247 ``<numBytes>`` is enough to encode a call to the target address on the
248 supported targets. If the call target is constant null, then there is
249 no minimum requirement. A zero-byte null target patchpoint is
252 The runtime may patch the code emitted for the patch point, including
253 the call sequence and nops. However, the runtime may not assume
254 anything about the code LLVM emits within the reserved space. Partial
255 patching is not allowed. The runtime must patch all reserved bytes,
256 padding with nops if necessary.
258 This example shows a patch point reserving 15 bytes, with one argument
259 in $rdi, and a return value in $rax per native calling convention:
263 %target = inttoptr i64 -281474976710654 to i8*
264 %val = call i64 (i64, i32, ...)*
265 @llvm.experimental.patchpoint.i64(i64 78, i32 15,
266 i8* %target, i32 1, i64* %ptr)
267 %add = add i64 %val, 3
274 0x00 movabsq $0xffff000000000002, %r11 <--- patch point address
277 0x0e nop <--- end of reserved 15-bytes
279 0x10 movl %rax, 8(%rsp)
281 Note that no stack map locations will be recorded. If the patched code
282 sequence does not need arguments fixed to specific calling convention
283 registers, then the ``anyregcc`` convention may be used:
287 %val = call anyregcc @llvm.experimental.patchpoint(i64 78, i32 15,
291 The stack map now indicates the location of the %ptr argument and
296 Stack Map: ID=78, Loc0=%r9 Loc1=%r8
298 The patch code sequence may now use the argument that happened to be
299 allocated in %r8 and return a value allocated in %r9:
303 0x00 movslq 4(%r8) %r9 <--- patched code at patch point address
306 0x0e nop <--- end of reserved 15-bytes
308 0x10 movl %r9, 8(%rsp)
315 The existence of a stack map or patch point intrinsic within an LLVM
316 Module forces code emission to create a :ref:`stackmap-section`. The
317 format of this section follows:
322 uint8 : Stack Map Version (current version is 1)
323 uint8 : Reserved (expected to be 0)
324 uint16 : Reserved (expected to be 0)
326 uint32 : NumFunctions
327 uint32 : NumConstants
329 StkSizeRecord[NumFunctions] {
330 uint64 : Function Address
333 Constants[NumConstants] {
334 uint64 : LargeConstant
336 StkMapRecord[NumRecords] {
337 uint64 : PatchPoint ID
338 uint32 : Instruction Offset
339 uint16 : Reserved (record flags)
340 uint16 : NumLocations
341 Location[NumLocations] {
342 uint8 : Register | Direct | Indirect | Constant | ConstantIndex
343 uint8 : Reserved (location flags)
344 uint16 : Dwarf RegNum
345 int32 : Offset or SmallConstant
349 LiveOuts[NumLiveOuts]
350 uint16 : Dwarf RegNum
352 uint8 : Size in Bytes
354 uint32 : Padding (only if required to align to 8 byte)
357 The first byte of each location encodes a type that indicates how to
358 interpret the ``RegNum`` and ``Offset`` fields as follows:
360 ======== ========== =================== ===========================
361 Encoding Type Value Description
362 -------- ---------- ------------------- ---------------------------
363 0x1 Register Reg Value in a register
364 0x2 Direct Reg + Offset Frame index value
365 0x3 Indirect [Reg + Offset] Spilled value
366 0x4 Constant Offset Small constant
367 0x5 ConstIndex Constants[Offset] Large constant
368 ======== ========== =================== ===========================
370 In the common case, a value is available in a register, and the
371 ``Offset`` field will be zero. Values spilled to the stack are encoded
372 as ``Indirect`` locations. The runtime must load those values from a
373 stack address, typically in the form ``[BP + Offset]``. If an
374 ``alloca`` value is passed directly to a stack map intrinsic, then
375 LLVM may fold the frame index into the stack map as an optimization to
376 avoid allocating a register or stack slot. These frame indices will be
377 encoded as ``Direct`` locations in the form ``BP + Offset``. LLVM may
378 also optimize constants by emitting them directly in the stack map,
379 either in the ``Offset`` of a ``Constant`` location or in the constant
380 pool, referred to by ``ConstantIndex`` locations.
382 At each callsite, a "liveout" register list is also recorded. These
383 are the registers that are live across the stackmap and therefore must
384 be saved by the runtime. This is an important optimization when the
385 patchpoint intrinsic is used with a calling convention that by default
386 preserves most registers as callee-save.
388 Each entry in the liveout register list contains a DWARF register
389 number and size in bytes. The stackmap format deliberately omits
390 specific subregister information. Instead the runtime must interpret
391 this information conservatively. For example, if the stackmap reports
392 one byte at ``%rax``, then the value may be in either ``%al`` or
393 ``%ah``. It doesn't matter in practice, because the runtime will
394 simply save ``%rax``. However, if the stackmap reports 16 bytes at
395 ``%ymm0``, then the runtime can safely optimize by saving only
398 The stack map format is a contract between an LLVM SVN revision and
399 the runtime. It is currently experimental and may change in the short
400 term, but minimizing the need to update the runtime is
401 important. Consequently, the stack map design is motivated by
402 simplicity and extensibility. Compactness of the representation is
403 secondary because the runtime is expected to parse the data
404 immediately after compiling a module and encode the information in its
405 own format. Since the runtime controls the allocation of sections, it
406 can reuse the same stack map space for multiple modules.
408 Stackmap support is currently only implemented for 64-bit
409 platforms. However, a 32-bit implementation should be able to use the
410 same format with an insignificant amount of wasted space.
412 .. _stackmap-section:
417 A JIT compiler can easily access this section by providing its own
418 memory manager via the LLVM C API
419 ``LLVMCreateSimpleMCJITMemoryManager()``. When creating the memory
420 manager, the JIT provides a callback:
421 ``LLVMMemoryManagerAllocateDataSectionCallback()``. When LLVM creates
422 this section, it invokes the callback and passes the section name. The
423 JIT can record the in-memory address of the section at this time and
424 later parse it to recover the stack map data.
426 On Darwin, the stack map section name is "__llvm_stackmaps". The
427 segment name is "__LLVM_STACKMAPS".
432 The stack map support described in this document can be used to
433 precisely determine the location of values at a specific position in
434 the code. LLVM does not maintain any mapping between those values and
435 any higher-level entity. The runtime must be able to interpret the
436 stack map record given only the ID, offset, and the order of the
437 locations, which LLVM preserves.
439 Note that this is quite different from the goal of debug information,
440 which is a best-effort attempt to track the location of named
441 variables at every instruction.
443 An important motivation for this design is to allow a runtime to
444 commandeer a stack frame when execution reaches an instruction address
445 associated with a stack map. The runtime must be able to rebuild a
446 stack frame and resume program execution using the information
447 provided by the stack map. For example, execution may resume in an
448 interpreter or a recompiled version of the same function.
450 This usage restricts LLVM optimization. Clearly, LLVM must not move
451 stores across a stack map. However, loads must also be handled
452 conservatively. If the load may trigger an exception, hoisting it
453 above a stack map could be invalid. For example, the runtime may
454 determine that a load is safe to execute without a type check given
455 the current state of the type system. If the type system changes while
456 some activation of the load's function exists on the stack, the load
457 becomes unsafe. The runtime can prevent subsequent execution of that
458 load by immediately patching any stack map location that lies between
459 the current call site and the load (typically, the runtime would
460 simply patch all stack map locations to invalidate the function). If
461 the compiler had hoisted the load above the stack map, then the
462 program could crash before the runtime could take back control.
464 To enforce these semantics, stackmap and patchpoint intrinsics are
465 considered to potentially read and write all memory. This may limit
466 optimization more than some clients desire. This limitation may be
467 avoided by marking the call site as "readonly". In the future we may
468 also allow meta-data to be added to the intrinsic call to express
469 aliasing, thereby allowing optimizations to hoist certain loads above
472 Direct Stack Map Entries
473 ^^^^^^^^^^^^^^^^^^^^^^^^
475 As shown in :ref:`stackmap-section`, a Direct stack map location
476 records the address of frame index. This address is itself the value
477 that the runtime requested. This differs from Indirect locations,
478 which refer to a stack locations from which the requested values must
479 be loaded. Direct locations can communicate the address if an alloca,
480 while Indirect locations handle register spills.
488 llvm.experimental.stackmap(i64 <ID>, i32 <shadowBytes>, i64* %a)
490 The runtime can determine this alloca's relative location on the
491 stack immediately after compilation, or at any time thereafter. This
492 differs from Register and Indirect locations, because the runtime can
493 only read the values in those locations when execution reaches the
494 instruction address of the stack map.
496 This functionality requires LLVM to treat entry-block allocas
497 specially when they are directly consumed by an intrinsics. (This is
498 the same requirement imposed by the llvm.gcroot intrinsic.) LLVM
499 transformations must not substitute the alloca with any intervening
500 value. This can be verified by the runtime simply by checking that the
501 stack map's location is a Direct location type.
504 Supported Architectures
505 =======================
507 Support for StackMap generation and the related intrinsics requires
508 some code for each backend. Today, only a subset of LLVM's backends
509 are supported. The currently supported architectures are X86_64,
510 PowerPC, and Aarch64.