xref: /llvm-project/llvm/docs/CodeGenerator.rst (revision 0e24179797faf5d309fcc5024131c2293accff26)
1==========================================
2The LLVM Target-Independent Code Generator
3==========================================
4
5.. role:: raw-html(raw)
6   :format: html
7
8.. raw:: html
9
10  <style>
11    .unknown { background-color: #C0C0C0; text-align: center; }
12    .unknown:before { content: "?" }
13    .no { background-color: #C11B17 }
14    .no:before { content: "N" }
15    .partial { background-color: #F88017 }
16    .yes { background-color: #0F0; }
17    .yes:before { content: "Y" }
18    .na { background-color: #6666FF; }
19    .na:before { content: "N/A" }
20  </style>
21
22.. contents::
23   :local:
24
25.. warning::
26  This is a work in progress.
27
28Introduction
29============
30
31The LLVM target-independent code generator is a framework that provides a suite
32of reusable components for translating the LLVM internal representation to the
33machine code for a specified target---either in assembly form (suitable for a
34static compiler) or in binary machine code format (usable for a JIT
35compiler). The LLVM target-independent code generator consists of six main
36components:
37
381. `Abstract target description`_ interfaces which capture important properties
39   about various aspects of the machine, independently of how they will be used.
40   These interfaces are defined in ``include/llvm/Target/``.
41
422. Classes used to represent the `code being generated`_ for a target.  These
43   classes are intended to be abstract enough to represent the machine code for
44   *any* target machine.  These classes are defined in
45   ``include/llvm/CodeGen/``. At this level, concepts like "constant pool
46   entries" and "jump tables" are explicitly exposed.
47
483. Classes and algorithms used to represent code at the object file level, the
49   `MC Layer`_.  These classes represent assembly level constructs like labels,
50   sections, and instructions.  At this level, concepts like "constant pool
51   entries" and "jump tables" don't exist.
52
534. `Target-independent algorithms`_ used to implement various phases of native
54   code generation (register allocation, scheduling, stack frame representation,
55   etc).  This code lives in ``lib/CodeGen/``.
56
575. `Implementations of the abstract target description interfaces`_ for
58   particular targets.  These machine descriptions make use of the components
59   provided by LLVM, and can optionally provide custom target-specific passes,
60   to build complete code generators for a specific target.  Target descriptions
61   live in ``lib/Target/``.
62
636. The target-independent JIT components.  The LLVM JIT is completely target
64   independent (it uses the ``TargetJITInfo`` structure to interface for
65   target-specific issues.  The code for the target-independent JIT lives in
66   ``lib/ExecutionEngine/JIT``.
67
68Depending on which part of the code generator you are interested in working on,
69different pieces of this will be useful to you.  In any case, you should be
70familiar with the `target description`_ and `machine code representation`_
71classes.  If you want to add a backend for a new target, you will need to
72`implement the target description`_ classes for your new target and understand
73the :doc:`LLVM code representation <LangRef>`.  If you are interested in
74implementing a new `code generation algorithm`_, it should only depend on the
75target-description and machine code representation classes, ensuring that it is
76portable.
77
78Required components in the code generator
79-----------------------------------------
80
81The two pieces of the LLVM code generator are the high-level interface to the
82code generator and the set of reusable components that can be used to build
83target-specific backends.  The two most important interfaces (:raw-html:`<tt>`
84`TargetMachine`_ :raw-html:`</tt>` and :raw-html:`<tt>` `DataLayout`_
85:raw-html:`</tt>`) are the only ones that are required to be defined for a
86backend to fit into the LLVM system, but the others must be defined if the
87reusable code generator components are going to be used.
88
89This design has two important implications.  The first is that LLVM can support
90completely non-traditional code generation targets.  For example, the C backend
91does not require register allocation, instruction selection, or any of the other
92standard components provided by the system.  As such, it only implements these
93two interfaces, and does its own thing. Note that C backend was removed from the
94trunk since LLVM 3.1 release. Another example of a code generator like this is a
95(purely hypothetical) backend that converts LLVM to the GCC RTL form and uses
96GCC to emit machine code for a target.
97
98This design also implies that it is possible to design and implement radically
99different code generators in the LLVM system that do not make use of any of the
100built-in components.  Doing so is not recommended at all, but could be required
101for radically different targets that do not fit into the LLVM machine
102description model: FPGAs for example.
103
104.. _high-level design of the code generator:
105
106The high-level design of the code generator
107-------------------------------------------
108
109The LLVM target-independent code generator is designed to support efficient and
110quality code generation for standard register-based microprocessors.  Code
111generation in this model is divided into the following stages:
112
1131. `Instruction Selection`_ --- This phase determines an efficient way to
114   express the input LLVM code in the target instruction set.  This stage
115   produces the initial code for the program in the target instruction set, then
116   makes use of virtual registers in SSA form and physical registers that
117   represent any required register assignments due to target constraints or
118   calling conventions.  This step turns the LLVM code into a DAG of target
119   instructions.
120
1212. `Scheduling and Formation`_ --- This phase takes the DAG of target
122   instructions produced by the instruction selection phase, determines an
123   ordering of the instructions, then emits the instructions as :raw-html:`<tt>`
124   `MachineInstr`_\s :raw-html:`</tt>` with that ordering.  Note that we
125   describe this in the `instruction selection section`_ because it operates on
126   a `SelectionDAG`_.
127
1283. `SSA-based Machine Code Optimizations`_ --- This optional stage consists of a
129   series of machine-code optimizations that operate on the SSA-form produced by
130   the instruction selector.  Optimizations like modulo-scheduling or peephole
131   optimization work here.
132
1334. `Register Allocation`_ --- The target code is transformed from an infinite
134   virtual register file in SSA form to the concrete register file used by the
135   target.  This phase introduces spill code and eliminates all virtual register
136   references from the program.
137
1385. `Prolog/Epilog Code Insertion`_ --- Once the machine code has been generated
139   for the function and the amount of stack space required is known (used for
140   LLVM alloca's and spill slots), the prolog and epilog code for the function
141   can be inserted and "abstract stack location references" can be eliminated.
142   This stage is responsible for implementing optimizations like frame-pointer
143   elimination and stack packing.
144
1456. `Late Machine Code Optimizations`_ --- Optimizations that operate on "final"
146   machine code can go here, such as spill code scheduling and peephole
147   optimizations.
148
1497. `Code Emission`_ --- The final stage actually puts out the code for the
150   current function, either in the target assembler format or in machine
151   code.
152
153The code generator is based on the assumption that the instruction selector will
154use an optimal pattern matching selector to create high-quality sequences of
155native instructions.  Alternative code generator designs based on pattern
156expansion and aggressive iterative peephole optimization are much slower.  This
157design permits efficient compilation (important for JIT environments) and
158aggressive optimization (used when generating code offline) by allowing
159components of varying levels of sophistication to be used for any step of
160compilation.
161
162In addition to these stages, target implementations can insert arbitrary
163target-specific passes into the flow.  For example, the X86 target uses a
164special pass to handle the 80x87 floating point stack architecture.  Other
165targets with unusual requirements can be supported with custom passes as needed.
166
167Using TableGen for target description
168-------------------------------------
169
170The target description classes require a detailed description of the target
171architecture.  These target descriptions often have a large amount of common
172information (e.g., an ``add`` instruction is almost identical to a ``sub``
173instruction).  In order to allow the maximum amount of commonality to be
174factored out, the LLVM code generator uses the
175:doc:`TableGen/index` tool to describe big chunks of the
176target machine, which allows the use of domain-specific and target-specific
177abstractions to reduce the amount of repetition.
178
179As LLVM continues to be developed and refined, we plan to move more and more of
180the target description to the ``.td`` form.  Doing so gives us a number of
181advantages.  The most important is that it makes it easier to port LLVM because
182it reduces the amount of C++ code that has to be written, and the surface area
183of the code generator that needs to be understood before someone can get
184something working.  Second, it makes it easier to change things. In particular,
185if tables and other things are all emitted by ``tblgen``, we only need a change
186in one place (``tblgen``) to update all of the targets to a new interface.
187
188.. _Abstract target description:
189.. _target description:
190
191Target description classes
192==========================
193
194The LLVM target description classes (located in the ``include/llvm/Target``
195directory) provide an abstract description of the target machine independent of
196any particular client.  These classes are designed to capture the *abstract*
197properties of the target (such as the instructions and registers it has), and do
198not incorporate any particular pieces of code generation algorithms.
199
200All of the target description classes (except the :raw-html:`<tt>` `DataLayout`_
201:raw-html:`</tt>` class) are designed to be subclassed by the concrete target
202implementation, and have virtual methods implemented.  To get to these
203implementations, the :raw-html:`<tt>` `TargetMachine`_ :raw-html:`</tt>` class
204provides accessors that should be implemented by the target.
205
206.. _TargetMachine:
207
208The ``TargetMachine`` class
209---------------------------
210
211The ``TargetMachine`` class provides virtual methods that are used to access the
212target-specific implementations of the various target description classes via
213the ``get*Info`` methods (``getInstrInfo``, ``getRegisterInfo``,
214``getFrameInfo``, etc.).  This class is designed to be specialized by a concrete
215target implementation (e.g., ``X86TargetMachine``) which implements the various
216virtual methods.  The only required target description class is the
217:raw-html:`<tt>` `DataLayout`_ :raw-html:`</tt>` class, but if the code
218generator components are to be used, the other interfaces should be implemented
219as well.
220
221.. _DataLayout:
222
223The ``DataLayout`` class
224------------------------
225
226The ``DataLayout`` class is the only required target description class, and it
227is the only class that is not extensible (you cannot derive a new class from
228it).  ``DataLayout`` specifies information about how the target lays out memory
229for structures, the alignment requirements for various data types, the size of
230pointers in the target, and whether the target is little-endian or
231big-endian.
232
233.. _TargetLowering:
234
235The ``TargetLowering`` class
236----------------------------
237
238The ``TargetLowering`` class is used by SelectionDAG based instruction selectors
239primarily to describe how LLVM code should be lowered to SelectionDAG
240operations.  Among other things, this class indicates:
241
242* an initial register class to use for various ``ValueType``\s,
243
244* which operations are natively supported by the target machine,
245
246* the return type of ``setcc`` operations,
247
248* the type to use for shift amounts, and
249
250* various high-level characteristics, like whether it is profitable to turn
251  division by a constant into a multiplication sequence.
252
253.. _TargetRegisterInfo:
254
255The ``TargetRegisterInfo`` class
256--------------------------------
257
258The ``TargetRegisterInfo`` class is used to describe the register file of the
259target and any interactions between the registers.
260
261Registers are represented in the code generator by unsigned integers.  Physical
262registers (those that actually exist in the target description) are unique
263small numbers, and virtual registers are generally large.  Note that
264register ``#0`` is reserved as a flag value.
265
266Each register in the processor description has an associated
267``TargetRegisterDesc`` entry, which provides a textual name for the register
268(used for assembly output and debugging dumps) and a set of aliases (used to
269indicate whether one register overlaps with another).
270
271In addition to the per-register description, the ``TargetRegisterInfo`` class
272exposes a set of processor specific register classes (instances of the
273``TargetRegisterClass`` class).  Each register class contains sets of registers
274that have the same properties (for example, they are all 32-bit integer
275registers).  Each SSA virtual register created by the instruction selector has
276an associated register class.  When the register allocator runs, it replaces
277virtual registers with a physical register in the set.
278
279The target-specific implementations of these classes is auto-generated from a
280:doc:`TableGen/index` description of the register file.
281
282.. _TargetInstrInfo:
283
284The ``TargetInstrInfo`` class
285-----------------------------
286
287The ``TargetInstrInfo`` class is used to describe the machine instructions
288supported by the target.  Descriptions define things like the mnemonic for
289the opcode, the number of operands, the list of implicit register uses and defs,
290whether the instruction has certain target-independent properties (accesses
291memory, is commutable, etc), and holds any target-specific flags.
292
293The ``TargetFrameLowering`` class
294---------------------------------
295
296The ``TargetFrameLowering`` class is used to provide information about the stack
297frame layout of the target. It holds the direction of stack growth, the known
298stack alignment on entry to each function, and the offset to the local area.
299The offset to the local area is the offset from the stack pointer on function
300entry to the first location where function data (local variables, spill
301locations) can be stored.
302
303The ``TargetSubtarget`` class
304-----------------------------
305
306The ``TargetSubtarget`` class is used to provide information about the specific
307chip set being targeted.  A sub-target informs code generation of which
308instructions are supported, instruction latencies and instruction execution
309itinerary; i.e., which processing units are used, in what order, and for how
310long.
311
312The ``TargetJITInfo`` class
313---------------------------
314
315The ``TargetJITInfo`` class exposes an abstract interface used by the
316Just-In-Time code generator to perform target-specific activities, such as
317emitting stubs.  If a ``TargetMachine`` supports JIT code generation, it should
318provide one of these objects through the ``getJITInfo`` method.
319
320.. _code being generated:
321.. _machine code representation:
322
323Machine code description classes
324================================
325
326At the high-level, LLVM code is translated to a machine specific representation
327formed out of :raw-html:`<tt>` `MachineFunction`_ :raw-html:`</tt>`,
328:raw-html:`<tt>` `MachineBasicBlock`_ :raw-html:`</tt>`, and :raw-html:`<tt>`
329`MachineInstr`_ :raw-html:`</tt>` instances (defined in
330``include/llvm/CodeGen``).  This representation is completely target agnostic,
331representing instructions in their most abstract form: an opcode and a series of
332operands.  This representation is designed to support both an SSA representation
333for machine code, as well as a register allocated, non-SSA form.
334
335.. _MachineInstr:
336
337The ``MachineInstr`` class
338--------------------------
339
340Target machine instructions are represented as instances of the ``MachineInstr``
341class.  This class is an extremely abstract way of representing machine
342instructions.  In particular, it only keeps track of an opcode number and a set
343of operands.
344
345The opcode number is a simple unsigned integer that only has meaning to a
346specific backend.  All of the instructions for a target should be defined in the
347``*InstrInfo.td`` file for the target. The opcode enum values are auto-generated
348from this description.  The ``MachineInstr`` class does not have any information
349about how to interpret the instruction (i.e., what the semantics of the
350instruction are); for that you must refer to the :raw-html:`<tt>`
351`TargetInstrInfo`_ :raw-html:`</tt>` class.
352
353The operands of a machine instruction can be of several different types: a
354register reference, a constant integer, a basic block reference, etc.  In
355addition, a machine operand should be marked as a def or a use of the value
356(though only registers are allowed to be defs).
357
358By convention, the LLVM code generator orders instruction operands so that all
359register definitions come before the register uses, even on architectures that
360are normally printed in other orders.  For example, the SPARC add instruction:
361"``add %i1, %i2, %i3``" adds the "%i1", and "%i2" registers and stores the
362result into the "%i3" register.  In the LLVM code generator, the operands should
363be stored as "``%i3, %i1, %i2``": with the destination first.
364
365Keeping destination (definition) operands at the beginning of the operand list
366has several advantages.  In particular, the debugging printer will print the
367instruction like this:
368
369.. code-block:: llvm
370
371  %r3 = add %i1, %i2
372
373Also if the first operand is a def, it is easier to `create instructions`_ whose
374only def is the first operand.
375
376.. _create instructions:
377
378Using the ``MachineInstrBuilder.h`` functions
379^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
380
381Machine instructions are created by using the ``BuildMI`` functions, located in
382the ``include/llvm/CodeGen/MachineInstrBuilder.h`` file.  The ``BuildMI``
383functions make it easy to build arbitrary machine instructions.  Usage of the
384``BuildMI`` functions look like this:
385
386.. code-block:: c++
387
388  // Create a 'DestReg = mov 42' (rendered in X86 assembly as 'mov DestReg, 42')
389  // instruction and insert it at the end of the given MachineBasicBlock.
390  const TargetInstrInfo &TII = ...
391  MachineBasicBlock &MBB = ...
392  DebugLoc DL;
393  MachineInstr *MI = BuildMI(MBB, DL, TII.get(X86::MOV32ri), DestReg).addImm(42);
394
395  // Create the same instr, but insert it before a specified iterator point.
396  MachineBasicBlock::iterator MBBI = ...
397  BuildMI(MBB, MBBI, DL, TII.get(X86::MOV32ri), DestReg).addImm(42);
398
399  // Create a 'cmp Reg, 0' instruction, no destination reg.
400  MI = BuildMI(MBB, DL, TII.get(X86::CMP32ri8)).addReg(Reg).addImm(42);
401
402  // Create an 'sahf' instruction which takes no operands and stores nothing.
403  MI = BuildMI(MBB, DL, TII.get(X86::SAHF));
404
405  // Create a self looping branch instruction.
406  BuildMI(MBB, DL, TII.get(X86::JNE)).addMBB(&MBB);
407
408If you need to add a definition operand (other than the optional destination
409register), you must explicitly mark it as such:
410
411.. code-block:: c++
412
413  MI.addReg(Reg, RegState::Define);
414
415Fixed (preassigned) registers
416^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
417
418One important issue that the code generator needs to be aware of is the presence
419of fixed registers.  In particular, there are often places in the instruction
420stream where the register allocator *must* arrange for a particular value to be
421in a particular register.  This can occur due to limitations of the instruction
422set (e.g., the X86 can only do a 32-bit divide with the ``EAX``/``EDX``
423registers), or external factors like calling conventions.  In any case, the
424instruction selector should emit code that copies a virtual register into or out
425of a physical register when needed.
426
427For example, consider this simple LLVM example:
428
429.. code-block:: llvm
430
431  define i32 @test(i32 %X, i32 %Y) {
432    %Z = sdiv i32 %X, %Y
433    ret i32 %Z
434  }
435
436The X86 instruction selector might produce this machine code for the ``div`` and
437``ret``:
438
439.. code-block:: text
440
441  ;; Start of div
442  %EAX = mov %reg1024           ;; Copy X (in reg1024) into EAX
443  %reg1027 = sar %reg1024, 31
444  %EDX = mov %reg1027           ;; Sign extend X into EDX
445  idiv %reg1025                 ;; Divide by Y (in reg1025)
446  %reg1026 = mov %EAX           ;; Read the result (Z) out of EAX
447
448  ;; Start of ret
449  %EAX = mov %reg1026           ;; 32-bit return value goes in EAX
450  ret
451
452By the end of code generation, the register allocator would coalesce the
453registers and delete the resultant identity moves producing the following
454code:
455
456.. code-block:: text
457
458  ;; X is in EAX, Y is in ECX
459  mov %EAX, %EDX
460  sar %EDX, 31
461  idiv %ECX
462  ret
463
464This approach is extremely general (if it can handle the X86 architecture, it
465can handle anything!) and allows all of the target specific knowledge about the
466instruction stream to be isolated in the instruction selector.  Note that
467physical registers should have a short lifetime for good code generation, and
468all physical registers are assumed dead on entry to and exit from basic blocks
469(before register allocation).  Thus, if you need a value to be live across basic
470block boundaries, it *must* live in a virtual register.
471
472Call-clobbered registers
473^^^^^^^^^^^^^^^^^^^^^^^^
474
475Some machine instructions, like calls, clobber a large number of physical
476registers.  Rather than adding ``<def,dead>`` operands for all of them, it is
477possible to use an ``MO_RegisterMask`` operand instead.  The register mask
478operand holds a bit mask of preserved registers, and everything else is
479considered to be clobbered by the instruction.
480
481Machine code in SSA form
482^^^^^^^^^^^^^^^^^^^^^^^^
483
484``MachineInstr``'s are initially selected in SSA-form, and are maintained in
485SSA-form until register allocation happens.  For the most part, this is
486trivially simple since LLVM is already in SSA form; LLVM PHI nodes become
487machine code PHI nodes, and virtual registers are only allowed to have a single
488definition.
489
490After register allocation, machine code is no longer in SSA-form because there
491are no virtual registers left in the code.
492
493.. _MachineBasicBlock:
494
495The ``MachineBasicBlock`` class
496-------------------------------
497
498The ``MachineBasicBlock`` class contains a list of machine instructions
499(:raw-html:`<tt>` `MachineInstr`_ :raw-html:`</tt>` instances).  It roughly
500corresponds to the LLVM code input to the instruction selector, but there can be
501a one-to-many mapping (i.e. one LLVM basic block can map to multiple machine
502basic blocks). The ``MachineBasicBlock`` class has a "``getBasicBlock``" method,
503which returns the LLVM basic block that it comes from.
504
505.. _MachineFunction:
506
507The ``MachineFunction`` class
508-----------------------------
509
510The ``MachineFunction`` class contains a list of machine basic blocks
511(:raw-html:`<tt>` `MachineBasicBlock`_ :raw-html:`</tt>` instances).  It
512corresponds one-to-one with the LLVM function input to the instruction selector.
513In addition to a list of basic blocks, the ``MachineFunction`` contains a
514``MachineConstantPool``, a ``MachineFrameInfo``, a ``MachineFunctionInfo``, and
515a ``MachineRegisterInfo``.  See ``include/llvm/CodeGen/MachineFunction.h`` for
516more information.
517
518``MachineInstr Bundles``
519------------------------
520
521LLVM code generator can model sequences of instructions as MachineInstr
522bundles. A MI bundle can model a VLIW group / pack which contains an arbitrary
523number of parallel instructions. It can also be used to model a sequential list
524of instructions (potentially with data dependencies) that cannot be legally
525separated (e.g. ARM Thumb2 IT blocks).
526
527Conceptually a MI bundle is a MI with a number of other MIs nested within:
528
529::
530
531  --------------
532  |   Bundle   | ---------
533  --------------          \
534         |           ----------------
535         |           |      MI      |
536         |           ----------------
537         |                   |
538         |           ----------------
539         |           |      MI      |
540         |           ----------------
541         |                   |
542         |           ----------------
543         |           |      MI      |
544         |           ----------------
545         |
546  --------------
547  |   Bundle   | --------
548  --------------         \
549         |           ----------------
550         |           |      MI      |
551         |           ----------------
552         |                   |
553         |           ----------------
554         |           |      MI      |
555         |           ----------------
556         |                   |
557         |                  ...
558         |
559  --------------
560  |   Bundle   | --------
561  --------------         \
562         |
563        ...
564
565MI bundle support does not change the physical representations of
566MachineBasicBlock and MachineInstr. All the MIs (including top level and nested
567ones) are stored as sequential list of MIs. The "bundled" MIs are marked with
568the 'InsideBundle' flag. A top level MI with the special BUNDLE opcode is used
569to represent the start of a bundle. It's legal to mix BUNDLE MIs with individual
570MIs that are not inside bundles nor represent bundles.
571
572MachineInstr passes should operate on a MI bundle as a single unit. Member
573methods have been taught to correctly handle bundles and MIs inside bundles.
574The MachineBasicBlock iterator has been modified to skip over bundled MIs to
575enforce the bundle-as-a-single-unit concept. An alternative iterator
576instr_iterator has been added to MachineBasicBlock to allow passes to iterate
577over all of the MIs in a MachineBasicBlock, including those which are nested
578inside bundles. The top level BUNDLE instruction must have the correct set of
579register MachineOperand's that represent the cumulative inputs and outputs of
580the bundled MIs.
581
582Packing / bundling of MachineInstrs for VLIW architectures should
583generally be done as part of the register allocation super-pass. More
584specifically, the pass which determines what MIs should be bundled
585together should be done after code generator exits SSA form
586(i.e. after two-address pass, PHI elimination, and copy coalescing).
587Such bundles should be finalized (i.e. adding BUNDLE MIs and input and
588output register MachineOperands) after virtual registers have been
589rewritten into physical registers. This eliminates the need to add
590virtual register operands to BUNDLE instructions which would
591effectively double the virtual register def and use lists. Bundles may
592use virtual registers and be formed in SSA form, but may not be
593appropriate for all use cases.
594
595.. _MC Layer:
596
597The "MC" Layer
598==============
599
600The MC Layer is used to represent and process code at the raw machine code
601level, devoid of "high level" information like "constant pools", "jump tables",
602"global variables" or anything like that.  At this level, LLVM handles things
603like label names, machine instructions, and sections in the object file.  The
604code in this layer is used for a number of important purposes: the tail end of
605the code generator uses it to write a .s or .o file, and it is also used by the
606llvm-mc tool to implement standalone machine code assemblers and disassemblers.
607
608This section describes some of the important classes.  There are also a number
609of important subsystems that interact at this layer, they are described later in
610this manual.
611
612.. _MCStreamer:
613
614The ``MCStreamer`` API
615----------------------
616
617MCStreamer is best thought of as an assembler API.  It is an abstract API which
618is *implemented* in different ways (e.g. to output a .s file, output an ELF .o
619file, etc) but whose API correspond directly to what you see in a .s file.
620MCStreamer has one method per directive, such as EmitLabel, EmitSymbolAttribute,
621switchSection, emitValue (for .byte, .word), etc, which directly correspond to
622assembly level directives.  It also has an EmitInstruction method, which is used
623to output an MCInst to the streamer.
624
625This API is most important for two clients: the llvm-mc stand-alone assembler is
626effectively a parser that parses a line, then invokes a method on MCStreamer. In
627the code generator, the `Code Emission`_ phase of the code generator lowers
628higher level LLVM IR and Machine* constructs down to the MC layer, emitting
629directives through MCStreamer.
630
631On the implementation side of MCStreamer, there are two major implementations:
632one for writing out a .s file (MCAsmStreamer), and one for writing out a .o
633file (MCObjectStreamer).  MCAsmStreamer is a straightforward implementation
634that prints out a directive for each method (e.g. ``EmitValue -> .byte``), but
635MCObjectStreamer implements a full assembler.
636
637For target specific directives, the MCStreamer has a MCTargetStreamer instance.
638Each target that needs it defines a class that inherits from it and is a lot
639like MCStreamer itself: It has one method per directive and two classes that
640inherit from it, a target object streamer and a target asm streamer. The target
641asm streamer just prints it (``emitFnStart -> .fnstart``), and the object
642streamer implement the assembler logic for it.
643
644To make llvm use these classes, the target initialization must call
645TargetRegistry::RegisterAsmStreamer and TargetRegistry::RegisterMCObjectStreamer
646passing callbacks that allocate the corresponding target streamer and pass it
647to createAsmStreamer or to the appropriate object streamer constructor.
648
649The ``MCContext`` class
650-----------------------
651
652The MCContext class is the owner of a variety of uniqued data structures at the
653MC layer, including symbols, sections, etc.  As such, this is the class that you
654interact with to create symbols and sections.  This class can not be subclassed.
655
656The ``MCSymbol`` class
657----------------------
658
659The MCSymbol class represents a symbol (aka label) in the assembly file.  There
660are two interesting kinds of symbols: assembler temporary symbols, and normal
661symbols.  Assembler temporary symbols are used and processed by the assembler
662but are discarded when the object file is produced.  The distinction is usually
663represented by adding a prefix to the label, for example "L" labels are
664assembler temporary labels in MachO.
665
666MCSymbols are created by MCContext and uniqued there.  This means that MCSymbols
667can be compared for pointer equivalence to find out if they are the same symbol.
668Note that pointer inequality does not guarantee the labels will end up at
669different addresses though.  It's perfectly legal to output something like this
670to the .s file:
671
672::
673
674  foo:
675  bar:
676    .byte 4
677
678In this case, both the foo and bar symbols will have the same address.
679
680The ``MCSection`` class
681-----------------------
682
683The ``MCSection`` class represents an object-file specific section. It is
684subclassed by object file specific implementations (e.g. ``MCSectionMachO``,
685``MCSectionCOFF``, ``MCSectionELF``) and these are created and uniqued by
686MCContext.  The MCStreamer has a notion of the current section, which can be
687changed with the SwitchToSection method (which corresponds to a ".section"
688directive in a .s file).
689
690.. _MCInst:
691
692The ``MCInst`` class
693--------------------
694
695The ``MCInst`` class is a target-independent representation of an instruction.
696It is a simple class (much more so than `MachineInstr`_) that holds a
697target-specific opcode and a vector of MCOperands.  MCOperand, in turn, is a
698simple discriminated union of three cases: 1) a simple immediate, 2) a target
699register ID, 3) a symbolic expression (e.g. "``Lfoo-Lbar+42``") as an MCExpr.
700
701MCInst is the common currency used to represent machine instructions at the MC
702layer.  It is the type used by the instruction encoder, the instruction printer,
703and the type generated by the assembly parser and disassembler.
704
705.. _ObjectFormats:
706
707Object File Format
708------------------
709
710The MC layer's object writers support a variety of object formats. Because of
711target-specific aspects of object formats each target only supports a subset of
712the formats supported by the MC layer. Most targets support emitting ELF
713objects. Other vendor-specific objects are generally supported only on targets
714that are supported by that vendor (i.e. MachO is only supported on targets
715supported by Darwin, and XCOFF is only supported on targets that support AIX).
716Additionally some targets have their own object formats (i.e. DirectX, SPIR-V
717and WebAssembly).
718
719The table below captures a snapshot of object file support in LLVM:
720
721  .. table:: Object File Formats
722
723     ================== ========================================================
724     Format              Supported Targets
725     ================== ========================================================
726     ``COFF``            AArch64, ARM, X86
727     ``DXContainer``     DirectX
728     ``ELF``             AArch64, AMDGPU, ARM, AVR, BPF, CSKY, Hexagon, Lanai, LoongArch, M86k, MSP430, MIPS, PowerPC, RISCV, SPARC, SystemZ, VE, X86
729     ``GOFF``            SystemZ
730     ``MachO``           AArch64, ARM, X86
731     ``SPIR-V``          SPIRV
732     ``WASM``            WebAssembly
733     ``XCOFF``           PowerPC
734     ================== ========================================================
735
736.. _Target-independent algorithms:
737.. _code generation algorithm:
738
739Target-independent code generation algorithms
740=============================================
741
742This section documents the phases described in the `high-level design of the
743code generator`_.  It explains how they work and some of the rationale behind
744their design.
745
746.. _Instruction Selection:
747.. _instruction selection section:
748
749Instruction Selection
750---------------------
751
752Instruction Selection is the process of translating LLVM code presented to the
753code generator into target-specific machine instructions.  There are several
754well-known ways to do this in the literature.  LLVM uses a SelectionDAG based
755instruction selector.
756
757Portions of the DAG instruction selector are generated from the target
758description (``*.td``) files.  Our goal is for the entire instruction selector
759to be generated from these ``.td`` files, though currently there are still
760things that require custom C++ code.
761
762`GlobalISel <https://llvm.org/docs/GlobalISel/index.html>`_ is another
763instruction selection framework.
764
765.. _SelectionDAG:
766
767Introduction to SelectionDAGs
768^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
769
770The SelectionDAG provides an abstraction for code representation in a way that
771is amenable to instruction selection using automatic techniques
772(e.g. dynamic-programming based optimal pattern matching selectors). It is also
773well-suited to other phases of code generation; in particular, instruction
774scheduling (SelectionDAG's are very close to scheduling DAGs post-selection).
775Additionally, the SelectionDAG provides a host representation where a large
776variety of very-low-level (but target-independent) `optimizations`_ may be
777performed; ones which require extensive information about the instructions
778efficiently supported by the target.
779
780The SelectionDAG is a Directed-Acyclic-Graph whose nodes are instances of the
781``SDNode`` class.  The primary payload of the ``SDNode`` is its operation code
782(Opcode) that indicates what operation the node performs and the operands to the
783operation.  The various operation node types are described at the top of the
784``include/llvm/CodeGen/ISDOpcodes.h`` file.
785
786Although most operations define a single value, each node in the graph may
787define multiple values.  For example, a combined div/rem operation will define
788both the dividend and the remainder. Many other situations require multiple
789values as well.  Each node also has some number of operands, which are edges to
790the node defining the used value.  Because nodes may define multiple values,
791edges are represented by instances of the ``SDValue`` class, which is a
792``<SDNode, unsigned>`` pair, indicating the node and result value being used,
793respectively.  Each value produced by an ``SDNode`` has an associated ``MVT``
794(Machine Value Type) indicating what the type of the value is.
795
796SelectionDAGs contain two different kinds of values: those that represent data
797flow and those that represent control flow dependencies.  Data values are simple
798edges with an integer or floating point value type.  Control edges are
799represented as "chain" edges which are of type ``MVT::Other``.  These edges
800provide an ordering between nodes that have side effects (such as loads, stores,
801calls, returns, etc).  All nodes that have side effects should take a token
802chain as input and produce a new one as output.  By convention, token chain
803inputs are always operand #0, and chain results are always the last value
804produced by an operation. However, after instruction selection, the
805machine nodes have their chain after the instruction's operands, and
806may be followed by glue nodes.
807
808A SelectionDAG has designated "Entry" and "Root" nodes.  The Entry node is
809always a marker node with an Opcode of ``ISD::EntryToken``.  The Root node is
810the final side-effecting node in the token chain. For example, in a single basic
811block function it would be the return node.
812
813One important concept for SelectionDAGs is the notion of a "legal" vs.
814"illegal" DAG.  A legal DAG for a target is one that only uses supported
815operations and supported types.  On a 32-bit PowerPC, for example, a DAG with a
816value of type i1, i8, i16, or i64 would be illegal, as would a DAG that uses a
817SREM or UREM operation.  The `legalize types`_ and `legalize operations`_ phases
818are responsible for turning an illegal DAG into a legal DAG.
819
820.. _SelectionDAG-Process:
821
822SelectionDAG Instruction Selection Process
823^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
824
825SelectionDAG-based instruction selection consists of the following steps:
826
827#. `Build initial DAG`_ --- This stage performs a simple translation from the
828   input LLVM code to an illegal SelectionDAG.
829
830#. `Optimize SelectionDAG`_ --- This stage performs simple optimizations on the
831   SelectionDAG to simplify it, and recognize meta instructions (like rotates
832   and ``div``/``rem`` pairs) for targets that support these meta operations.
833   This makes the resultant code more efficient and the `select instructions
834   from DAG`_ phase (below) simpler.
835
836#. `Legalize SelectionDAG Types`_ --- This stage transforms SelectionDAG nodes
837   to eliminate any types that are unsupported on the target.
838
839#. `Optimize SelectionDAG`_ --- The SelectionDAG optimizer is run to clean up
840   redundancies exposed by type legalization.
841
842#. `Legalize SelectionDAG Ops`_ --- This stage transforms SelectionDAG nodes to
843   eliminate any operations that are unsupported on the target.
844
845#. `Optimize SelectionDAG`_ --- The SelectionDAG optimizer is run to eliminate
846   inefficiencies introduced by operation legalization.
847
848#. `Select instructions from DAG`_ --- Finally, the target instruction selector
849   matches the DAG operations to target instructions.  This process translates
850   the target-independent input DAG into another DAG of target instructions.
851
852#. `SelectionDAG Scheduling and Formation`_ --- The last phase assigns a linear
853   order to the instructions in the target-instruction DAG and emits them into
854   the MachineFunction being compiled.  This step uses traditional prepass
855   scheduling techniques.
856
857After all of these steps are complete, the SelectionDAG is destroyed and the
858rest of the code generation passes are run.
859
860One of the most common ways to debug these steps is using ``-debug-only=isel``,
861which prints out the DAG, along with other information like debug info,
862after each of these steps. Alternatively, ``-debug-only=isel-dump`` shows only
863the DAG dumps, but the results can be filtered by function names using
864``-filter-print-funcs=<function names>``.
865
866One great way to visualize what is going on here is to take advantage of a few
867LLC command line options.  The following options pop up a window displaying the
868SelectionDAG at specific times (if you only get errors printed to the console
869while using this, you probably `need to configure your
870system <ProgrammersManual.html#viewing-graphs-while-debugging-code>`_ to add support for it).
871
872* ``-view-dag-combine1-dags`` displays the DAG after being built, before the
873  first optimization pass.
874
875* ``-view-legalize-dags`` displays the DAG before Legalization.
876
877* ``-view-dag-combine2-dags`` displays the DAG before the second optimization
878  pass.
879
880* ``-view-isel-dags`` displays the DAG before the Select phase.
881
882* ``-view-sched-dags`` displays the DAG before Scheduling.
883
884The ``-view-sunit-dags`` displays the Scheduler's dependency graph.  This graph
885is based on the final SelectionDAG, with nodes that must be scheduled together
886bundled into a single scheduling-unit node, and with immediate operands and
887other nodes that aren't relevant for scheduling omitted.
888
889The option ``-filter-view-dags`` allows to select the name of the basic block
890that you are interested to visualize and filters all the previous
891``view-*-dags`` options.
892
893.. _Build initial DAG:
894
895Initial SelectionDAG Construction
896^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
897
898The initial SelectionDAG is na\ :raw-html:`&iuml;`\ vely peephole expanded from
899the LLVM input by the ``SelectionDAGBuilder`` class.  The intent of this pass
900is to expose as much low-level, target-specific details to the SelectionDAG as
901possible.  This pass is mostly hard-coded (e.g. an LLVM ``add`` turns into an
902``SDNode add`` while a ``getelementptr`` is expanded into the obvious
903arithmetic). This pass requires target-specific hooks to lower calls, returns,
904varargs, etc.  For these features, the :raw-html:`<tt>` `TargetLowering`_
905:raw-html:`</tt>` interface is used.
906
907.. _legalize types:
908.. _Legalize SelectionDAG Types:
909.. _Legalize SelectionDAG Ops:
910
911SelectionDAG LegalizeTypes Phase
912^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
913
914The Legalize phase is in charge of converting a DAG to only use the types that
915are natively supported by the target.
916
917There are two main ways of converting values of unsupported scalar types to
918values of supported types: converting small types to larger types ("promoting"),
919and breaking up large integer types into smaller ones ("expanding").  For
920example, a target might require that all f32 values are promoted to f64 and that
921all i1/i8/i16 values are promoted to i32.  The same target might require that
922all i64 values be expanded into pairs of i32 values.  These changes can insert
923sign and zero extensions as needed to make sure that the final code has the same
924behavior as the input.
925
926There are two main ways of converting values of unsupported vector types to
927value of supported types: splitting vector types, multiple times if necessary,
928until a legal type is found, and extending vector types by adding elements to
929the end to round them out to legal types ("widening").  If a vector gets split
930all the way down to single-element parts with no supported vector type being
931found, the elements are converted to scalars ("scalarizing").
932
933A target implementation tells the legalizer which types are supported (and which
934register class to use for them) by calling the ``addRegisterClass`` method in
935its ``TargetLowering`` constructor.
936
937.. _legalize operations:
938.. _Legalizer:
939
940SelectionDAG Legalize Phase
941^^^^^^^^^^^^^^^^^^^^^^^^^^^
942
943The Legalize phase is in charge of converting a DAG to only use the operations
944that are natively supported by the target.
945
946Targets often have weird constraints, such as not supporting every operation on
947every supported datatype (e.g. X86 does not support byte conditional moves and
948PowerPC does not support sign-extending loads from a 16-bit memory location).
949Legalize takes care of this by open-coding another sequence of operations to
950emulate the operation ("expansion"), by promoting one type to a larger type that
951supports the operation ("promotion"), or by using a target-specific hook to
952implement the legalization ("custom").
953
954A target implementation tells the legalizer which operations are not supported
955(and which of the above three actions to take) by calling the
956``setOperationAction`` method in its ``TargetLowering`` constructor.
957
958If a target has legal vector types, it is expected to produce efficient machine
959code for common forms of the shufflevector IR instruction using those types.
960This may require custom legalization for SelectionDAG vector operations that
961are created from the shufflevector IR. The shufflevector forms that should be
962handled include:
963
964* Vector select --- Each element of the vector is chosen from either of the
965  corresponding elements of the 2 input vectors. This operation may also be
966  known as a "blend" or "bitwise select" in target assembly. This type of shuffle
967  maps directly to the ``shuffle_vector`` SelectionDAG node.
968
969* Insert subvector --- A vector is placed into a longer vector type starting
970  at index 0. This type of shuffle maps directly to the ``insert_subvector``
971  SelectionDAG node with the ``index`` operand set to 0.
972
973* Extract subvector --- A vector is pulled from a longer vector type starting
974  at index 0. This type of shuffle maps directly to the ``extract_subvector``
975  SelectionDAG node with the ``index`` operand set to 0.
976
977* Splat --- All elements of the vector have identical scalar elements. This
978  operation may also be known as a "broadcast" or "duplicate" in target assembly.
979  The shufflevector IR instruction may change the vector length, so this operation
980  may map to multiple SelectionDAG nodes including ``shuffle_vector``,
981  ``concat_vectors``, ``insert_subvector``, and ``extract_subvector``.
982
983Prior to the existence of the Legalize passes, we required that every target
984`selector`_ supported and handled every operator and type even if they are not
985natively supported.  The introduction of the Legalize phases allows all of the
986canonicalization patterns to be shared across targets, and makes it very easy to
987optimize the canonicalized code because it is still in the form of a DAG.
988
989.. _optimizations:
990.. _Optimize SelectionDAG:
991.. _selector:
992
993SelectionDAG Optimization Phase: the DAG Combiner
994^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
995
996The SelectionDAG optimization phase is run multiple times for code generation,
997immediately after the DAG is built and once after each legalization.  The first
998run of the pass allows the initial code to be cleaned up (e.g. performing
999optimizations that depend on knowing that the operators have restricted type
1000inputs).  Subsequent runs of the pass clean up the messy code generated by the
1001Legalize passes, which allows Legalize to be very simple (it can focus on making
1002code legal instead of focusing on generating *good* and legal code).
1003
1004One important class of optimizations performed is optimizing inserted sign and
1005zero extension instructions.  We currently use ad-hoc techniques, but could move
1006to more rigorous techniques in the future.  Here are some good papers on the
1007subject:
1008
1009"`Widening integer arithmetic <http://www.eecs.harvard.edu/~nr/pubs/widen-abstract.html>`_" :raw-html:`<br>`
1010Kevin Redwine and Norman Ramsey :raw-html:`<br>`
1011International Conference on Compiler Construction (CC) 2004
1012
1013"`Effective sign extension elimination <http://portal.acm.org/citation.cfm?doid=512529.512552>`_"  :raw-html:`<br>`
1014Motohiro Kawahito, Hideaki Komatsu, and Toshio Nakatani :raw-html:`<br>`
1015Proceedings of the ACM SIGPLAN 2002 Conference on Programming Language Design
1016and Implementation.
1017
1018.. _Select instructions from DAG:
1019
1020SelectionDAG Select Phase
1021^^^^^^^^^^^^^^^^^^^^^^^^^
1022
1023The Select phase is the bulk of the target-specific code for instruction
1024selection.  This phase takes a legal SelectionDAG as input, pattern matches the
1025instructions supported by the target to this DAG, and produces a new DAG of
1026target code.  For example, consider the following LLVM fragment:
1027
1028.. code-block:: llvm
1029
1030  %t1 = fadd float %W, %X
1031  %t2 = fmul float %t1, %Y
1032  %t3 = fadd float %t2, %Z
1033
1034This LLVM code corresponds to a SelectionDAG that looks basically like this:
1035
1036.. code-block:: text
1037
1038  (fadd:f32 (fmul:f32 (fadd:f32 W, X), Y), Z)
1039
1040If a target supports floating point multiply-and-add (FMA) operations, one of
1041the adds can be merged with the multiply.  On the PowerPC, for example, the
1042output of the instruction selector might look like this DAG:
1043
1044::
1045
1046  (FMADDS (FADDS W, X), Y, Z)
1047
1048The ``FMADDS`` instruction is a ternary instruction that multiplies its first
1049two operands and adds the third (as single-precision floating-point numbers).
1050The ``FADDS`` instruction is a simple binary single-precision add instruction.
1051To perform this pattern match, the PowerPC backend includes the following
1052instruction definitions:
1053
1054.. code-block:: text
1055  :emphasize-lines: 4-5,9
1056
1057  def FMADDS : AForm_1<59, 29,
1058                      (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
1059                      "fmadds $FRT, $FRA, $FRC, $FRB",
1060                      [(set F4RC:$FRT, (fadd (fmul F4RC:$FRA, F4RC:$FRC),
1061                                             F4RC:$FRB))]>;
1062  def FADDS : AForm_2<59, 21,
1063                      (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
1064                      "fadds $FRT, $FRA, $FRB",
1065                      [(set F4RC:$FRT, (fadd F4RC:$FRA, F4RC:$FRB))]>;
1066
1067The highlighted portion of the instruction definitions indicates the pattern
1068used to match the instructions. The DAG operators (like ``fmul``/``fadd``)
1069are defined in the ``include/llvm/Target/TargetSelectionDAG.td`` file.
1070"``F4RC``" is the register class of the input and result values.
1071
1072The TableGen DAG instruction selector generator reads the instruction patterns
1073in the ``.td`` file and automatically builds parts of the pattern matching code
1074for your target.  It has the following strengths:
1075
1076* At compiler-compile time, it analyzes your instruction patterns and tells you
1077  if your patterns make sense or not.
1078
1079* It can handle arbitrary constraints on operands for the pattern match.  In
1080  particular, it is straight-forward to say things like "match any immediate
1081  that is a 13-bit sign-extended value".  For examples, see the ``immSExt16``
1082  and related ``tblgen`` classes in the PowerPC backend.
1083
1084* It knows several important identities for the patterns defined.  For example,
1085  it knows that addition is commutative, so it allows the ``FMADDS`` pattern
1086  above to match "``(fadd X, (fmul Y, Z))``" as well as "``(fadd (fmul X, Y),
1087  Z)``", without the target author having to specially handle this case.
1088
1089* It has a full-featured type-inferencing system.  In particular, you should
1090  rarely have to explicitly tell the system what type parts of your patterns
1091  are.  In the ``FMADDS`` case above, we didn't have to tell ``tblgen`` that all
1092  of the nodes in the pattern are of type 'f32'.  It was able to infer and
1093  propagate this knowledge from the fact that ``F4RC`` has type 'f32'.
1094
1095* Targets can define their own (and rely on built-in) "pattern fragments".
1096  Pattern fragments are chunks of reusable patterns that get inlined into your
1097  patterns during compiler-compile time.  For example, the integer "``(not
1098  x)``" operation is actually defined as a pattern fragment that expands as
1099  "``(xor x, -1)``", since the SelectionDAG does not have a native '``not``'
1100  operation.  Targets can define their own short-hand fragments as they see fit.
1101  See the definition of '``not``' and '``ineg``' for examples.
1102
1103* In addition to instructions, targets can specify arbitrary patterns that map
1104  to one or more instructions using the 'Pat' class.  For example, the PowerPC
1105  has no way to load an arbitrary integer immediate into a register in one
1106  instruction. To tell tblgen how to do this, it defines:
1107
1108  ::
1109
1110    // Arbitrary immediate support.  Implement in terms of LIS/ORI.
1111    def : Pat<(i32 imm:$imm),
1112              (ORI (LIS (HI16 imm:$imm)), (LO16 imm:$imm))>;
1113
1114  If none of the single-instruction patterns for loading an immediate into a
1115  register match, this will be used.  This rule says "match an arbitrary i32
1116  immediate, turning it into an ``ORI`` ('or a 16-bit immediate') and an ``LIS``
1117  ('load 16-bit immediate, where the immediate is shifted to the left 16 bits')
1118  instruction".  To make this work, the ``LO16``/``HI16`` node transformations
1119  are used to manipulate the input immediate (in this case, take the high or low
1120  16-bits of the immediate).
1121
1122* When using the 'Pat' class to map a pattern to an instruction that has one
1123  or more complex operands (like e.g. `X86 addressing mode`_), the pattern may
1124  either specify the operand as a whole using a ``ComplexPattern``, or else it
1125  may specify the components of the complex operand separately.  The latter is
1126  done e.g. for pre-increment instructions by the PowerPC back end:
1127
1128  ::
1129
1130    def STWU  : DForm_1<37, (outs ptr_rc:$ea_res), (ins GPRC:$rS, memri:$dst),
1131                    "stwu $rS, $dst", LdStStoreUpd, []>,
1132                    RegConstraint<"$dst.reg = $ea_res">, NoEncode<"$ea_res">;
1133
1134    def : Pat<(pre_store GPRC:$rS, ptr_rc:$ptrreg, iaddroff:$ptroff),
1135              (STWU GPRC:$rS, iaddroff:$ptroff, ptr_rc:$ptrreg)>;
1136
1137  Here, the pair of ``ptroff`` and ``ptrreg`` operands is matched onto the
1138  complex operand ``dst`` of class ``memri`` in the ``STWU`` instruction.
1139
1140* While the system does automate a lot, it still allows you to write custom C++
1141  code to match special cases if there is something that is hard to
1142  express.
1143
1144While it has many strengths, the system currently has some limitations,
1145primarily because it is a work in progress and is not yet finished:
1146
1147* Overall, there is no way to define or match SelectionDAG nodes that define
1148  multiple values (e.g. ``SMUL_LOHI``, ``LOAD``, ``CALL``, etc).  This is the
1149  biggest reason that you currently still *have to* write custom C++ code
1150  for your instruction selector.
1151
1152* There is no great way to support matching complex addressing modes yet.  In
1153  the future, we will extend pattern fragments to allow them to define multiple
1154  values (e.g. the four operands of the `X86 addressing mode`_, which are
1155  currently matched with custom C++ code).  In addition, we'll extend fragments
1156  so that a fragment can match multiple different patterns.
1157
1158* We don't automatically infer flags like ``isStore``/``isLoad`` yet.
1159
1160* We don't automatically generate the set of supported registers and operations
1161  for the `Legalizer`_ yet.
1162
1163* We don't have a way of tying in custom legalized nodes yet.
1164
1165Despite these limitations, the instruction selector generator is still quite
1166useful for most of the binary and logical operations in typical instruction
1167sets.  If you run into any problems or can't figure out how to do something,
1168please let Chris know!
1169
1170.. _Scheduling and Formation:
1171.. _SelectionDAG Scheduling and Formation:
1172
1173SelectionDAG Scheduling and Formation Phase
1174^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1175
1176The scheduling phase takes the DAG of target instructions from the selection
1177phase and assigns an order.  The scheduler can pick an order depending on
1178various constraints of the machines (i.e. order for minimal register pressure or
1179try to cover instruction latencies).  Once an order is established, the DAG is
1180converted to a list of :raw-html:`<tt>` `MachineInstr`_\s :raw-html:`</tt>` and
1181the SelectionDAG is destroyed.
1182
1183Note that this phase is logically separate from the instruction selection phase,
1184but is tied to it closely in the code because it operates on SelectionDAGs.
1185
1186Future directions for the SelectionDAG
1187^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1188
1189#. Optional function-at-a-time selection.
1190
1191#. Auto-generate entire selector from ``.td`` file.
1192
1193.. _SSA-based Machine Code Optimizations:
1194
1195SSA-based Machine Code Optimizations
1196------------------------------------
1197
1198To Be Written
1199
1200Live Intervals
1201--------------
1202
1203Live Intervals are the ranges (intervals) where a variable is *live*.  They are
1204used by some `register allocator`_ passes to determine if two or more virtual
1205registers which require the same physical register are live at the same point in
1206the program (i.e., they conflict).  When this situation occurs, one virtual
1207register must be *spilled*.
1208
1209Live Variable Analysis
1210^^^^^^^^^^^^^^^^^^^^^^
1211
1212The first step in determining the live intervals of variables is to calculate
1213the set of registers that are immediately dead after the instruction (i.e., the
1214instruction calculates the value, but it is never used) and the set of registers
1215that are used by the instruction, but are never used after the instruction
1216(i.e., they are killed). Live variable information is computed for
1217each *virtual* register and *register allocatable* physical register
1218in the function.  This is done in a very efficient manner because it uses SSA to
1219sparsely compute lifetime information for virtual registers (which are in SSA
1220form) and only has to track physical registers within a block.  Before register
1221allocation, LLVM can assume that physical registers are only live within a
1222single basic block.  This allows it to do a single, local analysis to resolve
1223physical register lifetimes within each basic block. If a physical register is
1224not register allocatable (e.g., a stack pointer or condition codes), it is not
1225tracked.
1226
1227Physical registers may be live in to or out of a function. Live in values are
1228typically arguments in registers. Live out values are typically return values in
1229registers. Live in values are marked as such, and are given a dummy "defining"
1230instruction during live intervals analysis. If the last basic block of a
1231function is a ``return``, then it's marked as using all live out values in the
1232function.
1233
1234``PHI`` nodes need to be handled specially, because the calculation of the live
1235variable information from a depth first traversal of the CFG of the function
1236won't guarantee that a virtual register used by the ``PHI`` node is defined
1237before it's used. When a ``PHI`` node is encountered, only the definition is
1238handled, because the uses will be handled in other basic blocks.
1239
1240For each ``PHI`` node of the current basic block, we simulate an assignment at
1241the end of the current basic block and traverse the successor basic blocks. If a
1242successor basic block has a ``PHI`` node and one of the ``PHI`` node's operands
1243is coming from the current basic block, then the variable is marked as *alive*
1244within the current basic block and all of its predecessor basic blocks, until
1245the basic block with the defining instruction is encountered.
1246
1247Live Intervals Analysis
1248^^^^^^^^^^^^^^^^^^^^^^^
1249
1250We now have the information available to perform the live intervals analysis and
1251build the live intervals themselves.  We start off by numbering the basic blocks
1252and machine instructions.  We then handle the "live-in" values.  These are in
1253physical registers, so the physical register is assumed to be killed by the end
1254of the basic block.  Live intervals for virtual registers are computed for some
1255ordering of the machine instructions ``[1, N]``.  A live interval is an interval
1256``[i, j)``, where ``1 >= i >= j > N``, for which a variable is live.
1257
1258.. note::
1259  More to come...
1260
1261.. _Register Allocation:
1262.. _register allocator:
1263
1264Register Allocation
1265-------------------
1266
1267The *Register Allocation problem* consists in mapping a program
1268:raw-html:`<b><tt>` P\ :sub:`v`\ :raw-html:`</tt></b>`, that can use an unbounded
1269number of virtual registers, to a program :raw-html:`<b><tt>` P\ :sub:`p`\
1270:raw-html:`</tt></b>` that contains a finite (possibly small) number of physical
1271registers. Each target architecture has a different number of physical
1272registers. If the number of physical registers is not enough to accommodate all
1273the virtual registers, some of them will have to be mapped into memory. These
1274virtuals are called *spilled virtuals*.
1275
1276How registers are represented in LLVM
1277^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1278
1279In LLVM, physical registers are denoted by integer numbers that normally range
1280from 1 to 1023. To see how this numbering is defined for a particular
1281architecture, you can read the ``GenRegisterNames.inc`` file for that
1282architecture. For instance, by inspecting
1283``lib/Target/X86/X86GenRegisterInfo.inc`` we see that the 32-bit register
1284``EAX`` is denoted by 43, and the MMX register ``MM0`` is mapped to 65.
1285
1286Some architectures contain registers that share the same physical location. A
1287notable example is the X86 platform. For instance, in the X86 architecture, the
1288registers ``EAX``, ``AX`` and ``AL`` share the first eight bits. These physical
1289registers are marked as *aliased* in LLVM. Given a particular architecture, you
1290can check which registers are aliased by inspecting its ``RegisterInfo.td``
1291file. Moreover, the class ``MCRegAliasIterator`` enumerates all the physical
1292registers aliased to a register.
1293
1294Physical registers, in LLVM, are grouped in *Register Classes*.  Elements in the
1295same register class are functionally equivalent, and can be interchangeably
1296used. Each virtual register can only be mapped to physical registers of a
1297particular class. For instance, in the X86 architecture, some virtuals can only
1298be allocated to 8 bit registers.  A register class is described by
1299``TargetRegisterClass`` objects.  To discover if a virtual register is
1300compatible with a given physical, this code can be used:
1301
1302.. code-block:: c++
1303
1304  bool RegMapping_Fer::compatible_class(MachineFunction &mf,
1305                                        unsigned v_reg,
1306                                        unsigned p_reg) {
1307    assert(TargetRegisterInfo::isPhysicalRegister(p_reg) &&
1308           "Target register must be physical");
1309    const TargetRegisterClass *trc = mf.getRegInfo().getRegClass(v_reg);
1310    return trc->contains(p_reg);
1311  }
1312
1313Sometimes, mostly for debugging purposes, it is useful to change the number of
1314physical registers available in the target architecture. This must be done
1315statically, inside the ``TargetRegisterInfo.td`` file. Just ``grep`` for
1316``RegisterClass``, the last parameter of which is a list of registers. Just
1317commenting some out is one simple way to avoid them being used. A more polite
1318way is to explicitly exclude some registers from the *allocation order*. See the
1319definition of the ``GR8`` register class in
1320``lib/Target/X86/X86RegisterInfo.td`` for an example of this.
1321
1322Virtual registers are also denoted by integer numbers. Contrary to physical
1323registers, different virtual registers never share the same number. Whereas
1324physical registers are statically defined in a ``TargetRegisterInfo.td`` file
1325and cannot be created by the application developer, that is not the case with
1326virtual registers. In order to create new virtual registers, use the method
1327``MachineRegisterInfo::createVirtualRegister()``. This method will return a new
1328virtual register. Use an ``IndexedMap<Foo, VirtReg2IndexFunctor>`` to hold
1329information per virtual register. If you need to enumerate all virtual
1330registers, use the function ``TargetRegisterInfo::index2VirtReg()`` to find the
1331virtual register numbers:
1332
1333.. code-block:: c++
1334
1335    for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
1336      unsigned VirtReg = TargetRegisterInfo::index2VirtReg(i);
1337      stuff(VirtReg);
1338    }
1339
1340Before register allocation, the operands of an instruction are mostly virtual
1341registers, although physical registers may also be used. In order to check if a
1342given machine operand is a register, use the boolean function
1343``MachineOperand::isRegister()``. To obtain the integer code of a register, use
1344``MachineOperand::getReg()``. An instruction may define or use a register. For
1345instance, ``ADD reg:1026 := reg:1025 reg:1024`` defines the registers 1024, and
1346uses registers 1025 and 1026. Given a register operand, the method
1347``MachineOperand::isUse()`` informs if that register is being used by the
1348instruction. The method ``MachineOperand::isDef()`` informs if that registers is
1349being defined.
1350
1351We will call physical registers present in the LLVM bitcode before register
1352allocation *pre-colored registers*. Pre-colored registers are used in many
1353different situations, for instance, to pass parameters of functions calls, and
1354to store results of particular instructions. There are two types of pre-colored
1355registers: the ones *implicitly* defined, and those *explicitly*
1356defined. Explicitly defined registers are normal operands, and can be accessed
1357with ``MachineInstr::getOperand(int)::getReg()``.  In order to check which
1358registers are implicitly defined by an instruction, use the
1359``TargetInstrInfo::get(opcode)::ImplicitDefs``, where ``opcode`` is the opcode
1360of the target instruction. One important difference between explicit and
1361implicit physical registers is that the latter are defined statically for each
1362instruction, whereas the former may vary depending on the program being
1363compiled. For example, an instruction that represents a function call will
1364always implicitly define or use the same set of physical registers. To read the
1365registers implicitly used by an instruction, use
1366``TargetInstrInfo::get(opcode)::ImplicitUses``. Pre-colored registers impose
1367constraints on any register allocation algorithm. The register allocator must
1368make sure that none of them are overwritten by the values of virtual registers
1369while still alive.
1370
1371Mapping virtual registers to physical registers
1372^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1373
1374There are two ways to map virtual registers to physical registers (or to memory
1375slots). The first way, that we will call *direct mapping*, is based on the use
1376of methods of the classes ``TargetRegisterInfo``, and ``MachineOperand``. The
1377second way, that we will call *indirect mapping*, relies on the ``VirtRegMap``
1378class in order to insert loads and stores sending and getting values to and from
1379memory.
1380
1381The direct mapping provides more flexibility to the developer of the register
1382allocator; however, it is more error prone, and demands more implementation
1383work.  Basically, the programmer will have to specify where load and store
1384instructions should be inserted in the target function being compiled in order
1385to get and store values in memory. To assign a physical register to a virtual
1386register present in a given operand, use ``MachineOperand::setReg(p_reg)``. To
1387insert a store instruction, use ``TargetInstrInfo::storeRegToStackSlot(...)``,
1388and to insert a load instruction, use ``TargetInstrInfo::loadRegFromStackSlot``.
1389
1390The indirect mapping shields the application developer from the complexities of
1391inserting load and store instructions. In order to map a virtual register to a
1392physical one, use ``VirtRegMap::assignVirt2Phys(vreg, preg)``.  In order to map
1393a certain virtual register to memory, use
1394``VirtRegMap::assignVirt2StackSlot(vreg)``. This method will return the stack
1395slot where ``vreg``'s value will be located.  If it is necessary to map another
1396virtual register to the same stack slot, use
1397``VirtRegMap::assignVirt2StackSlot(vreg, stack_location)``. One important point
1398to consider when using the indirect mapping, is that even if a virtual register
1399is mapped to memory, it still needs to be mapped to a physical register. This
1400physical register is the location where the virtual register is supposed to be
1401found before being stored or after being reloaded.
1402
1403If the indirect strategy is used, after all the virtual registers have been
1404mapped to physical registers or stack slots, it is necessary to use a spiller
1405object to place load and store instructions in the code. Every virtual that has
1406been mapped to a stack slot will be stored to memory after being defined and will
1407be loaded before being used. The implementation of the spiller tries to recycle
1408load/store instructions, avoiding unnecessary instructions. For an example of
1409how to invoke the spiller, see ``RegAllocLinearScan::runOnMachineFunction`` in
1410``lib/CodeGen/RegAllocLinearScan.cpp``.
1411
1412Handling two address instructions
1413^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1414
1415With very rare exceptions (e.g., function calls), the LLVM machine code
1416instructions are three address instructions. That is, each instruction is
1417expected to define at most one register, and to use at most two registers.
1418However, some architectures use two address instructions. In this case, the
1419defined register is also one of the used registers. For instance, an instruction
1420such as ``ADD %EAX, %EBX``, in X86 is actually equivalent to ``%EAX = %EAX +
1421%EBX``.
1422
1423In order to produce correct code, LLVM must convert three address instructions
1424that represent two address instructions into true two address instructions. LLVM
1425provides the pass ``TwoAddressInstructionPass`` for this specific purpose. It
1426must be run before register allocation takes place. After its execution, the
1427resulting code may no longer be in SSA form. This happens, for instance, in
1428situations where an instruction such as ``%a = ADD %b %c`` is converted to two
1429instructions such as:
1430
1431::
1432
1433  %a = MOVE %b
1434  %a = ADD %a %c
1435
1436Notice that, internally, the second instruction is represented as ``ADD
1437%a[def/use] %c``. I.e., the register operand ``%a`` is both used and defined by
1438the instruction.
1439
1440The SSA deconstruction phase
1441^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1442
1443An important transformation that happens during register allocation is called
1444the *SSA Deconstruction Phase*. The SSA form simplifies many analyses that are
1445performed on the control flow graph of programs. However, traditional
1446instruction sets do not implement PHI instructions. Thus, in order to generate
1447executable code, compilers must replace PHI instructions with other instructions
1448that preserve their semantics.
1449
1450There are many ways in which PHI instructions can safely be removed from the
1451target code. The most traditional PHI deconstruction algorithm replaces PHI
1452instructions with copy instructions. That is the strategy adopted by LLVM. The
1453SSA deconstruction algorithm is implemented in
1454``lib/CodeGen/PHIElimination.cpp``. In order to invoke this pass, the identifier
1455``PHIEliminationID`` must be marked as required in the code of the register
1456allocator.
1457
1458Instruction folding
1459^^^^^^^^^^^^^^^^^^^
1460
1461*Instruction folding* is an optimization performed during register allocation
1462that removes unnecessary copy instructions. For instance, a sequence of
1463instructions such as:
1464
1465::
1466
1467  %EBX = LOAD %mem_address
1468  %EAX = COPY %EBX
1469
1470can be safely substituted by the single instruction:
1471
1472::
1473
1474  %EAX = LOAD %mem_address
1475
1476Instructions can be folded with the
1477``TargetRegisterInfo::foldMemoryOperand(...)`` method. Care must be taken when
1478folding instructions; a folded instruction can be quite different from the
1479original instruction. See ``LiveIntervals::addIntervalsForSpills`` in
1480``lib/CodeGen/LiveIntervalAnalysis.cpp`` for an example of its use.
1481
1482Built in register allocators
1483^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1484
1485The LLVM infrastructure provides the application developer with three different
1486register allocators:
1487
1488* *Fast* --- This register allocator is the default for debug builds. It
1489  allocates registers on a basic block level, attempting to keep values in
1490  registers and reusing registers as appropriate.
1491
1492* *Basic* --- This is an incremental approach to register allocation. Live
1493  ranges are assigned to registers one at a time in an order that is driven by
1494  heuristics. Since code can be rewritten on-the-fly during allocation, this
1495  framework allows interesting allocators to be developed as extensions. It is
1496  not itself a production register allocator but is a potentially useful
1497  stand-alone mode for triaging bugs and as a performance baseline.
1498
1499* *Greedy* --- *The default allocator*. This is a highly tuned implementation of
1500  the *Basic* allocator that incorporates global live range splitting. This
1501  allocator works hard to minimize the cost of spill code.
1502
1503* *PBQP* --- A Partitioned Boolean Quadratic Programming (PBQP) based register
1504  allocator. This allocator works by constructing a PBQP problem representing
1505  the register allocation problem under consideration, solving this using a PBQP
1506  solver, and mapping the solution back to a register assignment.
1507
1508The type of register allocator used in ``llc`` can be chosen with the command
1509line option ``-regalloc=...``:
1510
1511.. code-block:: bash
1512
1513  $ llc -regalloc=linearscan file.bc -o ln.s
1514  $ llc -regalloc=fast file.bc -o fa.s
1515  $ llc -regalloc=pbqp file.bc -o pbqp.s
1516
1517.. _Prolog/Epilog Code Insertion:
1518
1519Prolog/Epilog Code Insertion
1520----------------------------
1521
1522.. note::
1523
1524  To Be Written
1525
1526Compact Unwind
1527--------------
1528
1529Throwing an exception requires *unwinding* out of a function. The information on
1530how to unwind a given function is traditionally expressed in DWARF unwind
1531(a.k.a. frame) info. But that format was originally developed for debuggers to
1532backtrace, and each Frame Description Entry (FDE) requires ~20-30 bytes per
1533function. There is also the cost of mapping from an address in a function to the
1534corresponding FDE at runtime. An alternative unwind encoding is called *compact
1535unwind* and requires just 4-bytes per function.
1536
1537The compact unwind encoding is a 32-bit value, which is encoded in an
1538architecture-specific way. It specifies which registers to restore and from
1539where, and how to unwind out of the function. When the linker creates a final
1540linked image, it will create a ``__TEXT,__unwind_info`` section. This section is
1541a small and fast way for the runtime to access unwind info for any given
1542function. If we emit compact unwind info for the function, that compact unwind
1543info will be encoded in the ``__TEXT,__unwind_info`` section. If we emit DWARF
1544unwind info, the ``__TEXT,__unwind_info`` section will contain the offset of the
1545FDE in the ``__TEXT,__eh_frame`` section in the final linked image.
1546
1547For X86, there are three modes for the compact unwind encoding:
1548
1549*Function with a Frame Pointer (``EBP`` or ``RBP``)*
1550  ``EBP/RBP``-based frame, where ``EBP/RBP`` is pushed onto the stack
1551  immediately after the return address, then ``ESP/RSP`` is moved to
1552  ``EBP/RBP``. Thus to unwind, ``ESP/RSP`` is restored with the current
1553  ``EBP/RBP`` value, then ``EBP/RBP`` is restored by popping the stack, and the
1554  return is done by popping the stack once more into the PC. All non-volatile
1555  registers that need to be restored must have been saved in a small range on
1556  the stack that starts ``EBP-4`` to ``EBP-1020`` (``RBP-8`` to
1557  ``RBP-1020``). The offset (divided by 4 in 32-bit mode and 8 in 64-bit mode)
1558  is encoded in bits 16-23 (mask: ``0x00FF0000``).  The registers saved are
1559  encoded in bits 0-14 (mask: ``0x00007FFF``) as five 3-bit entries from the
1560  following table:
1561
1562    ==============  =============  ===============
1563    Compact Number  i386 Register  x86-64 Register
1564    ==============  =============  ===============
1565    1               ``EBX``        ``RBX``
1566    2               ``ECX``        ``R12``
1567    3               ``EDX``        ``R13``
1568    4               ``EDI``        ``R14``
1569    5               ``ESI``        ``R15``
1570    6               ``EBP``        ``RBP``
1571    ==============  =============  ===============
1572
1573*Frameless with a Small Constant Stack Size (``EBP`` or ``RBP`` is not used as a frame pointer)*
1574  To return, a constant (encoded in the compact unwind encoding) is added to the
1575  ``ESP/RSP``.  Then the return is done by popping the stack into the PC. All
1576  non-volatile registers that need to be restored must have been saved on the
1577  stack immediately after the return address. The stack size (divided by 4 in
1578  32-bit mode and 8 in 64-bit mode) is encoded in bits 16-23 (mask:
1579  ``0x00FF0000``). There is a maximum stack size of 1024 bytes in 32-bit mode
1580  and 2048 in 64-bit mode. The number of registers saved is encoded in bits 9-12
1581  (mask: ``0x00001C00``). Bits 0-9 (mask: ``0x000003FF``) contain which
1582  registers were saved and their order. (See the
1583  ``encodeCompactUnwindRegistersWithoutFrame()`` function in
1584  ``lib/Target/X86FrameLowering.cpp`` for the encoding algorithm.)
1585
1586*Frameless with a Large Constant Stack Size (``EBP`` or ``RBP`` is not used as a frame pointer)*
1587  This case is like the "Frameless with a Small Constant Stack Size" case, but
1588  the stack size is too large to encode in the compact unwind encoding. Instead
1589  it requires that the function contains "``subl $nnnnnn, %esp``" in its
1590  prolog. The compact encoding contains the offset to the ``$nnnnnn`` value in
1591  the function in bits 9-12 (mask: ``0x00001C00``).
1592
1593.. _Late Machine Code Optimizations:
1594
1595Late Machine Code Optimizations
1596-------------------------------
1597
1598.. note::
1599
1600  To Be Written
1601
1602.. _Code Emission:
1603
1604Code Emission
1605-------------
1606
1607The code emission step of code generation is responsible for lowering from the
1608code generator abstractions (like `MachineFunction`_, `MachineInstr`_, etc) down
1609to the abstractions used by the MC layer (`MCInst`_, `MCStreamer`_, etc).  This
1610is done with a combination of several different classes: the (misnamed)
1611target-independent AsmPrinter class, target-specific subclasses of AsmPrinter
1612(such as SparcAsmPrinter), and the TargetLoweringObjectFile class.
1613
1614Since the MC layer works at the level of abstraction of object files, it doesn't
1615have a notion of functions, global variables etc.  Instead, it thinks about
1616labels, directives, and instructions.  A key class used at this time is the
1617MCStreamer class.  This is an abstract API that is implemented in different ways
1618(e.g. to output a .s file, output an ELF .o file, etc) that is effectively an
1619"assembler API".  MCStreamer has one method per directive, such as EmitLabel,
1620EmitSymbolAttribute, switchSection, etc, which directly correspond to assembly
1621level directives.
1622
1623If you are interested in implementing a code generator for a target, there are
1624three important things that you have to implement for your target:
1625
1626#. First, you need a subclass of AsmPrinter for your target.  This class
1627   implements the general lowering process converting MachineFunction's into MC
1628   label constructs.  The AsmPrinter base class provides a number of useful
1629   methods and routines, and also allows you to override the lowering process in
1630   some important ways.  You should get much of the lowering for free if you are
1631   implementing an ELF, COFF, or MachO target, because the
1632   TargetLoweringObjectFile class implements much of the common logic.
1633
1634#. Second, you need to implement an instruction printer for your target.  The
1635   instruction printer takes an `MCInst`_ and renders it to a raw_ostream as
1636   text.  Most of this is automatically generated from the .td file (when you
1637   specify something like "``add $dst, $src1, $src2``" in the instructions), but
1638   you need to implement routines to print operands.
1639
1640#. Third, you need to implement code that lowers a `MachineInstr`_ to an MCInst,
1641   usually implemented in "<target>MCInstLower.cpp".  This lowering process is
1642   often target specific, and is responsible for turning jump table entries,
1643   constant pool indices, global variable addresses, etc into MCLabels as
1644   appropriate.  This translation layer is also responsible for expanding pseudo
1645   ops used by the code generator into the actual machine instructions they
1646   correspond to. The MCInsts that are generated by this are fed into the
1647   instruction printer or the encoder.
1648
1649Finally, at your choosing, you can also implement a subclass of MCCodeEmitter
1650which lowers MCInst's into machine code bytes and relocations.  This is
1651important if you want to support direct .o file emission, or would like to
1652implement an assembler for your target.
1653
1654Emitting function stack size information
1655^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1656
1657A section containing metadata on function stack sizes will be emitted when
1658``TargetLoweringObjectFile::StackSizesSection`` is not null, and
1659``TargetOptions::EmitStackSizeSection`` is set (-stack-size-section). The
1660section will contain an array of pairs of function symbol values (pointer size)
1661and stack sizes (unsigned LEB128). The stack size values only include the space
1662allocated in the function prologue. Functions with dynamic stack allocations are
1663not included.
1664
1665VLIW Packetizer
1666---------------
1667
1668In a Very Long Instruction Word (VLIW) architecture, the compiler is responsible
1669for mapping instructions to functional-units available on the architecture. To
1670that end, the compiler creates groups of instructions called *packets* or
1671*bundles*. The VLIW packetizer in LLVM is a target-independent mechanism to
1672enable the packetization of machine instructions.
1673
1674Mapping from instructions to functional units
1675^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1676
1677Instructions in a VLIW target can typically be mapped to multiple functional
1678units. During the process of packetizing, the compiler must be able to reason
1679about whether an instruction can be added to a packet. This decision can be
1680complex since the compiler has to examine all possible mappings of instructions
1681to functional units. Therefore to alleviate compilation-time complexity, the
1682VLIW packetizer parses the instruction classes of a target and generates tables
1683at compiler build time. These tables can then be queried by the provided
1684machine-independent API to determine if an instruction can be accommodated in a
1685packet.
1686
1687How the packetization tables are generated and used
1688^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1689
1690The packetizer reads instruction classes from a target's itineraries and creates
1691a deterministic finite automaton (DFA) to represent the state of a packet. A DFA
1692consists of three major elements: inputs, states, and transitions. The set of
1693inputs for the generated DFA represents the instruction being added to a
1694packet. The states represent the possible consumption of functional units by
1695instructions in a packet. In the DFA, transitions from one state to another
1696occur on the addition of an instruction to an existing packet. If there is a
1697legal mapping of functional units to instructions, then the DFA contains a
1698corresponding transition. The absence of a transition indicates that a legal
1699mapping does not exist and that the instruction cannot be added to the packet.
1700
1701To generate tables for a VLIW target, add *Target*\ GenDFAPacketizer.inc as a
1702target to the Makefile in the target directory. The exported API provides three
1703functions: ``DFAPacketizer::clearResources()``,
1704``DFAPacketizer::reserveResources(MachineInstr *MI)``, and
1705``DFAPacketizer::canReserveResources(MachineInstr *MI)``. These functions allow
1706a target packetizer to add an instruction to an existing packet and to check
1707whether an instruction can be added to a packet. See
1708``llvm/CodeGen/DFAPacketizer.h`` for more information.
1709
1710Implementing a Native Assembler
1711===============================
1712
1713Though you're probably reading this because you want to write or maintain a
1714compiler backend, LLVM also fully supports building a native assembler.
1715We've tried hard to automate the generation of the assembler from the .td files
1716(in particular the instruction syntax and encodings), which means that a large
1717part of the manual and repetitive data entry can be factored and shared with the
1718compiler.
1719
1720Instruction Parsing
1721-------------------
1722
1723.. note::
1724
1725  To Be Written
1726
1727
1728Instruction Alias Processing
1729----------------------------
1730
1731Once the instruction is parsed, it enters the MatchInstructionImpl function.
1732The MatchInstructionImpl function performs alias processing and then does actual
1733matching.
1734
1735Alias processing is the phase that canonicalizes different lexical forms of the
1736same instructions down to one representation.  There are several different kinds
1737of alias that are possible to implement and they are listed below in the order
1738that they are processed (which is in order from simplest/weakest to most
1739complex/powerful).  Generally you want to use the first alias mechanism that
1740meets the needs of your instruction, because it will allow a more concise
1741description.
1742
1743Mnemonic Aliases
1744^^^^^^^^^^^^^^^^
1745
1746The first phase of alias processing is simple instruction mnemonic remapping for
1747classes of instructions which are allowed with two different mnemonics.  This
1748phase is a simple and unconditionally remapping from one input mnemonic to one
1749output mnemonic.  It isn't possible for this form of alias to look at the
1750operands at all, so the remapping must apply for all forms of a given mnemonic.
1751Mnemonic aliases are defined simply, for example X86 has:
1752
1753::
1754
1755  def : MnemonicAlias<"cbw",     "cbtw">;
1756  def : MnemonicAlias<"smovq",   "movsq">;
1757  def : MnemonicAlias<"fldcww",  "fldcw">;
1758  def : MnemonicAlias<"fucompi", "fucomip">;
1759  def : MnemonicAlias<"ud2a",    "ud2">;
1760
1761... and many others.  With a MnemonicAlias definition, the mnemonic is remapped
1762simply and directly.  Though MnemonicAlias's can't look at any aspect of the
1763instruction (such as the operands) they can depend on global modes (the same
1764ones supported by the matcher), through a Requires clause:
1765
1766::
1767
1768  def : MnemonicAlias<"pushf", "pushfq">, Requires<[In64BitMode]>;
1769  def : MnemonicAlias<"pushf", "pushfl">, Requires<[In32BitMode]>;
1770
1771In this example, the mnemonic gets mapped into a different one depending on
1772the current instruction set.
1773
1774Instruction Aliases
1775^^^^^^^^^^^^^^^^^^^
1776
1777The most general phase of alias processing occurs while matching is happening:
1778it provides new forms for the matcher to match along with a specific instruction
1779to generate.  An instruction alias has two parts: the string to match and the
1780instruction to generate.  For example:
1781
1782::
1783
1784  def : InstAlias<"movsx $src, $dst", (MOVSX16rr8W GR16:$dst, GR8  :$src)>;
1785  def : InstAlias<"movsx $src, $dst", (MOVSX16rm8W GR16:$dst, i8mem:$src)>;
1786  def : InstAlias<"movsx $src, $dst", (MOVSX32rr8  GR32:$dst, GR8  :$src)>;
1787  def : InstAlias<"movsx $src, $dst", (MOVSX32rr16 GR32:$dst, GR16 :$src)>;
1788  def : InstAlias<"movsx $src, $dst", (MOVSX64rr8  GR64:$dst, GR8  :$src)>;
1789  def : InstAlias<"movsx $src, $dst", (MOVSX64rr16 GR64:$dst, GR16 :$src)>;
1790  def : InstAlias<"movsx $src, $dst", (MOVSX64rr32 GR64:$dst, GR32 :$src)>;
1791
1792This shows a powerful example of the instruction aliases, matching the same
1793mnemonic in multiple different ways depending on what operands are present in
1794the assembly.  The result of instruction aliases can include operands in a
1795different order than the destination instruction, and can use an input multiple
1796times, for example:
1797
1798::
1799
1800  def : InstAlias<"clrb $reg", (XOR8rr  GR8 :$reg, GR8 :$reg)>;
1801  def : InstAlias<"clrw $reg", (XOR16rr GR16:$reg, GR16:$reg)>;
1802  def : InstAlias<"clrl $reg", (XOR32rr GR32:$reg, GR32:$reg)>;
1803  def : InstAlias<"clrq $reg", (XOR64rr GR64:$reg, GR64:$reg)>;
1804
1805This example also shows that tied operands are only listed once.  In the X86
1806backend, XOR8rr has two input GR8's and one output GR8 (where an input is tied
1807to the output).  InstAliases take a flattened operand list without duplicates
1808for tied operands.  The result of an instruction alias can also use immediates
1809and fixed physical registers which are added as simple immediate operands in the
1810result, for example:
1811
1812::
1813
1814  // Fixed Immediate operand.
1815  def : InstAlias<"aad", (AAD8i8 10)>;
1816
1817  // Fixed register operand.
1818  def : InstAlias<"fcomi", (COM_FIr ST1)>;
1819
1820  // Simple alias.
1821  def : InstAlias<"fcomi $reg", (COM_FIr RST:$reg)>;
1822
1823Instruction aliases can also have a Requires clause to make them subtarget
1824specific.
1825
1826If the back-end supports it, the instruction printer can automatically emit the
1827alias rather than what's being aliased. It typically leads to better, more
1828readable code. If it's better to print out what's being aliased, then pass a '0'
1829as the third parameter to the InstAlias definition.
1830
1831Instruction Matching
1832--------------------
1833
1834.. note::
1835
1836  To Be Written
1837
1838.. _Implementations of the abstract target description interfaces:
1839.. _implement the target description:
1840
1841Target-specific Implementation Notes
1842====================================
1843
1844This section of the document explains features or design decisions that are
1845specific to the code generator for a particular target.
1846
1847.. _tail call section:
1848
1849Tail call optimization
1850----------------------
1851
1852Tail call optimization, callee reusing the stack of the caller, is currently
1853supported on x86/x86-64, PowerPC, AArch64, and WebAssembly. It is performed on
1854x86/x86-64, PowerPC, and AArch64 if:
1855
1856* Caller and callee have the calling convention ``fastcc``, ``cc 10`` (GHC
1857  calling convention), ``cc 11`` (HiPE calling convention), ``tailcc``, or
1858  ``swifttailcc``.
1859
1860* The call is a tail call - in tail position (ret immediately follows call and
1861  ret uses value of call or is void).
1862
1863* Option ``-tailcallopt`` is enabled or the calling convention is ``tailcc``.
1864
1865* Platform-specific constraints are met.
1866
1867x86/x86-64 constraints:
1868
1869* No variable argument lists are used.
1870
1871* On x86-64 when generating GOT/PIC code only module-local calls (visibility =
1872  hidden or protected) are supported.
1873
1874PowerPC constraints:
1875
1876* No variable argument lists are used.
1877
1878* No byval parameters are used.
1879
1880* On ppc32/64 GOT/PIC only module-local calls (visibility = hidden or protected)
1881  are supported.
1882
1883WebAssembly constraints:
1884
1885* No variable argument lists are used
1886
1887* The 'tail-call' target attribute is enabled.
1888
1889* The caller and callee's return types must match. The caller cannot
1890  be void unless the callee is, too.
1891
1892AArch64 constraints:
1893
1894* No variable argument lists are used.
1895
1896Example:
1897
1898Call as ``llc -tailcallopt test.ll``.
1899
1900.. code-block:: llvm
1901
1902  declare fastcc i32 @tailcallee(i32 inreg %a1, i32 inreg %a2, i32 %a3, i32 %a4)
1903
1904  define fastcc i32 @tailcaller(i32 %in1, i32 %in2) {
1905    %l1 = add i32 %in1, %in2
1906    %tmp = tail call fastcc i32 @tailcallee(i32 inreg %in1, i32 inreg %in2, i32 %in1, i32 %l1)
1907    ret i32 %tmp
1908  }
1909
1910Implications of ``-tailcallopt``:
1911
1912To support tail call optimization in situations where the callee has more
1913arguments than the caller a 'callee pops arguments' convention is used. This
1914currently causes each ``fastcc`` call that is not tail call optimized (because
1915one or more of above constraints are not met) to be followed by a readjustment
1916of the stack. So performance might be worse in such cases.
1917
1918Sibling call optimization
1919-------------------------
1920
1921Sibling call optimization is a restricted form of tail call optimization.
1922Unlike tail call optimization described in the previous section, it can be
1923performed automatically on any tail calls when ``-tailcallopt`` option is not
1924specified.
1925
1926Sibling call optimization is currently performed on x86/x86-64 when the
1927following constraints are met:
1928
1929* Caller and callee have the same calling convention. It can be either ``c`` or
1930  ``fastcc``.
1931
1932* The call is a tail call - in tail position (ret immediately follows call and
1933  ret uses value of call or is void).
1934
1935* Caller and callee have matching return type or the callee result is not used.
1936
1937* If any of the callee arguments are being passed in stack, they must be
1938  available in caller's own incoming argument stack and the frame offsets must
1939  be the same.
1940
1941Example:
1942
1943.. code-block:: llvm
1944
1945  declare i32 @bar(i32, i32)
1946
1947  define i32 @foo(i32 %a, i32 %b, i32 %c) {
1948  entry:
1949    %0 = tail call i32 @bar(i32 %a, i32 %b)
1950    ret i32 %0
1951  }
1952
1953The X86 backend
1954---------------
1955
1956The X86 code generator lives in the ``lib/Target/X86`` directory.  This code
1957generator is capable of targeting a variety of x86-32 and x86-64 processors, and
1958includes support for ISA extensions such as MMX and SSE.
1959
1960X86 Target Triples supported
1961^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1962
1963The following are the known target triples that are supported by the X86
1964backend.  This is not an exhaustive list, and it would be useful to add those
1965that people test.
1966
1967* **i686-pc-linux-gnu** --- Linux
1968
1969* **i386-unknown-freebsd5.3** --- FreeBSD 5.3
1970
1971* **i686-pc-cygwin** --- Cygwin on Win32
1972
1973* **i686-pc-mingw32** --- MingW on Win32
1974
1975* **i386-pc-mingw32msvc** --- MingW crosscompiler on Linux
1976
1977* **i686-apple-darwin*** --- Apple Darwin on X86
1978
1979* **x86_64-unknown-linux-gnu** --- Linux
1980
1981X86 Calling Conventions supported
1982^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1983
1984The following target-specific calling conventions are known to backend:
1985
1986* **x86_StdCall** --- stdcall calling convention seen on Microsoft Windows
1987  platform (CC ID = 64).
1988
1989* **x86_FastCall** --- fastcall calling convention seen on Microsoft Windows
1990  platform (CC ID = 65).
1991
1992* **x86_ThisCall** --- Similar to X86_StdCall. Passes first argument in ECX,
1993  others via stack. Callee is responsible for stack cleaning. This convention is
1994  used by MSVC by default for methods in its ABI (CC ID = 70).
1995
1996.. _X86 addressing mode:
1997
1998Representing X86 addressing modes in MachineInstrs
1999^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2000
2001The x86 has a very flexible way of accessing memory.  It is capable of forming
2002memory addresses of the following expression directly in integer instructions
2003(which use ModR/M addressing):
2004
2005::
2006
2007  SegmentReg: Base + [1,2,4,8] * IndexReg + Disp32
2008
2009In order to represent this, LLVM tracks no less than 5 operands for each memory
2010operand of this form.  This means that the "load" form of '``mov``' has the
2011following ``MachineOperand``\s in this order:
2012
2013::
2014
2015  Index:        0     |    1        2       3           4          5
2016  Meaning:   DestReg, | BaseReg,  Scale, IndexReg, Displacement Segment
2017  OperandTy: VirtReg, | VirtReg, UnsImm, VirtReg,   SignExtImm  PhysReg
2018
2019Stores, and all other instructions, treat the four memory operands in the same
2020way and in the same order.  If the segment register is unspecified (regno = 0),
2021then no segment override is generated.  "Lea" operations do not have a segment
2022register specified, so they only have 4 operands for their memory reference.
2023
2024X86 address spaces supported
2025^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2026
2027x86 has a feature which provides the ability to perform loads and stores to
2028different address spaces via the x86 segment registers.  A segment override
2029prefix byte on an instruction causes the instruction's memory access to go to
2030the specified segment.  LLVM address space 0 is the default address space, which
2031includes the stack, and any unqualified memory accesses in a program.  Address
2032spaces 1-255 are currently reserved for user-defined code.  The GS-segment is
2033represented by address space 256, the FS-segment is represented by address space
2034257, and the SS-segment is represented by address space 258. Other x86 segments
2035have yet to be allocated address space numbers.
2036
2037While these address spaces may seem similar to TLS via the ``thread_local``
2038keyword, and often use the same underlying hardware, there are some fundamental
2039differences.
2040
2041The ``thread_local`` keyword applies to global variables and specifies that they
2042are to be allocated in thread-local memory. There are no type qualifiers
2043involved, and these variables can be pointed to with normal pointers and
2044accessed with normal loads and stores.  The ``thread_local`` keyword is
2045target-independent at the LLVM IR level (though LLVM doesn't yet have
2046implementations of it for some configurations)
2047
2048Special address spaces, in contrast, apply to static types. Every load and store
2049has a particular address space in its address operand type, and this is what
2050determines which address space is accessed.  LLVM ignores these special address
2051space qualifiers on global variables, and does not provide a way to directly
2052allocate storage in them.  At the LLVM IR level, the behavior of these special
2053address spaces depends in part on the underlying OS or runtime environment, and
2054they are specific to x86 (and LLVM doesn't yet handle them correctly in some
2055cases).
2056
2057Some operating systems and runtime environments use (or may in the future use)
2058the FS/GS-segment registers for various low-level purposes, so care should be
2059taken when considering them.
2060
2061Instruction naming
2062^^^^^^^^^^^^^^^^^^
2063
2064An instruction name consists of the base name, a default operand size, and a
2065character per operand with an optional special size. For example:
2066
2067::
2068
2069  ADD8rr      -> add, 8-bit register, 8-bit register
2070  IMUL16rmi   -> imul, 16-bit register, 16-bit memory, 16-bit immediate
2071  IMUL16rmi8  -> imul, 16-bit register, 16-bit memory, 8-bit immediate
2072  MOVSX32rm16 -> movsx, 32-bit register, 16-bit memory
2073
2074The PowerPC backend
2075-------------------
2076
2077The PowerPC code generator lives in the lib/Target/PowerPC directory.  The code
2078generation is retargetable to several variations or *subtargets* of the PowerPC
2079ISA; including ppc32, ppc64 and altivec.
2080
2081LLVM PowerPC ABI
2082^^^^^^^^^^^^^^^^
2083
2084LLVM follows the AIX PowerPC ABI, with two deviations. LLVM uses a PC relative
2085(PIC) or static addressing for accessing global values, so no TOC (r2) is
2086used. Second, r31 is used as a frame pointer to allow dynamic growth of a stack
2087frame.  LLVM takes advantage of having no TOC to provide space to save the frame
2088pointer in the PowerPC linkage area of the caller frame.  Other details of
2089PowerPC ABI can be found at `PowerPC ABI
2090<http://developer.apple.com/documentation/DeveloperTools/Conceptual/LowLevelABI/Articles/32bitPowerPC.html>`_\
2091. Note: This link describes the 32 bit ABI.  The 64 bit ABI is similar except
2092space for GPRs are 8 bytes wide (not 4) and r13 is reserved for system use.
2093
2094Frame Layout
2095^^^^^^^^^^^^
2096
2097The size of a PowerPC frame is usually fixed for the duration of a function's
2098invocation.  Since the frame is fixed size, all references into the frame can be
2099accessed via fixed offsets from the stack pointer.  The exception to this is
2100when dynamic alloca or variable sized arrays are present, then a base pointer
2101(r31) is used as a proxy for the stack pointer and stack pointer is free to grow
2102or shrink.  A base pointer is also used if llvm-gcc is not passed the
2103-fomit-frame-pointer flag. The stack pointer is always aligned to 16 bytes, so
2104that space allocated for altivec vectors will be properly aligned.
2105
2106An invocation frame is laid out as follows (low memory at top):
2107
2108:raw-html:`<table border="1" cellspacing="0">`
2109:raw-html:`<tr>`
2110:raw-html:`<td>Linkage<br><br></td>`
2111:raw-html:`</tr>`
2112:raw-html:`<tr>`
2113:raw-html:`<td>Parameter area<br><br></td>`
2114:raw-html:`</tr>`
2115:raw-html:`<tr>`
2116:raw-html:`<td>Dynamic area<br><br></td>`
2117:raw-html:`</tr>`
2118:raw-html:`<tr>`
2119:raw-html:`<td>Locals area<br><br></td>`
2120:raw-html:`</tr>`
2121:raw-html:`<tr>`
2122:raw-html:`<td>Saved registers area<br><br></td>`
2123:raw-html:`</tr>`
2124:raw-html:`<tr style="border-style: none hidden none hidden;">`
2125:raw-html:`<td><br></td>`
2126:raw-html:`</tr>`
2127:raw-html:`<tr>`
2128:raw-html:`<td>Previous Frame<br><br></td>`
2129:raw-html:`</tr>`
2130:raw-html:`</table>`
2131
2132The *linkage* area is used by a callee to save special registers prior to
2133allocating its own frame.  Only three entries are relevant to LLVM. The first
2134entry is the previous stack pointer (sp), aka link.  This allows probing tools
2135like gdb or exception handlers to quickly scan the frames in the stack.  A
2136function epilog can also use the link to pop the frame from the stack.  The
2137third entry in the linkage area is used to save the return address from the lr
2138register. Finally, as mentioned above, the last entry is used to save the
2139previous frame pointer (r31.)  The entries in the linkage area are the size of a
2140GPR, thus the linkage area is 24 bytes long in 32 bit mode and 48 bytes in 64
2141bit mode.
2142
214332 bit linkage area:
2144
2145:raw-html:`<table  border="1" cellspacing="0">`
2146:raw-html:`<tr>`
2147:raw-html:`<td>0</td>`
2148:raw-html:`<td>Saved SP (r1)</td>`
2149:raw-html:`</tr>`
2150:raw-html:`<tr>`
2151:raw-html:`<td>4</td>`
2152:raw-html:`<td>Saved CR</td>`
2153:raw-html:`</tr>`
2154:raw-html:`<tr>`
2155:raw-html:`<td>8</td>`
2156:raw-html:`<td>Saved LR</td>`
2157:raw-html:`</tr>`
2158:raw-html:`<tr>`
2159:raw-html:`<td>12</td>`
2160:raw-html:`<td>Reserved</td>`
2161:raw-html:`</tr>`
2162:raw-html:`<tr>`
2163:raw-html:`<td>16</td>`
2164:raw-html:`<td>Reserved</td>`
2165:raw-html:`</tr>`
2166:raw-html:`<tr>`
2167:raw-html:`<td>20</td>`
2168:raw-html:`<td>Saved FP (r31)</td>`
2169:raw-html:`</tr>`
2170:raw-html:`</table>`
2171
217264 bit linkage area:
2173
2174:raw-html:`<table border="1" cellspacing="0">`
2175:raw-html:`<tr>`
2176:raw-html:`<td>0</td>`
2177:raw-html:`<td>Saved SP (r1)</td>`
2178:raw-html:`</tr>`
2179:raw-html:`<tr>`
2180:raw-html:`<td>8</td>`
2181:raw-html:`<td>Saved CR</td>`
2182:raw-html:`</tr>`
2183:raw-html:`<tr>`
2184:raw-html:`<td>16</td>`
2185:raw-html:`<td>Saved LR</td>`
2186:raw-html:`</tr>`
2187:raw-html:`<tr>`
2188:raw-html:`<td>24</td>`
2189:raw-html:`<td>Reserved</td>`
2190:raw-html:`</tr>`
2191:raw-html:`<tr>`
2192:raw-html:`<td>32</td>`
2193:raw-html:`<td>Reserved</td>`
2194:raw-html:`</tr>`
2195:raw-html:`<tr>`
2196:raw-html:`<td>40</td>`
2197:raw-html:`<td>Saved FP (r31)</td>`
2198:raw-html:`</tr>`
2199:raw-html:`</table>`
2200
2201The *parameter area* is used to store arguments being passed to a callee
2202function.  Following the PowerPC ABI, the first few arguments are actually
2203passed in registers, with the space in the parameter area unused.  However, if
2204there are not enough registers or the callee is a thunk or vararg function,
2205these register arguments can be spilled into the parameter area.  Thus, the
2206parameter area must be large enough to store all the parameters for the largest
2207call sequence made by the caller.  The size must also be minimally large enough
2208to spill registers r3-r10.  This allows callees blind to the call signature,
2209such as thunks and vararg functions, enough space to cache the argument
2210registers.  Therefore, the parameter area is minimally 32 bytes (64 bytes in 64
2211bit mode.)  Also note that since the parameter area is a fixed offset from the
2212top of the frame, that a callee can access its split arguments using fixed
2213offsets from the stack pointer (or base pointer.)
2214
2215Combining the information about the linkage, parameter areas and alignment. A
2216stack frame is minimally 64 bytes in 32 bit mode and 128 bytes in 64 bit mode.
2217
2218The *dynamic area* starts out as size zero.  If a function uses dynamic alloca
2219then space is added to the stack, the linkage and parameter areas are shifted to
2220top of stack, and the new space is available immediately below the linkage and
2221parameter areas.  The cost of shifting the linkage and parameter areas is minor
2222since only the link value needs to be copied.  The link value can be easily
2223fetched by adding the original frame size to the base pointer.  Note that
2224allocations in the dynamic space need to observe 16 byte alignment.
2225
2226The *locals area* is where the llvm compiler reserves space for local variables.
2227
2228The *saved registers area* is where the llvm compiler spills callee saved
2229registers on entry to the callee.
2230
2231Prolog/Epilog
2232^^^^^^^^^^^^^
2233
2234The llvm prolog and epilog are the same as described in the PowerPC ABI, with
2235the following exceptions.  Callee saved registers are spilled after the frame is
2236created.  This allows the llvm epilog/prolog support to be common with other
2237targets.  The base pointer callee saved register r31 is saved in the TOC slot of
2238linkage area.  This simplifies allocation of space for the base pointer and
2239makes it convenient to locate programmatically and during debugging.
2240
2241Dynamic Allocation
2242^^^^^^^^^^^^^^^^^^
2243
2244.. note::
2245
2246  TODO - More to come.
2247
2248The NVPTX backend
2249-----------------
2250
2251The NVPTX code generator under lib/Target/NVPTX is an open-source version of
2252the NVIDIA NVPTX code generator for LLVM.  It is contributed by NVIDIA and is
2253a port of the code generator used in the CUDA compiler (nvcc).  It targets the
2254PTX 3.0/3.1 ISA and can target any compute capability greater than or equal to
22552.0 (Fermi).
2256
2257This target is of production quality and should be completely compatible with
2258the official NVIDIA toolchain.
2259
2260Code Generator Options:
2261
2262:raw-html:`<table border="1" cellspacing="0">`
2263:raw-html:`<tr>`
2264:raw-html:`<th>Option</th>`
2265:raw-html:`<th>Description</th>`
2266:raw-html:`</tr>`
2267:raw-html:`<tr>`
2268:raw-html:`<td>sm_20</td>`
2269:raw-html:`<td align="left">Set shader model/compute capability to 2.0</td>`
2270:raw-html:`</tr>`
2271:raw-html:`<tr>`
2272:raw-html:`<td>sm_21</td>`
2273:raw-html:`<td align="left">Set shader model/compute capability to 2.1</td>`
2274:raw-html:`</tr>`
2275:raw-html:`<tr>`
2276:raw-html:`<td>sm_30</td>`
2277:raw-html:`<td align="left">Set shader model/compute capability to 3.0</td>`
2278:raw-html:`</tr>`
2279:raw-html:`<tr>`
2280:raw-html:`<td>sm_35</td>`
2281:raw-html:`<td align="left">Set shader model/compute capability to 3.5</td>`
2282:raw-html:`</tr>`
2283:raw-html:`<tr>`
2284:raw-html:`<td>ptx30</td>`
2285:raw-html:`<td align="left">Target PTX 3.0</td>`
2286:raw-html:`</tr>`
2287:raw-html:`<tr>`
2288:raw-html:`<td>ptx31</td>`
2289:raw-html:`<td align="left">Target PTX 3.1</td>`
2290:raw-html:`</tr>`
2291:raw-html:`</table>`
2292
2293The extended Berkeley Packet Filter (eBPF) backend
2294--------------------------------------------------
2295
2296Extended BPF (or eBPF) is similar to the original ("classic") BPF (cBPF) used
2297to filter network packets.  The
2298`bpf() system call <http://man7.org/linux/man-pages/man2/bpf.2.html>`_
2299performs a range of operations related to eBPF.  For both cBPF and eBPF
2300programs, the Linux kernel statically analyzes the programs before loading
2301them, in order to ensure that they cannot harm the running system.  eBPF is
2302a 64-bit RISC instruction set designed for one to one mapping to 64-bit CPUs.
2303Opcodes are 8-bit encoded, and 87 instructions are defined.  There are 10
2304registers, grouped by function as outlined below.
2305
2306::
2307
2308  R0        return value from in-kernel functions; exit value for eBPF program
2309  R1 - R5   function call arguments to in-kernel functions
2310  R6 - R9   callee-saved registers preserved by in-kernel functions
2311  R10       stack frame pointer (read only)
2312
2313Instruction encoding (arithmetic and jump)
2314^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2315eBPF is reusing most of the opcode encoding from classic to simplify conversion
2316of classic BPF to eBPF.  For arithmetic and jump instructions the 8-bit 'code'
2317field is divided into three parts:
2318
2319::
2320
2321  +----------------+--------+--------------------+
2322  |   4 bits       |  1 bit |   3 bits           |
2323  | operation code | source | instruction class  |
2324  +----------------+--------+--------------------+
2325  (MSB)                                      (LSB)
2326
2327Three LSB bits store instruction class which is one of:
2328
2329::
2330
2331  BPF_LD     0x0
2332  BPF_LDX    0x1
2333  BPF_ST     0x2
2334  BPF_STX    0x3
2335  BPF_ALU    0x4
2336  BPF_JMP    0x5
2337  (unused)   0x6
2338  BPF_ALU64  0x7
2339
2340When BPF_CLASS(code) == BPF_ALU or BPF_ALU64 or BPF_JMP,
23414th bit encodes source operand
2342
2343::
2344
2345  BPF_X     0x1  use src_reg register as source operand
2346  BPF_K     0x0  use 32 bit immediate as source operand
2347
2348and four MSB bits store operation code
2349
2350::
2351
2352  BPF_ADD   0x0  add
2353  BPF_SUB   0x1  subtract
2354  BPF_MUL   0x2  multiply
2355  BPF_DIV   0x3  divide
2356  BPF_OR    0x4  bitwise logical OR
2357  BPF_AND   0x5  bitwise logical AND
2358  BPF_LSH   0x6  left shift
2359  BPF_RSH   0x7  right shift (zero extended)
2360  BPF_NEG   0x8  arithmetic negation
2361  BPF_MOD   0x9  modulo
2362  BPF_XOR   0xa  bitwise logical XOR
2363  BPF_MOV   0xb  move register to register
2364  BPF_ARSH  0xc  right shift (sign extended)
2365  BPF_END   0xd  endianness conversion
2366
2367If BPF_CLASS(code) == BPF_JMP, BPF_OP(code) is one of
2368
2369::
2370
2371  BPF_JA    0x0  unconditional jump
2372  BPF_JEQ   0x1  jump ==
2373  BPF_JGT   0x2  jump >
2374  BPF_JGE   0x3  jump >=
2375  BPF_JSET  0x4  jump if (DST & SRC)
2376  BPF_JNE   0x5  jump !=
2377  BPF_JSGT  0x6  jump signed >
2378  BPF_JSGE  0x7  jump signed >=
2379  BPF_CALL  0x8  function call
2380  BPF_EXIT  0x9  function return
2381
2382Instruction encoding (load, store)
2383^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2384For load and store instructions the 8-bit 'code' field is divided as:
2385
2386::
2387
2388  +--------+--------+-------------------+
2389  | 3 bits | 2 bits |   3 bits          |
2390  |  mode  |  size  | instruction class |
2391  +--------+--------+-------------------+
2392  (MSB)                             (LSB)
2393
2394Size modifier is one of
2395
2396::
2397
2398  BPF_W       0x0  word
2399  BPF_H       0x1  half word
2400  BPF_B       0x2  byte
2401  BPF_DW      0x3  double word
2402
2403Mode modifier is one of
2404
2405::
2406
2407  BPF_IMM     0x0  immediate
2408  BPF_ABS     0x1  used to access packet data
2409  BPF_IND     0x2  used to access packet data
2410  BPF_MEM     0x3  memory
2411  (reserved)  0x4
2412  (reserved)  0x5
2413  BPF_XADD    0x6  exclusive add
2414
2415
2416Packet data access (BPF_ABS, BPF_IND)
2417^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2418
2419Two non-generic instructions: (BPF_ABS | <size> | BPF_LD) and
2420(BPF_IND | <size> | BPF_LD) which are used to access packet data.
2421Register R6 is an implicit input that must contain pointer to sk_buff.
2422Register R0 is an implicit output which contains the data fetched
2423from the packet.  Registers R1-R5 are scratch registers and must not
2424be used to store the data across BPF_ABS | BPF_LD or BPF_IND | BPF_LD
2425instructions.  These instructions have implicit program exit condition
2426as well.  When eBPF program is trying to access the data beyond
2427the packet boundary, the interpreter will abort the execution of the program.
2428
2429BPF_IND | BPF_W | BPF_LD is equivalent to:
2430  R0 = ntohl(\*(u32 \*) (((struct sk_buff \*) R6)->data + src_reg + imm32))
2431
2432eBPF maps
2433^^^^^^^^^
2434
2435eBPF maps are provided for sharing data between kernel and user-space.
2436Currently implemented types are hash and array, with potential extension to
2437support bloom filters, radix trees, etc.  A map is defined by its type,
2438maximum number of elements, key size and value size in bytes.  eBPF syscall
2439supports create, update, find and delete functions on maps.
2440
2441Function calls
2442^^^^^^^^^^^^^^
2443
2444Function call arguments are passed using up to five registers (R1 - R5).
2445The return value is passed in a dedicated register (R0).  Four additional
2446registers (R6 - R9) are callee-saved, and the values in these registers
2447are preserved within kernel functions.  R0 - R5 are scratch registers within
2448kernel functions, and eBPF programs must therefor store/restore values in
2449these registers if needed across function calls.  The stack can be accessed
2450using the read-only frame pointer R10.  eBPF registers map 1:1 to hardware
2451registers on x86_64 and other 64-bit architectures.  For example, x86_64
2452in-kernel JIT maps them as
2453
2454::
2455
2456  R0 - rax
2457  R1 - rdi
2458  R2 - rsi
2459  R3 - rdx
2460  R4 - rcx
2461  R5 - r8
2462  R6 - rbx
2463  R7 - r13
2464  R8 - r14
2465  R9 - r15
2466  R10 - rbp
2467
2468since x86_64 ABI mandates rdi, rsi, rdx, rcx, r8, r9 for argument passing
2469and rbx, r12 - r15 are callee saved.
2470
2471Program start
2472^^^^^^^^^^^^^
2473
2474An eBPF program receives a single argument and contains
2475a single eBPF main routine; the program does not contain eBPF functions.
2476Function calls are limited to a predefined set of kernel functions.  The size
2477of a program is limited to 4K instructions:  this ensures fast termination and
2478a limited number of kernel function calls.  Prior to running an eBPF program,
2479a verifier performs static analysis to prevent loops in the code and
2480to ensure valid register usage and operand types.
2481
2482The AMDGPU backend
2483------------------
2484
2485The AMDGPU code generator lives in the ``lib/Target/AMDGPU``
2486directory. This code generator is capable of targeting a variety of
2487AMD GPU processors. Refer to :doc:`AMDGPUUsage` for more information.
2488