1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <meta http-equiv="content-type" content="text/html; charset=utf-8">
6 <title>The LLVM Target-Independent Code Generator</title>
7 <link rel="stylesheet" href="_static/llvm.css" type="text/css">
9 <style type="text/css">
10 .unknown { background-color: #C0C0C0; text-align: center; }
11 .unknown:before { content: "?" }
12 .no { background-color: #C11B17 }
13 .no:before { content: "N" }
14 .partial { background-color: #F88017 }
15 .yes { background-color: #0F0; }
16 .yes:before { content: "Y" }
23 The LLVM Target-Independent Code Generator
27 <li><a href="#introduction">Introduction</a>
29 <li><a href="#required">Required components in the code generator</a></li>
30 <li><a href="#high-level-design">The high-level design of the code
32 <li><a href="#tablegen">Using TableGen for target description</a></li>
35 <li><a href="#targetdesc">Target description classes</a>
37 <li><a href="#targetmachine">The <tt>TargetMachine</tt> class</a></li>
38 <li><a href="#targetdata">The <tt>TargetData</tt> class</a></li>
39 <li><a href="#targetlowering">The <tt>TargetLowering</tt> class</a></li>
40 <li><a href="#targetregisterinfo">The <tt>TargetRegisterInfo</tt> class</a></li>
41 <li><a href="#targetinstrinfo">The <tt>TargetInstrInfo</tt> class</a></li>
42 <li><a href="#targetframeinfo">The <tt>TargetFrameInfo</tt> class</a></li>
43 <li><a href="#targetsubtarget">The <tt>TargetSubtarget</tt> class</a></li>
44 <li><a href="#targetjitinfo">The <tt>TargetJITInfo</tt> class</a></li>
47 <li><a href="#codegendesc">The "Machine" Code Generator classes</a>
49 <li><a href="#machineinstr">The <tt>MachineInstr</tt> class</a></li>
50 <li><a href="#machinebasicblock">The <tt>MachineBasicBlock</tt>
52 <li><a href="#machinefunction">The <tt>MachineFunction</tt> class</a></li>
53 <li><a href="#machineinstrbundle"><tt>MachineInstr Bundles</tt></a></li>
56 <li><a href="#mc">The "MC" Layer</a>
58 <li><a href="#mcstreamer">The <tt>MCStreamer</tt> API</a></li>
59 <li><a href="#mccontext">The <tt>MCContext</tt> class</a>
60 <li><a href="#mcsymbol">The <tt>MCSymbol</tt> class</a></li>
61 <li><a href="#mcsection">The <tt>MCSection</tt> class</a></li>
62 <li><a href="#mcinst">The <tt>MCInst</tt> class</a></li>
65 <li><a href="#codegenalgs">Target-independent code generation algorithms</a>
67 <li><a href="#instselect">Instruction Selection</a>
69 <li><a href="#selectiondag_intro">Introduction to SelectionDAGs</a></li>
70 <li><a href="#selectiondag_process">SelectionDAG Code Generation
72 <li><a href="#selectiondag_build">Initial SelectionDAG
74 <li><a href="#selectiondag_legalize_types">SelectionDAG LegalizeTypes Phase</a></li>
75 <li><a href="#selectiondag_legalize">SelectionDAG Legalize Phase</a></li>
76 <li><a href="#selectiondag_optimize">SelectionDAG Optimization
77 Phase: the DAG Combiner</a></li>
78 <li><a href="#selectiondag_select">SelectionDAG Select Phase</a></li>
79 <li><a href="#selectiondag_sched">SelectionDAG Scheduling and Formation
81 <li><a href="#selectiondag_future">Future directions for the
84 <li><a href="#liveintervals">Live Intervals</a>
86 <li><a href="#livevariable_analysis">Live Variable Analysis</a></li>
87 <li><a href="#liveintervals_analysis">Live Intervals Analysis</a></li>
89 <li><a href="#regalloc">Register Allocation</a>
91 <li><a href="#regAlloc_represent">How registers are represented in
93 <li><a href="#regAlloc_howTo">Mapping virtual registers to physical
95 <li><a href="#regAlloc_twoAddr">Handling two address instructions</a></li>
96 <li><a href="#regAlloc_ssaDecon">The SSA deconstruction phase</a></li>
97 <li><a href="#regAlloc_fold">Instruction folding</a></li>
98 <li><a href="#regAlloc_builtIn">Built in register allocators</a></li>
100 <li><a href="#codeemit">Code Emission</a></li>
101 <li><a href="#vliw_packetizer">VLIW Packetizer</a>
103 <li><a href="#vliw_mapping">Mapping from instructions to functional
105 <li><a href="#vliw_repr">How the packetization tables are
106 generated and used</a></li>
111 <li><a href="#nativeassembler">Implementing a Native Assembler</a></li>
113 <li><a href="#targetimpls">Target-specific Implementation Notes</a>
115 <li><a href="#targetfeatures">Target Feature Matrix</a></li>
116 <li><a href="#tailcallopt">Tail call optimization</a></li>
117 <li><a href="#sibcallopt">Sibling call optimization</a></li>
118 <li><a href="#x86">The X86 backend</a></li>
119 <li><a href="#ppc">The PowerPC backend</a>
121 <li><a href="#ppc_abi">LLVM PowerPC ABI</a></li>
122 <li><a href="#ppc_frame">Frame Layout</a></li>
123 <li><a href="#ppc_prolog">Prolog/Epilog</a></li>
124 <li><a href="#ppc_dynamic">Dynamic Allocation</a></li>
126 <li><a href="#ptx">The PTX backend</a></li>
131 <div class="doc_author">
132 <p>Written by the LLVM Team.</p>
135 <div class="doc_warning">
136 <p>Warning: This is a work in progress.</p>
139 <!-- *********************************************************************** -->
141 <a name="introduction">Introduction</a>
143 <!-- *********************************************************************** -->
147 <p>The LLVM target-independent code generator is a framework that provides a
148 suite of reusable components for translating the LLVM internal representation
149 to the machine code for a specified target—either in assembly form
150 (suitable for a static compiler) or in binary machine code format (usable for
151 a JIT compiler). The LLVM target-independent code generator consists of six
155 <li><a href="#targetdesc">Abstract target description</a> interfaces which
156 capture important properties about various aspects of the machine,
157 independently of how they will be used. These interfaces are defined in
158 <tt>include/llvm/Target/</tt>.</li>
160 <li>Classes used to represent the <a href="#codegendesc">code being
161 generated</a> for a target. These classes are intended to be abstract
162 enough to represent the machine code for <i>any</i> target machine. These
163 classes are defined in <tt>include/llvm/CodeGen/</tt>. At this level,
164 concepts like "constant pool entries" and "jump tables" are explicitly
167 <li>Classes and algorithms used to represent code as the object file level,
168 the <a href="#mc">MC Layer</a>. These classes represent assembly level
169 constructs like labels, sections, and instructions. At this level,
170 concepts like "constant pool entries" and "jump tables" don't exist.</li>
172 <li><a href="#codegenalgs">Target-independent algorithms</a> used to implement
173 various phases of native code generation (register allocation, scheduling,
174 stack frame representation, etc). This code lives
175 in <tt>lib/CodeGen/</tt>.</li>
177 <li><a href="#targetimpls">Implementations of the abstract target description
178 interfaces</a> for particular targets. These machine descriptions make
179 use of the components provided by LLVM, and can optionally provide custom
180 target-specific passes, to build complete code generators for a specific
181 target. Target descriptions live in <tt>lib/Target/</tt>.</li>
183 <li><a href="#jit">The target-independent JIT components</a>. The LLVM JIT is
184 completely target independent (it uses the <tt>TargetJITInfo</tt>
185 structure to interface for target-specific issues. The code for the
186 target-independent JIT lives in <tt>lib/ExecutionEngine/JIT</tt>.</li>
189 <p>Depending on which part of the code generator you are interested in working
190 on, different pieces of this will be useful to you. In any case, you should
191 be familiar with the <a href="#targetdesc">target description</a>
192 and <a href="#codegendesc">machine code representation</a> classes. If you
193 want to add a backend for a new target, you will need
194 to <a href="#targetimpls">implement the target description</a> classes for
195 your new target and understand the <a href="LangRef.html">LLVM code
196 representation</a>. If you are interested in implementing a
197 new <a href="#codegenalgs">code generation algorithm</a>, it should only
198 depend on the target-description and machine code representation classes,
199 ensuring that it is portable.</p>
201 <!-- ======================================================================= -->
203 <a name="required">Required components in the code generator</a>
208 <p>The two pieces of the LLVM code generator are the high-level interface to the
209 code generator and the set of reusable components that can be used to build
210 target-specific backends. The two most important interfaces
211 (<a href="#targetmachine"><tt>TargetMachine</tt></a>
212 and <a href="#targetdata"><tt>TargetData</tt></a>) are the only ones that are
213 required to be defined for a backend to fit into the LLVM system, but the
214 others must be defined if the reusable code generator components are going to
217 <p>This design has two important implications. The first is that LLVM can
218 support completely non-traditional code generation targets. For example, the
219 C backend does not require register allocation, instruction selection, or any
220 of the other standard components provided by the system. As such, it only
221 implements these two interfaces, and does its own thing. Note that C backend
222 was removed from the trunk since LLVM 3.1 release. Another example of
223 a code generator like this is a (purely hypothetical) backend that converts
224 LLVM to the GCC RTL form and uses GCC to emit machine code for a target.</p>
226 <p>This design also implies that it is possible to design and implement
227 radically different code generators in the LLVM system that do not make use
228 of any of the built-in components. Doing so is not recommended at all, but
229 could be required for radically different targets that do not fit into the
230 LLVM machine description model: FPGAs for example.</p>
234 <!-- ======================================================================= -->
236 <a name="high-level-design">The high-level design of the code generator</a>
241 <p>The LLVM target-independent code generator is designed to support efficient
242 and quality code generation for standard register-based microprocessors.
243 Code generation in this model is divided into the following stages:</p>
246 <li><b><a href="#instselect">Instruction Selection</a></b> — This phase
247 determines an efficient way to express the input LLVM code in the target
248 instruction set. This stage produces the initial code for the program in
249 the target instruction set, then makes use of virtual registers in SSA
250 form and physical registers that represent any required register
251 assignments due to target constraints or calling conventions. This step
252 turns the LLVM code into a DAG of target instructions.</li>
254 <li><b><a href="#selectiondag_sched">Scheduling and Formation</a></b> —
255 This phase takes the DAG of target instructions produced by the
256 instruction selection phase, determines an ordering of the instructions,
257 then emits the instructions
258 as <tt><a href="#machineinstr">MachineInstr</a></tt>s with that ordering.
259 Note that we describe this in the <a href="#instselect">instruction
260 selection section</a> because it operates on
261 a <a href="#selectiondag_intro">SelectionDAG</a>.</li>
263 <li><b><a href="#ssamco">SSA-based Machine Code Optimizations</a></b> —
264 This optional stage consists of a series of machine-code optimizations
265 that operate on the SSA-form produced by the instruction selector.
266 Optimizations like modulo-scheduling or peephole optimization work
269 <li><b><a href="#regalloc">Register Allocation</a></b> — The target code
270 is transformed from an infinite virtual register file in SSA form to the
271 concrete register file used by the target. This phase introduces spill
272 code and eliminates all virtual register references from the program.</li>
274 <li><b><a href="#proepicode">Prolog/Epilog Code Insertion</a></b> — Once
275 the machine code has been generated for the function and the amount of
276 stack space required is known (used for LLVM alloca's and spill slots),
277 the prolog and epilog code for the function can be inserted and "abstract
278 stack location references" can be eliminated. This stage is responsible
279 for implementing optimizations like frame-pointer elimination and stack
282 <li><b><a href="#latemco">Late Machine Code Optimizations</a></b> —
283 Optimizations that operate on "final" machine code can go here, such as
284 spill code scheduling and peephole optimizations.</li>
286 <li><b><a href="#codeemit">Code Emission</a></b> — The final stage
287 actually puts out the code for the current function, either in the target
288 assembler format or in machine code.</li>
291 <p>The code generator is based on the assumption that the instruction selector
292 will use an optimal pattern matching selector to create high-quality
293 sequences of native instructions. Alternative code generator designs based
294 on pattern expansion and aggressive iterative peephole optimization are much
295 slower. This design permits efficient compilation (important for JIT
296 environments) and aggressive optimization (used when generating code offline)
297 by allowing components of varying levels of sophistication to be used for any
298 step of compilation.</p>
300 <p>In addition to these stages, target implementations can insert arbitrary
301 target-specific passes into the flow. For example, the X86 target uses a
302 special pass to handle the 80x87 floating point stack architecture. Other
303 targets with unusual requirements can be supported with custom passes as
308 <!-- ======================================================================= -->
310 <a name="tablegen">Using TableGen for target description</a>
315 <p>The target description classes require a detailed description of the target
316 architecture. These target descriptions often have a large amount of common
317 information (e.g., an <tt>add</tt> instruction is almost identical to a
318 <tt>sub</tt> instruction). In order to allow the maximum amount of
319 commonality to be factored out, the LLVM code generator uses
320 the <a href="TableGenFundamentals.html">TableGen</a> tool to describe big
321 chunks of the target machine, which allows the use of domain-specific and
322 target-specific abstractions to reduce the amount of repetition.</p>
324 <p>As LLVM continues to be developed and refined, we plan to move more and more
325 of the target description to the <tt>.td</tt> form. Doing so gives us a
326 number of advantages. The most important is that it makes it easier to port
327 LLVM because it reduces the amount of C++ code that has to be written, and
328 the surface area of the code generator that needs to be understood before
329 someone can get something working. Second, it makes it easier to change
330 things. In particular, if tables and other things are all emitted
331 by <tt>tblgen</tt>, we only need a change in one place (<tt>tblgen</tt>) to
332 update all of the targets to a new interface.</p>
338 <!-- *********************************************************************** -->
340 <a name="targetdesc">Target description classes</a>
342 <!-- *********************************************************************** -->
346 <p>The LLVM target description classes (located in the
347 <tt>include/llvm/Target</tt> directory) provide an abstract description of
348 the target machine independent of any particular client. These classes are
349 designed to capture the <i>abstract</i> properties of the target (such as the
350 instructions and registers it has), and do not incorporate any particular
351 pieces of code generation algorithms.</p>
353 <p>All of the target description classes (except the
354 <tt><a href="#targetdata">TargetData</a></tt> class) are designed to be
355 subclassed by the concrete target implementation, and have virtual methods
356 implemented. To get to these implementations, the
357 <tt><a href="#targetmachine">TargetMachine</a></tt> class provides accessors
358 that should be implemented by the target.</p>
360 <!-- ======================================================================= -->
362 <a name="targetmachine">The <tt>TargetMachine</tt> class</a>
367 <p>The <tt>TargetMachine</tt> class provides virtual methods that are used to
368 access the target-specific implementations of the various target description
369 classes via the <tt>get*Info</tt> methods (<tt>getInstrInfo</tt>,
370 <tt>getRegisterInfo</tt>, <tt>getFrameInfo</tt>, etc.). This class is
371 designed to be specialized by a concrete target implementation
372 (e.g., <tt>X86TargetMachine</tt>) which implements the various virtual
373 methods. The only required target description class is
374 the <a href="#targetdata"><tt>TargetData</tt></a> class, but if the code
375 generator components are to be used, the other interfaces should be
376 implemented as well.</p>
380 <!-- ======================================================================= -->
382 <a name="targetdata">The <tt>TargetData</tt> class</a>
387 <p>The <tt>TargetData</tt> class is the only required target description class,
388 and it is the only class that is not extensible (you cannot derived a new
389 class from it). <tt>TargetData</tt> specifies information about how the
390 target lays out memory for structures, the alignment requirements for various
391 data types, the size of pointers in the target, and whether the target is
392 little-endian or big-endian.</p>
396 <!-- ======================================================================= -->
398 <a name="targetlowering">The <tt>TargetLowering</tt> class</a>
403 <p>The <tt>TargetLowering</tt> class is used by SelectionDAG based instruction
404 selectors primarily to describe how LLVM code should be lowered to
405 SelectionDAG operations. Among other things, this class indicates:</p>
408 <li>an initial register class to use for various <tt>ValueType</tt>s,</li>
410 <li>which operations are natively supported by the target machine,</li>
412 <li>the return type of <tt>setcc</tt> operations,</li>
414 <li>the type to use for shift amounts, and</li>
416 <li>various high-level characteristics, like whether it is profitable to turn
417 division by a constant into a multiplication sequence</li>
422 <!-- ======================================================================= -->
424 <a name="targetregisterinfo">The <tt>TargetRegisterInfo</tt> class</a>
429 <p>The <tt>TargetRegisterInfo</tt> class is used to describe the register file
430 of the target and any interactions between the registers.</p>
432 <p>Registers in the code generator are represented in the code generator by
433 unsigned integers. Physical registers (those that actually exist in the
434 target description) are unique small numbers, and virtual registers are
435 generally large. Note that register #0 is reserved as a flag value.</p>
437 <p>Each register in the processor description has an associated
438 <tt>TargetRegisterDesc</tt> entry, which provides a textual name for the
439 register (used for assembly output and debugging dumps) and a set of aliases
440 (used to indicate whether one register overlaps with another).</p>
442 <p>In addition to the per-register description, the <tt>TargetRegisterInfo</tt>
443 class exposes a set of processor specific register classes (instances of the
444 <tt>TargetRegisterClass</tt> class). Each register class contains sets of
445 registers that have the same properties (for example, they are all 32-bit
446 integer registers). Each SSA virtual register created by the instruction
447 selector has an associated register class. When the register allocator runs,
448 it replaces virtual registers with a physical register in the set.</p>
450 <p>The target-specific implementations of these classes is auto-generated from
451 a <a href="TableGenFundamentals.html">TableGen</a> description of the
456 <!-- ======================================================================= -->
458 <a name="targetinstrinfo">The <tt>TargetInstrInfo</tt> class</a>
463 <p>The <tt>TargetInstrInfo</tt> class is used to describe the machine
464 instructions supported by the target. It is essentially an array of
465 <tt>TargetInstrDescriptor</tt> objects, each of which describes one
466 instruction the target supports. Descriptors define things like the mnemonic
467 for the opcode, the number of operands, the list of implicit register uses
468 and defs, whether the instruction has certain target-independent properties
469 (accesses memory, is commutable, etc), and holds any target-specific
474 <!-- ======================================================================= -->
476 <a name="targetframeinfo">The <tt>TargetFrameInfo</tt> class</a>
481 <p>The <tt>TargetFrameInfo</tt> class is used to provide information about the
482 stack frame layout of the target. It holds the direction of stack growth, the
483 known stack alignment on entry to each function, and the offset to the local
484 area. The offset to the local area is the offset from the stack pointer on
485 function entry to the first location where function data (local variables,
486 spill locations) can be stored.</p>
490 <!-- ======================================================================= -->
492 <a name="targetsubtarget">The <tt>TargetSubtarget</tt> class</a>
497 <p>The <tt>TargetSubtarget</tt> class is used to provide information about the
498 specific chip set being targeted. A sub-target informs code generation of
499 which instructions are supported, instruction latencies and instruction
500 execution itinerary; i.e., which processing units are used, in what order,
501 and for how long.</p>
506 <!-- ======================================================================= -->
508 <a name="targetjitinfo">The <tt>TargetJITInfo</tt> class</a>
513 <p>The <tt>TargetJITInfo</tt> class exposes an abstract interface used by the
514 Just-In-Time code generator to perform target-specific activities, such as
515 emitting stubs. If a <tt>TargetMachine</tt> supports JIT code generation, it
516 should provide one of these objects through the <tt>getJITInfo</tt>
523 <!-- *********************************************************************** -->
525 <a name="codegendesc">Machine code description classes</a>
527 <!-- *********************************************************************** -->
531 <p>At the high-level, LLVM code is translated to a machine specific
532 representation formed out of
533 <a href="#machinefunction"><tt>MachineFunction</tt></a>,
534 <a href="#machinebasicblock"><tt>MachineBasicBlock</tt></a>,
535 and <a href="#machineinstr"><tt>MachineInstr</tt></a> instances (defined
536 in <tt>include/llvm/CodeGen</tt>). This representation is completely target
537 agnostic, representing instructions in their most abstract form: an opcode
538 and a series of operands. This representation is designed to support both an
539 SSA representation for machine code, as well as a register allocated, non-SSA
542 <!-- ======================================================================= -->
544 <a name="machineinstr">The <tt>MachineInstr</tt> class</a>
549 <p>Target machine instructions are represented as instances of the
550 <tt>MachineInstr</tt> class. This class is an extremely abstract way of
551 representing machine instructions. In particular, it only keeps track of an
552 opcode number and a set of operands.</p>
554 <p>The opcode number is a simple unsigned integer that only has meaning to a
555 specific backend. All of the instructions for a target should be defined in
556 the <tt>*InstrInfo.td</tt> file for the target. The opcode enum values are
557 auto-generated from this description. The <tt>MachineInstr</tt> class does
558 not have any information about how to interpret the instruction (i.e., what
559 the semantics of the instruction are); for that you must refer to the
560 <tt><a href="#targetinstrinfo">TargetInstrInfo</a></tt> class.</p>
562 <p>The operands of a machine instruction can be of several different types: a
563 register reference, a constant integer, a basic block reference, etc. In
564 addition, a machine operand should be marked as a def or a use of the value
565 (though only registers are allowed to be defs).</p>
567 <p>By convention, the LLVM code generator orders instruction operands so that
568 all register definitions come before the register uses, even on architectures
569 that are normally printed in other orders. For example, the SPARC add
570 instruction: "<tt>add %i1, %i2, %i3</tt>" adds the "%i1", and "%i2" registers
571 and stores the result into the "%i3" register. In the LLVM code generator,
572 the operands should be stored as "<tt>%i3, %i1, %i2</tt>": with the
573 destination first.</p>
575 <p>Keeping destination (definition) operands at the beginning of the operand
576 list has several advantages. In particular, the debugging printer will print
577 the instruction like this:</p>
579 <div class="doc_code">
585 <p>Also if the first operand is a def, it is easier to <a href="#buildmi">create
586 instructions</a> whose only def is the first operand.</p>
588 <!-- _______________________________________________________________________ -->
590 <a name="buildmi">Using the <tt>MachineInstrBuilder.h</tt> functions</a>
595 <p>Machine instructions are created by using the <tt>BuildMI</tt> functions,
596 located in the <tt>include/llvm/CodeGen/MachineInstrBuilder.h</tt> file. The
597 <tt>BuildMI</tt> functions make it easy to build arbitrary machine
598 instructions. Usage of the <tt>BuildMI</tt> functions look like this:</p>
600 <div class="doc_code">
602 // Create a 'DestReg = mov 42' (rendered in X86 assembly as 'mov DestReg, 42')
603 // instruction. The '1' specifies how many operands will be added.
604 MachineInstr *MI = BuildMI(X86::MOV32ri, 1, DestReg).addImm(42);
606 // Create the same instr, but insert it at the end of a basic block.
607 MachineBasicBlock &MBB = ...
608 BuildMI(MBB, X86::MOV32ri, 1, DestReg).addImm(42);
610 // Create the same instr, but insert it before a specified iterator point.
611 MachineBasicBlock::iterator MBBI = ...
612 BuildMI(MBB, MBBI, X86::MOV32ri, 1, DestReg).addImm(42);
614 // Create a 'cmp Reg, 0' instruction, no destination reg.
615 MI = BuildMI(X86::CMP32ri, 2).addReg(Reg).addImm(0);
616 // Create an 'sahf' instruction which takes no operands and stores nothing.
617 MI = BuildMI(X86::SAHF, 0);
619 // Create a self looping branch instruction.
620 BuildMI(MBB, X86::JNE, 1).addMBB(&MBB);
624 <p>The key thing to remember with the <tt>BuildMI</tt> functions is that you
625 have to specify the number of operands that the machine instruction will
626 take. This allows for efficient memory allocation. You also need to specify
627 if operands default to be uses of values, not definitions. If you need to
628 add a definition operand (other than the optional destination register), you
629 must explicitly mark it as such:</p>
631 <div class="doc_code">
633 MI.addReg(Reg, RegState::Define);
639 <!-- _______________________________________________________________________ -->
641 <a name="fixedregs">Fixed (preassigned) registers</a>
646 <p>One important issue that the code generator needs to be aware of is the
647 presence of fixed registers. In particular, there are often places in the
648 instruction stream where the register allocator <em>must</em> arrange for a
649 particular value to be in a particular register. This can occur due to
650 limitations of the instruction set (e.g., the X86 can only do a 32-bit divide
651 with the <tt>EAX</tt>/<tt>EDX</tt> registers), or external factors like
652 calling conventions. In any case, the instruction selector should emit code
653 that copies a virtual register into or out of a physical register when
656 <p>For example, consider this simple LLVM example:</p>
658 <div class="doc_code">
660 define i32 @test(i32 %X, i32 %Y) {
667 <p>The X86 instruction selector produces this machine code for the <tt>div</tt>
668 and <tt>ret</tt> (use "<tt>llc X.bc -march=x86 -print-machineinstrs</tt>" to
671 <div class="doc_code">
674 %EAX = mov %reg1024 ;; Copy X (in reg1024) into EAX
675 %reg1027 = sar %reg1024, 31
676 %EDX = mov %reg1027 ;; Sign extend X into EDX
677 idiv %reg1025 ;; Divide by Y (in reg1025)
678 %reg1026 = mov %EAX ;; Read the result (Z) out of EAX
681 %EAX = mov %reg1026 ;; 32-bit return value goes in EAX
686 <p>By the end of code generation, the register allocator has coalesced the
687 registers and deleted the resultant identity moves producing the following
690 <div class="doc_code">
692 ;; X is in EAX, Y is in ECX
700 <p>This approach is extremely general (if it can handle the X86 architecture, it
701 can handle anything!) and allows all of the target specific knowledge about
702 the instruction stream to be isolated in the instruction selector. Note that
703 physical registers should have a short lifetime for good code generation, and
704 all physical registers are assumed dead on entry to and exit from basic
705 blocks (before register allocation). Thus, if you need a value to be live
706 across basic block boundaries, it <em>must</em> live in a virtual
711 <!-- _______________________________________________________________________ -->
713 <a name="callclobber">Call-clobbered registers</a>
718 <p>Some machine instructions, like calls, clobber a large number of physical
719 registers. Rather than adding <code><def,dead></code> operands for
720 all of them, it is possible to use an <code>MO_RegisterMask</code> operand
721 instead. The register mask operand holds a bit mask of preserved registers,
722 and everything else is considered to be clobbered by the instruction. </p>
726 <!-- _______________________________________________________________________ -->
728 <a name="ssa">Machine code in SSA form</a>
733 <p><tt>MachineInstr</tt>'s are initially selected in SSA-form, and are
734 maintained in SSA-form until register allocation happens. For the most part,
735 this is trivially simple since LLVM is already in SSA form; LLVM PHI nodes
736 become machine code PHI nodes, and virtual registers are only allowed to have
737 a single definition.</p>
739 <p>After register allocation, machine code is no longer in SSA-form because
740 there are no virtual registers left in the code.</p>
746 <!-- ======================================================================= -->
748 <a name="machinebasicblock">The <tt>MachineBasicBlock</tt> class</a>
753 <p>The <tt>MachineBasicBlock</tt> class contains a list of machine instructions
754 (<tt><a href="#machineinstr">MachineInstr</a></tt> instances). It roughly
755 corresponds to the LLVM code input to the instruction selector, but there can
756 be a one-to-many mapping (i.e. one LLVM basic block can map to multiple
757 machine basic blocks). The <tt>MachineBasicBlock</tt> class has a
758 "<tt>getBasicBlock</tt>" method, which returns the LLVM basic block that it
763 <!-- ======================================================================= -->
765 <a name="machinefunction">The <tt>MachineFunction</tt> class</a>
770 <p>The <tt>MachineFunction</tt> class contains a list of machine basic blocks
771 (<tt><a href="#machinebasicblock">MachineBasicBlock</a></tt> instances). It
772 corresponds one-to-one with the LLVM function input to the instruction
773 selector. In addition to a list of basic blocks,
774 the <tt>MachineFunction</tt> contains a a <tt>MachineConstantPool</tt>,
775 a <tt>MachineFrameInfo</tt>, a <tt>MachineFunctionInfo</tt>, and a
776 <tt>MachineRegisterInfo</tt>. See
777 <tt>include/llvm/CodeGen/MachineFunction.h</tt> for more information.</p>
781 <!-- ======================================================================= -->
783 <a name="machineinstrbundle"><tt>MachineInstr Bundles</tt></a>
788 <p>LLVM code generator can model sequences of instructions as MachineInstr
789 bundles. A MI bundle can model a VLIW group / pack which contains an
790 arbitrary number of parallel instructions. It can also be used to model
791 a sequential list of instructions (potentially with data dependencies) that
792 cannot be legally separated (e.g. ARM Thumb2 IT blocks).</p>
794 <p>Conceptually a MI bundle is a MI with a number of other MIs nested within:
797 <div class="doc_code">
835 <p> MI bundle support does not change the physical representations of
836 MachineBasicBlock and MachineInstr. All the MIs (including top level and
837 nested ones) are stored as sequential list of MIs. The "bundled" MIs are
838 marked with the 'InsideBundle' flag. A top level MI with the special BUNDLE
839 opcode is used to represent the start of a bundle. It's legal to mix BUNDLE
840 MIs with indiviual MIs that are not inside bundles nor represent bundles.
843 <p> MachineInstr passes should operate on a MI bundle as a single unit. Member
844 methods have been taught to correctly handle bundles and MIs inside bundles.
845 The MachineBasicBlock iterator has been modified to skip over bundled MIs to
846 enforce the bundle-as-a-single-unit concept. An alternative iterator
847 instr_iterator has been added to MachineBasicBlock to allow passes to
848 iterate over all of the MIs in a MachineBasicBlock, including those which
849 are nested inside bundles. The top level BUNDLE instruction must have the
850 correct set of register MachineOperand's that represent the cumulative
851 inputs and outputs of the bundled MIs.</p>
853 <p> Packing / bundling of MachineInstr's should be done as part of the register
854 allocation super-pass. More specifically, the pass which determines what
855 MIs should be bundled together must be done after code generator exits SSA
856 form (i.e. after two-address pass, PHI elimination, and copy coalescing).
857 Bundles should only be finalized (i.e. adding BUNDLE MIs and input and
858 output register MachineOperands) after virtual registers have been
859 rewritten into physical registers. This requirement eliminates the need to
860 add virtual register operands to BUNDLE instructions which would effectively
861 double the virtual register def and use lists.</p>
867 <!-- *********************************************************************** -->
869 <a name="mc">The "MC" Layer</a>
871 <!-- *********************************************************************** -->
876 The MC Layer is used to represent and process code at the raw machine code
877 level, devoid of "high level" information like "constant pools", "jump tables",
878 "global variables" or anything like that. At this level, LLVM handles things
879 like label names, machine instructions, and sections in the object file. The
880 code in this layer is used for a number of important purposes: the tail end of
881 the code generator uses it to write a .s or .o file, and it is also used by the
882 llvm-mc tool to implement standalone machine code assemblers and disassemblers.
886 This section describes some of the important classes. There are also a number
887 of important subsystems that interact at this layer, they are described later
891 <!-- ======================================================================= -->
893 <a name="mcstreamer">The <tt>MCStreamer</tt> API</a>
899 MCStreamer is best thought of as an assembler API. It is an abstract API which
900 is <em>implemented</em> in different ways (e.g. to output a .s file, output an
901 ELF .o file, etc) but whose API correspond directly to what you see in a .s
902 file. MCStreamer has one method per directive, such as EmitLabel,
903 EmitSymbolAttribute, SwitchSection, EmitValue (for .byte, .word), etc, which
904 directly correspond to assembly level directives. It also has an
905 EmitInstruction method, which is used to output an MCInst to the streamer.
909 This API is most important for two clients: the llvm-mc stand-alone assembler is
910 effectively a parser that parses a line, then invokes a method on MCStreamer. In
911 the code generator, the <a href="#codeemit">Code Emission</a> phase of the code
912 generator lowers higher level LLVM IR and Machine* constructs down to the MC
913 layer, emitting directives through MCStreamer.</p>
916 On the implementation side of MCStreamer, there are two major implementations:
917 one for writing out a .s file (MCAsmStreamer), and one for writing out a .o
918 file (MCObjectStreamer). MCAsmStreamer is a straight-forward implementation
919 that prints out a directive for each method (e.g. EmitValue -> .byte), but
920 MCObjectStreamer implements a full assembler.
925 <!-- ======================================================================= -->
927 <a name="mccontext">The <tt>MCContext</tt> class</a>
933 The MCContext class is the owner of a variety of uniqued data structures at the
934 MC layer, including symbols, sections, etc. As such, this is the class that you
935 interact with to create symbols and sections. This class can not be subclassed.
940 <!-- ======================================================================= -->
942 <a name="mcsymbol">The <tt>MCSymbol</tt> class</a>
948 The MCSymbol class represents a symbol (aka label) in the assembly file. There
949 are two interesting kinds of symbols: assembler temporary symbols, and normal
950 symbols. Assembler temporary symbols are used and processed by the assembler
951 but are discarded when the object file is produced. The distinction is usually
952 represented by adding a prefix to the label, for example "L" labels are
953 assembler temporary labels in MachO.
956 <p>MCSymbols are created by MCContext and uniqued there. This means that
957 MCSymbols can be compared for pointer equivalence to find out if they are the
958 same symbol. Note that pointer inequality does not guarantee the labels will
959 end up at different addresses though. It's perfectly legal to output something
960 like this to the .s file:<p>
968 <p>In this case, both the foo and bar symbols will have the same address.</p>
972 <!-- ======================================================================= -->
974 <a name="mcsection">The <tt>MCSection</tt> class</a>
980 The MCSection class represents an object-file specific section. It is subclassed
981 by object file specific implementations (e.g. <tt>MCSectionMachO</tt>,
982 <tt>MCSectionCOFF</tt>, <tt>MCSectionELF</tt>) and these are created and uniqued
983 by MCContext. The MCStreamer has a notion of the current section, which can be
984 changed with the SwitchToSection method (which corresponds to a ".section"
985 directive in a .s file).
990 <!-- ======================================================================= -->
992 <a name="mcinst">The <tt>MCInst</tt> class</a>
998 The MCInst class is a target-independent representation of an instruction. It
999 is a simple class (much more so than <a href="#machineinstr">MachineInstr</a>)
1000 that holds a target-specific opcode and a vector of MCOperands. MCOperand, in
1001 turn, is a simple discriminated union of three cases: 1) a simple immediate,
1002 2) a target register ID, 3) a symbolic expression (e.g. "Lfoo-Lbar+42") as an
1006 <p>MCInst is the common currency used to represent machine instructions at the
1007 MC layer. It is the type used by the instruction encoder, the instruction
1008 printer, and the type generated by the assembly parser and disassembler.
1015 <!-- *********************************************************************** -->
1017 <a name="codegenalgs">Target-independent code generation algorithms</a>
1019 <!-- *********************************************************************** -->
1023 <p>This section documents the phases described in the
1024 <a href="#high-level-design">high-level design of the code generator</a>.
1025 It explains how they work and some of the rationale behind their design.</p>
1027 <!-- ======================================================================= -->
1029 <a name="instselect">Instruction Selection</a>
1034 <p>Instruction Selection is the process of translating LLVM code presented to
1035 the code generator into target-specific machine instructions. There are
1036 several well-known ways to do this in the literature. LLVM uses a
1037 SelectionDAG based instruction selector.</p>
1039 <p>Portions of the DAG instruction selector are generated from the target
1040 description (<tt>*.td</tt>) files. Our goal is for the entire instruction
1041 selector to be generated from these <tt>.td</tt> files, though currently
1042 there are still things that require custom C++ code.</p>
1044 <!-- _______________________________________________________________________ -->
1046 <a name="selectiondag_intro">Introduction to SelectionDAGs</a>
1051 <p>The SelectionDAG provides an abstraction for code representation in a way
1052 that is amenable to instruction selection using automatic techniques
1053 (e.g. dynamic-programming based optimal pattern matching selectors). It is
1054 also well-suited to other phases of code generation; in particular,
1055 instruction scheduling (SelectionDAG's are very close to scheduling DAGs
1056 post-selection). Additionally, the SelectionDAG provides a host
1057 representation where a large variety of very-low-level (but
1058 target-independent) <a href="#selectiondag_optimize">optimizations</a> may be
1059 performed; ones which require extensive information about the instructions
1060 efficiently supported by the target.</p>
1062 <p>The SelectionDAG is a Directed-Acyclic-Graph whose nodes are instances of the
1063 <tt>SDNode</tt> class. The primary payload of the <tt>SDNode</tt> is its
1064 operation code (Opcode) that indicates what operation the node performs and
1065 the operands to the operation. The various operation node types are
1066 described at the top of the <tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt>
1069 <p>Although most operations define a single value, each node in the graph may
1070 define multiple values. For example, a combined div/rem operation will
1071 define both the dividend and the remainder. Many other situations require
1072 multiple values as well. Each node also has some number of operands, which
1073 are edges to the node defining the used value. Because nodes may define
1074 multiple values, edges are represented by instances of the <tt>SDValue</tt>
1075 class, which is a <tt><SDNode, unsigned></tt> pair, indicating the node
1076 and result value being used, respectively. Each value produced by
1077 an <tt>SDNode</tt> has an associated <tt>MVT</tt> (Machine Value Type)
1078 indicating what the type of the value is.</p>
1080 <p>SelectionDAGs contain two different kinds of values: those that represent
1081 data flow and those that represent control flow dependencies. Data values
1082 are simple edges with an integer or floating point value type. Control edges
1083 are represented as "chain" edges which are of type <tt>MVT::Other</tt>.
1084 These edges provide an ordering between nodes that have side effects (such as
1085 loads, stores, calls, returns, etc). All nodes that have side effects should
1086 take a token chain as input and produce a new one as output. By convention,
1087 token chain inputs are always operand #0, and chain results are always the
1088 last value produced by an operation.</p>
1090 <p>A SelectionDAG has designated "Entry" and "Root" nodes. The Entry node is
1091 always a marker node with an Opcode of <tt>ISD::EntryToken</tt>. The Root
1092 node is the final side-effecting node in the token chain. For example, in a
1093 single basic block function it would be the return node.</p>
1095 <p>One important concept for SelectionDAGs is the notion of a "legal" vs.
1096 "illegal" DAG. A legal DAG for a target is one that only uses supported
1097 operations and supported types. On a 32-bit PowerPC, for example, a DAG with
1098 a value of type i1, i8, i16, or i64 would be illegal, as would a DAG that
1099 uses a SREM or UREM operation. The
1100 <a href="#selectinodag_legalize_types">legalize types</a> and
1101 <a href="#selectiondag_legalize">legalize operations</a> phases are
1102 responsible for turning an illegal DAG into a legal DAG.</p>
1106 <!-- _______________________________________________________________________ -->
1108 <a name="selectiondag_process">SelectionDAG Instruction Selection Process</a>
1113 <p>SelectionDAG-based instruction selection consists of the following steps:</p>
1116 <li><a href="#selectiondag_build">Build initial DAG</a> — This stage
1117 performs a simple translation from the input LLVM code to an illegal
1120 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> — This
1121 stage performs simple optimizations on the SelectionDAG to simplify it,
1122 and recognize meta instructions (like rotates
1123 and <tt>div</tt>/<tt>rem</tt> pairs) for targets that support these meta
1124 operations. This makes the resultant code more efficient and
1125 the <a href="#selectiondag_select">select instructions from DAG</a> phase
1126 (below) simpler.</li>
1128 <li><a href="#selectiondag_legalize_types">Legalize SelectionDAG Types</a>
1129 — This stage transforms SelectionDAG nodes to eliminate any types
1130 that are unsupported on the target.</li>
1132 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> — The
1133 SelectionDAG optimizer is run to clean up redundancies exposed by type
1136 <li><a href="#selectiondag_legalize">Legalize SelectionDAG Ops</a> —
1137 This stage transforms SelectionDAG nodes to eliminate any operations
1138 that are unsupported on the target.</li>
1140 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> — The
1141 SelectionDAG optimizer is run to eliminate inefficiencies introduced by
1142 operation legalization.</li>
1144 <li><a href="#selectiondag_select">Select instructions from DAG</a> —
1145 Finally, the target instruction selector matches the DAG operations to
1146 target instructions. This process translates the target-independent input
1147 DAG into another DAG of target instructions.</li>
1149 <li><a href="#selectiondag_sched">SelectionDAG Scheduling and Formation</a>
1150 — The last phase assigns a linear order to the instructions in the
1151 target-instruction DAG and emits them into the MachineFunction being
1152 compiled. This step uses traditional prepass scheduling techniques.</li>
1155 <p>After all of these steps are complete, the SelectionDAG is destroyed and the
1156 rest of the code generation passes are run.</p>
1158 <p>One great way to visualize what is going on here is to take advantage of a
1159 few LLC command line options. The following options pop up a window
1160 displaying the SelectionDAG at specific times (if you only get errors printed
1161 to the console while using this, you probably
1162 <a href="ProgrammersManual.html#ViewGraph">need to configure your system</a>
1163 to add support for it).</p>
1166 <li><tt>-view-dag-combine1-dags</tt> displays the DAG after being built,
1167 before the first optimization pass.</li>
1169 <li><tt>-view-legalize-dags</tt> displays the DAG before Legalization.</li>
1171 <li><tt>-view-dag-combine2-dags</tt> displays the DAG before the second
1172 optimization pass.</li>
1174 <li><tt>-view-isel-dags</tt> displays the DAG before the Select phase.</li>
1176 <li><tt>-view-sched-dags</tt> displays the DAG before Scheduling.</li>
1179 <p>The <tt>-view-sunit-dags</tt> displays the Scheduler's dependency graph.
1180 This graph is based on the final SelectionDAG, with nodes that must be
1181 scheduled together bundled into a single scheduling-unit node, and with
1182 immediate operands and other nodes that aren't relevant for scheduling
1187 <!-- _______________________________________________________________________ -->
1189 <a name="selectiondag_build">Initial SelectionDAG Construction</a>
1194 <p>The initial SelectionDAG is naïvely peephole expanded from the LLVM
1195 input by the <tt>SelectionDAGLowering</tt> class in the
1196 <tt>lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp</tt> file. The intent of
1197 this pass is to expose as much low-level, target-specific details to the
1198 SelectionDAG as possible. This pass is mostly hard-coded (e.g. an
1199 LLVM <tt>add</tt> turns into an <tt>SDNode add</tt> while a
1200 <tt>getelementptr</tt> is expanded into the obvious arithmetic). This pass
1201 requires target-specific hooks to lower calls, returns, varargs, etc. For
1202 these features, the <tt><a href="#targetlowering">TargetLowering</a></tt>
1203 interface is used.</p>
1207 <!-- _______________________________________________________________________ -->
1209 <a name="selectiondag_legalize_types">SelectionDAG LegalizeTypes Phase</a>
1214 <p>The Legalize phase is in charge of converting a DAG to only use the types
1215 that are natively supported by the target.</p>
1217 <p>There are two main ways of converting values of unsupported scalar types to
1218 values of supported types: converting small types to larger types
1219 ("promoting"), and breaking up large integer types into smaller ones
1220 ("expanding"). For example, a target might require that all f32 values are
1221 promoted to f64 and that all i1/i8/i16 values are promoted to i32. The same
1222 target might require that all i64 values be expanded into pairs of i32
1223 values. These changes can insert sign and zero extensions as needed to make
1224 sure that the final code has the same behavior as the input.</p>
1226 <p>There are two main ways of converting values of unsupported vector types to
1227 value of supported types: splitting vector types, multiple times if
1228 necessary, until a legal type is found, and extending vector types by adding
1229 elements to the end to round them out to legal types ("widening"). If a
1230 vector gets split all the way down to single-element parts with no supported
1231 vector type being found, the elements are converted to scalars
1232 ("scalarizing").</p>
1234 <p>A target implementation tells the legalizer which types are supported (and
1235 which register class to use for them) by calling the
1236 <tt>addRegisterClass</tt> method in its TargetLowering constructor.</p>
1240 <!-- _______________________________________________________________________ -->
1242 <a name="selectiondag_legalize">SelectionDAG Legalize Phase</a>
1247 <p>The Legalize phase is in charge of converting a DAG to only use the
1248 operations that are natively supported by the target.</p>
1250 <p>Targets often have weird constraints, such as not supporting every operation
1251 on every supported datatype (e.g. X86 does not support byte conditional moves
1252 and PowerPC does not support sign-extending loads from a 16-bit memory
1253 location). Legalize takes care of this by open-coding another sequence of
1254 operations to emulate the operation ("expansion"), by promoting one type to a
1255 larger type that supports the operation ("promotion"), or by using a
1256 target-specific hook to implement the legalization ("custom").</p>
1258 <p>A target implementation tells the legalizer which operations are not
1259 supported (and which of the above three actions to take) by calling the
1260 <tt>setOperationAction</tt> method in its <tt>TargetLowering</tt>
1263 <p>Prior to the existence of the Legalize passes, we required that every target
1264 <a href="#selectiondag_optimize">selector</a> supported and handled every
1265 operator and type even if they are not natively supported. The introduction
1266 of the Legalize phases allows all of the canonicalization patterns to be
1267 shared across targets, and makes it very easy to optimize the canonicalized
1268 code because it is still in the form of a DAG.</p>
1272 <!-- _______________________________________________________________________ -->
1274 <a name="selectiondag_optimize">
1275 SelectionDAG Optimization Phase: the DAG Combiner
1281 <p>The SelectionDAG optimization phase is run multiple times for code
1282 generation, immediately after the DAG is built and once after each
1283 legalization. The first run of the pass allows the initial code to be
1284 cleaned up (e.g. performing optimizations that depend on knowing that the
1285 operators have restricted type inputs). Subsequent runs of the pass clean up
1286 the messy code generated by the Legalize passes, which allows Legalize to be
1287 very simple (it can focus on making code legal instead of focusing on
1288 generating <em>good</em> and legal code).</p>
1290 <p>One important class of optimizations performed is optimizing inserted sign
1291 and zero extension instructions. We currently use ad-hoc techniques, but
1292 could move to more rigorous techniques in the future. Here are some good
1293 papers on the subject:</p>
1295 <p>"<a href="http://www.eecs.harvard.edu/~nr/pubs/widen-abstract.html">Widening
1296 integer arithmetic</a>"<br>
1297 Kevin Redwine and Norman Ramsey<br>
1298 International Conference on Compiler Construction (CC) 2004</p>
1300 <p>"<a href="http://portal.acm.org/citation.cfm?doid=512529.512552">Effective
1301 sign extension elimination</a>"<br>
1302 Motohiro Kawahito, Hideaki Komatsu, and Toshio Nakatani<br>
1303 Proceedings of the ACM SIGPLAN 2002 Conference on Programming Language Design
1304 and Implementation.</p>
1308 <!-- _______________________________________________________________________ -->
1310 <a name="selectiondag_select">SelectionDAG Select Phase</a>
1315 <p>The Select phase is the bulk of the target-specific code for instruction
1316 selection. This phase takes a legal SelectionDAG as input, pattern matches
1317 the instructions supported by the target to this DAG, and produces a new DAG
1318 of target code. For example, consider the following LLVM fragment:</p>
1320 <div class="doc_code">
1322 %t1 = fadd float %W, %X
1323 %t2 = fmul float %t1, %Y
1324 %t3 = fadd float %t2, %Z
1328 <p>This LLVM code corresponds to a SelectionDAG that looks basically like
1331 <div class="doc_code">
1333 (fadd:f32 (fmul:f32 (fadd:f32 W, X), Y), Z)
1337 <p>If a target supports floating point multiply-and-add (FMA) operations, one of
1338 the adds can be merged with the multiply. On the PowerPC, for example, the
1339 output of the instruction selector might look like this DAG:</p>
1341 <div class="doc_code">
1343 (FMADDS (FADDS W, X), Y, Z)
1347 <p>The <tt>FMADDS</tt> instruction is a ternary instruction that multiplies its
1348 first two operands and adds the third (as single-precision floating-point
1349 numbers). The <tt>FADDS</tt> instruction is a simple binary single-precision
1350 add instruction. To perform this pattern match, the PowerPC backend includes
1351 the following instruction definitions:</p>
1353 <div class="doc_code">
1355 def FMADDS : AForm_1<59, 29,
1356 (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
1357 "fmadds $FRT, $FRA, $FRC, $FRB",
1358 [<b>(set F4RC:$FRT, (fadd (fmul F4RC:$FRA, F4RC:$FRC),
1359 F4RC:$FRB))</b>]>;
1360 def FADDS : AForm_2<59, 21,
1361 (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
1362 "fadds $FRT, $FRA, $FRB",
1363 [<b>(set F4RC:$FRT, (fadd F4RC:$FRA, F4RC:$FRB))</b>]>;
1367 <p>The portion of the instruction definition in bold indicates the pattern used
1368 to match the instruction. The DAG operators
1369 (like <tt>fmul</tt>/<tt>fadd</tt>) are defined in
1370 the <tt>include/llvm/Target/TargetSelectionDAG.td</tt> file. "
1371 <tt>F4RC</tt>" is the register class of the input and result values.</p>
1373 <p>The TableGen DAG instruction selector generator reads the instruction
1374 patterns in the <tt>.td</tt> file and automatically builds parts of the
1375 pattern matching code for your target. It has the following strengths:</p>
1378 <li>At compiler-compiler time, it analyzes your instruction patterns and tells
1379 you if your patterns make sense or not.</li>
1381 <li>It can handle arbitrary constraints on operands for the pattern match. In
1382 particular, it is straight-forward to say things like "match any immediate
1383 that is a 13-bit sign-extended value". For examples, see the
1384 <tt>immSExt16</tt> and related <tt>tblgen</tt> classes in the PowerPC
1387 <li>It knows several important identities for the patterns defined. For
1388 example, it knows that addition is commutative, so it allows the
1389 <tt>FMADDS</tt> pattern above to match "<tt>(fadd X, (fmul Y, Z))</tt>" as
1390 well as "<tt>(fadd (fmul X, Y), Z)</tt>", without the target author having
1391 to specially handle this case.</li>
1393 <li>It has a full-featured type-inferencing system. In particular, you should
1394 rarely have to explicitly tell the system what type parts of your patterns
1395 are. In the <tt>FMADDS</tt> case above, we didn't have to tell
1396 <tt>tblgen</tt> that all of the nodes in the pattern are of type 'f32'.
1397 It was able to infer and propagate this knowledge from the fact that
1398 <tt>F4RC</tt> has type 'f32'.</li>
1400 <li>Targets can define their own (and rely on built-in) "pattern fragments".
1401 Pattern fragments are chunks of reusable patterns that get inlined into
1402 your patterns during compiler-compiler time. For example, the integer
1403 "<tt>(not x)</tt>" operation is actually defined as a pattern fragment
1404 that expands as "<tt>(xor x, -1)</tt>", since the SelectionDAG does not
1405 have a native '<tt>not</tt>' operation. Targets can define their own
1406 short-hand fragments as they see fit. See the definition of
1407 '<tt>not</tt>' and '<tt>ineg</tt>' for examples.</li>
1409 <li>In addition to instructions, targets can specify arbitrary patterns that
1410 map to one or more instructions using the 'Pat' class. For example, the
1411 PowerPC has no way to load an arbitrary integer immediate into a register
1412 in one instruction. To tell tblgen how to do this, it defines:
1415 <div class="doc_code">
1417 // Arbitrary immediate support. Implement in terms of LIS/ORI.
1418 def : Pat<(i32 imm:$imm),
1419 (ORI (LIS (HI16 imm:$imm)), (LO16 imm:$imm))>;
1423 If none of the single-instruction patterns for loading an immediate into a
1424 register match, this will be used. This rule says "match an arbitrary i32
1425 immediate, turning it into an <tt>ORI</tt> ('or a 16-bit immediate') and
1426 an <tt>LIS</tt> ('load 16-bit immediate, where the immediate is shifted to
1427 the left 16 bits') instruction". To make this work, the
1428 <tt>LO16</tt>/<tt>HI16</tt> node transformations are used to manipulate
1429 the input immediate (in this case, take the high or low 16-bits of the
1432 <li>While the system does automate a lot, it still allows you to write custom
1433 C++ code to match special cases if there is something that is hard to
1437 <p>While it has many strengths, the system currently has some limitations,
1438 primarily because it is a work in progress and is not yet finished:</p>
1441 <li>Overall, there is no way to define or match SelectionDAG nodes that define
1442 multiple values (e.g. <tt>SMUL_LOHI</tt>, <tt>LOAD</tt>, <tt>CALL</tt>,
1443 etc). This is the biggest reason that you currently still <em>have
1444 to</em> write custom C++ code for your instruction selector.</li>
1446 <li>There is no great way to support matching complex addressing modes yet.
1447 In the future, we will extend pattern fragments to allow them to define
1448 multiple values (e.g. the four operands of the <a href="#x86_memory">X86
1449 addressing mode</a>, which are currently matched with custom C++ code).
1450 In addition, we'll extend fragments so that a fragment can match multiple
1451 different patterns.</li>
1453 <li>We don't automatically infer flags like isStore/isLoad yet.</li>
1455 <li>We don't automatically generate the set of supported registers and
1456 operations for the <a href="#selectiondag_legalize">Legalizer</a>
1459 <li>We don't have a way of tying in custom legalized nodes yet.</li>
1462 <p>Despite these limitations, the instruction selector generator is still quite
1463 useful for most of the binary and logical operations in typical instruction
1464 sets. If you run into any problems or can't figure out how to do something,
1465 please let Chris know!</p>
1469 <!-- _______________________________________________________________________ -->
1471 <a name="selectiondag_sched">SelectionDAG Scheduling and Formation Phase</a>
1476 <p>The scheduling phase takes the DAG of target instructions from the selection
1477 phase and assigns an order. The scheduler can pick an order depending on
1478 various constraints of the machines (i.e. order for minimal register pressure
1479 or try to cover instruction latencies). Once an order is established, the
1480 DAG is converted to a list
1481 of <tt><a href="#machineinstr">MachineInstr</a></tt>s and the SelectionDAG is
1484 <p>Note that this phase is logically separate from the instruction selection
1485 phase, but is tied to it closely in the code because it operates on
1490 <!-- _______________________________________________________________________ -->
1492 <a name="selectiondag_future">Future directions for the SelectionDAG</a>
1498 <li>Optional function-at-a-time selection.</li>
1500 <li>Auto-generate entire selector from <tt>.td</tt> file.</li>
1507 <!-- ======================================================================= -->
1509 <a name="ssamco">SSA-based Machine Code Optimizations</a>
1511 <div><p>To Be Written</p></div>
1513 <!-- ======================================================================= -->
1515 <a name="liveintervals">Live Intervals</a>
1520 <p>Live Intervals are the ranges (intervals) where a variable is <i>live</i>.
1521 They are used by some <a href="#regalloc">register allocator</a> passes to
1522 determine if two or more virtual registers which require the same physical
1523 register are live at the same point in the program (i.e., they conflict).
1524 When this situation occurs, one virtual register must be <i>spilled</i>.</p>
1526 <!-- _______________________________________________________________________ -->
1528 <a name="livevariable_analysis">Live Variable Analysis</a>
1533 <p>The first step in determining the live intervals of variables is to calculate
1534 the set of registers that are immediately dead after the instruction (i.e.,
1535 the instruction calculates the value, but it is never used) and the set of
1536 registers that are used by the instruction, but are never used after the
1537 instruction (i.e., they are killed). Live variable information is computed
1538 for each <i>virtual</i> register and <i>register allocatable</i> physical
1539 register in the function. This is done in a very efficient manner because it
1540 uses SSA to sparsely compute lifetime information for virtual registers
1541 (which are in SSA form) and only has to track physical registers within a
1542 block. Before register allocation, LLVM can assume that physical registers
1543 are only live within a single basic block. This allows it to do a single,
1544 local analysis to resolve physical register lifetimes within each basic
1545 block. If a physical register is not register allocatable (e.g., a stack
1546 pointer or condition codes), it is not tracked.</p>
1548 <p>Physical registers may be live in to or out of a function. Live in values are
1549 typically arguments in registers. Live out values are typically return values
1550 in registers. Live in values are marked as such, and are given a dummy
1551 "defining" instruction during live intervals analysis. If the last basic
1552 block of a function is a <tt>return</tt>, then it's marked as using all live
1553 out values in the function.</p>
1555 <p><tt>PHI</tt> nodes need to be handled specially, because the calculation of
1556 the live variable information from a depth first traversal of the CFG of the
1557 function won't guarantee that a virtual register used by the <tt>PHI</tt>
1558 node is defined before it's used. When a <tt>PHI</tt> node is encountered,
1559 only the definition is handled, because the uses will be handled in other
1562 <p>For each <tt>PHI</tt> node of the current basic block, we simulate an
1563 assignment at the end of the current basic block and traverse the successor
1564 basic blocks. If a successor basic block has a <tt>PHI</tt> node and one of
1565 the <tt>PHI</tt> node's operands is coming from the current basic block, then
1566 the variable is marked as <i>alive</i> within the current basic block and all
1567 of its predecessor basic blocks, until the basic block with the defining
1568 instruction is encountered.</p>
1572 <!-- _______________________________________________________________________ -->
1574 <a name="liveintervals_analysis">Live Intervals Analysis</a>
1579 <p>We now have the information available to perform the live intervals analysis
1580 and build the live intervals themselves. We start off by numbering the basic
1581 blocks and machine instructions. We then handle the "live-in" values. These
1582 are in physical registers, so the physical register is assumed to be killed
1583 by the end of the basic block. Live intervals for virtual registers are
1584 computed for some ordering of the machine instructions <tt>[1, N]</tt>. A
1585 live interval is an interval <tt>[i, j)</tt>, where <tt>1 <= i <= j
1586 < N</tt>, for which a variable is live.</p>
1588 <p><i><b>More to come...</b></i></p>
1594 <!-- ======================================================================= -->
1596 <a name="regalloc">Register Allocation</a>
1601 <p>The <i>Register Allocation problem</i> consists in mapping a program
1602 <i>P<sub>v</sub></i>, that can use an unbounded number of virtual registers,
1603 to a program <i>P<sub>p</sub></i> that contains a finite (possibly small)
1604 number of physical registers. Each target architecture has a different number
1605 of physical registers. If the number of physical registers is not enough to
1606 accommodate all the virtual registers, some of them will have to be mapped
1607 into memory. These virtuals are called <i>spilled virtuals</i>.</p>
1609 <!-- _______________________________________________________________________ -->
1612 <a name="regAlloc_represent">How registers are represented in LLVM</a>
1617 <p>In LLVM, physical registers are denoted by integer numbers that normally
1618 range from 1 to 1023. To see how this numbering is defined for a particular
1619 architecture, you can read the <tt>GenRegisterNames.inc</tt> file for that
1620 architecture. For instance, by
1621 inspecting <tt>lib/Target/X86/X86GenRegisterInfo.inc</tt> we see that the
1622 32-bit register <tt>EAX</tt> is denoted by 43, and the MMX register
1623 <tt>MM0</tt> is mapped to 65.</p>
1625 <p>Some architectures contain registers that share the same physical location. A
1626 notable example is the X86 platform. For instance, in the X86 architecture,
1627 the registers <tt>EAX</tt>, <tt>AX</tt> and <tt>AL</tt> share the first eight
1628 bits. These physical registers are marked as <i>aliased</i> in LLVM. Given a
1629 particular architecture, you can check which registers are aliased by
1630 inspecting its <tt>RegisterInfo.td</tt> file. Moreover, the class
1631 <tt>MCRegAliasIterator</tt> enumerates all the physical registers aliased to
1634 <p>Physical registers, in LLVM, are grouped in <i>Register Classes</i>.
1635 Elements in the same register class are functionally equivalent, and can be
1636 interchangeably used. Each virtual register can only be mapped to physical
1637 registers of a particular class. For instance, in the X86 architecture, some
1638 virtuals can only be allocated to 8 bit registers. A register class is
1639 described by <tt>TargetRegisterClass</tt> objects. To discover if a virtual
1640 register is compatible with a given physical, this code can be used:</p>
1642 <div class="doc_code">
1644 bool RegMapping_Fer::compatible_class(MachineFunction &mf,
1647 assert(TargetRegisterInfo::isPhysicalRegister(p_reg) &&
1648 "Target register must be physical");
1649 const TargetRegisterClass *trc = mf.getRegInfo().getRegClass(v_reg);
1650 return trc->contains(p_reg);
1655 <p>Sometimes, mostly for debugging purposes, it is useful to change the number
1656 of physical registers available in the target architecture. This must be done
1657 statically, inside the <tt>TargetRegsterInfo.td</tt> file. Just <tt>grep</tt>
1658 for <tt>RegisterClass</tt>, the last parameter of which is a list of
1659 registers. Just commenting some out is one simple way to avoid them being
1660 used. A more polite way is to explicitly exclude some registers from
1661 the <i>allocation order</i>. See the definition of the <tt>GR8</tt> register
1662 class in <tt>lib/Target/X86/X86RegisterInfo.td</tt> for an example of this.
1665 <p>Virtual registers are also denoted by integer numbers. Contrary to physical
1666 registers, different virtual registers never share the same number. Whereas
1667 physical registers are statically defined in a <tt>TargetRegisterInfo.td</tt>
1668 file and cannot be created by the application developer, that is not the case
1669 with virtual registers. In order to create new virtual registers, use the
1670 method <tt>MachineRegisterInfo::createVirtualRegister()</tt>. This method
1671 will return a new virtual register. Use an <tt>IndexedMap<Foo,
1672 VirtReg2IndexFunctor></tt> to hold information per virtual register. If you
1673 need to enumerate all virtual registers, use the function
1674 <tt>TargetRegisterInfo::index2VirtReg()</tt> to find the virtual register
1677 <div class="doc_code">
1679 for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
1680 unsigned VirtReg = TargetRegisterInfo::index2VirtReg(i);
1686 <p>Before register allocation, the operands of an instruction are mostly virtual
1687 registers, although physical registers may also be used. In order to check if
1688 a given machine operand is a register, use the boolean
1689 function <tt>MachineOperand::isRegister()</tt>. To obtain the integer code of
1690 a register, use <tt>MachineOperand::getReg()</tt>. An instruction may define
1691 or use a register. For instance, <tt>ADD reg:1026 := reg:1025 reg:1024</tt>
1692 defines the registers 1024, and uses registers 1025 and 1026. Given a
1693 register operand, the method <tt>MachineOperand::isUse()</tt> informs if that
1694 register is being used by the instruction. The
1695 method <tt>MachineOperand::isDef()</tt> informs if that registers is being
1698 <p>We will call physical registers present in the LLVM bitcode before register
1699 allocation <i>pre-colored registers</i>. Pre-colored registers are used in
1700 many different situations, for instance, to pass parameters of functions
1701 calls, and to store results of particular instructions. There are two types
1702 of pre-colored registers: the ones <i>implicitly</i> defined, and
1703 those <i>explicitly</i> defined. Explicitly defined registers are normal
1704 operands, and can be accessed
1705 with <tt>MachineInstr::getOperand(int)::getReg()</tt>. In order to check
1706 which registers are implicitly defined by an instruction, use
1707 the <tt>TargetInstrInfo::get(opcode)::ImplicitDefs</tt>,
1708 where <tt>opcode</tt> is the opcode of the target instruction. One important
1709 difference between explicit and implicit physical registers is that the
1710 latter are defined statically for each instruction, whereas the former may
1711 vary depending on the program being compiled. For example, an instruction
1712 that represents a function call will always implicitly define or use the same
1713 set of physical registers. To read the registers implicitly used by an
1715 use <tt>TargetInstrInfo::get(opcode)::ImplicitUses</tt>. Pre-colored
1716 registers impose constraints on any register allocation algorithm. The
1717 register allocator must make sure that none of them are overwritten by
1718 the values of virtual registers while still alive.</p>
1722 <!-- _______________________________________________________________________ -->
1725 <a name="regAlloc_howTo">Mapping virtual registers to physical registers</a>
1730 <p>There are two ways to map virtual registers to physical registers (or to
1731 memory slots). The first way, that we will call <i>direct mapping</i>, is
1732 based on the use of methods of the classes <tt>TargetRegisterInfo</tt>,
1733 and <tt>MachineOperand</tt>. The second way, that we will call <i>indirect
1734 mapping</i>, relies on the <tt>VirtRegMap</tt> class in order to insert loads
1735 and stores sending and getting values to and from memory.</p>
1737 <p>The direct mapping provides more flexibility to the developer of the register
1738 allocator; however, it is more error prone, and demands more implementation
1739 work. Basically, the programmer will have to specify where load and store
1740 instructions should be inserted in the target function being compiled in
1741 order to get and store values in memory. To assign a physical register to a
1742 virtual register present in a given operand,
1743 use <tt>MachineOperand::setReg(p_reg)</tt>. To insert a store instruction,
1744 use <tt>TargetInstrInfo::storeRegToStackSlot(...)</tt>, and to insert a
1745 load instruction, use <tt>TargetInstrInfo::loadRegFromStackSlot</tt>.</p>
1747 <p>The indirect mapping shields the application developer from the complexities
1748 of inserting load and store instructions. In order to map a virtual register
1749 to a physical one, use <tt>VirtRegMap::assignVirt2Phys(vreg, preg)</tt>. In
1750 order to map a certain virtual register to memory,
1751 use <tt>VirtRegMap::assignVirt2StackSlot(vreg)</tt>. This method will return
1752 the stack slot where <tt>vreg</tt>'s value will be located. If it is
1753 necessary to map another virtual register to the same stack slot,
1754 use <tt>VirtRegMap::assignVirt2StackSlot(vreg, stack_location)</tt>. One
1755 important point to consider when using the indirect mapping, is that even if
1756 a virtual register is mapped to memory, it still needs to be mapped to a
1757 physical register. This physical register is the location where the virtual
1758 register is supposed to be found before being stored or after being
1761 <p>If the indirect strategy is used, after all the virtual registers have been
1762 mapped to physical registers or stack slots, it is necessary to use a spiller
1763 object to place load and store instructions in the code. Every virtual that
1764 has been mapped to a stack slot will be stored to memory after been defined
1765 and will be loaded before being used. The implementation of the spiller tries
1766 to recycle load/store instructions, avoiding unnecessary instructions. For an
1767 example of how to invoke the spiller,
1768 see <tt>RegAllocLinearScan::runOnMachineFunction</tt>
1769 in <tt>lib/CodeGen/RegAllocLinearScan.cpp</tt>.</p>
1773 <!-- _______________________________________________________________________ -->
1775 <a name="regAlloc_twoAddr">Handling two address instructions</a>
1780 <p>With very rare exceptions (e.g., function calls), the LLVM machine code
1781 instructions are three address instructions. That is, each instruction is
1782 expected to define at most one register, and to use at most two registers.
1783 However, some architectures use two address instructions. In this case, the
1784 defined register is also one of the used register. For instance, an
1785 instruction such as <tt>ADD %EAX, %EBX</tt>, in X86 is actually equivalent
1786 to <tt>%EAX = %EAX + %EBX</tt>.</p>
1788 <p>In order to produce correct code, LLVM must convert three address
1789 instructions that represent two address instructions into true two address
1790 instructions. LLVM provides the pass <tt>TwoAddressInstructionPass</tt> for
1791 this specific purpose. It must be run before register allocation takes
1792 place. After its execution, the resulting code may no longer be in SSA
1793 form. This happens, for instance, in situations where an instruction such
1794 as <tt>%a = ADD %b %c</tt> is converted to two instructions such as:</p>
1796 <div class="doc_code">
1803 <p>Notice that, internally, the second instruction is represented as
1804 <tt>ADD %a[def/use] %c</tt>. I.e., the register operand <tt>%a</tt> is both
1805 used and defined by the instruction.</p>
1809 <!-- _______________________________________________________________________ -->
1811 <a name="regAlloc_ssaDecon">The SSA deconstruction phase</a>
1816 <p>An important transformation that happens during register allocation is called
1817 the <i>SSA Deconstruction Phase</i>. The SSA form simplifies many analyses
1818 that are performed on the control flow graph of programs. However,
1819 traditional instruction sets do not implement PHI instructions. Thus, in
1820 order to generate executable code, compilers must replace PHI instructions
1821 with other instructions that preserve their semantics.</p>
1823 <p>There are many ways in which PHI instructions can safely be removed from the
1824 target code. The most traditional PHI deconstruction algorithm replaces PHI
1825 instructions with copy instructions. That is the strategy adopted by
1826 LLVM. The SSA deconstruction algorithm is implemented
1827 in <tt>lib/CodeGen/PHIElimination.cpp</tt>. In order to invoke this pass, the
1828 identifier <tt>PHIEliminationID</tt> must be marked as required in the code
1829 of the register allocator.</p>
1833 <!-- _______________________________________________________________________ -->
1835 <a name="regAlloc_fold">Instruction folding</a>
1840 <p><i>Instruction folding</i> is an optimization performed during register
1841 allocation that removes unnecessary copy instructions. For instance, a
1842 sequence of instructions such as:</p>
1844 <div class="doc_code">
1846 %EBX = LOAD %mem_address
1851 <p>can be safely substituted by the single instruction:</p>
1853 <div class="doc_code">
1855 %EAX = LOAD %mem_address
1859 <p>Instructions can be folded with
1860 the <tt>TargetRegisterInfo::foldMemoryOperand(...)</tt> method. Care must be
1861 taken when folding instructions; a folded instruction can be quite different
1863 instruction. See <tt>LiveIntervals::addIntervalsForSpills</tt>
1864 in <tt>lib/CodeGen/LiveIntervalAnalysis.cpp</tt> for an example of its
1869 <!-- _______________________________________________________________________ -->
1872 <a name="regAlloc_builtIn">Built in register allocators</a>
1877 <p>The LLVM infrastructure provides the application developer with three
1878 different register allocators:</p>
1881 <li><i>Fast</i> — This register allocator is the default for debug
1882 builds. It allocates registers on a basic block level, attempting to keep
1883 values in registers and reusing registers as appropriate.</li>
1885 <li><i>Basic</i> — This is an incremental approach to register
1886 allocation. Live ranges are assigned to registers one at a time in
1887 an order that is driven by heuristics. Since code can be rewritten
1888 on-the-fly during allocation, this framework allows interesting
1889 allocators to be developed as extensions. It is not itself a
1890 production register allocator but is a potentially useful
1891 stand-alone mode for triaging bugs and as a performance baseline.
1893 <li><i>Greedy</i> — <i>The default allocator</i>. This is a
1894 highly tuned implementation of the <i>Basic</i> allocator that
1895 incorporates global live range splitting. This allocator works hard
1896 to minimize the cost of spill code.
1898 <li><i>PBQP</i> — A Partitioned Boolean Quadratic Programming (PBQP)
1899 based register allocator. This allocator works by constructing a PBQP
1900 problem representing the register allocation problem under consideration,
1901 solving this using a PBQP solver, and mapping the solution back to a
1902 register assignment.</li>
1905 <p>The type of register allocator used in <tt>llc</tt> can be chosen with the
1906 command line option <tt>-regalloc=...</tt>:</p>
1908 <div class="doc_code">
1910 $ llc -regalloc=linearscan file.bc -o ln.s;
1911 $ llc -regalloc=fast file.bc -o fa.s;
1912 $ llc -regalloc=pbqp file.bc -o pbqp.s;
1920 <!-- ======================================================================= -->
1922 <a name="proepicode">Prolog/Epilog Code Insertion</a>
1927 <!-- _______________________________________________________________________ -->
1929 <a name="compact_unwind">Compact Unwind</a>
1934 <p>Throwing an exception requires <em>unwinding</em> out of a function. The
1935 information on how to unwind a given function is traditionally expressed in
1936 DWARF unwind (a.k.a. frame) info. But that format was originally developed
1937 for debuggers to backtrace, and each Frame Description Entry (FDE) requires
1938 ~20-30 bytes per function. There is also the cost of mapping from an address
1939 in a function to the corresponding FDE at runtime. An alternative unwind
1940 encoding is called <em>compact unwind</em> and requires just 4-bytes per
1943 <p>The compact unwind encoding is a 32-bit value, which is encoded in an
1944 architecture-specific way. It specifies which registers to restore and from
1945 where, and how to unwind out of the function. When the linker creates a final
1946 linked image, it will create a <code>__TEXT,__unwind_info</code>
1947 section. This section is a small and fast way for the runtime to access
1948 unwind info for any given function. If we emit compact unwind info for the
1949 function, that compact unwind info will be encoded in
1950 the <code>__TEXT,__unwind_info</code> section. If we emit DWARF unwind info,
1951 the <code>__TEXT,__unwind_info</code> section will contain the offset of the
1952 FDE in the <code>__TEXT,__eh_frame</code> section in the final linked
1955 <p>For X86, there are three modes for the compact unwind encoding:</p>
1958 <dt><i>Function with a Frame Pointer (<code>EBP</code> or <code>RBP</code>)</i></dt>
1959 <dd><p><code>EBP/RBP</code>-based frame, where <code>EBP/RBP</code> is pushed
1960 onto the stack immediately after the return address,
1961 then <code>ESP/RSP</code> is moved to <code>EBP/RBP</code>. Thus to
1962 unwind, <code>ESP/RSP</code> is restored with the
1963 current <code>EBP/RBP</code> value, then <code>EBP/RBP</code> is restored
1964 by popping the stack, and the return is done by popping the stack once
1965 more into the PC. All non-volatile registers that need to be restored must
1966 have been saved in a small range on the stack that
1967 starts <code>EBP-4</code> to <code>EBP-1020</code> (<code>RBP-8</code>
1968 to <code>RBP-1020</code>). The offset (divided by 4 in 32-bit mode and 8
1969 in 64-bit mode) is encoded in bits 16-23 (mask: <code>0x00FF0000</code>).
1970 The registers saved are encoded in bits 0-14
1971 (mask: <code>0x00007FFF</code>) as five 3-bit entries from the following
1973 <table border="1" cellspacing="0">
1975 <th>Compact Number</th>
1976 <th>i386 Register</th>
1977 <th>x86-64 Regiser</th>
1981 <td><code>EBX</code></td>
1982 <td><code>RBX</code></td>
1986 <td><code>ECX</code></td>
1987 <td><code>R12</code></td>
1991 <td><code>EDX</code></td>
1992 <td><code>R13</code></td>
1996 <td><code>EDI</code></td>
1997 <td><code>R14</code></td>
2001 <td><code>ESI</code></td>
2002 <td><code>R15</code></td>
2006 <td><code>EBP</code></td>
2007 <td><code>RBP</code></td>
2013 <dt><i>Frameless with a Small Constant Stack Size (<code>EBP</code>
2014 or <code>RBP</code> is not used as a frame pointer)</i></dt>
2015 <dd><p>To return, a constant (encoded in the compact unwind encoding) is added
2016 to the <code>ESP/RSP</code>. Then the return is done by popping the stack
2017 into the PC. All non-volatile registers that need to be restored must have
2018 been saved on the stack immediately after the return address. The stack
2019 size (divided by 4 in 32-bit mode and 8 in 64-bit mode) is encoded in bits
2020 16-23 (mask: <code>0x00FF0000</code>). There is a maximum stack size of
2021 1024 bytes in 32-bit mode and 2048 in 64-bit mode. The number of registers
2022 saved is encoded in bits 9-12 (mask: <code>0x00001C00</code>). Bits 0-9
2023 (mask: <code>0x000003FF</code>) contain which registers were saved and
2025 the <code>encodeCompactUnwindRegistersWithoutFrame()</code> function
2026 in <code>lib/Target/X86FrameLowering.cpp</code> for the encoding
2027 algorithm.)</p></dd>
2029 <dt><i>Frameless with a Large Constant Stack Size (<code>EBP</code>
2030 or <code>RBP</code> is not used as a frame pointer)</i></dt>
2031 <dd><p>This case is like the "Frameless with a Small Constant Stack Size"
2032 case, but the stack size is too large to encode in the compact unwind
2033 encoding. Instead it requires that the function contains "<code>subl
2034 $nnnnnn, %esp</code>" in its prolog. The compact encoding contains the
2035 offset to the <code>$nnnnnn</code> value in the function in bits 9-12
2036 (mask: <code>0x00001C00</code>).</p></dd>
2043 <!-- ======================================================================= -->
2045 <a name="latemco">Late Machine Code Optimizations</a>
2047 <div><p>To Be Written</p></div>
2049 <!-- ======================================================================= -->
2051 <a name="codeemit">Code Emission</a>
2056 <p>The code emission step of code generation is responsible for lowering from
2057 the code generator abstractions (like <a
2058 href="#machinefunction">MachineFunction</a>, <a
2059 href="#machineinstr">MachineInstr</a>, etc) down
2060 to the abstractions used by the MC layer (<a href="#mcinst">MCInst</a>,
2061 <a href="#mcstreamer">MCStreamer</a>, etc). This is
2062 done with a combination of several different classes: the (misnamed)
2063 target-independent AsmPrinter class, target-specific subclasses of AsmPrinter
2064 (such as SparcAsmPrinter), and the TargetLoweringObjectFile class.</p>
2066 <p>Since the MC layer works at the level of abstraction of object files, it
2067 doesn't have a notion of functions, global variables etc. Instead, it thinks
2068 about labels, directives, and instructions. A key class used at this time is
2069 the MCStreamer class. This is an abstract API that is implemented in different
2070 ways (e.g. to output a .s file, output an ELF .o file, etc) that is effectively
2071 an "assembler API". MCStreamer has one method per directive, such as EmitLabel,
2072 EmitSymbolAttribute, SwitchSection, etc, which directly correspond to assembly
2076 <p>If you are interested in implementing a code generator for a target, there
2077 are three important things that you have to implement for your target:</p>
2080 <li>First, you need a subclass of AsmPrinter for your target. This class
2081 implements the general lowering process converting MachineFunction's into MC
2082 label constructs. The AsmPrinter base class provides a number of useful methods
2083 and routines, and also allows you to override the lowering process in some
2084 important ways. You should get much of the lowering for free if you are
2085 implementing an ELF, COFF, or MachO target, because the TargetLoweringObjectFile
2086 class implements much of the common logic.</li>
2088 <li>Second, you need to implement an instruction printer for your target. The
2089 instruction printer takes an <a href="#mcinst">MCInst</a> and renders it to a
2090 raw_ostream as text. Most of this is automatically generated from the .td file
2091 (when you specify something like "<tt>add $dst, $src1, $src2</tt>" in the
2092 instructions), but you need to implement routines to print operands.</li>
2094 <li>Third, you need to implement code that lowers a <a
2095 href="#machineinstr">MachineInstr</a> to an MCInst, usually implemented in
2096 "<target>MCInstLower.cpp". This lowering process is often target
2097 specific, and is responsible for turning jump table entries, constant pool
2098 indices, global variable addresses, etc into MCLabels as appropriate. This
2099 translation layer is also responsible for expanding pseudo ops used by the code
2100 generator into the actual machine instructions they correspond to. The MCInsts
2101 that are generated by this are fed into the instruction printer or the encoder.
2106 <p>Finally, at your choosing, you can also implement an subclass of
2107 MCCodeEmitter which lowers MCInst's into machine code bytes and relocations.
2108 This is important if you want to support direct .o file emission, or would like
2109 to implement an assembler for your target.</p>
2113 <!-- ======================================================================= -->
2115 <a name="vliw_packetizer">VLIW Packetizer</a>
2120 <p>In a Very Long Instruction Word (VLIW) architecture, the compiler is
2121 responsible for mapping instructions to functional-units available on
2122 the architecture. To that end, the compiler creates groups of instructions
2123 called <i>packets</i> or <i>bundles</i>. The VLIW packetizer in LLVM is
2124 a target-independent mechanism to enable the packetization of machine
2127 <!-- _______________________________________________________________________ -->
2130 <a name="vliw_mapping">Mapping from instructions to functional units</a>
2135 <p>Instructions in a VLIW target can typically be mapped to multiple functional
2136 units. During the process of packetizing, the compiler must be able to reason
2137 about whether an instruction can be added to a packet. This decision can be
2138 complex since the compiler has to examine all possible mappings of instructions
2139 to functional units. Therefore to alleviate compilation-time complexity, the
2140 VLIW packetizer parses the instruction classes of a target and generates tables
2141 at compiler build time. These tables can then be queried by the provided
2142 machine-independent API to determine if an instruction can be accommodated in a
2146 <!-- ======================================================================= -->
2148 <a name="vliw_repr">
2149 How the packetization tables are generated and used
2155 <p>The packetizer reads instruction classes from a target's itineraries and
2156 creates a deterministic finite automaton (DFA) to represent the state of a
2157 packet. A DFA consists of three major elements: inputs, states, and
2158 transitions. The set of inputs for the generated DFA represents the instruction
2159 being added to a packet. The states represent the possible consumption
2160 of functional units by instructions in a packet. In the DFA, transitions from
2161 one state to another occur on the addition of an instruction to an existing
2162 packet. If there is a legal mapping of functional units to instructions, then
2163 the DFA contains a corresponding transition. The absence of a transition
2164 indicates that a legal mapping does not exist and that the instruction cannot
2165 be added to the packet.</p>
2167 <p>To generate tables for a VLIW target, add <i>Target</i>GenDFAPacketizer.inc
2168 as a target to the Makefile in the target directory. The exported API provides
2169 three functions: <tt>DFAPacketizer::clearResources()</tt>,
2170 <tt>DFAPacketizer::reserveResources(MachineInstr *MI)</tt>, and
2171 <tt>DFAPacketizer::canReserveResources(MachineInstr *MI)</tt>. These functions
2172 allow a target packetizer to add an instruction to an existing packet and to
2173 check whether an instruction can be added to a packet. See
2174 <tt>llvm/CodeGen/DFAPacketizer.h</tt> for more information.</p>
2182 <!-- *********************************************************************** -->
2184 <a name="nativeassembler">Implementing a Native Assembler</a>
2186 <!-- *********************************************************************** -->
2190 <p>Though you're probably reading this because you want to write or maintain a
2191 compiler backend, LLVM also fully supports building a native assemblers too.
2192 We've tried hard to automate the generation of the assembler from the .td files
2193 (in particular the instruction syntax and encodings), which means that a large
2194 part of the manual and repetitive data entry can be factored and shared with the
2197 <!-- ======================================================================= -->
2198 <h3 id="na_instparsing">Instruction Parsing</h3>
2200 <div><p>To Be Written</p></div>
2203 <!-- ======================================================================= -->
2204 <h3 id="na_instaliases">
2205 Instruction Alias Processing
2209 <p>Once the instruction is parsed, it enters the MatchInstructionImpl function.
2210 The MatchInstructionImpl function performs alias processing and then does
2211 actual matching.</p>
2213 <p>Alias processing is the phase that canonicalizes different lexical forms of
2214 the same instructions down to one representation. There are several different
2215 kinds of alias that are possible to implement and they are listed below in the
2216 order that they are processed (which is in order from simplest/weakest to most
2217 complex/powerful). Generally you want to use the first alias mechanism that
2218 meets the needs of your instruction, because it will allow a more concise
2221 <!-- _______________________________________________________________________ -->
2222 <h4>Mnemonic Aliases</h4>
2226 <p>The first phase of alias processing is simple instruction mnemonic
2227 remapping for classes of instructions which are allowed with two different
2228 mnemonics. This phase is a simple and unconditionally remapping from one input
2229 mnemonic to one output mnemonic. It isn't possible for this form of alias to
2230 look at the operands at all, so the remapping must apply for all forms of a
2231 given mnemonic. Mnemonic aliases are defined simply, for example X86 has:
2234 <div class="doc_code">
2236 def : MnemonicAlias<"cbw", "cbtw">;
2237 def : MnemonicAlias<"smovq", "movsq">;
2238 def : MnemonicAlias<"fldcww", "fldcw">;
2239 def : MnemonicAlias<"fucompi", "fucomip">;
2240 def : MnemonicAlias<"ud2a", "ud2">;
2244 <p>... and many others. With a MnemonicAlias definition, the mnemonic is
2245 remapped simply and directly. Though MnemonicAlias's can't look at any aspect
2246 of the instruction (such as the operands) they can depend on global modes (the
2247 same ones supported by the matcher), through a Requires clause:</p>
2249 <div class="doc_code">
2251 def : MnemonicAlias<"pushf", "pushfq">, Requires<[In64BitMode]>;
2252 def : MnemonicAlias<"pushf", "pushfl">, Requires<[In32BitMode]>;
2256 <p>In this example, the mnemonic gets mapped into different a new one depending
2257 on the current instruction set.</p>
2261 <!-- _______________________________________________________________________ -->
2262 <h4>Instruction Aliases</h4>
2266 <p>The most general phase of alias processing occurs while matching is
2267 happening: it provides new forms for the matcher to match along with a specific
2268 instruction to generate. An instruction alias has two parts: the string to
2269 match and the instruction to generate. For example:
2272 <div class="doc_code">
2274 def : InstAlias<"movsx $src, $dst", (MOVSX16rr8W GR16:$dst, GR8 :$src)>;
2275 def : InstAlias<"movsx $src, $dst", (MOVSX16rm8W GR16:$dst, i8mem:$src)>;
2276 def : InstAlias<"movsx $src, $dst", (MOVSX32rr8 GR32:$dst, GR8 :$src)>;
2277 def : InstAlias<"movsx $src, $dst", (MOVSX32rr16 GR32:$dst, GR16 :$src)>;
2278 def : InstAlias<"movsx $src, $dst", (MOVSX64rr8 GR64:$dst, GR8 :$src)>;
2279 def : InstAlias<"movsx $src, $dst", (MOVSX64rr16 GR64:$dst, GR16 :$src)>;
2280 def : InstAlias<"movsx $src, $dst", (MOVSX64rr32 GR64:$dst, GR32 :$src)>;
2284 <p>This shows a powerful example of the instruction aliases, matching the
2285 same mnemonic in multiple different ways depending on what operands are present
2286 in the assembly. The result of instruction aliases can include operands in a
2287 different order than the destination instruction, and can use an input
2288 multiple times, for example:</p>
2290 <div class="doc_code">
2292 def : InstAlias<"clrb $reg", (XOR8rr GR8 :$reg, GR8 :$reg)>;
2293 def : InstAlias<"clrw $reg", (XOR16rr GR16:$reg, GR16:$reg)>;
2294 def : InstAlias<"clrl $reg", (XOR32rr GR32:$reg, GR32:$reg)>;
2295 def : InstAlias<"clrq $reg", (XOR64rr GR64:$reg, GR64:$reg)>;
2299 <p>This example also shows that tied operands are only listed once. In the X86
2300 backend, XOR8rr has two input GR8's and one output GR8 (where an input is tied
2301 to the output). InstAliases take a flattened operand list without duplicates
2302 for tied operands. The result of an instruction alias can also use immediates
2303 and fixed physical registers which are added as simple immediate operands in the
2304 result, for example:</p>
2306 <div class="doc_code">
2308 // Fixed Immediate operand.
2309 def : InstAlias<"aad", (AAD8i8 10)>;
2311 // Fixed register operand.
2312 def : InstAlias<"fcomi", (COM_FIr ST1)>;
2315 def : InstAlias<"fcomi $reg", (COM_FIr RST:$reg)>;
2320 <p>Instruction aliases can also have a Requires clause to make them
2321 subtarget specific.</p>
2323 <p>If the back-end supports it, the instruction printer can automatically emit
2324 the alias rather than what's being aliased. It typically leads to better,
2325 more readable code. If it's better to print out what's being aliased, then
2326 pass a '0' as the third parameter to the InstAlias definition.</p>
2332 <!-- ======================================================================= -->
2333 <h3 id="na_matching">Instruction Matching</h3>
2335 <div><p>To Be Written</p></div>
2339 <!-- *********************************************************************** -->
2341 <a name="targetimpls">Target-specific Implementation Notes</a>
2343 <!-- *********************************************************************** -->
2347 <p>This section of the document explains features or design decisions that are
2348 specific to the code generator for a particular target. First we start
2349 with a table that summarizes what features are supported by each target.</p>
2351 <!-- ======================================================================= -->
2353 <a name="targetfeatures">Target Feature Matrix</a>
2358 <p>Note that this table does not include the C backend or Cpp backends, since
2359 they do not use the target independent code generator infrastructure. It also
2360 doesn't list features that are not supported fully by any target yet. It
2361 considers a feature to be supported if at least one subtarget supports it. A
2362 feature being supported means that it is useful and works for most cases, it
2363 does not indicate that there are zero known bugs in the implementation. Here
2367 <table border="1" cellspacing="0">
2371 <th>Partial Support</th>
2372 <th>Complete Support</th>
2375 <td class="unknown"></td>
2376 <td class="no"></td>
2377 <td class="partial"></td>
2378 <td class="yes"></td>
2382 <p>Here is the table:</p>
2384 <table width="689" border="1" cellspacing="0">
2386 <td colspan="13" align="center" style="background-color:#ffc">Target</td>
2404 <td><a href="#feat_reliable">is generally reliable</a></td>
2405 <td class="yes"></td> <!-- ARM -->
2406 <td class="no"></td> <!-- CellSPU -->
2407 <td class="yes"></td> <!-- Hexagon -->
2408 <td class="no"></td> <!-- MBlaze -->
2409 <td class="unknown"></td> <!-- MSP430 -->
2410 <td class="yes"></td> <!-- Mips -->
2411 <td class="no"></td> <!-- PTX -->
2412 <td class="yes"></td> <!-- PowerPC -->
2413 <td class="yes"></td> <!-- Sparc -->
2414 <td class="yes"></td> <!-- X86 -->
2415 <td class="unknown"></td> <!-- XCore -->
2419 <td><a href="#feat_asmparser">assembly parser</a></td>
2420 <td class="no"></td> <!-- ARM -->
2421 <td class="no"></td> <!-- CellSPU -->
2422 <td class="no"></td> <!-- Hexagon -->
2423 <td class="yes"></td> <!-- MBlaze -->
2424 <td class="no"></td> <!-- MSP430 -->
2425 <td class="no"></td> <!-- Mips -->
2426 <td class="no"></td> <!-- PTX -->
2427 <td class="no"></td> <!-- PowerPC -->
2428 <td class="no"></td> <!-- Sparc -->
2429 <td class="yes"></td> <!-- X86 -->
2430 <td class="no"></td> <!-- XCore -->
2434 <td><a href="#feat_disassembler">disassembler</a></td>
2435 <td class="yes"></td> <!-- ARM -->
2436 <td class="no"></td> <!-- CellSPU -->
2437 <td class="no"></td> <!-- Hexagon -->
2438 <td class="yes"></td> <!-- MBlaze -->
2439 <td class="no"></td> <!-- MSP430 -->
2440 <td class="no"></td> <!-- Mips -->
2441 <td class="no"></td> <!-- PTX -->
2442 <td class="no"></td> <!-- PowerPC -->
2443 <td class="no"></td> <!-- Sparc -->
2444 <td class="yes"></td> <!-- X86 -->
2445 <td class="no"></td> <!-- XCore -->
2449 <td><a href="#feat_inlineasm">inline asm</a></td>
2450 <td class="yes"></td> <!-- ARM -->
2451 <td class="no"></td> <!-- CellSPU -->
2452 <td class="yes"></td> <!-- Hexagon -->
2453 <td class="yes"></td> <!-- MBlaze -->
2454 <td class="unknown"></td> <!-- MSP430 -->
2455 <td class="no"></td> <!-- Mips -->
2456 <td class="unknown"></td> <!-- PTX -->
2457 <td class="yes"></td> <!-- PowerPC -->
2458 <td class="unknown"></td> <!-- Sparc -->
2459 <td class="yes"></td> <!-- X86 -->
2460 <td class="unknown"></td> <!-- XCore -->
2464 <td><a href="#feat_jit">jit</a></td>
2465 <td class="partial"><a href="#feat_jit_arm">*</a></td> <!-- ARM -->
2466 <td class="no"></td> <!-- CellSPU -->
2467 <td class="no"></td> <!-- Hexagon -->
2468 <td class="no"></td> <!-- MBlaze -->
2469 <td class="unknown"></td> <!-- MSP430 -->
2470 <td class="yes"></td> <!-- Mips -->
2471 <td class="unknown"></td> <!-- PTX -->
2472 <td class="yes"></td> <!-- PowerPC -->
2473 <td class="unknown"></td> <!-- Sparc -->
2474 <td class="yes"></td> <!-- X86 -->
2475 <td class="unknown"></td> <!-- XCore -->
2479 <td><a href="#feat_objectwrite">.o file writing</a></td>
2480 <td class="no"></td> <!-- ARM -->
2481 <td class="no"></td> <!-- CellSPU -->
2482 <td class="no"></td> <!-- Hexagon -->
2483 <td class="yes"></td> <!-- MBlaze -->
2484 <td class="no"></td> <!-- MSP430 -->
2485 <td class="no"></td> <!-- Mips -->
2486 <td class="no"></td> <!-- PTX -->
2487 <td class="no"></td> <!-- PowerPC -->
2488 <td class="no"></td> <!-- Sparc -->
2489 <td class="yes"></td> <!-- X86 -->
2490 <td class="no"></td> <!-- XCore -->
2494 <td><a href="#feat_tailcall">tail calls</a></td>
2495 <td class="yes"></td> <!-- ARM -->
2496 <td class="no"></td> <!-- CellSPU -->
2497 <td class="yes"></td> <!-- Hexagon -->
2498 <td class="no"></td> <!-- MBlaze -->
2499 <td class="unknown"></td> <!-- MSP430 -->
2500 <td class="no"></td> <!-- Mips -->
2501 <td class="unknown"></td> <!-- PTX -->
2502 <td class="yes"></td> <!-- PowerPC -->
2503 <td class="unknown"></td> <!-- Sparc -->
2504 <td class="yes"></td> <!-- X86 -->
2505 <td class="unknown"></td> <!-- XCore -->
2509 <td><a href="#feat_segstacks">segmented stacks</a></td>
2510 <td class="no"></td> <!-- ARM -->
2511 <td class="no"></td> <!-- CellSPU -->
2512 <td class="no"></td> <!-- Hexagon -->
2513 <td class="no"></td> <!-- MBlaze -->
2514 <td class="no"></td> <!-- MSP430 -->
2515 <td class="no"></td> <!-- Mips -->
2516 <td class="no"></td> <!-- PTX -->
2517 <td class="no"></td> <!-- PowerPC -->
2518 <td class="no"></td> <!-- Sparc -->
2519 <td class="partial"><a href="#feat_segstacks_x86">*</a></td> <!-- X86 -->
2520 <td class="no"></td> <!-- XCore -->
2526 <!-- _______________________________________________________________________ -->
2527 <h4 id="feat_reliable">Is Generally Reliable</h4>
2530 <p>This box indicates whether the target is considered to be production quality.
2531 This indicates that the target has been used as a static compiler to
2532 compile large amounts of code by a variety of different people and is in
2536 <!-- _______________________________________________________________________ -->
2537 <h4 id="feat_asmparser">Assembly Parser</h4>
2540 <p>This box indicates whether the target supports parsing target specific .s
2541 files by implementing the MCAsmParser interface. This is required for llvm-mc
2542 to be able to act as a native assembler and is required for inline assembly
2543 support in the native .o file writer.</p>
2548 <!-- _______________________________________________________________________ -->
2549 <h4 id="feat_disassembler">Disassembler</h4>
2552 <p>This box indicates whether the target supports the MCDisassembler API for
2553 disassembling machine opcode bytes into MCInst's.</p>
2557 <!-- _______________________________________________________________________ -->
2558 <h4 id="feat_inlineasm">Inline Asm</h4>
2561 <p>This box indicates whether the target supports most popular inline assembly
2562 constraints and modifiers.</p>
2566 <!-- _______________________________________________________________________ -->
2567 <h4 id="feat_jit">JIT Support</h4>
2570 <p>This box indicates whether the target supports the JIT compiler through
2571 the ExecutionEngine interface.</p>
2573 <p id="feat_jit_arm">The ARM backend has basic support for integer code
2574 in ARM codegen mode, but lacks NEON and full Thumb support.</p>
2578 <!-- _______________________________________________________________________ -->
2579 <h4 id="feat_objectwrite">.o File Writing</h4>
2583 <p>This box indicates whether the target supports writing .o files (e.g. MachO,
2584 ELF, and/or COFF) files directly from the target. Note that the target also
2585 must include an assembly parser and general inline assembly support for full
2586 inline assembly support in the .o writer.</p>
2588 <p>Targets that don't support this feature can obviously still write out .o
2589 files, they just rely on having an external assembler to translate from a .s
2590 file to a .o file (as is the case for many C compilers).</p>
2594 <!-- _______________________________________________________________________ -->
2595 <h4 id="feat_tailcall">Tail Calls</h4>
2599 <p>This box indicates whether the target supports guaranteed tail calls. These
2600 are calls marked "<a href="LangRef.html#i_call">tail</a>" and use the fastcc
2601 calling convention. Please see the <a href="#tailcallopt">tail call section
2602 more more details</a>.</p>
2606 <!-- _______________________________________________________________________ -->
2607 <h4 id="feat_segstacks">Segmented Stacks</h4>
2611 <p>This box indicates whether the target supports segmented stacks. This
2612 replaces the traditional large C stack with many linked segments. It
2613 is compatible with the <a href="http://gcc.gnu.org/wiki/SplitStacks">gcc
2614 implementation</a> used by the Go front end.</p>
2616 <p id="feat_segstacks_x86">Basic support exists on the X86 backend. Currently
2617 vararg doesn't work and the object files are not marked the way the gold
2618 linker expects, but simple Go programs can be built by dragonegg.</p>
2624 <!-- ======================================================================= -->
2626 <a name="tailcallopt">Tail call optimization</a>
2631 <p>Tail call optimization, callee reusing the stack of the caller, is currently
2632 supported on x86/x86-64 and PowerPC. It is performed if:</p>
2635 <li>Caller and callee have the calling convention <tt>fastcc</tt> or
2636 <tt>cc 10</tt> (GHC call convention).</li>
2638 <li>The call is a tail call - in tail position (ret immediately follows call
2639 and ret uses value of call or is void).</li>
2641 <li>Option <tt>-tailcallopt</tt> is enabled.</li>
2643 <li>Platform specific constraints are met.</li>
2646 <p>x86/x86-64 constraints:</p>
2649 <li>No variable argument lists are used.</li>
2651 <li>On x86-64 when generating GOT/PIC code only module-local calls (visibility
2652 = hidden or protected) are supported.</li>
2655 <p>PowerPC constraints:</p>
2658 <li>No variable argument lists are used.</li>
2660 <li>No byval parameters are used.</li>
2662 <li>On ppc32/64 GOT/PIC only module-local calls (visibility = hidden or protected) are supported.</li>
2667 <p>Call as <tt>llc -tailcallopt test.ll</tt>.</p>
2669 <div class="doc_code">
2671 declare fastcc i32 @tailcallee(i32 inreg %a1, i32 inreg %a2, i32 %a3, i32 %a4)
2673 define fastcc i32 @tailcaller(i32 %in1, i32 %in2) {
2674 %l1 = add i32 %in1, %in2
2675 %tmp = tail call fastcc i32 @tailcallee(i32 %in1 inreg, i32 %in2 inreg, i32 %in1, i32 %l1)
2681 <p>Implications of <tt>-tailcallopt</tt>:</p>
2683 <p>To support tail call optimization in situations where the callee has more
2684 arguments than the caller a 'callee pops arguments' convention is used. This
2685 currently causes each <tt>fastcc</tt> call that is not tail call optimized
2686 (because one or more of above constraints are not met) to be followed by a
2687 readjustment of the stack. So performance might be worse in such cases.</p>
2690 <!-- ======================================================================= -->
2692 <a name="sibcallopt">Sibling call optimization</a>
2697 <p>Sibling call optimization is a restricted form of tail call optimization.
2698 Unlike tail call optimization described in the previous section, it can be
2699 performed automatically on any tail calls when <tt>-tailcallopt</tt> option
2700 is not specified.</p>
2702 <p>Sibling call optimization is currently performed on x86/x86-64 when the
2703 following constraints are met:</p>
2706 <li>Caller and callee have the same calling convention. It can be either
2707 <tt>c</tt> or <tt>fastcc</tt>.
2709 <li>The call is a tail call - in tail position (ret immediately follows call
2710 and ret uses value of call or is void).</li>
2712 <li>Caller and callee have matching return type or the callee result is not
2715 <li>If any of the callee arguments are being passed in stack, they must be
2716 available in caller's own incoming argument stack and the frame offsets
2721 <div class="doc_code">
2723 declare i32 @bar(i32, i32)
2725 define i32 @foo(i32 %a, i32 %b, i32 %c) {
2727 %0 = tail call i32 @bar(i32 %a, i32 %b)
2734 <!-- ======================================================================= -->
2736 <a name="x86">The X86 backend</a>
2741 <p>The X86 code generator lives in the <tt>lib/Target/X86</tt> directory. This
2742 code generator is capable of targeting a variety of x86-32 and x86-64
2743 processors, and includes support for ISA extensions such as MMX and SSE.</p>
2745 <!-- _______________________________________________________________________ -->
2747 <a name="x86_tt">X86 Target Triples supported</a>
2752 <p>The following are the known target triples that are supported by the X86
2753 backend. This is not an exhaustive list, and it would be useful to add those
2754 that people test.</p>
2757 <li><b>i686-pc-linux-gnu</b> — Linux</li>
2759 <li><b>i386-unknown-freebsd5.3</b> — FreeBSD 5.3</li>
2761 <li><b>i686-pc-cygwin</b> — Cygwin on Win32</li>
2763 <li><b>i686-pc-mingw32</b> — MingW on Win32</li>
2765 <li><b>i386-pc-mingw32msvc</b> — MingW crosscompiler on Linux</li>
2767 <li><b>i686-apple-darwin*</b> — Apple Darwin on X86</li>
2769 <li><b>x86_64-unknown-linux-gnu</b> — Linux</li>
2774 <!-- _______________________________________________________________________ -->
2776 <a name="x86_cc">X86 Calling Conventions supported</a>
2782 <p>The following target-specific calling conventions are known to backend:</p>
2785 <li><b>x86_StdCall</b> — stdcall calling convention seen on Microsoft
2786 Windows platform (CC ID = 64).</li>
2787 <li><b>x86_FastCall</b> — fastcall calling convention seen on Microsoft
2788 Windows platform (CC ID = 65).</li>
2789 <li><b>x86_ThisCall</b> — Similar to X86_StdCall. Passes first argument
2790 in ECX, others via stack. Callee is responsible for stack cleaning. This
2791 convention is used by MSVC by default for methods in its ABI
2797 <!-- _______________________________________________________________________ -->
2799 <a name="x86_memory">Representing X86 addressing modes in MachineInstrs</a>
2804 <p>The x86 has a very flexible way of accessing memory. It is capable of
2805 forming memory addresses of the following expression directly in integer
2806 instructions (which use ModR/M addressing):</p>
2808 <div class="doc_code">
2810 SegmentReg: Base + [1,2,4,8] * IndexReg + Disp32
2814 <p>In order to represent this, LLVM tracks no less than 5 operands for each
2815 memory operand of this form. This means that the "load" form of
2816 '<tt>mov</tt>' has the following <tt>MachineOperand</tt>s in this order:</p>
2818 <div class="doc_code">
2820 Index: 0 | 1 2 3 4 5
2821 Meaning: DestReg, | BaseReg, Scale, IndexReg, Displacement Segment
2822 OperandTy: VirtReg, | VirtReg, UnsImm, VirtReg, SignExtImm PhysReg
2826 <p>Stores, and all other instructions, treat the four memory operands in the
2827 same way and in the same order. If the segment register is unspecified
2828 (regno = 0), then no segment override is generated. "Lea" operations do not
2829 have a segment register specified, so they only have 4 operands for their
2830 memory reference.</p>
2834 <!-- _______________________________________________________________________ -->
2836 <a name="x86_memory">X86 address spaces supported</a>
2841 <p>x86 has a feature which provides
2842 the ability to perform loads and stores to different address spaces
2843 via the x86 segment registers. A segment override prefix byte on an
2844 instruction causes the instruction's memory access to go to the specified
2845 segment. LLVM address space 0 is the default address space, which includes
2846 the stack, and any unqualified memory accesses in a program. Address spaces
2847 1-255 are currently reserved for user-defined code. The GS-segment is
2848 represented by address space 256, while the FS-segment is represented by
2849 address space 257. Other x86 segments have yet to be allocated address space
2852 <p>While these address spaces may seem similar to TLS via the
2853 <tt>thread_local</tt> keyword, and often use the same underlying hardware,
2854 there are some fundamental differences.</p>
2856 <p>The <tt>thread_local</tt> keyword applies to global variables and
2857 specifies that they are to be allocated in thread-local memory. There are
2858 no type qualifiers involved, and these variables can be pointed to with
2859 normal pointers and accessed with normal loads and stores.
2860 The <tt>thread_local</tt> keyword is target-independent at the LLVM IR
2861 level (though LLVM doesn't yet have implementations of it for some
2864 <p>Special address spaces, in contrast, apply to static types. Every
2865 load and store has a particular address space in its address operand type,
2866 and this is what determines which address space is accessed.
2867 LLVM ignores these special address space qualifiers on global variables,
2868 and does not provide a way to directly allocate storage in them.
2869 At the LLVM IR level, the behavior of these special address spaces depends
2870 in part on the underlying OS or runtime environment, and they are specific
2871 to x86 (and LLVM doesn't yet handle them correctly in some cases).</p>
2873 <p>Some operating systems and runtime environments use (or may in the future
2874 use) the FS/GS-segment registers for various low-level purposes, so care
2875 should be taken when considering them.</p>
2879 <!-- _______________________________________________________________________ -->
2881 <a name="x86_names">Instruction naming</a>
2886 <p>An instruction name consists of the base name, a default operand size, and a
2887 a character per operand with an optional special size. For example:</p>
2889 <div class="doc_code">
2891 ADD8rr -> add, 8-bit register, 8-bit register
2892 IMUL16rmi -> imul, 16-bit register, 16-bit memory, 16-bit immediate
2893 IMUL16rmi8 -> imul, 16-bit register, 16-bit memory, 8-bit immediate
2894 MOVSX32rm16 -> movsx, 32-bit register, 16-bit memory
2902 <!-- ======================================================================= -->
2904 <a name="ppc">The PowerPC backend</a>
2909 <p>The PowerPC code generator lives in the lib/Target/PowerPC directory. The
2910 code generation is retargetable to several variations or <i>subtargets</i> of
2911 the PowerPC ISA; including ppc32, ppc64 and altivec.</p>
2913 <!-- _______________________________________________________________________ -->
2915 <a name="ppc_abi">LLVM PowerPC ABI</a>
2920 <p>LLVM follows the AIX PowerPC ABI, with two deviations. LLVM uses a PC
2921 relative (PIC) or static addressing for accessing global values, so no TOC
2922 (r2) is used. Second, r31 is used as a frame pointer to allow dynamic growth
2923 of a stack frame. LLVM takes advantage of having no TOC to provide space to
2924 save the frame pointer in the PowerPC linkage area of the caller frame.
2925 Other details of PowerPC ABI can be found at <a href=
2926 "http://developer.apple.com/documentation/DeveloperTools/Conceptual/LowLevelABI/Articles/32bitPowerPC.html"
2927 >PowerPC ABI.</a> Note: This link describes the 32 bit ABI. The 64 bit ABI
2928 is similar except space for GPRs are 8 bytes wide (not 4) and r13 is reserved
2933 <!-- _______________________________________________________________________ -->
2935 <a name="ppc_frame">Frame Layout</a>
2940 <p>The size of a PowerPC frame is usually fixed for the duration of a
2941 function's invocation. Since the frame is fixed size, all references
2942 into the frame can be accessed via fixed offsets from the stack pointer. The
2943 exception to this is when dynamic alloca or variable sized arrays are
2944 present, then a base pointer (r31) is used as a proxy for the stack pointer
2945 and stack pointer is free to grow or shrink. A base pointer is also used if
2946 llvm-gcc is not passed the -fomit-frame-pointer flag. The stack pointer is
2947 always aligned to 16 bytes, so that space allocated for altivec vectors will
2948 be properly aligned.</p>
2950 <p>An invocation frame is laid out as follows (low memory at top);</p>
2952 <table class="layout">
2954 <td>Linkage<br><br></td>
2957 <td>Parameter area<br><br></td>
2960 <td>Dynamic area<br><br></td>
2963 <td>Locals area<br><br></td>
2966 <td>Saved registers area<br><br></td>
2968 <tr style="border-style: none hidden none hidden;">
2972 <td>Previous Frame<br><br></td>
2976 <p>The <i>linkage</i> area is used by a callee to save special registers prior
2977 to allocating its own frame. Only three entries are relevant to LLVM. The
2978 first entry is the previous stack pointer (sp), aka link. This allows
2979 probing tools like gdb or exception handlers to quickly scan the frames in
2980 the stack. A function epilog can also use the link to pop the frame from the
2981 stack. The third entry in the linkage area is used to save the return
2982 address from the lr register. Finally, as mentioned above, the last entry is
2983 used to save the previous frame pointer (r31.) The entries in the linkage
2984 area are the size of a GPR, thus the linkage area is 24 bytes long in 32 bit
2985 mode and 48 bytes in 64 bit mode.</p>
2987 <p>32 bit linkage area</p>
2989 <table class="layout">
2992 <td>Saved SP (r1)</td>
3012 <td>Saved FP (r31)</td>
3016 <p>64 bit linkage area</p>
3018 <table class="layout">
3021 <td>Saved SP (r1)</td>
3041 <td>Saved FP (r31)</td>
3045 <p>The <i>parameter area</i> is used to store arguments being passed to a callee
3046 function. Following the PowerPC ABI, the first few arguments are actually
3047 passed in registers, with the space in the parameter area unused. However,
3048 if there are not enough registers or the callee is a thunk or vararg
3049 function, these register arguments can be spilled into the parameter area.
3050 Thus, the parameter area must be large enough to store all the parameters for
3051 the largest call sequence made by the caller. The size must also be
3052 minimally large enough to spill registers r3-r10. This allows callees blind
3053 to the call signature, such as thunks and vararg functions, enough space to
3054 cache the argument registers. Therefore, the parameter area is minimally 32
3055 bytes (64 bytes in 64 bit mode.) Also note that since the parameter area is
3056 a fixed offset from the top of the frame, that a callee can access its spilt
3057 arguments using fixed offsets from the stack pointer (or base pointer.)</p>
3059 <p>Combining the information about the linkage, parameter areas and alignment. A
3060 stack frame is minimally 64 bytes in 32 bit mode and 128 bytes in 64 bit
3063 <p>The <i>dynamic area</i> starts out as size zero. If a function uses dynamic
3064 alloca then space is added to the stack, the linkage and parameter areas are
3065 shifted to top of stack, and the new space is available immediately below the
3066 linkage and parameter areas. The cost of shifting the linkage and parameter
3067 areas is minor since only the link value needs to be copied. The link value
3068 can be easily fetched by adding the original frame size to the base pointer.
3069 Note that allocations in the dynamic space need to observe 16 byte
3072 <p>The <i>locals area</i> is where the llvm compiler reserves space for local
3075 <p>The <i>saved registers area</i> is where the llvm compiler spills callee
3076 saved registers on entry to the callee.</p>
3080 <!-- _______________________________________________________________________ -->
3082 <a name="ppc_prolog">Prolog/Epilog</a>
3087 <p>The llvm prolog and epilog are the same as described in the PowerPC ABI, with
3088 the following exceptions. Callee saved registers are spilled after the frame
3089 is created. This allows the llvm epilog/prolog support to be common with
3090 other targets. The base pointer callee saved register r31 is saved in the
3091 TOC slot of linkage area. This simplifies allocation of space for the base
3092 pointer and makes it convenient to locate programatically and during
3097 <!-- _______________________________________________________________________ -->
3099 <a name="ppc_dynamic">Dynamic Allocation</a>
3104 <p><i>TODO - More to come.</i></p>
3110 <!-- ======================================================================= -->
3112 <a name="ptx">The PTX backend</a>
3117 <p>The PTX code generator lives in the lib/Target/PTX directory. It is
3118 currently a work-in-progress, but already supports most of the code
3119 generation functionality needed to generate correct PTX kernels for
3122 <p>The code generator can target PTX 2.0+, and shader model 1.0+. The
3123 PTX ISA Reference Manual is used as the primary source of ISA
3124 information, though an effort is made to make the output of the code
3125 generator match the output of the NVidia nvcc compiler, whenever
3128 <p>Code Generator Options:</p>
3129 <table border="1" cellspacing="0">
3132 <th>Description</th>
3135 <td><code>double</code></td>
3136 <td align="left">If enabled, the map_f64_to_f32 directive is
3137 disabled in the PTX output, allowing native double-precision
3141 <td><code>no-fma</code></td>
3142 <td align="left">Disable generation of Fused-Multiply Add
3143 instructions, which may be beneficial for some devices</td>
3146 <td><code>smxy / computexy</code></td>
3147 <td align="left">Set shader model/compute capability to x.y,
3148 e.g. sm20 or compute13</td>
3154 <li>Arithmetic instruction selection (including combo FMA)</li>
3155 <li>Bitwise instruction selection</li>
3156 <li>Control-flow instruction selection</li>
3157 <li>Function calls (only on SM 2.0+ and no return arguments)</li>
3158 <li>Addresses spaces (0 = global, 1 = constant, 2 = local, 4 =
3160 <li>Thread synchronization (bar.sync)</li>
3161 <li>Special register reads ([N]TID, [N]CTAID, PMx, CLOCK, etc.)</li>
3166 <li>Robust call instruction selection</li>
3167 <li>Stack frame allocation</li>
3168 <li>Device-specific instruction scheduling optimizations</li>
3176 <!-- *********************************************************************** -->
3179 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
3180 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
3181 <a href="http://validator.w3.org/check/referer"><img
3182 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
3184 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
3185 <a href="http://llvm.org/">The LLVM Compiler Infrastructure</a><br>
3186 Last modified: $Date$