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="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. Another example of
222 a code generator like this is a (purely hypothetical) backend that converts
223 LLVM to the GCC RTL form and uses GCC to emit machine code for a target.</p>
225 <p>This design also implies that it is possible to design and implement
226 radically different code generators in the LLVM system that do not make use
227 of any of the built-in components. Doing so is not recommended at all, but
228 could be required for radically different targets that do not fit into the
229 LLVM machine description model: FPGAs for example.</p>
233 <!-- ======================================================================= -->
235 <a name="high-level-design">The high-level design of the code generator</a>
240 <p>The LLVM target-independent code generator is designed to support efficient
241 and quality code generation for standard register-based microprocessors.
242 Code generation in this model is divided into the following stages:</p>
245 <li><b><a href="#instselect">Instruction Selection</a></b> — This phase
246 determines an efficient way to express the input LLVM code in the target
247 instruction set. This stage produces the initial code for the program in
248 the target instruction set, then makes use of virtual registers in SSA
249 form and physical registers that represent any required register
250 assignments due to target constraints or calling conventions. This step
251 turns the LLVM code into a DAG of target instructions.</li>
253 <li><b><a href="#selectiondag_sched">Scheduling and Formation</a></b> —
254 This phase takes the DAG of target instructions produced by the
255 instruction selection phase, determines an ordering of the instructions,
256 then emits the instructions
257 as <tt><a href="#machineinstr">MachineInstr</a></tt>s with that ordering.
258 Note that we describe this in the <a href="#instselect">instruction
259 selection section</a> because it operates on
260 a <a href="#selectiondag_intro">SelectionDAG</a>.</li>
262 <li><b><a href="#ssamco">SSA-based Machine Code Optimizations</a></b> —
263 This optional stage consists of a series of machine-code optimizations
264 that operate on the SSA-form produced by the instruction selector.
265 Optimizations like modulo-scheduling or peephole optimization work
268 <li><b><a href="#regalloc">Register Allocation</a></b> — The target code
269 is transformed from an infinite virtual register file in SSA form to the
270 concrete register file used by the target. This phase introduces spill
271 code and eliminates all virtual register references from the program.</li>
273 <li><b><a href="#proepicode">Prolog/Epilog Code Insertion</a></b> — Once
274 the machine code has been generated for the function and the amount of
275 stack space required is known (used for LLVM alloca's and spill slots),
276 the prolog and epilog code for the function can be inserted and "abstract
277 stack location references" can be eliminated. This stage is responsible
278 for implementing optimizations like frame-pointer elimination and stack
281 <li><b><a href="#latemco">Late Machine Code Optimizations</a></b> —
282 Optimizations that operate on "final" machine code can go here, such as
283 spill code scheduling and peephole optimizations.</li>
285 <li><b><a href="#codeemit">Code Emission</a></b> — The final stage
286 actually puts out the code for the current function, either in the target
287 assembler format or in machine code.</li>
290 <p>The code generator is based on the assumption that the instruction selector
291 will use an optimal pattern matching selector to create high-quality
292 sequences of native instructions. Alternative code generator designs based
293 on pattern expansion and aggressive iterative peephole optimization are much
294 slower. This design permits efficient compilation (important for JIT
295 environments) and aggressive optimization (used when generating code offline)
296 by allowing components of varying levels of sophistication to be used for any
297 step of compilation.</p>
299 <p>In addition to these stages, target implementations can insert arbitrary
300 target-specific passes into the flow. For example, the X86 target uses a
301 special pass to handle the 80x87 floating point stack architecture. Other
302 targets with unusual requirements can be supported with custom passes as
307 <!-- ======================================================================= -->
309 <a name="tablegen">Using TableGen for target description</a>
314 <p>The target description classes require a detailed description of the target
315 architecture. These target descriptions often have a large amount of common
316 information (e.g., an <tt>add</tt> instruction is almost identical to a
317 <tt>sub</tt> instruction). In order to allow the maximum amount of
318 commonality to be factored out, the LLVM code generator uses
319 the <a href="TableGenFundamentals.html">TableGen</a> tool to describe big
320 chunks of the target machine, which allows the use of domain-specific and
321 target-specific abstractions to reduce the amount of repetition.</p>
323 <p>As LLVM continues to be developed and refined, we plan to move more and more
324 of the target description to the <tt>.td</tt> form. Doing so gives us a
325 number of advantages. The most important is that it makes it easier to port
326 LLVM because it reduces the amount of C++ code that has to be written, and
327 the surface area of the code generator that needs to be understood before
328 someone can get something working. Second, it makes it easier to change
329 things. In particular, if tables and other things are all emitted
330 by <tt>tblgen</tt>, we only need a change in one place (<tt>tblgen</tt>) to
331 update all of the targets to a new interface.</p>
337 <!-- *********************************************************************** -->
339 <a name="targetdesc">Target description classes</a>
341 <!-- *********************************************************************** -->
345 <p>The LLVM target description classes (located in the
346 <tt>include/llvm/Target</tt> directory) provide an abstract description of
347 the target machine independent of any particular client. These classes are
348 designed to capture the <i>abstract</i> properties of the target (such as the
349 instructions and registers it has), and do not incorporate any particular
350 pieces of code generation algorithms.</p>
352 <p>All of the target description classes (except the
353 <tt><a href="#targetdata">TargetData</a></tt> class) are designed to be
354 subclassed by the concrete target implementation, and have virtual methods
355 implemented. To get to these implementations, the
356 <tt><a href="#targetmachine">TargetMachine</a></tt> class provides accessors
357 that should be implemented by the target.</p>
359 <!-- ======================================================================= -->
361 <a name="targetmachine">The <tt>TargetMachine</tt> class</a>
366 <p>The <tt>TargetMachine</tt> class provides virtual methods that are used to
367 access the target-specific implementations of the various target description
368 classes via the <tt>get*Info</tt> methods (<tt>getInstrInfo</tt>,
369 <tt>getRegisterInfo</tt>, <tt>getFrameInfo</tt>, etc.). This class is
370 designed to be specialized by a concrete target implementation
371 (e.g., <tt>X86TargetMachine</tt>) which implements the various virtual
372 methods. The only required target description class is
373 the <a href="#targetdata"><tt>TargetData</tt></a> class, but if the code
374 generator components are to be used, the other interfaces should be
375 implemented as well.</p>
379 <!-- ======================================================================= -->
381 <a name="targetdata">The <tt>TargetData</tt> class</a>
386 <p>The <tt>TargetData</tt> class is the only required target description class,
387 and it is the only class that is not extensible (you cannot derived a new
388 class from it). <tt>TargetData</tt> specifies information about how the
389 target lays out memory for structures, the alignment requirements for various
390 data types, the size of pointers in the target, and whether the target is
391 little-endian or big-endian.</p>
395 <!-- ======================================================================= -->
397 <a name="targetlowering">The <tt>TargetLowering</tt> class</a>
402 <p>The <tt>TargetLowering</tt> class is used by SelectionDAG based instruction
403 selectors primarily to describe how LLVM code should be lowered to
404 SelectionDAG operations. Among other things, this class indicates:</p>
407 <li>an initial register class to use for various <tt>ValueType</tt>s,</li>
409 <li>which operations are natively supported by the target machine,</li>
411 <li>the return type of <tt>setcc</tt> operations,</li>
413 <li>the type to use for shift amounts, and</li>
415 <li>various high-level characteristics, like whether it is profitable to turn
416 division by a constant into a multiplication sequence</li>
421 <!-- ======================================================================= -->
423 <a name="targetregisterinfo">The <tt>TargetRegisterInfo</tt> class</a>
428 <p>The <tt>TargetRegisterInfo</tt> class is used to describe the register file
429 of the target and any interactions between the registers.</p>
431 <p>Registers in the code generator are represented in the code generator by
432 unsigned integers. Physical registers (those that actually exist in the
433 target description) are unique small numbers, and virtual registers are
434 generally large. Note that register #0 is reserved as a flag value.</p>
436 <p>Each register in the processor description has an associated
437 <tt>TargetRegisterDesc</tt> entry, which provides a textual name for the
438 register (used for assembly output and debugging dumps) and a set of aliases
439 (used to indicate whether one register overlaps with another).</p>
441 <p>In addition to the per-register description, the <tt>TargetRegisterInfo</tt>
442 class exposes a set of processor specific register classes (instances of the
443 <tt>TargetRegisterClass</tt> class). Each register class contains sets of
444 registers that have the same properties (for example, they are all 32-bit
445 integer registers). Each SSA virtual register created by the instruction
446 selector has an associated register class. When the register allocator runs,
447 it replaces virtual registers with a physical register in the set.</p>
449 <p>The target-specific implementations of these classes is auto-generated from
450 a <a href="TableGenFundamentals.html">TableGen</a> description of the
455 <!-- ======================================================================= -->
457 <a name="targetinstrinfo">The <tt>TargetInstrInfo</tt> class</a>
462 <p>The <tt>TargetInstrInfo</tt> class is used to describe the machine
463 instructions supported by the target. It is essentially an array of
464 <tt>TargetInstrDescriptor</tt> objects, each of which describes one
465 instruction the target supports. Descriptors define things like the mnemonic
466 for the opcode, the number of operands, the list of implicit register uses
467 and defs, whether the instruction has certain target-independent properties
468 (accesses memory, is commutable, etc), and holds any target-specific
473 <!-- ======================================================================= -->
475 <a name="targetframeinfo">The <tt>TargetFrameInfo</tt> class</a>
480 <p>The <tt>TargetFrameInfo</tt> class is used to provide information about the
481 stack frame layout of the target. It holds the direction of stack growth, the
482 known stack alignment on entry to each function, and the offset to the local
483 area. The offset to the local area is the offset from the stack pointer on
484 function entry to the first location where function data (local variables,
485 spill locations) can be stored.</p>
489 <!-- ======================================================================= -->
491 <a name="targetsubtarget">The <tt>TargetSubtarget</tt> class</a>
496 <p>The <tt>TargetSubtarget</tt> class is used to provide information about the
497 specific chip set being targeted. A sub-target informs code generation of
498 which instructions are supported, instruction latencies and instruction
499 execution itinerary; i.e., which processing units are used, in what order,
500 and for how long.</p>
505 <!-- ======================================================================= -->
507 <a name="targetjitinfo">The <tt>TargetJITInfo</tt> class</a>
512 <p>The <tt>TargetJITInfo</tt> class exposes an abstract interface used by the
513 Just-In-Time code generator to perform target-specific activities, such as
514 emitting stubs. If a <tt>TargetMachine</tt> supports JIT code generation, it
515 should provide one of these objects through the <tt>getJITInfo</tt>
522 <!-- *********************************************************************** -->
524 <a name="codegendesc">Machine code description classes</a>
526 <!-- *********************************************************************** -->
530 <p>At the high-level, LLVM code is translated to a machine specific
531 representation formed out of
532 <a href="#machinefunction"><tt>MachineFunction</tt></a>,
533 <a href="#machinebasicblock"><tt>MachineBasicBlock</tt></a>,
534 and <a href="#machineinstr"><tt>MachineInstr</tt></a> instances (defined
535 in <tt>include/llvm/CodeGen</tt>). This representation is completely target
536 agnostic, representing instructions in their most abstract form: an opcode
537 and a series of operands. This representation is designed to support both an
538 SSA representation for machine code, as well as a register allocated, non-SSA
541 <!-- ======================================================================= -->
543 <a name="machineinstr">The <tt>MachineInstr</tt> class</a>
548 <p>Target machine instructions are represented as instances of the
549 <tt>MachineInstr</tt> class. This class is an extremely abstract way of
550 representing machine instructions. In particular, it only keeps track of an
551 opcode number and a set of operands.</p>
553 <p>The opcode number is a simple unsigned integer that only has meaning to a
554 specific backend. All of the instructions for a target should be defined in
555 the <tt>*InstrInfo.td</tt> file for the target. The opcode enum values are
556 auto-generated from this description. The <tt>MachineInstr</tt> class does
557 not have any information about how to interpret the instruction (i.e., what
558 the semantics of the instruction are); for that you must refer to the
559 <tt><a href="#targetinstrinfo">TargetInstrInfo</a></tt> class.</p>
561 <p>The operands of a machine instruction can be of several different types: a
562 register reference, a constant integer, a basic block reference, etc. In
563 addition, a machine operand should be marked as a def or a use of the value
564 (though only registers are allowed to be defs).</p>
566 <p>By convention, the LLVM code generator orders instruction operands so that
567 all register definitions come before the register uses, even on architectures
568 that are normally printed in other orders. For example, the SPARC add
569 instruction: "<tt>add %i1, %i2, %i3</tt>" adds the "%i1", and "%i2" registers
570 and stores the result into the "%i3" register. In the LLVM code generator,
571 the operands should be stored as "<tt>%i3, %i1, %i2</tt>": with the
572 destination first.</p>
574 <p>Keeping destination (definition) operands at the beginning of the operand
575 list has several advantages. In particular, the debugging printer will print
576 the instruction like this:</p>
578 <div class="doc_code">
584 <p>Also if the first operand is a def, it is easier to <a href="#buildmi">create
585 instructions</a> whose only def is the first operand.</p>
587 <!-- _______________________________________________________________________ -->
589 <a name="buildmi">Using the <tt>MachineInstrBuilder.h</tt> functions</a>
594 <p>Machine instructions are created by using the <tt>BuildMI</tt> functions,
595 located in the <tt>include/llvm/CodeGen/MachineInstrBuilder.h</tt> file. The
596 <tt>BuildMI</tt> functions make it easy to build arbitrary machine
597 instructions. Usage of the <tt>BuildMI</tt> functions look like this:</p>
599 <div class="doc_code">
601 // Create a 'DestReg = mov 42' (rendered in X86 assembly as 'mov DestReg, 42')
602 // instruction. The '1' specifies how many operands will be added.
603 MachineInstr *MI = BuildMI(X86::MOV32ri, 1, DestReg).addImm(42);
605 // Create the same instr, but insert it at the end of a basic block.
606 MachineBasicBlock &MBB = ...
607 BuildMI(MBB, X86::MOV32ri, 1, DestReg).addImm(42);
609 // Create the same instr, but insert it before a specified iterator point.
610 MachineBasicBlock::iterator MBBI = ...
611 BuildMI(MBB, MBBI, X86::MOV32ri, 1, DestReg).addImm(42);
613 // Create a 'cmp Reg, 0' instruction, no destination reg.
614 MI = BuildMI(X86::CMP32ri, 2).addReg(Reg).addImm(0);
615 // Create an 'sahf' instruction which takes no operands and stores nothing.
616 MI = BuildMI(X86::SAHF, 0);
618 // Create a self looping branch instruction.
619 BuildMI(MBB, X86::JNE, 1).addMBB(&MBB);
623 <p>The key thing to remember with the <tt>BuildMI</tt> functions is that you
624 have to specify the number of operands that the machine instruction will
625 take. This allows for efficient memory allocation. You also need to specify
626 if operands default to be uses of values, not definitions. If you need to
627 add a definition operand (other than the optional destination register), you
628 must explicitly mark it as such:</p>
630 <div class="doc_code">
632 MI.addReg(Reg, RegState::Define);
638 <!-- _______________________________________________________________________ -->
640 <a name="fixedregs">Fixed (preassigned) registers</a>
645 <p>One important issue that the code generator needs to be aware of is the
646 presence of fixed registers. In particular, there are often places in the
647 instruction stream where the register allocator <em>must</em> arrange for a
648 particular value to be in a particular register. This can occur due to
649 limitations of the instruction set (e.g., the X86 can only do a 32-bit divide
650 with the <tt>EAX</tt>/<tt>EDX</tt> registers), or external factors like
651 calling conventions. In any case, the instruction selector should emit code
652 that copies a virtual register into or out of a physical register when
655 <p>For example, consider this simple LLVM example:</p>
657 <div class="doc_code">
659 define i32 @test(i32 %X, i32 %Y) {
666 <p>The X86 instruction selector produces this machine code for the <tt>div</tt>
667 and <tt>ret</tt> (use "<tt>llc X.bc -march=x86 -print-machineinstrs</tt>" to
670 <div class="doc_code">
673 %EAX = mov %reg1024 ;; Copy X (in reg1024) into EAX
674 %reg1027 = sar %reg1024, 31
675 %EDX = mov %reg1027 ;; Sign extend X into EDX
676 idiv %reg1025 ;; Divide by Y (in reg1025)
677 %reg1026 = mov %EAX ;; Read the result (Z) out of EAX
680 %EAX = mov %reg1026 ;; 32-bit return value goes in EAX
685 <p>By the end of code generation, the register allocator has coalesced the
686 registers and deleted the resultant identity moves producing the following
689 <div class="doc_code">
691 ;; X is in EAX, Y is in ECX
699 <p>This approach is extremely general (if it can handle the X86 architecture, it
700 can handle anything!) and allows all of the target specific knowledge about
701 the instruction stream to be isolated in the instruction selector. Note that
702 physical registers should have a short lifetime for good code generation, and
703 all physical registers are assumed dead on entry to and exit from basic
704 blocks (before register allocation). Thus, if you need a value to be live
705 across basic block boundaries, it <em>must</em> live in a virtual
710 <!-- _______________________________________________________________________ -->
712 <a name="callclobber">Call-clobbered registers</a>
717 <p>Some machine instructions, like calls, clobber a large number of physical
718 registers. Rather than adding <code><def,dead></code> operands for
719 all of them, it is possible to use an <code>MO_RegisterMask</code> operand
720 instead. The register mask operand holds a bit mask of preserved registers,
721 and everything else is considered to be clobbered by the instruction. </p>
725 <!-- _______________________________________________________________________ -->
727 <a name="ssa">Machine code in SSA form</a>
732 <p><tt>MachineInstr</tt>'s are initially selected in SSA-form, and are
733 maintained in SSA-form until register allocation happens. For the most part,
734 this is trivially simple since LLVM is already in SSA form; LLVM PHI nodes
735 become machine code PHI nodes, and virtual registers are only allowed to have
736 a single definition.</p>
738 <p>After register allocation, machine code is no longer in SSA-form because
739 there are no virtual registers left in the code.</p>
745 <!-- ======================================================================= -->
747 <a name="machinebasicblock">The <tt>MachineBasicBlock</tt> class</a>
752 <p>The <tt>MachineBasicBlock</tt> class contains a list of machine instructions
753 (<tt><a href="#machineinstr">MachineInstr</a></tt> instances). It roughly
754 corresponds to the LLVM code input to the instruction selector, but there can
755 be a one-to-many mapping (i.e. one LLVM basic block can map to multiple
756 machine basic blocks). The <tt>MachineBasicBlock</tt> class has a
757 "<tt>getBasicBlock</tt>" method, which returns the LLVM basic block that it
762 <!-- ======================================================================= -->
764 <a name="machinefunction">The <tt>MachineFunction</tt> class</a>
769 <p>The <tt>MachineFunction</tt> class contains a list of machine basic blocks
770 (<tt><a href="#machinebasicblock">MachineBasicBlock</a></tt> instances). It
771 corresponds one-to-one with the LLVM function input to the instruction
772 selector. In addition to a list of basic blocks,
773 the <tt>MachineFunction</tt> contains a a <tt>MachineConstantPool</tt>,
774 a <tt>MachineFrameInfo</tt>, a <tt>MachineFunctionInfo</tt>, and a
775 <tt>MachineRegisterInfo</tt>. See
776 <tt>include/llvm/CodeGen/MachineFunction.h</tt> for more information.</p>
780 <!-- ======================================================================= -->
782 <a name="machineinstrbundle"><tt>MachineInstr Bundles</tt></a>
787 <p>LLVM code generator can model sequences of instructions as MachineInstr
788 bundles. A MI bundle can model a VLIW group / pack which contains an
789 arbitrary number of parallel instructions. It can also be used to model
790 a sequential list of instructions (potentially with data dependencies) that
791 cannot be legally separated (e.g. ARM Thumb2 IT blocks).</p>
793 <p>Conceptually a MI bundle is a MI with a number of other MIs nested within:
796 <div class="doc_code">
834 <p> MI bundle support does not change the physical representations of
835 MachineBasicBlock and MachineInstr. All the MIs (including top level and
836 nested ones) are stored as sequential list of MIs. The "bundled" MIs are
837 marked with the 'InsideBundle' flag. A top level MI with the special BUNDLE
838 opcode is used to represent the start of a bundle. It's legal to mix BUNDLE
839 MIs with indiviual MIs that are not inside bundles nor represent bundles.
842 <p> MachineInstr passes should operate on a MI bundle as a single unit. Member
843 methods have been taught to correctly handle bundles and MIs inside bundles.
844 The MachineBasicBlock iterator has been modified to skip over bundled MIs to
845 enforce the bundle-as-a-single-unit concept. An alternative iterator
846 instr_iterator has been added to MachineBasicBlock to allow passes to
847 iterate over all of the MIs in a MachineBasicBlock, including those which
848 are nested inside bundles. The top level BUNDLE instruction must have the
849 correct set of register MachineOperand's that represent the cumulative
850 inputs and outputs of the bundled MIs.</p>
852 <p> Packing / bundling of MachineInstr's should be done as part of the register
853 allocation super-pass. More specifically, the pass which determines what
854 MIs should be bundled together must be done after code generator exits SSA
855 form (i.e. after two-address pass, PHI elimination, and copy coalescing).
856 Bundles should only be finalized (i.e. adding BUNDLE MIs and input and
857 output register MachineOperands) after virtual registers have been
858 rewritten into physical registers. This requirement eliminates the need to
859 add virtual register operands to BUNDLE instructions which would effectively
860 double the virtual register def and use lists.</p>
866 <!-- *********************************************************************** -->
868 <a name="mc">The "MC" Layer</a>
870 <!-- *********************************************************************** -->
875 The MC Layer is used to represent and process code at the raw machine code
876 level, devoid of "high level" information like "constant pools", "jump tables",
877 "global variables" or anything like that. At this level, LLVM handles things
878 like label names, machine instructions, and sections in the object file. The
879 code in this layer is used for a number of important purposes: the tail end of
880 the code generator uses it to write a .s or .o file, and it is also used by the
881 llvm-mc tool to implement standalone machine code assemblers and disassemblers.
885 This section describes some of the important classes. There are also a number
886 of important subsystems that interact at this layer, they are described later
890 <!-- ======================================================================= -->
892 <a name="mcstreamer">The <tt>MCStreamer</tt> API</a>
898 MCStreamer is best thought of as an assembler API. It is an abstract API which
899 is <em>implemented</em> in different ways (e.g. to output a .s file, output an
900 ELF .o file, etc) but whose API correspond directly to what you see in a .s
901 file. MCStreamer has one method per directive, such as EmitLabel,
902 EmitSymbolAttribute, SwitchSection, EmitValue (for .byte, .word), etc, which
903 directly correspond to assembly level directives. It also has an
904 EmitInstruction method, which is used to output an MCInst to the streamer.
908 This API is most important for two clients: the llvm-mc stand-alone assembler is
909 effectively a parser that parses a line, then invokes a method on MCStreamer. In
910 the code generator, the <a href="#codeemit">Code Emission</a> phase of the code
911 generator lowers higher level LLVM IR and Machine* constructs down to the MC
912 layer, emitting directives through MCStreamer.</p>
915 On the implementation side of MCStreamer, there are two major implementations:
916 one for writing out a .s file (MCAsmStreamer), and one for writing out a .o
917 file (MCObjectStreamer). MCAsmStreamer is a straight-forward implementation
918 that prints out a directive for each method (e.g. EmitValue -> .byte), but
919 MCObjectStreamer implements a full assembler.
924 <!-- ======================================================================= -->
926 <a name="mccontext">The <tt>MCContext</tt> class</a>
932 The MCContext class is the owner of a variety of uniqued data structures at the
933 MC layer, including symbols, sections, etc. As such, this is the class that you
934 interact with to create symbols and sections. This class can not be subclassed.
939 <!-- ======================================================================= -->
941 <a name="mcsymbol">The <tt>MCSymbol</tt> class</a>
947 The MCSymbol class represents a symbol (aka label) in the assembly file. There
948 are two interesting kinds of symbols: assembler temporary symbols, and normal
949 symbols. Assembler temporary symbols are used and processed by the assembler
950 but are discarded when the object file is produced. The distinction is usually
951 represented by adding a prefix to the label, for example "L" labels are
952 assembler temporary labels in MachO.
955 <p>MCSymbols are created by MCContext and uniqued there. This means that
956 MCSymbols can be compared for pointer equivalence to find out if they are the
957 same symbol. Note that pointer inequality does not guarantee the labels will
958 end up at different addresses though. It's perfectly legal to output something
959 like this to the .s file:<p>
967 <p>In this case, both the foo and bar symbols will have the same address.</p>
971 <!-- ======================================================================= -->
973 <a name="mcsection">The <tt>MCSection</tt> class</a>
979 The MCSection class represents an object-file specific section. It is subclassed
980 by object file specific implementations (e.g. <tt>MCSectionMachO</tt>,
981 <tt>MCSectionCOFF</tt>, <tt>MCSectionELF</tt>) and these are created and uniqued
982 by MCContext. The MCStreamer has a notion of the current section, which can be
983 changed with the SwitchToSection method (which corresponds to a ".section"
984 directive in a .s file).
989 <!-- ======================================================================= -->
991 <a name="mcinst">The <tt>MCInst</tt> class</a>
997 The MCInst class is a target-independent representation of an instruction. It
998 is a simple class (much more so than <a href="#machineinstr">MachineInstr</a>)
999 that holds a target-specific opcode and a vector of MCOperands. MCOperand, in
1000 turn, is a simple discriminated union of three cases: 1) a simple immediate,
1001 2) a target register ID, 3) a symbolic expression (e.g. "Lfoo-Lbar+42") as an
1005 <p>MCInst is the common currency used to represent machine instructions at the
1006 MC layer. It is the type used by the instruction encoder, the instruction
1007 printer, and the type generated by the assembly parser and disassembler.
1014 <!-- *********************************************************************** -->
1016 <a name="codegenalgs">Target-independent code generation algorithms</a>
1018 <!-- *********************************************************************** -->
1022 <p>This section documents the phases described in the
1023 <a href="#high-level-design">high-level design of the code generator</a>.
1024 It explains how they work and some of the rationale behind their design.</p>
1026 <!-- ======================================================================= -->
1028 <a name="instselect">Instruction Selection</a>
1033 <p>Instruction Selection is the process of translating LLVM code presented to
1034 the code generator into target-specific machine instructions. There are
1035 several well-known ways to do this in the literature. LLVM uses a
1036 SelectionDAG based instruction selector.</p>
1038 <p>Portions of the DAG instruction selector are generated from the target
1039 description (<tt>*.td</tt>) files. Our goal is for the entire instruction
1040 selector to be generated from these <tt>.td</tt> files, though currently
1041 there are still things that require custom C++ code.</p>
1043 <!-- _______________________________________________________________________ -->
1045 <a name="selectiondag_intro">Introduction to SelectionDAGs</a>
1050 <p>The SelectionDAG provides an abstraction for code representation in a way
1051 that is amenable to instruction selection using automatic techniques
1052 (e.g. dynamic-programming based optimal pattern matching selectors). It is
1053 also well-suited to other phases of code generation; in particular,
1054 instruction scheduling (SelectionDAG's are very close to scheduling DAGs
1055 post-selection). Additionally, the SelectionDAG provides a host
1056 representation where a large variety of very-low-level (but
1057 target-independent) <a href="#selectiondag_optimize">optimizations</a> may be
1058 performed; ones which require extensive information about the instructions
1059 efficiently supported by the target.</p>
1061 <p>The SelectionDAG is a Directed-Acyclic-Graph whose nodes are instances of the
1062 <tt>SDNode</tt> class. The primary payload of the <tt>SDNode</tt> is its
1063 operation code (Opcode) that indicates what operation the node performs and
1064 the operands to the operation. The various operation node types are
1065 described at the top of the <tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt>
1068 <p>Although most operations define a single value, each node in the graph may
1069 define multiple values. For example, a combined div/rem operation will
1070 define both the dividend and the remainder. Many other situations require
1071 multiple values as well. Each node also has some number of operands, which
1072 are edges to the node defining the used value. Because nodes may define
1073 multiple values, edges are represented by instances of the <tt>SDValue</tt>
1074 class, which is a <tt><SDNode, unsigned></tt> pair, indicating the node
1075 and result value being used, respectively. Each value produced by
1076 an <tt>SDNode</tt> has an associated <tt>MVT</tt> (Machine Value Type)
1077 indicating what the type of the value is.</p>
1079 <p>SelectionDAGs contain two different kinds of values: those that represent
1080 data flow and those that represent control flow dependencies. Data values
1081 are simple edges with an integer or floating point value type. Control edges
1082 are represented as "chain" edges which are of type <tt>MVT::Other</tt>.
1083 These edges provide an ordering between nodes that have side effects (such as
1084 loads, stores, calls, returns, etc). All nodes that have side effects should
1085 take a token chain as input and produce a new one as output. By convention,
1086 token chain inputs are always operand #0, and chain results are always the
1087 last value produced by an operation.</p>
1089 <p>A SelectionDAG has designated "Entry" and "Root" nodes. The Entry node is
1090 always a marker node with an Opcode of <tt>ISD::EntryToken</tt>. The Root
1091 node is the final side-effecting node in the token chain. For example, in a
1092 single basic block function it would be the return node.</p>
1094 <p>One important concept for SelectionDAGs is the notion of a "legal" vs.
1095 "illegal" DAG. A legal DAG for a target is one that only uses supported
1096 operations and supported types. On a 32-bit PowerPC, for example, a DAG with
1097 a value of type i1, i8, i16, or i64 would be illegal, as would a DAG that
1098 uses a SREM or UREM operation. The
1099 <a href="#selectinodag_legalize_types">legalize types</a> and
1100 <a href="#selectiondag_legalize">legalize operations</a> phases are
1101 responsible for turning an illegal DAG into a legal DAG.</p>
1105 <!-- _______________________________________________________________________ -->
1107 <a name="selectiondag_process">SelectionDAG Instruction Selection Process</a>
1112 <p>SelectionDAG-based instruction selection consists of the following steps:</p>
1115 <li><a href="#selectiondag_build">Build initial DAG</a> — This stage
1116 performs a simple translation from the input LLVM code to an illegal
1119 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> — This
1120 stage performs simple optimizations on the SelectionDAG to simplify it,
1121 and recognize meta instructions (like rotates
1122 and <tt>div</tt>/<tt>rem</tt> pairs) for targets that support these meta
1123 operations. This makes the resultant code more efficient and
1124 the <a href="#selectiondag_select">select instructions from DAG</a> phase
1125 (below) simpler.</li>
1127 <li><a href="#selectiondag_legalize_types">Legalize SelectionDAG Types</a>
1128 — This stage transforms SelectionDAG nodes to eliminate any types
1129 that are unsupported on the target.</li>
1131 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> — The
1132 SelectionDAG optimizer is run to clean up redundancies exposed by type
1135 <li><a href="#selectiondag_legalize">Legalize SelectionDAG Ops</a> —
1136 This stage transforms SelectionDAG nodes to eliminate any operations
1137 that are unsupported on the target.</li>
1139 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> — The
1140 SelectionDAG optimizer is run to eliminate inefficiencies introduced by
1141 operation legalization.</li>
1143 <li><a href="#selectiondag_select">Select instructions from DAG</a> —
1144 Finally, the target instruction selector matches the DAG operations to
1145 target instructions. This process translates the target-independent input
1146 DAG into another DAG of target instructions.</li>
1148 <li><a href="#selectiondag_sched">SelectionDAG Scheduling and Formation</a>
1149 — The last phase assigns a linear order to the instructions in the
1150 target-instruction DAG and emits them into the MachineFunction being
1151 compiled. This step uses traditional prepass scheduling techniques.</li>
1154 <p>After all of these steps are complete, the SelectionDAG is destroyed and the
1155 rest of the code generation passes are run.</p>
1157 <p>One great way to visualize what is going on here is to take advantage of a
1158 few LLC command line options. The following options pop up a window
1159 displaying the SelectionDAG at specific times (if you only get errors printed
1160 to the console while using this, you probably
1161 <a href="ProgrammersManual.html#ViewGraph">need to configure your system</a>
1162 to add support for it).</p>
1165 <li><tt>-view-dag-combine1-dags</tt> displays the DAG after being built,
1166 before the first optimization pass.</li>
1168 <li><tt>-view-legalize-dags</tt> displays the DAG before Legalization.</li>
1170 <li><tt>-view-dag-combine2-dags</tt> displays the DAG before the second
1171 optimization pass.</li>
1173 <li><tt>-view-isel-dags</tt> displays the DAG before the Select phase.</li>
1175 <li><tt>-view-sched-dags</tt> displays the DAG before Scheduling.</li>
1178 <p>The <tt>-view-sunit-dags</tt> displays the Scheduler's dependency graph.
1179 This graph is based on the final SelectionDAG, with nodes that must be
1180 scheduled together bundled into a single scheduling-unit node, and with
1181 immediate operands and other nodes that aren't relevant for scheduling
1186 <!-- _______________________________________________________________________ -->
1188 <a name="selectiondag_build">Initial SelectionDAG Construction</a>
1193 <p>The initial SelectionDAG is naïvely peephole expanded from the LLVM
1194 input by the <tt>SelectionDAGLowering</tt> class in the
1195 <tt>lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp</tt> file. The intent of
1196 this pass is to expose as much low-level, target-specific details to the
1197 SelectionDAG as possible. This pass is mostly hard-coded (e.g. an
1198 LLVM <tt>add</tt> turns into an <tt>SDNode add</tt> while a
1199 <tt>getelementptr</tt> is expanded into the obvious arithmetic). This pass
1200 requires target-specific hooks to lower calls, returns, varargs, etc. For
1201 these features, the <tt><a href="#targetlowering">TargetLowering</a></tt>
1202 interface is used.</p>
1206 <!-- _______________________________________________________________________ -->
1208 <a name="selectiondag_legalize_types">SelectionDAG LegalizeTypes Phase</a>
1213 <p>The Legalize phase is in charge of converting a DAG to only use the types
1214 that are natively supported by the target.</p>
1216 <p>There are two main ways of converting values of unsupported scalar types to
1217 values of supported types: converting small types to larger types
1218 ("promoting"), and breaking up large integer types into smaller ones
1219 ("expanding"). For example, a target might require that all f32 values are
1220 promoted to f64 and that all i1/i8/i16 values are promoted to i32. The same
1221 target might require that all i64 values be expanded into pairs of i32
1222 values. These changes can insert sign and zero extensions as needed to make
1223 sure that the final code has the same behavior as the input.</p>
1225 <p>There are two main ways of converting values of unsupported vector types to
1226 value of supported types: splitting vector types, multiple times if
1227 necessary, until a legal type is found, and extending vector types by adding
1228 elements to the end to round them out to legal types ("widening"). If a
1229 vector gets split all the way down to single-element parts with no supported
1230 vector type being found, the elements are converted to scalars
1231 ("scalarizing").</p>
1233 <p>A target implementation tells the legalizer which types are supported (and
1234 which register class to use for them) by calling the
1235 <tt>addRegisterClass</tt> method in its TargetLowering constructor.</p>
1239 <!-- _______________________________________________________________________ -->
1241 <a name="selectiondag_legalize">SelectionDAG Legalize Phase</a>
1246 <p>The Legalize phase is in charge of converting a DAG to only use the
1247 operations that are natively supported by the target.</p>
1249 <p>Targets often have weird constraints, such as not supporting every operation
1250 on every supported datatype (e.g. X86 does not support byte conditional moves
1251 and PowerPC does not support sign-extending loads from a 16-bit memory
1252 location). Legalize takes care of this by open-coding another sequence of
1253 operations to emulate the operation ("expansion"), by promoting one type to a
1254 larger type that supports the operation ("promotion"), or by using a
1255 target-specific hook to implement the legalization ("custom").</p>
1257 <p>A target implementation tells the legalizer which operations are not
1258 supported (and which of the above three actions to take) by calling the
1259 <tt>setOperationAction</tt> method in its <tt>TargetLowering</tt>
1262 <p>Prior to the existence of the Legalize passes, we required that every target
1263 <a href="#selectiondag_optimize">selector</a> supported and handled every
1264 operator and type even if they are not natively supported. The introduction
1265 of the Legalize phases allows all of the canonicalization patterns to be
1266 shared across targets, and makes it very easy to optimize the canonicalized
1267 code because it is still in the form of a DAG.</p>
1271 <!-- _______________________________________________________________________ -->
1273 <a name="selectiondag_optimize">
1274 SelectionDAG Optimization Phase: the DAG Combiner
1280 <p>The SelectionDAG optimization phase is run multiple times for code
1281 generation, immediately after the DAG is built and once after each
1282 legalization. The first run of the pass allows the initial code to be
1283 cleaned up (e.g. performing optimizations that depend on knowing that the
1284 operators have restricted type inputs). Subsequent runs of the pass clean up
1285 the messy code generated by the Legalize passes, which allows Legalize to be
1286 very simple (it can focus on making code legal instead of focusing on
1287 generating <em>good</em> and legal code).</p>
1289 <p>One important class of optimizations performed is optimizing inserted sign
1290 and zero extension instructions. We currently use ad-hoc techniques, but
1291 could move to more rigorous techniques in the future. Here are some good
1292 papers on the subject:</p>
1294 <p>"<a href="http://www.eecs.harvard.edu/~nr/pubs/widen-abstract.html">Widening
1295 integer arithmetic</a>"<br>
1296 Kevin Redwine and Norman Ramsey<br>
1297 International Conference on Compiler Construction (CC) 2004</p>
1299 <p>"<a href="http://portal.acm.org/citation.cfm?doid=512529.512552">Effective
1300 sign extension elimination</a>"<br>
1301 Motohiro Kawahito, Hideaki Komatsu, and Toshio Nakatani<br>
1302 Proceedings of the ACM SIGPLAN 2002 Conference on Programming Language Design
1303 and Implementation.</p>
1307 <!-- _______________________________________________________________________ -->
1309 <a name="selectiondag_select">SelectionDAG Select Phase</a>
1314 <p>The Select phase is the bulk of the target-specific code for instruction
1315 selection. This phase takes a legal SelectionDAG as input, pattern matches
1316 the instructions supported by the target to this DAG, and produces a new DAG
1317 of target code. For example, consider the following LLVM fragment:</p>
1319 <div class="doc_code">
1321 %t1 = fadd float %W, %X
1322 %t2 = fmul float %t1, %Y
1323 %t3 = fadd float %t2, %Z
1327 <p>This LLVM code corresponds to a SelectionDAG that looks basically like
1330 <div class="doc_code">
1332 (fadd:f32 (fmul:f32 (fadd:f32 W, X), Y), Z)
1336 <p>If a target supports floating point multiply-and-add (FMA) operations, one of
1337 the adds can be merged with the multiply. On the PowerPC, for example, the
1338 output of the instruction selector might look like this DAG:</p>
1340 <div class="doc_code">
1342 (FMADDS (FADDS W, X), Y, Z)
1346 <p>The <tt>FMADDS</tt> instruction is a ternary instruction that multiplies its
1347 first two operands and adds the third (as single-precision floating-point
1348 numbers). The <tt>FADDS</tt> instruction is a simple binary single-precision
1349 add instruction. To perform this pattern match, the PowerPC backend includes
1350 the following instruction definitions:</p>
1352 <div class="doc_code">
1354 def FMADDS : AForm_1<59, 29,
1355 (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
1356 "fmadds $FRT, $FRA, $FRC, $FRB",
1357 [<b>(set F4RC:$FRT, (fadd (fmul F4RC:$FRA, F4RC:$FRC),
1358 F4RC:$FRB))</b>]>;
1359 def FADDS : AForm_2<59, 21,
1360 (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
1361 "fadds $FRT, $FRA, $FRB",
1362 [<b>(set F4RC:$FRT, (fadd F4RC:$FRA, F4RC:$FRB))</b>]>;
1366 <p>The portion of the instruction definition in bold indicates the pattern used
1367 to match the instruction. The DAG operators
1368 (like <tt>fmul</tt>/<tt>fadd</tt>) are defined in
1369 the <tt>include/llvm/Target/TargetSelectionDAG.td</tt> file. "
1370 <tt>F4RC</tt>" is the register class of the input and result values.</p>
1372 <p>The TableGen DAG instruction selector generator reads the instruction
1373 patterns in the <tt>.td</tt> file and automatically builds parts of the
1374 pattern matching code for your target. It has the following strengths:</p>
1377 <li>At compiler-compiler time, it analyzes your instruction patterns and tells
1378 you if your patterns make sense or not.</li>
1380 <li>It can handle arbitrary constraints on operands for the pattern match. In
1381 particular, it is straight-forward to say things like "match any immediate
1382 that is a 13-bit sign-extended value". For examples, see the
1383 <tt>immSExt16</tt> and related <tt>tblgen</tt> classes in the PowerPC
1386 <li>It knows several important identities for the patterns defined. For
1387 example, it knows that addition is commutative, so it allows the
1388 <tt>FMADDS</tt> pattern above to match "<tt>(fadd X, (fmul Y, Z))</tt>" as
1389 well as "<tt>(fadd (fmul X, Y), Z)</tt>", without the target author having
1390 to specially handle this case.</li>
1392 <li>It has a full-featured type-inferencing system. In particular, you should
1393 rarely have to explicitly tell the system what type parts of your patterns
1394 are. In the <tt>FMADDS</tt> case above, we didn't have to tell
1395 <tt>tblgen</tt> that all of the nodes in the pattern are of type 'f32'.
1396 It was able to infer and propagate this knowledge from the fact that
1397 <tt>F4RC</tt> has type 'f32'.</li>
1399 <li>Targets can define their own (and rely on built-in) "pattern fragments".
1400 Pattern fragments are chunks of reusable patterns that get inlined into
1401 your patterns during compiler-compiler time. For example, the integer
1402 "<tt>(not x)</tt>" operation is actually defined as a pattern fragment
1403 that expands as "<tt>(xor x, -1)</tt>", since the SelectionDAG does not
1404 have a native '<tt>not</tt>' operation. Targets can define their own
1405 short-hand fragments as they see fit. See the definition of
1406 '<tt>not</tt>' and '<tt>ineg</tt>' for examples.</li>
1408 <li>In addition to instructions, targets can specify arbitrary patterns that
1409 map to one or more instructions using the 'Pat' class. For example, the
1410 PowerPC has no way to load an arbitrary integer immediate into a register
1411 in one instruction. To tell tblgen how to do this, it defines:
1414 <div class="doc_code">
1416 // Arbitrary immediate support. Implement in terms of LIS/ORI.
1417 def : Pat<(i32 imm:$imm),
1418 (ORI (LIS (HI16 imm:$imm)), (LO16 imm:$imm))>;
1422 If none of the single-instruction patterns for loading an immediate into a
1423 register match, this will be used. This rule says "match an arbitrary i32
1424 immediate, turning it into an <tt>ORI</tt> ('or a 16-bit immediate') and
1425 an <tt>LIS</tt> ('load 16-bit immediate, where the immediate is shifted to
1426 the left 16 bits') instruction". To make this work, the
1427 <tt>LO16</tt>/<tt>HI16</tt> node transformations are used to manipulate
1428 the input immediate (in this case, take the high or low 16-bits of the
1431 <li>While the system does automate a lot, it still allows you to write custom
1432 C++ code to match special cases if there is something that is hard to
1436 <p>While it has many strengths, the system currently has some limitations,
1437 primarily because it is a work in progress and is not yet finished:</p>
1440 <li>Overall, there is no way to define or match SelectionDAG nodes that define
1441 multiple values (e.g. <tt>SMUL_LOHI</tt>, <tt>LOAD</tt>, <tt>CALL</tt>,
1442 etc). This is the biggest reason that you currently still <em>have
1443 to</em> write custom C++ code for your instruction selector.</li>
1445 <li>There is no great way to support matching complex addressing modes yet.
1446 In the future, we will extend pattern fragments to allow them to define
1447 multiple values (e.g. the four operands of the <a href="#x86_memory">X86
1448 addressing mode</a>, which are currently matched with custom C++ code).
1449 In addition, we'll extend fragments so that a fragment can match multiple
1450 different patterns.</li>
1452 <li>We don't automatically infer flags like isStore/isLoad yet.</li>
1454 <li>We don't automatically generate the set of supported registers and
1455 operations for the <a href="#selectiondag_legalize">Legalizer</a>
1458 <li>We don't have a way of tying in custom legalized nodes yet.</li>
1461 <p>Despite these limitations, the instruction selector generator is still quite
1462 useful for most of the binary and logical operations in typical instruction
1463 sets. If you run into any problems or can't figure out how to do something,
1464 please let Chris know!</p>
1468 <!-- _______________________________________________________________________ -->
1470 <a name="selectiondag_sched">SelectionDAG Scheduling and Formation Phase</a>
1475 <p>The scheduling phase takes the DAG of target instructions from the selection
1476 phase and assigns an order. The scheduler can pick an order depending on
1477 various constraints of the machines (i.e. order for minimal register pressure
1478 or try to cover instruction latencies). Once an order is established, the
1479 DAG is converted to a list
1480 of <tt><a href="#machineinstr">MachineInstr</a></tt>s and the SelectionDAG is
1483 <p>Note that this phase is logically separate from the instruction selection
1484 phase, but is tied to it closely in the code because it operates on
1489 <!-- _______________________________________________________________________ -->
1491 <a name="selectiondag_future">Future directions for the SelectionDAG</a>
1497 <li>Optional function-at-a-time selection.</li>
1499 <li>Auto-generate entire selector from <tt>.td</tt> file.</li>
1506 <!-- ======================================================================= -->
1508 <a name="ssamco">SSA-based Machine Code Optimizations</a>
1510 <div><p>To Be Written</p></div>
1512 <!-- ======================================================================= -->
1514 <a name="liveintervals">Live Intervals</a>
1519 <p>Live Intervals are the ranges (intervals) where a variable is <i>live</i>.
1520 They are used by some <a href="#regalloc">register allocator</a> passes to
1521 determine if two or more virtual registers which require the same physical
1522 register are live at the same point in the program (i.e., they conflict).
1523 When this situation occurs, one virtual register must be <i>spilled</i>.</p>
1525 <!-- _______________________________________________________________________ -->
1527 <a name="livevariable_analysis">Live Variable Analysis</a>
1532 <p>The first step in determining the live intervals of variables is to calculate
1533 the set of registers that are immediately dead after the instruction (i.e.,
1534 the instruction calculates the value, but it is never used) and the set of
1535 registers that are used by the instruction, but are never used after the
1536 instruction (i.e., they are killed). Live variable information is computed
1537 for each <i>virtual</i> register and <i>register allocatable</i> physical
1538 register in the function. This is done in a very efficient manner because it
1539 uses SSA to sparsely compute lifetime information for virtual registers
1540 (which are in SSA form) and only has to track physical registers within a
1541 block. Before register allocation, LLVM can assume that physical registers
1542 are only live within a single basic block. This allows it to do a single,
1543 local analysis to resolve physical register lifetimes within each basic
1544 block. If a physical register is not register allocatable (e.g., a stack
1545 pointer or condition codes), it is not tracked.</p>
1547 <p>Physical registers may be live in to or out of a function. Live in values are
1548 typically arguments in registers. Live out values are typically return values
1549 in registers. Live in values are marked as such, and are given a dummy
1550 "defining" instruction during live intervals analysis. If the last basic
1551 block of a function is a <tt>return</tt>, then it's marked as using all live
1552 out values in the function.</p>
1554 <p><tt>PHI</tt> nodes need to be handled specially, because the calculation of
1555 the live variable information from a depth first traversal of the CFG of the
1556 function won't guarantee that a virtual register used by the <tt>PHI</tt>
1557 node is defined before it's used. When a <tt>PHI</tt> node is encountered,
1558 only the definition is handled, because the uses will be handled in other
1561 <p>For each <tt>PHI</tt> node of the current basic block, we simulate an
1562 assignment at the end of the current basic block and traverse the successor
1563 basic blocks. If a successor basic block has a <tt>PHI</tt> node and one of
1564 the <tt>PHI</tt> node's operands is coming from the current basic block, then
1565 the variable is marked as <i>alive</i> within the current basic block and all
1566 of its predecessor basic blocks, until the basic block with the defining
1567 instruction is encountered.</p>
1571 <!-- _______________________________________________________________________ -->
1573 <a name="liveintervals_analysis">Live Intervals Analysis</a>
1578 <p>We now have the information available to perform the live intervals analysis
1579 and build the live intervals themselves. We start off by numbering the basic
1580 blocks and machine instructions. We then handle the "live-in" values. These
1581 are in physical registers, so the physical register is assumed to be killed
1582 by the end of the basic block. Live intervals for virtual registers are
1583 computed for some ordering of the machine instructions <tt>[1, N]</tt>. A
1584 live interval is an interval <tt>[i, j)</tt>, where <tt>1 <= i <= j
1585 < N</tt>, for which a variable is live.</p>
1587 <p><i><b>More to come...</b></i></p>
1593 <!-- ======================================================================= -->
1595 <a name="regalloc">Register Allocation</a>
1600 <p>The <i>Register Allocation problem</i> consists in mapping a program
1601 <i>P<sub>v</sub></i>, that can use an unbounded number of virtual registers,
1602 to a program <i>P<sub>p</sub></i> that contains a finite (possibly small)
1603 number of physical registers. Each target architecture has a different number
1604 of physical registers. If the number of physical registers is not enough to
1605 accommodate all the virtual registers, some of them will have to be mapped
1606 into memory. These virtuals are called <i>spilled virtuals</i>.</p>
1608 <!-- _______________________________________________________________________ -->
1611 <a name="regAlloc_represent">How registers are represented in LLVM</a>
1616 <p>In LLVM, physical registers are denoted by integer numbers that normally
1617 range from 1 to 1023. To see how this numbering is defined for a particular
1618 architecture, you can read the <tt>GenRegisterNames.inc</tt> file for that
1619 architecture. For instance, by
1620 inspecting <tt>lib/Target/X86/X86GenRegisterNames.inc</tt> we see that the
1621 32-bit register <tt>EAX</tt> is denoted by 15, and the MMX register
1622 <tt>MM0</tt> is mapped to 48.</p>
1624 <p>Some architectures contain registers that share the same physical location. A
1625 notable example is the X86 platform. For instance, in the X86 architecture,
1626 the registers <tt>EAX</tt>, <tt>AX</tt> and <tt>AL</tt> share the first eight
1627 bits. These physical registers are marked as <i>aliased</i> in LLVM. Given a
1628 particular architecture, you can check which registers are aliased by
1629 inspecting its <tt>RegisterInfo.td</tt> file. Moreover, the method
1630 <tt>TargetRegisterInfo::getAliasSet(p_reg)</tt> returns an array containing
1631 all the physical registers aliased to the register <tt>p_reg</tt>.</p>
1633 <p>Physical registers, in LLVM, are grouped in <i>Register Classes</i>.
1634 Elements in the same register class are functionally equivalent, and can be
1635 interchangeably used. Each virtual register can only be mapped to physical
1636 registers of a particular class. For instance, in the X86 architecture, some
1637 virtuals can only be allocated to 8 bit registers. A register class is
1638 described by <tt>TargetRegisterClass</tt> objects. To discover if a virtual
1639 register is compatible with a given physical, this code can be used:</p>
1641 <div class="doc_code">
1643 bool RegMapping_Fer::compatible_class(MachineFunction &mf,
1646 assert(TargetRegisterInfo::isPhysicalRegister(p_reg) &&
1647 "Target register must be physical");
1648 const TargetRegisterClass *trc = mf.getRegInfo().getRegClass(v_reg);
1649 return trc->contains(p_reg);
1654 <p>Sometimes, mostly for debugging purposes, it is useful to change the number
1655 of physical registers available in the target architecture. This must be done
1656 statically, inside the <tt>TargetRegsterInfo.td</tt> file. Just <tt>grep</tt>
1657 for <tt>RegisterClass</tt>, the last parameter of which is a list of
1658 registers. Just commenting some out is one simple way to avoid them being
1659 used. A more polite way is to explicitly exclude some registers from
1660 the <i>allocation order</i>. See the definition of the <tt>GR8</tt> register
1661 class in <tt>lib/Target/X86/X86RegisterInfo.td</tt> for an example of this.
1664 <p>Virtual registers are also denoted by integer numbers. Contrary to physical
1665 registers, different virtual registers never share the same number. Whereas
1666 physical registers are statically defined in a <tt>TargetRegisterInfo.td</tt>
1667 file and cannot be created by the application developer, that is not the case
1668 with virtual registers. In order to create new virtual registers, use the
1669 method <tt>MachineRegisterInfo::createVirtualRegister()</tt>. This method
1670 will return a new virtual register. Use an <tt>IndexedMap<Foo,
1671 VirtReg2IndexFunctor></tt> to hold information per virtual register. If you
1672 need to enumerate all virtual registers, use the function
1673 <tt>TargetRegisterInfo::index2VirtReg()</tt> to find the virtual register
1676 <div class="doc_code">
1678 for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
1679 unsigned VirtReg = TargetRegisterInfo::index2VirtReg(i);
1685 <p>Before register allocation, the operands of an instruction are mostly virtual
1686 registers, although physical registers may also be used. In order to check if
1687 a given machine operand is a register, use the boolean
1688 function <tt>MachineOperand::isRegister()</tt>. To obtain the integer code of
1689 a register, use <tt>MachineOperand::getReg()</tt>. An instruction may define
1690 or use a register. For instance, <tt>ADD reg:1026 := reg:1025 reg:1024</tt>
1691 defines the registers 1024, and uses registers 1025 and 1026. Given a
1692 register operand, the method <tt>MachineOperand::isUse()</tt> informs if that
1693 register is being used by the instruction. The
1694 method <tt>MachineOperand::isDef()</tt> informs if that registers is being
1697 <p>We will call physical registers present in the LLVM bitcode before register
1698 allocation <i>pre-colored registers</i>. Pre-colored registers are used in
1699 many different situations, for instance, to pass parameters of functions
1700 calls, and to store results of particular instructions. There are two types
1701 of pre-colored registers: the ones <i>implicitly</i> defined, and
1702 those <i>explicitly</i> defined. Explicitly defined registers are normal
1703 operands, and can be accessed
1704 with <tt>MachineInstr::getOperand(int)::getReg()</tt>. In order to check
1705 which registers are implicitly defined by an instruction, use
1706 the <tt>TargetInstrInfo::get(opcode)::ImplicitDefs</tt>,
1707 where <tt>opcode</tt> is the opcode of the target instruction. One important
1708 difference between explicit and implicit physical registers is that the
1709 latter are defined statically for each instruction, whereas the former may
1710 vary depending on the program being compiled. For example, an instruction
1711 that represents a function call will always implicitly define or use the same
1712 set of physical registers. To read the registers implicitly used by an
1714 use <tt>TargetInstrInfo::get(opcode)::ImplicitUses</tt>. Pre-colored
1715 registers impose constraints on any register allocation algorithm. The
1716 register allocator must make sure that none of them are overwritten by
1717 the values of virtual registers while still alive.</p>
1721 <!-- _______________________________________________________________________ -->
1724 <a name="regAlloc_howTo">Mapping virtual registers to physical registers</a>
1729 <p>There are two ways to map virtual registers to physical registers (or to
1730 memory slots). The first way, that we will call <i>direct mapping</i>, is
1731 based on the use of methods of the classes <tt>TargetRegisterInfo</tt>,
1732 and <tt>MachineOperand</tt>. The second way, that we will call <i>indirect
1733 mapping</i>, relies on the <tt>VirtRegMap</tt> class in order to insert loads
1734 and stores sending and getting values to and from memory.</p>
1736 <p>The direct mapping provides more flexibility to the developer of the register
1737 allocator; however, it is more error prone, and demands more implementation
1738 work. Basically, the programmer will have to specify where load and store
1739 instructions should be inserted in the target function being compiled in
1740 order to get and store values in memory. To assign a physical register to a
1741 virtual register present in a given operand,
1742 use <tt>MachineOperand::setReg(p_reg)</tt>. To insert a store instruction,
1743 use <tt>TargetInstrInfo::storeRegToStackSlot(...)</tt>, and to insert a
1744 load instruction, use <tt>TargetInstrInfo::loadRegFromStackSlot</tt>.</p>
1746 <p>The indirect mapping shields the application developer from the complexities
1747 of inserting load and store instructions. In order to map a virtual register
1748 to a physical one, use <tt>VirtRegMap::assignVirt2Phys(vreg, preg)</tt>. In
1749 order to map a certain virtual register to memory,
1750 use <tt>VirtRegMap::assignVirt2StackSlot(vreg)</tt>. This method will return
1751 the stack slot where <tt>vreg</tt>'s value will be located. If it is
1752 necessary to map another virtual register to the same stack slot,
1753 use <tt>VirtRegMap::assignVirt2StackSlot(vreg, stack_location)</tt>. One
1754 important point to consider when using the indirect mapping, is that even if
1755 a virtual register is mapped to memory, it still needs to be mapped to a
1756 physical register. This physical register is the location where the virtual
1757 register is supposed to be found before being stored or after being
1760 <p>If the indirect strategy is used, after all the virtual registers have been
1761 mapped to physical registers or stack slots, it is necessary to use a spiller
1762 object to place load and store instructions in the code. Every virtual that
1763 has been mapped to a stack slot will be stored to memory after been defined
1764 and will be loaded before being used. The implementation of the spiller tries
1765 to recycle load/store instructions, avoiding unnecessary instructions. For an
1766 example of how to invoke the spiller,
1767 see <tt>RegAllocLinearScan::runOnMachineFunction</tt>
1768 in <tt>lib/CodeGen/RegAllocLinearScan.cpp</tt>.</p>
1772 <!-- _______________________________________________________________________ -->
1774 <a name="regAlloc_twoAddr">Handling two address instructions</a>
1779 <p>With very rare exceptions (e.g., function calls), the LLVM machine code
1780 instructions are three address instructions. That is, each instruction is
1781 expected to define at most one register, and to use at most two registers.
1782 However, some architectures use two address instructions. In this case, the
1783 defined register is also one of the used register. For instance, an
1784 instruction such as <tt>ADD %EAX, %EBX</tt>, in X86 is actually equivalent
1785 to <tt>%EAX = %EAX + %EBX</tt>.</p>
1787 <p>In order to produce correct code, LLVM must convert three address
1788 instructions that represent two address instructions into true two address
1789 instructions. LLVM provides the pass <tt>TwoAddressInstructionPass</tt> for
1790 this specific purpose. It must be run before register allocation takes
1791 place. After its execution, the resulting code may no longer be in SSA
1792 form. This happens, for instance, in situations where an instruction such
1793 as <tt>%a = ADD %b %c</tt> is converted to two instructions such as:</p>
1795 <div class="doc_code">
1802 <p>Notice that, internally, the second instruction is represented as
1803 <tt>ADD %a[def/use] %c</tt>. I.e., the register operand <tt>%a</tt> is both
1804 used and defined by the instruction.</p>
1808 <!-- _______________________________________________________________________ -->
1810 <a name="regAlloc_ssaDecon">The SSA deconstruction phase</a>
1815 <p>An important transformation that happens during register allocation is called
1816 the <i>SSA Deconstruction Phase</i>. The SSA form simplifies many analyses
1817 that are performed on the control flow graph of programs. However,
1818 traditional instruction sets do not implement PHI instructions. Thus, in
1819 order to generate executable code, compilers must replace PHI instructions
1820 with other instructions that preserve their semantics.</p>
1822 <p>There are many ways in which PHI instructions can safely be removed from the
1823 target code. The most traditional PHI deconstruction algorithm replaces PHI
1824 instructions with copy instructions. That is the strategy adopted by
1825 LLVM. The SSA deconstruction algorithm is implemented
1826 in <tt>lib/CodeGen/PHIElimination.cpp</tt>. In order to invoke this pass, the
1827 identifier <tt>PHIEliminationID</tt> must be marked as required in the code
1828 of the register allocator.</p>
1832 <!-- _______________________________________________________________________ -->
1834 <a name="regAlloc_fold">Instruction folding</a>
1839 <p><i>Instruction folding</i> is an optimization performed during register
1840 allocation that removes unnecessary copy instructions. For instance, a
1841 sequence of instructions such as:</p>
1843 <div class="doc_code">
1845 %EBX = LOAD %mem_address
1850 <p>can be safely substituted by the single instruction:</p>
1852 <div class="doc_code">
1854 %EAX = LOAD %mem_address
1858 <p>Instructions can be folded with
1859 the <tt>TargetRegisterInfo::foldMemoryOperand(...)</tt> method. Care must be
1860 taken when folding instructions; a folded instruction can be quite different
1862 instruction. See <tt>LiveIntervals::addIntervalsForSpills</tt>
1863 in <tt>lib/CodeGen/LiveIntervalAnalysis.cpp</tt> for an example of its
1868 <!-- _______________________________________________________________________ -->
1871 <a name="regAlloc_builtIn">Built in register allocators</a>
1876 <p>The LLVM infrastructure provides the application developer with three
1877 different register allocators:</p>
1880 <li><i>Fast</i> — This register allocator is the default for debug
1881 builds. It allocates registers on a basic block level, attempting to keep
1882 values in registers and reusing registers as appropriate.</li>
1884 <li><i>Basic</i> — This is an incremental approach to register
1885 allocation. Live ranges are assigned to registers one at a time in
1886 an order that is driven by heuristics. Since code can be rewritten
1887 on-the-fly during allocation, this framework allows interesting
1888 allocators to be developed as extensions. It is not itself a
1889 production register allocator but is a potentially useful
1890 stand-alone mode for triaging bugs and as a performance baseline.
1892 <li><i>Greedy</i> — <i>The default allocator</i>. This is a
1893 highly tuned implementation of the <i>Basic</i> allocator that
1894 incorporates global live range splitting. This allocator works hard
1895 to minimize the cost of spill code.
1897 <li><i>PBQP</i> — A Partitioned Boolean Quadratic Programming (PBQP)
1898 based register allocator. This allocator works by constructing a PBQP
1899 problem representing the register allocation problem under consideration,
1900 solving this using a PBQP solver, and mapping the solution back to a
1901 register assignment.</li>
1904 <p>The type of register allocator used in <tt>llc</tt> can be chosen with the
1905 command line option <tt>-regalloc=...</tt>:</p>
1907 <div class="doc_code">
1909 $ llc -regalloc=linearscan file.bc -o ln.s;
1910 $ llc -regalloc=fast file.bc -o fa.s;
1911 $ llc -regalloc=pbqp file.bc -o pbqp.s;
1919 <!-- ======================================================================= -->
1921 <a name="proepicode">Prolog/Epilog Code Insertion</a>
1926 <!-- _______________________________________________________________________ -->
1928 <a name="compact_unwind">Compact Unwind</a>
1933 <p>Throwing an exception requires <em>unwinding</em> out of a function. The
1934 information on how to unwind a given function is traditionally expressed in
1935 DWARF unwind (a.k.a. frame) info. But that format was originally developed
1936 for debuggers to backtrace, and each Frame Description Entry (FDE) requires
1937 ~20-30 bytes per function. There is also the cost of mapping from an address
1938 in a function to the corresponding FDE at runtime. An alternative unwind
1939 encoding is called <em>compact unwind</em> and requires just 4-bytes per
1942 <p>The compact unwind encoding is a 32-bit value, which is encoded in an
1943 architecture-specific way. It specifies which registers to restore and from
1944 where, and how to unwind out of the function. When the linker creates a final
1945 linked image, it will create a <code>__TEXT,__unwind_info</code>
1946 section. This section is a small and fast way for the runtime to access
1947 unwind info for any given function. If we emit compact unwind info for the
1948 function, that compact unwind info will be encoded in
1949 the <code>__TEXT,__unwind_info</code> section. If we emit DWARF unwind info,
1950 the <code>__TEXT,__unwind_info</code> section will contain the offset of the
1951 FDE in the <code>__TEXT,__eh_frame</code> section in the final linked
1954 <p>For X86, there are three modes for the compact unwind encoding:</p>
1957 <dt><i>Function with a Frame Pointer (<code>EBP</code> or <code>RBP</code>)</i></dt>
1958 <dd><p><code>EBP/RBP</code>-based frame, where <code>EBP/RBP</code> is pushed
1959 onto the stack immediately after the return address,
1960 then <code>ESP/RSP</code> is moved to <code>EBP/RBP</code>. Thus to
1961 unwind, <code>ESP/RSP</code> is restored with the
1962 current <code>EBP/RBP</code> value, then <code>EBP/RBP</code> is restored
1963 by popping the stack, and the return is done by popping the stack once
1964 more into the PC. All non-volatile registers that need to be restored must
1965 have been saved in a small range on the stack that
1966 starts <code>EBP-4</code> to <code>EBP-1020</code> (<code>RBP-8</code>
1967 to <code>RBP-1020</code>). The offset (divided by 4 in 32-bit mode and 8
1968 in 64-bit mode) is encoded in bits 16-23 (mask: <code>0x00FF0000</code>).
1969 The registers saved are encoded in bits 0-14
1970 (mask: <code>0x00007FFF</code>) as five 3-bit entries from the following
1972 <table border="1" cellspacing="0">
1974 <th>Compact Number</th>
1975 <th>i386 Register</th>
1976 <th>x86-64 Regiser</th>
1980 <td><code>EBX</code></td>
1981 <td><code>RBX</code></td>
1985 <td><code>ECX</code></td>
1986 <td><code>R12</code></td>
1990 <td><code>EDX</code></td>
1991 <td><code>R13</code></td>
1995 <td><code>EDI</code></td>
1996 <td><code>R14</code></td>
2000 <td><code>ESI</code></td>
2001 <td><code>R15</code></td>
2005 <td><code>EBP</code></td>
2006 <td><code>RBP</code></td>
2012 <dt><i>Frameless with a Small Constant Stack Size (<code>EBP</code>
2013 or <code>RBP</code> is not used as a frame pointer)</i></dt>
2014 <dd><p>To return, a constant (encoded in the compact unwind encoding) is added
2015 to the <code>ESP/RSP</code>. Then the return is done by popping the stack
2016 into the PC. All non-volatile registers that need to be restored must have
2017 been saved on the stack immediately after the return address. The stack
2018 size (divided by 4 in 32-bit mode and 8 in 64-bit mode) is encoded in bits
2019 16-23 (mask: <code>0x00FF0000</code>). There is a maximum stack size of
2020 1024 bytes in 32-bit mode and 2048 in 64-bit mode. The number of registers
2021 saved is encoded in bits 9-12 (mask: <code>0x00001C00</code>). Bits 0-9
2022 (mask: <code>0x000003FF</code>) contain which registers were saved and
2024 the <code>encodeCompactUnwindRegistersWithoutFrame()</code> function
2025 in <code>lib/Target/X86FrameLowering.cpp</code> for the encoding
2026 algorithm.)</p></dd>
2028 <dt><i>Frameless with a Large Constant Stack Size (<code>EBP</code>
2029 or <code>RBP</code> is not used as a frame pointer)</i></dt>
2030 <dd><p>This case is like the "Frameless with a Small Constant Stack Size"
2031 case, but the stack size is too large to encode in the compact unwind
2032 encoding. Instead it requires that the function contains "<code>subl
2033 $nnnnnn, %esp</code>" in its prolog. The compact encoding contains the
2034 offset to the <code>$nnnnnn</code> value in the function in bits 9-12
2035 (mask: <code>0x00001C00</code>).</p></dd>
2042 <!-- ======================================================================= -->
2044 <a name="latemco">Late Machine Code Optimizations</a>
2046 <div><p>To Be Written</p></div>
2048 <!-- ======================================================================= -->
2050 <a name="codeemit">Code Emission</a>
2055 <p>The code emission step of code generation is responsible for lowering from
2056 the code generator abstractions (like <a
2057 href="#machinefunction">MachineFunction</a>, <a
2058 href="#machineinstr">MachineInstr</a>, etc) down
2059 to the abstractions used by the MC layer (<a href="#mcinst">MCInst</a>,
2060 <a href="#mcstreamer">MCStreamer</a>, etc). This is
2061 done with a combination of several different classes: the (misnamed)
2062 target-independent AsmPrinter class, target-specific subclasses of AsmPrinter
2063 (such as SparcAsmPrinter), and the TargetLoweringObjectFile class.</p>
2065 <p>Since the MC layer works at the level of abstraction of object files, it
2066 doesn't have a notion of functions, global variables etc. Instead, it thinks
2067 about labels, directives, and instructions. A key class used at this time is
2068 the MCStreamer class. This is an abstract API that is implemented in different
2069 ways (e.g. to output a .s file, output an ELF .o file, etc) that is effectively
2070 an "assembler API". MCStreamer has one method per directive, such as EmitLabel,
2071 EmitSymbolAttribute, SwitchSection, etc, which directly correspond to assembly
2075 <p>If you are interested in implementing a code generator for a target, there
2076 are three important things that you have to implement for your target:</p>
2079 <li>First, you need a subclass of AsmPrinter for your target. This class
2080 implements the general lowering process converting MachineFunction's into MC
2081 label constructs. The AsmPrinter base class provides a number of useful methods
2082 and routines, and also allows you to override the lowering process in some
2083 important ways. You should get much of the lowering for free if you are
2084 implementing an ELF, COFF, or MachO target, because the TargetLoweringObjectFile
2085 class implements much of the common logic.</li>
2087 <li>Second, you need to implement an instruction printer for your target. The
2088 instruction printer takes an <a href="#mcinst">MCInst</a> and renders it to a
2089 raw_ostream as text. Most of this is automatically generated from the .td file
2090 (when you specify something like "<tt>add $dst, $src1, $src2</tt>" in the
2091 instructions), but you need to implement routines to print operands.</li>
2093 <li>Third, you need to implement code that lowers a <a
2094 href="#machineinstr">MachineInstr</a> to an MCInst, usually implemented in
2095 "<target>MCInstLower.cpp". This lowering process is often target
2096 specific, and is responsible for turning jump table entries, constant pool
2097 indices, global variable addresses, etc into MCLabels as appropriate. This
2098 translation layer is also responsible for expanding pseudo ops used by the code
2099 generator into the actual machine instructions they correspond to. The MCInsts
2100 that are generated by this are fed into the instruction printer or the encoder.
2105 <p>Finally, at your choosing, you can also implement an subclass of
2106 MCCodeEmitter which lowers MCInst's into machine code bytes and relocations.
2107 This is important if you want to support direct .o file emission, or would like
2108 to implement an assembler for your target.</p>
2112 <!-- ======================================================================= -->
2114 <a name="vliw_packetizer">VLIW Packetizer</a>
2119 <p>In a Very Long Instruction Word (VLIW) architecture, the compiler is
2120 responsible for mapping instructions to functional-units available on
2121 the architecture. To that end, the compiler creates groups of instructions
2122 called <i>packets</i> or <i>bundles</i>. The VLIW packetizer in LLVM is
2123 a target-independent mechanism to enable the packetization of machine
2126 <!-- _______________________________________________________________________ -->
2129 <a name="vliw_mapping">Mapping from instructions to functional units</a>
2134 <p>Instructions in a VLIW target can typically be mapped to multiple functional
2135 units. During the process of packetizing, the compiler must be able to reason
2136 about whether an instruction can be added to a packet. This decision can be
2137 complex since the compiler has to examine all possible mappings of instructions
2138 to functional units. Therefore to alleviate compilation-time complexity, the
2139 VLIW packetizer parses the instruction classes of a target and generates tables
2140 at compiler build time. These tables can then be queried by the provided
2141 machine-independent API to determine if an instruction can be accommodated in a
2145 <!-- ======================================================================= -->
2147 <a name="vliw_repr">
2148 How the packetization tables are generated and used
2154 <p>The packetizer reads instruction classes from a target's itineraries and
2155 creates a deterministic finite automaton (DFA) to represent the state of a
2156 packet. A DFA consists of three major elements: inputs, states, and
2157 transitions. The set of inputs for the generated DFA represents the instruction
2158 being added to a packet. The states represent the possible consumption
2159 of functional units by instructions in a packet. In the DFA, transitions from
2160 one state to another occur on the addition of an instruction to an existing
2161 packet. If there is a legal mapping of functional units to instructions, then
2162 the DFA contains a corresponding transition. The absence of a transition
2163 indicates that a legal mapping does not exist and that the instruction cannot
2164 be added to the packet.</p>
2166 <p>To generate tables for a VLIW target, add <i>Target</i>GenDFAPacketizer.inc
2167 as a target to the Makefile in the target directory. The exported API provides
2168 three functions: <tt>DFAPacketizer::clearResources()</tt>,
2169 <tt>DFAPacketizer::reserveResources(MachineInstr *MI)</tt>, and
2170 <tt>DFAPacketizer::canReserveResources(MachineInstr *MI)</tt>. These functions
2171 allow a target packetizer to add an instruction to an existing packet and to
2172 check whether an instruction can be added to a packet. See
2173 <tt>llvm/CodeGen/DFAPacketizer.h</tt> for more information.</p>
2181 <!-- *********************************************************************** -->
2183 <a name="nativeassembler">Implementing a Native Assembler</a>
2185 <!-- *********************************************************************** -->
2189 <p>Though you're probably reading this because you want to write or maintain a
2190 compiler backend, LLVM also fully supports building a native assemblers too.
2191 We've tried hard to automate the generation of the assembler from the .td files
2192 (in particular the instruction syntax and encodings), which means that a large
2193 part of the manual and repetitive data entry can be factored and shared with the
2196 <!-- ======================================================================= -->
2197 <h3 id="na_instparsing">Instruction Parsing</h3>
2199 <div><p>To Be Written</p></div>
2202 <!-- ======================================================================= -->
2203 <h3 id="na_instaliases">
2204 Instruction Alias Processing
2208 <p>Once the instruction is parsed, it enters the MatchInstructionImpl function.
2209 The MatchInstructionImpl function performs alias processing and then does
2210 actual matching.</p>
2212 <p>Alias processing is the phase that canonicalizes different lexical forms of
2213 the same instructions down to one representation. There are several different
2214 kinds of alias that are possible to implement and they are listed below in the
2215 order that they are processed (which is in order from simplest/weakest to most
2216 complex/powerful). Generally you want to use the first alias mechanism that
2217 meets the needs of your instruction, because it will allow a more concise
2220 <!-- _______________________________________________________________________ -->
2221 <h4>Mnemonic Aliases</h4>
2225 <p>The first phase of alias processing is simple instruction mnemonic
2226 remapping for classes of instructions which are allowed with two different
2227 mnemonics. This phase is a simple and unconditionally remapping from one input
2228 mnemonic to one output mnemonic. It isn't possible for this form of alias to
2229 look at the operands at all, so the remapping must apply for all forms of a
2230 given mnemonic. Mnemonic aliases are defined simply, for example X86 has:
2233 <div class="doc_code">
2235 def : MnemonicAlias<"cbw", "cbtw">;
2236 def : MnemonicAlias<"smovq", "movsq">;
2237 def : MnemonicAlias<"fldcww", "fldcw">;
2238 def : MnemonicAlias<"fucompi", "fucomip">;
2239 def : MnemonicAlias<"ud2a", "ud2">;
2243 <p>... and many others. With a MnemonicAlias definition, the mnemonic is
2244 remapped simply and directly. Though MnemonicAlias's can't look at any aspect
2245 of the instruction (such as the operands) they can depend on global modes (the
2246 same ones supported by the matcher), through a Requires clause:</p>
2248 <div class="doc_code">
2250 def : MnemonicAlias<"pushf", "pushfq">, Requires<[In64BitMode]>;
2251 def : MnemonicAlias<"pushf", "pushfl">, Requires<[In32BitMode]>;
2255 <p>In this example, the mnemonic gets mapped into different a new one depending
2256 on the current instruction set.</p>
2260 <!-- _______________________________________________________________________ -->
2261 <h4>Instruction Aliases</h4>
2265 <p>The most general phase of alias processing occurs while matching is
2266 happening: it provides new forms for the matcher to match along with a specific
2267 instruction to generate. An instruction alias has two parts: the string to
2268 match and the instruction to generate. For example:
2271 <div class="doc_code">
2273 def : InstAlias<"movsx $src, $dst", (MOVSX16rr8W GR16:$dst, GR8 :$src)>;
2274 def : InstAlias<"movsx $src, $dst", (MOVSX16rm8W GR16:$dst, i8mem:$src)>;
2275 def : InstAlias<"movsx $src, $dst", (MOVSX32rr8 GR32:$dst, GR8 :$src)>;
2276 def : InstAlias<"movsx $src, $dst", (MOVSX32rr16 GR32:$dst, GR16 :$src)>;
2277 def : InstAlias<"movsx $src, $dst", (MOVSX64rr8 GR64:$dst, GR8 :$src)>;
2278 def : InstAlias<"movsx $src, $dst", (MOVSX64rr16 GR64:$dst, GR16 :$src)>;
2279 def : InstAlias<"movsx $src, $dst", (MOVSX64rr32 GR64:$dst, GR32 :$src)>;
2283 <p>This shows a powerful example of the instruction aliases, matching the
2284 same mnemonic in multiple different ways depending on what operands are present
2285 in the assembly. The result of instruction aliases can include operands in a
2286 different order than the destination instruction, and can use an input
2287 multiple times, for example:</p>
2289 <div class="doc_code">
2291 def : InstAlias<"clrb $reg", (XOR8rr GR8 :$reg, GR8 :$reg)>;
2292 def : InstAlias<"clrw $reg", (XOR16rr GR16:$reg, GR16:$reg)>;
2293 def : InstAlias<"clrl $reg", (XOR32rr GR32:$reg, GR32:$reg)>;
2294 def : InstAlias<"clrq $reg", (XOR64rr GR64:$reg, GR64:$reg)>;
2298 <p>This example also shows that tied operands are only listed once. In the X86
2299 backend, XOR8rr has two input GR8's and one output GR8 (where an input is tied
2300 to the output). InstAliases take a flattened operand list without duplicates
2301 for tied operands. The result of an instruction alias can also use immediates
2302 and fixed physical registers which are added as simple immediate operands in the
2303 result, for example:</p>
2305 <div class="doc_code">
2307 // Fixed Immediate operand.
2308 def : InstAlias<"aad", (AAD8i8 10)>;
2310 // Fixed register operand.
2311 def : InstAlias<"fcomi", (COM_FIr ST1)>;
2314 def : InstAlias<"fcomi $reg", (COM_FIr RST:$reg)>;
2319 <p>Instruction aliases can also have a Requires clause to make them
2320 subtarget specific.</p>
2322 <p>If the back-end supports it, the instruction printer can automatically emit
2323 the alias rather than what's being aliased. It typically leads to better,
2324 more readable code. If it's better to print out what's being aliased, then
2325 pass a '0' as the third parameter to the InstAlias definition.</p>
2331 <!-- ======================================================================= -->
2332 <h3 id="na_matching">Instruction Matching</h3>
2334 <div><p>To Be Written</p></div>
2338 <!-- *********************************************************************** -->
2340 <a name="targetimpls">Target-specific Implementation Notes</a>
2342 <!-- *********************************************************************** -->
2346 <p>This section of the document explains features or design decisions that are
2347 specific to the code generator for a particular target. First we start
2348 with a table that summarizes what features are supported by each target.</p>
2350 <!-- ======================================================================= -->
2352 <a name="targetfeatures">Target Feature Matrix</a>
2357 <p>Note that this table does not include the C backend or Cpp backends, since
2358 they do not use the target independent code generator infrastructure. It also
2359 doesn't list features that are not supported fully by any target yet. It
2360 considers a feature to be supported if at least one subtarget supports it. A
2361 feature being supported means that it is useful and works for most cases, it
2362 does not indicate that there are zero known bugs in the implementation. Here
2366 <table border="1" cellspacing="0">
2370 <th>Partial Support</th>
2371 <th>Complete Support</th>
2374 <td class="unknown"></td>
2375 <td class="no"></td>
2376 <td class="partial"></td>
2377 <td class="yes"></td>
2381 <p>Here is the table:</p>
2383 <table width="689" border="1" cellspacing="0">
2385 <td colspan="13" align="center" style="background-color:#ffc">Target</td>
2403 <td><a href="#feat_reliable">is generally reliable</a></td>
2404 <td class="yes"></td> <!-- ARM -->
2405 <td class="no"></td> <!-- CellSPU -->
2406 <td class="yes"></td> <!-- Hexagon -->
2407 <td class="no"></td> <!-- MBlaze -->
2408 <td class="unknown"></td> <!-- MSP430 -->
2409 <td class="yes"></td> <!-- Mips -->
2410 <td class="no"></td> <!-- PTX -->
2411 <td class="yes"></td> <!-- PowerPC -->
2412 <td class="yes"></td> <!-- Sparc -->
2413 <td class="yes"></td> <!-- X86 -->
2414 <td class="unknown"></td> <!-- XCore -->
2418 <td><a href="#feat_asmparser">assembly parser</a></td>
2419 <td class="no"></td> <!-- ARM -->
2420 <td class="no"></td> <!-- CellSPU -->
2421 <td class="no"></td> <!-- Hexagon -->
2422 <td class="yes"></td> <!-- MBlaze -->
2423 <td class="no"></td> <!-- MSP430 -->
2424 <td class="no"></td> <!-- Mips -->
2425 <td class="no"></td> <!-- PTX -->
2426 <td class="no"></td> <!-- PowerPC -->
2427 <td class="no"></td> <!-- Sparc -->
2428 <td class="yes"></td> <!-- X86 -->
2429 <td class="no"></td> <!-- XCore -->
2433 <td><a href="#feat_disassembler">disassembler</a></td>
2434 <td class="yes"></td> <!-- ARM -->
2435 <td class="no"></td> <!-- CellSPU -->
2436 <td class="no"></td> <!-- Hexagon -->
2437 <td class="yes"></td> <!-- MBlaze -->
2438 <td class="no"></td> <!-- MSP430 -->
2439 <td class="no"></td> <!-- Mips -->
2440 <td class="no"></td> <!-- PTX -->
2441 <td class="no"></td> <!-- PowerPC -->
2442 <td class="no"></td> <!-- Sparc -->
2443 <td class="yes"></td> <!-- X86 -->
2444 <td class="no"></td> <!-- XCore -->
2448 <td><a href="#feat_inlineasm">inline asm</a></td>
2449 <td class="yes"></td> <!-- ARM -->
2450 <td class="no"></td> <!-- CellSPU -->
2451 <td class="yes"></td> <!-- Hexagon -->
2452 <td class="yes"></td> <!-- MBlaze -->
2453 <td class="unknown"></td> <!-- MSP430 -->
2454 <td class="no"></td> <!-- Mips -->
2455 <td class="unknown"></td> <!-- PTX -->
2456 <td class="yes"></td> <!-- PowerPC -->
2457 <td class="unknown"></td> <!-- Sparc -->
2458 <td class="yes"></td> <!-- X86 -->
2459 <td class="unknown"></td> <!-- XCore -->
2463 <td><a href="#feat_jit">jit</a></td>
2464 <td class="partial"><a href="#feat_jit_arm">*</a></td> <!-- ARM -->
2465 <td class="no"></td> <!-- CellSPU -->
2466 <td class="no"></td> <!-- Hexagon -->
2467 <td class="no"></td> <!-- MBlaze -->
2468 <td class="unknown"></td> <!-- MSP430 -->
2469 <td class="yes"></td> <!-- Mips -->
2470 <td class="unknown"></td> <!-- PTX -->
2471 <td class="yes"></td> <!-- PowerPC -->
2472 <td class="unknown"></td> <!-- Sparc -->
2473 <td class="yes"></td> <!-- X86 -->
2474 <td class="unknown"></td> <!-- XCore -->
2478 <td><a href="#feat_objectwrite">.o file writing</a></td>
2479 <td class="no"></td> <!-- ARM -->
2480 <td class="no"></td> <!-- CellSPU -->
2481 <td class="no"></td> <!-- Hexagon -->
2482 <td class="yes"></td> <!-- MBlaze -->
2483 <td class="no"></td> <!-- MSP430 -->
2484 <td class="no"></td> <!-- Mips -->
2485 <td class="no"></td> <!-- PTX -->
2486 <td class="no"></td> <!-- PowerPC -->
2487 <td class="no"></td> <!-- Sparc -->
2488 <td class="yes"></td> <!-- X86 -->
2489 <td class="no"></td> <!-- XCore -->
2493 <td><a href="#feat_tailcall">tail calls</a></td>
2494 <td class="yes"></td> <!-- ARM -->
2495 <td class="no"></td> <!-- CellSPU -->
2496 <td class="yes"></td> <!-- Hexagon -->
2497 <td class="no"></td> <!-- MBlaze -->
2498 <td class="unknown"></td> <!-- MSP430 -->
2499 <td class="no"></td> <!-- Mips -->
2500 <td class="unknown"></td> <!-- PTX -->
2501 <td class="yes"></td> <!-- PowerPC -->
2502 <td class="unknown"></td> <!-- Sparc -->
2503 <td class="yes"></td> <!-- X86 -->
2504 <td class="unknown"></td> <!-- XCore -->
2508 <td><a href="#feat_segstacks">segmented stacks</a></td>
2509 <td class="no"></td> <!-- ARM -->
2510 <td class="no"></td> <!-- CellSPU -->
2511 <td class="no"></td> <!-- Hexagon -->
2512 <td class="no"></td> <!-- MBlaze -->
2513 <td class="no"></td> <!-- MSP430 -->
2514 <td class="no"></td> <!-- Mips -->
2515 <td class="no"></td> <!-- PTX -->
2516 <td class="no"></td> <!-- PowerPC -->
2517 <td class="no"></td> <!-- Sparc -->
2518 <td class="partial"><a href="#feat_segstacks_x86">*</a></td> <!-- X86 -->
2519 <td class="no"></td> <!-- XCore -->
2525 <!-- _______________________________________________________________________ -->
2526 <h4 id="feat_reliable">Is Generally Reliable</h4>
2529 <p>This box indicates whether the target is considered to be production quality.
2530 This indicates that the target has been used as a static compiler to
2531 compile large amounts of code by a variety of different people and is in
2535 <!-- _______________________________________________________________________ -->
2536 <h4 id="feat_asmparser">Assembly Parser</h4>
2539 <p>This box indicates whether the target supports parsing target specific .s
2540 files by implementing the MCAsmParser interface. This is required for llvm-mc
2541 to be able to act as a native assembler and is required for inline assembly
2542 support in the native .o file writer.</p>
2547 <!-- _______________________________________________________________________ -->
2548 <h4 id="feat_disassembler">Disassembler</h4>
2551 <p>This box indicates whether the target supports the MCDisassembler API for
2552 disassembling machine opcode bytes into MCInst's.</p>
2556 <!-- _______________________________________________________________________ -->
2557 <h4 id="feat_inlineasm">Inline Asm</h4>
2560 <p>This box indicates whether the target supports most popular inline assembly
2561 constraints and modifiers.</p>
2565 <!-- _______________________________________________________________________ -->
2566 <h4 id="feat_jit">JIT Support</h4>
2569 <p>This box indicates whether the target supports the JIT compiler through
2570 the ExecutionEngine interface.</p>
2572 <p id="feat_jit_arm">The ARM backend has basic support for integer code
2573 in ARM codegen mode, but lacks NEON and full Thumb support.</p>
2577 <!-- _______________________________________________________________________ -->
2578 <h4 id="feat_objectwrite">.o File Writing</h4>
2582 <p>This box indicates whether the target supports writing .o files (e.g. MachO,
2583 ELF, and/or COFF) files directly from the target. Note that the target also
2584 must include an assembly parser and general inline assembly support for full
2585 inline assembly support in the .o writer.</p>
2587 <p>Targets that don't support this feature can obviously still write out .o
2588 files, they just rely on having an external assembler to translate from a .s
2589 file to a .o file (as is the case for many C compilers).</p>
2593 <!-- _______________________________________________________________________ -->
2594 <h4 id="feat_tailcall">Tail Calls</h4>
2598 <p>This box indicates whether the target supports guaranteed tail calls. These
2599 are calls marked "<a href="LangRef.html#i_call">tail</a>" and use the fastcc
2600 calling convention. Please see the <a href="#tailcallopt">tail call section
2601 more more details</a>.</p>
2605 <!-- _______________________________________________________________________ -->
2606 <h4 id="feat_segstacks">Segmented Stacks</h4>
2610 <p>This box indicates whether the target supports segmented stacks. This
2611 replaces the traditional large C stack with many linked segments. It
2612 is compatible with the <a href="http://gcc.gnu.org/wiki/SplitStacks">gcc
2613 implementation</a> used by the Go front end.</p>
2615 <p id="feat_segstacks_x86">Basic support exists on the X86 backend. Currently
2616 vararg doesn't work and the object files are not marked the way the gold
2617 linker expects, but simple Go programs can be built by dragonegg.</p>
2623 <!-- ======================================================================= -->
2625 <a name="tailcallopt">Tail call optimization</a>
2630 <p>Tail call optimization, callee reusing the stack of the caller, is currently
2631 supported on x86/x86-64 and PowerPC. It is performed if:</p>
2634 <li>Caller and callee have the calling convention <tt>fastcc</tt> or
2635 <tt>cc 10</tt> (GHC call convention).</li>
2637 <li>The call is a tail call - in tail position (ret immediately follows call
2638 and ret uses value of call or is void).</li>
2640 <li>Option <tt>-tailcallopt</tt> is enabled.</li>
2642 <li>Platform specific constraints are met.</li>
2645 <p>x86/x86-64 constraints:</p>
2648 <li>No variable argument lists are used.</li>
2650 <li>On x86-64 when generating GOT/PIC code only module-local calls (visibility
2651 = hidden or protected) are supported.</li>
2654 <p>PowerPC constraints:</p>
2657 <li>No variable argument lists are used.</li>
2659 <li>No byval parameters are used.</li>
2661 <li>On ppc32/64 GOT/PIC only module-local calls (visibility = hidden or protected) are supported.</li>
2666 <p>Call as <tt>llc -tailcallopt test.ll</tt>.</p>
2668 <div class="doc_code">
2670 declare fastcc i32 @tailcallee(i32 inreg %a1, i32 inreg %a2, i32 %a3, i32 %a4)
2672 define fastcc i32 @tailcaller(i32 %in1, i32 %in2) {
2673 %l1 = add i32 %in1, %in2
2674 %tmp = tail call fastcc i32 @tailcallee(i32 %in1 inreg, i32 %in2 inreg, i32 %in1, i32 %l1)
2680 <p>Implications of <tt>-tailcallopt</tt>:</p>
2682 <p>To support tail call optimization in situations where the callee has more
2683 arguments than the caller a 'callee pops arguments' convention is used. This
2684 currently causes each <tt>fastcc</tt> call that is not tail call optimized
2685 (because one or more of above constraints are not met) to be followed by a
2686 readjustment of the stack. So performance might be worse in such cases.</p>
2689 <!-- ======================================================================= -->
2691 <a name="sibcallopt">Sibling call optimization</a>
2696 <p>Sibling call optimization is a restricted form of tail call optimization.
2697 Unlike tail call optimization described in the previous section, it can be
2698 performed automatically on any tail calls when <tt>-tailcallopt</tt> option
2699 is not specified.</p>
2701 <p>Sibling call optimization is currently performed on x86/x86-64 when the
2702 following constraints are met:</p>
2705 <li>Caller and callee have the same calling convention. It can be either
2706 <tt>c</tt> or <tt>fastcc</tt>.
2708 <li>The call is a tail call - in tail position (ret immediately follows call
2709 and ret uses value of call or is void).</li>
2711 <li>Caller and callee have matching return type or the callee result is not
2714 <li>If any of the callee arguments are being passed in stack, they must be
2715 available in caller's own incoming argument stack and the frame offsets
2720 <div class="doc_code">
2722 declare i32 @bar(i32, i32)
2724 define i32 @foo(i32 %a, i32 %b, i32 %c) {
2726 %0 = tail call i32 @bar(i32 %a, i32 %b)
2733 <!-- ======================================================================= -->
2735 <a name="x86">The X86 backend</a>
2740 <p>The X86 code generator lives in the <tt>lib/Target/X86</tt> directory. This
2741 code generator is capable of targeting a variety of x86-32 and x86-64
2742 processors, and includes support for ISA extensions such as MMX and SSE.</p>
2744 <!-- _______________________________________________________________________ -->
2746 <a name="x86_tt">X86 Target Triples supported</a>
2751 <p>The following are the known target triples that are supported by the X86
2752 backend. This is not an exhaustive list, and it would be useful to add those
2753 that people test.</p>
2756 <li><b>i686-pc-linux-gnu</b> — Linux</li>
2758 <li><b>i386-unknown-freebsd5.3</b> — FreeBSD 5.3</li>
2760 <li><b>i686-pc-cygwin</b> — Cygwin on Win32</li>
2762 <li><b>i686-pc-mingw32</b> — MingW on Win32</li>
2764 <li><b>i386-pc-mingw32msvc</b> — MingW crosscompiler on Linux</li>
2766 <li><b>i686-apple-darwin*</b> — Apple Darwin on X86</li>
2768 <li><b>x86_64-unknown-linux-gnu</b> — Linux</li>
2773 <!-- _______________________________________________________________________ -->
2775 <a name="x86_cc">X86 Calling Conventions supported</a>
2781 <p>The following target-specific calling conventions are known to backend:</p>
2784 <li><b>x86_StdCall</b> — stdcall calling convention seen on Microsoft
2785 Windows platform (CC ID = 64).</li>
2786 <li><b>x86_FastCall</b> — fastcall calling convention seen on Microsoft
2787 Windows platform (CC ID = 65).</li>
2788 <li><b>x86_ThisCall</b> — Similar to X86_StdCall. Passes first argument
2789 in ECX, others via stack. Callee is responsible for stack cleaning. This
2790 convention is used by MSVC by default for methods in its ABI
2796 <!-- _______________________________________________________________________ -->
2798 <a name="x86_memory">Representing X86 addressing modes in MachineInstrs</a>
2803 <p>The x86 has a very flexible way of accessing memory. It is capable of
2804 forming memory addresses of the following expression directly in integer
2805 instructions (which use ModR/M addressing):</p>
2807 <div class="doc_code">
2809 SegmentReg: Base + [1,2,4,8] * IndexReg + Disp32
2813 <p>In order to represent this, LLVM tracks no less than 5 operands for each
2814 memory operand of this form. This means that the "load" form of
2815 '<tt>mov</tt>' has the following <tt>MachineOperand</tt>s in this order:</p>
2817 <div class="doc_code">
2819 Index: 0 | 1 2 3 4 5
2820 Meaning: DestReg, | BaseReg, Scale, IndexReg, Displacement Segment
2821 OperandTy: VirtReg, | VirtReg, UnsImm, VirtReg, SignExtImm PhysReg
2825 <p>Stores, and all other instructions, treat the four memory operands in the
2826 same way and in the same order. If the segment register is unspecified
2827 (regno = 0), then no segment override is generated. "Lea" operations do not
2828 have a segment register specified, so they only have 4 operands for their
2829 memory reference.</p>
2833 <!-- _______________________________________________________________________ -->
2835 <a name="x86_memory">X86 address spaces supported</a>
2840 <p>x86 has a feature which provides
2841 the ability to perform loads and stores to different address spaces
2842 via the x86 segment registers. A segment override prefix byte on an
2843 instruction causes the instruction's memory access to go to the specified
2844 segment. LLVM address space 0 is the default address space, which includes
2845 the stack, and any unqualified memory accesses in a program. Address spaces
2846 1-255 are currently reserved for user-defined code. The GS-segment is
2847 represented by address space 256, while the FS-segment is represented by
2848 address space 257. Other x86 segments have yet to be allocated address space
2851 <p>While these address spaces may seem similar to TLS via the
2852 <tt>thread_local</tt> keyword, and often use the same underlying hardware,
2853 there are some fundamental differences.</p>
2855 <p>The <tt>thread_local</tt> keyword applies to global variables and
2856 specifies that they are to be allocated in thread-local memory. There are
2857 no type qualifiers involved, and these variables can be pointed to with
2858 normal pointers and accessed with normal loads and stores.
2859 The <tt>thread_local</tt> keyword is target-independent at the LLVM IR
2860 level (though LLVM doesn't yet have implementations of it for some
2863 <p>Special address spaces, in contrast, apply to static types. Every
2864 load and store has a particular address space in its address operand type,
2865 and this is what determines which address space is accessed.
2866 LLVM ignores these special address space qualifiers on global variables,
2867 and does not provide a way to directly allocate storage in them.
2868 At the LLVM IR level, the behavior of these special address spaces depends
2869 in part on the underlying OS or runtime environment, and they are specific
2870 to x86 (and LLVM doesn't yet handle them correctly in some cases).</p>
2872 <p>Some operating systems and runtime environments use (or may in the future
2873 use) the FS/GS-segment registers for various low-level purposes, so care
2874 should be taken when considering them.</p>
2878 <!-- _______________________________________________________________________ -->
2880 <a name="x86_names">Instruction naming</a>
2885 <p>An instruction name consists of the base name, a default operand size, and a
2886 a character per operand with an optional special size. For example:</p>
2888 <div class="doc_code">
2890 ADD8rr -> add, 8-bit register, 8-bit register
2891 IMUL16rmi -> imul, 16-bit register, 16-bit memory, 16-bit immediate
2892 IMUL16rmi8 -> imul, 16-bit register, 16-bit memory, 8-bit immediate
2893 MOVSX32rm16 -> movsx, 32-bit register, 16-bit memory
2901 <!-- ======================================================================= -->
2903 <a name="ppc">The PowerPC backend</a>
2908 <p>The PowerPC code generator lives in the lib/Target/PowerPC directory. The
2909 code generation is retargetable to several variations or <i>subtargets</i> of
2910 the PowerPC ISA; including ppc32, ppc64 and altivec.</p>
2912 <!-- _______________________________________________________________________ -->
2914 <a name="ppc_abi">LLVM PowerPC ABI</a>
2919 <p>LLVM follows the AIX PowerPC ABI, with two deviations. LLVM uses a PC
2920 relative (PIC) or static addressing for accessing global values, so no TOC
2921 (r2) is used. Second, r31 is used as a frame pointer to allow dynamic growth
2922 of a stack frame. LLVM takes advantage of having no TOC to provide space to
2923 save the frame pointer in the PowerPC linkage area of the caller frame.
2924 Other details of PowerPC ABI can be found at <a href=
2925 "http://developer.apple.com/documentation/DeveloperTools/Conceptual/LowLevelABI/Articles/32bitPowerPC.html"
2926 >PowerPC ABI.</a> Note: This link describes the 32 bit ABI. The 64 bit ABI
2927 is similar except space for GPRs are 8 bytes wide (not 4) and r13 is reserved
2932 <!-- _______________________________________________________________________ -->
2934 <a name="ppc_frame">Frame Layout</a>
2939 <p>The size of a PowerPC frame is usually fixed for the duration of a
2940 function's invocation. Since the frame is fixed size, all references
2941 into the frame can be accessed via fixed offsets from the stack pointer. The
2942 exception to this is when dynamic alloca or variable sized arrays are
2943 present, then a base pointer (r31) is used as a proxy for the stack pointer
2944 and stack pointer is free to grow or shrink. A base pointer is also used if
2945 llvm-gcc is not passed the -fomit-frame-pointer flag. The stack pointer is
2946 always aligned to 16 bytes, so that space allocated for altivec vectors will
2947 be properly aligned.</p>
2949 <p>An invocation frame is laid out as follows (low memory at top);</p>
2951 <table class="layout">
2953 <td>Linkage<br><br></td>
2956 <td>Parameter area<br><br></td>
2959 <td>Dynamic area<br><br></td>
2962 <td>Locals area<br><br></td>
2965 <td>Saved registers area<br><br></td>
2967 <tr style="border-style: none hidden none hidden;">
2971 <td>Previous Frame<br><br></td>
2975 <p>The <i>linkage</i> area is used by a callee to save special registers prior
2976 to allocating its own frame. Only three entries are relevant to LLVM. The
2977 first entry is the previous stack pointer (sp), aka link. This allows
2978 probing tools like gdb or exception handlers to quickly scan the frames in
2979 the stack. A function epilog can also use the link to pop the frame from the
2980 stack. The third entry in the linkage area is used to save the return
2981 address from the lr register. Finally, as mentioned above, the last entry is
2982 used to save the previous frame pointer (r31.) The entries in the linkage
2983 area are the size of a GPR, thus the linkage area is 24 bytes long in 32 bit
2984 mode and 48 bytes in 64 bit mode.</p>
2986 <p>32 bit linkage area</p>
2988 <table class="layout">
2991 <td>Saved SP (r1)</td>
3011 <td>Saved FP (r31)</td>
3015 <p>64 bit linkage area</p>
3017 <table class="layout">
3020 <td>Saved SP (r1)</td>
3040 <td>Saved FP (r31)</td>
3044 <p>The <i>parameter area</i> is used to store arguments being passed to a callee
3045 function. Following the PowerPC ABI, the first few arguments are actually
3046 passed in registers, with the space in the parameter area unused. However,
3047 if there are not enough registers or the callee is a thunk or vararg
3048 function, these register arguments can be spilled into the parameter area.
3049 Thus, the parameter area must be large enough to store all the parameters for
3050 the largest call sequence made by the caller. The size must also be
3051 minimally large enough to spill registers r3-r10. This allows callees blind
3052 to the call signature, such as thunks and vararg functions, enough space to
3053 cache the argument registers. Therefore, the parameter area is minimally 32
3054 bytes (64 bytes in 64 bit mode.) Also note that since the parameter area is
3055 a fixed offset from the top of the frame, that a callee can access its spilt
3056 arguments using fixed offsets from the stack pointer (or base pointer.)</p>
3058 <p>Combining the information about the linkage, parameter areas and alignment. A
3059 stack frame is minimally 64 bytes in 32 bit mode and 128 bytes in 64 bit
3062 <p>The <i>dynamic area</i> starts out as size zero. If a function uses dynamic
3063 alloca then space is added to the stack, the linkage and parameter areas are
3064 shifted to top of stack, and the new space is available immediately below the
3065 linkage and parameter areas. The cost of shifting the linkage and parameter
3066 areas is minor since only the link value needs to be copied. The link value
3067 can be easily fetched by adding the original frame size to the base pointer.
3068 Note that allocations in the dynamic space need to observe 16 byte
3071 <p>The <i>locals area</i> is where the llvm compiler reserves space for local
3074 <p>The <i>saved registers area</i> is where the llvm compiler spills callee
3075 saved registers on entry to the callee.</p>
3079 <!-- _______________________________________________________________________ -->
3081 <a name="ppc_prolog">Prolog/Epilog</a>
3086 <p>The llvm prolog and epilog are the same as described in the PowerPC ABI, with
3087 the following exceptions. Callee saved registers are spilled after the frame
3088 is created. This allows the llvm epilog/prolog support to be common with
3089 other targets. The base pointer callee saved register r31 is saved in the
3090 TOC slot of linkage area. This simplifies allocation of space for the base
3091 pointer and makes it convenient to locate programatically and during
3096 <!-- _______________________________________________________________________ -->
3098 <a name="ppc_dynamic">Dynamic Allocation</a>
3103 <p><i>TODO - More to come.</i></p>
3109 <!-- ======================================================================= -->
3111 <a name="ptx">The PTX backend</a>
3116 <p>The PTX code generator lives in the lib/Target/PTX directory. It is
3117 currently a work-in-progress, but already supports most of the code
3118 generation functionality needed to generate correct PTX kernels for
3121 <p>The code generator can target PTX 2.0+, and shader model 1.0+. The
3122 PTX ISA Reference Manual is used as the primary source of ISA
3123 information, though an effort is made to make the output of the code
3124 generator match the output of the NVidia nvcc compiler, whenever
3127 <p>Code Generator Options:</p>
3128 <table border="1" cellspacing="0">
3131 <th>Description</th>
3134 <td><code>double</code></td>
3135 <td align="left">If enabled, the map_f64_to_f32 directive is
3136 disabled in the PTX output, allowing native double-precision
3140 <td><code>no-fma</code></td>
3141 <td align="left">Disable generation of Fused-Multiply Add
3142 instructions, which may be beneficial for some devices</td>
3145 <td><code>smxy / computexy</code></td>
3146 <td align="left">Set shader model/compute capability to x.y,
3147 e.g. sm20 or compute13</td>
3153 <li>Arithmetic instruction selection (including combo FMA)</li>
3154 <li>Bitwise instruction selection</li>
3155 <li>Control-flow instruction selection</li>
3156 <li>Function calls (only on SM 2.0+ and no return arguments)</li>
3157 <li>Addresses spaces (0 = global, 1 = constant, 2 = local, 4 =
3159 <li>Thread synchronization (bar.sync)</li>
3160 <li>Special register reads ([N]TID, [N]CTAID, PMx, CLOCK, etc.)</li>
3165 <li>Robust call instruction selection</li>
3166 <li>Stack frame allocation</li>
3167 <li>Device-specific instruction scheduling optimizations</li>
3175 <!-- *********************************************************************** -->
3178 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
3179 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
3180 <a href="http://validator.w3.org/check/referer"><img
3181 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
3183 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
3184 <a href="http://llvm.org/">The LLVM Compiler Infrastructure</a><br>
3185 Last modified: $Date$