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>
55 <li><a href="#mc">The "MC" Layer</a>
57 <li><a href="#mcstreamer">The <tt>MCStreamer</tt> API</a></li>
58 <li><a href="#mccontext">The <tt>MCContext</tt> class</a>
59 <li><a href="#mcsymbol">The <tt>MCSymbol</tt> class</a></li>
60 <li><a href="#mcsection">The <tt>MCSection</tt> class</a></li>
61 <li><a href="#mcinst">The <tt>MCInst</tt> class</a></li>
64 <li><a href="#codegenalgs">Target-independent code generation algorithms</a>
66 <li><a href="#instselect">Instruction Selection</a>
68 <li><a href="#selectiondag_intro">Introduction to SelectionDAGs</a></li>
69 <li><a href="#selectiondag_process">SelectionDAG Code Generation
71 <li><a href="#selectiondag_build">Initial SelectionDAG
73 <li><a href="#selectiondag_legalize_types">SelectionDAG LegalizeTypes Phase</a></li>
74 <li><a href="#selectiondag_legalize">SelectionDAG Legalize Phase</a></li>
75 <li><a href="#selectiondag_optimize">SelectionDAG Optimization
76 Phase: the DAG Combiner</a></li>
77 <li><a href="#selectiondag_select">SelectionDAG Select Phase</a></li>
78 <li><a href="#selectiondag_sched">SelectionDAG Scheduling and Formation
80 <li><a href="#selectiondag_future">Future directions for the
83 <li><a href="#liveintervals">Live Intervals</a>
85 <li><a href="#livevariable_analysis">Live Variable Analysis</a></li>
86 <li><a href="#liveintervals_analysis">Live Intervals Analysis</a></li>
88 <li><a href="#regalloc">Register Allocation</a>
90 <li><a href="#regAlloc_represent">How registers are represented in
92 <li><a href="#regAlloc_howTo">Mapping virtual registers to physical
94 <li><a href="#regAlloc_twoAddr">Handling two address instructions</a></li>
95 <li><a href="#regAlloc_ssaDecon">The SSA deconstruction phase</a></li>
96 <li><a href="#regAlloc_fold">Instruction folding</a></li>
97 <li><a href="#regAlloc_builtIn">Built in register allocators</a></li>
99 <li><a href="#codeemit">Code Emission</a></li>
100 <li><a href="#vliw_packetizer">VLIW Packetizer</a>
102 <li><a href="#vliw_mapping">Mapping from instructions to functional
104 <li><a href="#vliw_repr">How the packetization tables are
105 generated and used</a></li>
110 <li><a href="#nativeassembler">Implementing a Native Assembler</a></li>
112 <li><a href="#targetimpls">Target-specific Implementation Notes</a>
114 <li><a href="#targetfeatures">Target Feature Matrix</a></li>
115 <li><a href="#tailcallopt">Tail call optimization</a></li>
116 <li><a href="#sibcallopt">Sibling call optimization</a></li>
117 <li><a href="#x86">The X86 backend</a></li>
118 <li><a href="#ppc">The PowerPC backend</a>
120 <li><a href="#ppc_abi">LLVM PowerPC ABI</a></li>
121 <li><a href="#ppc_frame">Frame Layout</a></li>
122 <li><a href="#ppc_prolog">Prolog/Epilog</a></li>
123 <li><a href="#ppc_dynamic">Dynamic Allocation</a></li>
125 <li><a href="#ptx">The PTX backend</a></li>
130 <div class="doc_author">
131 <p>Written by the LLVM Team.</p>
134 <div class="doc_warning">
135 <p>Warning: This is a work in progress.</p>
138 <!-- *********************************************************************** -->
140 <a name="introduction">Introduction</a>
142 <!-- *********************************************************************** -->
146 <p>The LLVM target-independent code generator is a framework that provides a
147 suite of reusable components for translating the LLVM internal representation
148 to the machine code for a specified target—either in assembly form
149 (suitable for a static compiler) or in binary machine code format (usable for
150 a JIT compiler). The LLVM target-independent code generator consists of six
154 <li><a href="#targetdesc">Abstract target description</a> interfaces which
155 capture important properties about various aspects of the machine,
156 independently of how they will be used. These interfaces are defined in
157 <tt>include/llvm/Target/</tt>.</li>
159 <li>Classes used to represent the <a href="#codegendesc">code being
160 generated</a> for a target. These classes are intended to be abstract
161 enough to represent the machine code for <i>any</i> target machine. These
162 classes are defined in <tt>include/llvm/CodeGen/</tt>. At this level,
163 concepts like "constant pool entries" and "jump tables" are explicitly
166 <li>Classes and algorithms used to represent code as the object file level,
167 the <a href="#mc">MC Layer</a>. These classes represent assembly level
168 constructs like labels, sections, and instructions. At this level,
169 concepts like "constant pool entries" and "jump tables" don't exist.</li>
171 <li><a href="#codegenalgs">Target-independent algorithms</a> used to implement
172 various phases of native code generation (register allocation, scheduling,
173 stack frame representation, etc). This code lives
174 in <tt>lib/CodeGen/</tt>.</li>
176 <li><a href="#targetimpls">Implementations of the abstract target description
177 interfaces</a> for particular targets. These machine descriptions make
178 use of the components provided by LLVM, and can optionally provide custom
179 target-specific passes, to build complete code generators for a specific
180 target. Target descriptions live in <tt>lib/Target/</tt>.</li>
182 <li><a href="#jit">The target-independent JIT components</a>. The LLVM JIT is
183 completely target independent (it uses the <tt>TargetJITInfo</tt>
184 structure to interface for target-specific issues. The code for the
185 target-independent JIT lives in <tt>lib/ExecutionEngine/JIT</tt>.</li>
188 <p>Depending on which part of the code generator you are interested in working
189 on, different pieces of this will be useful to you. In any case, you should
190 be familiar with the <a href="#targetdesc">target description</a>
191 and <a href="#codegendesc">machine code representation</a> classes. If you
192 want to add a backend for a new target, you will need
193 to <a href="#targetimpls">implement the target description</a> classes for
194 your new target and understand the <a href="LangRef.html">LLVM code
195 representation</a>. If you are interested in implementing a
196 new <a href="#codegenalgs">code generation algorithm</a>, it should only
197 depend on the target-description and machine code representation classes,
198 ensuring that it is portable.</p>
200 <!-- ======================================================================= -->
202 <a name="required">Required components in the code generator</a>
207 <p>The two pieces of the LLVM code generator are the high-level interface to the
208 code generator and the set of reusable components that can be used to build
209 target-specific backends. The two most important interfaces
210 (<a href="#targetmachine"><tt>TargetMachine</tt></a>
211 and <a href="#targetdata"><tt>TargetData</tt></a>) are the only ones that are
212 required to be defined for a backend to fit into the LLVM system, but the
213 others must be defined if the reusable code generator components are going to
216 <p>This design has two important implications. The first is that LLVM can
217 support completely non-traditional code generation targets. For example, the
218 C backend does not require register allocation, instruction selection, or any
219 of the other standard components provided by the system. As such, it only
220 implements these two interfaces, and does its own thing. Another example of
221 a code generator like this is a (purely hypothetical) backend that converts
222 LLVM to the GCC RTL form and uses GCC to emit machine code for a target.</p>
224 <p>This design also implies that it is possible to design and implement
225 radically different code generators in the LLVM system that do not make use
226 of any of the built-in components. Doing so is not recommended at all, but
227 could be required for radically different targets that do not fit into the
228 LLVM machine description model: FPGAs for example.</p>
232 <!-- ======================================================================= -->
234 <a name="high-level-design">The high-level design of the code generator</a>
239 <p>The LLVM target-independent code generator is designed to support efficient
240 and quality code generation for standard register-based microprocessors.
241 Code generation in this model is divided into the following stages:</p>
244 <li><b><a href="#instselect">Instruction Selection</a></b> — This phase
245 determines an efficient way to express the input LLVM code in the target
246 instruction set. This stage produces the initial code for the program in
247 the target instruction set, then makes use of virtual registers in SSA
248 form and physical registers that represent any required register
249 assignments due to target constraints or calling conventions. This step
250 turns the LLVM code into a DAG of target instructions.</li>
252 <li><b><a href="#selectiondag_sched">Scheduling and Formation</a></b> —
253 This phase takes the DAG of target instructions produced by the
254 instruction selection phase, determines an ordering of the instructions,
255 then emits the instructions
256 as <tt><a href="#machineinstr">MachineInstr</a></tt>s with that ordering.
257 Note that we describe this in the <a href="#instselect">instruction
258 selection section</a> because it operates on
259 a <a href="#selectiondag_intro">SelectionDAG</a>.</li>
261 <li><b><a href="#ssamco">SSA-based Machine Code Optimizations</a></b> —
262 This optional stage consists of a series of machine-code optimizations
263 that operate on the SSA-form produced by the instruction selector.
264 Optimizations like modulo-scheduling or peephole optimization work
267 <li><b><a href="#regalloc">Register Allocation</a></b> — The target code
268 is transformed from an infinite virtual register file in SSA form to the
269 concrete register file used by the target. This phase introduces spill
270 code and eliminates all virtual register references from the program.</li>
272 <li><b><a href="#proepicode">Prolog/Epilog Code Insertion</a></b> — Once
273 the machine code has been generated for the function and the amount of
274 stack space required is known (used for LLVM alloca's and spill slots),
275 the prolog and epilog code for the function can be inserted and "abstract
276 stack location references" can be eliminated. This stage is responsible
277 for implementing optimizations like frame-pointer elimination and stack
280 <li><b><a href="#latemco">Late Machine Code Optimizations</a></b> —
281 Optimizations that operate on "final" machine code can go here, such as
282 spill code scheduling and peephole optimizations.</li>
284 <li><b><a href="#codeemit">Code Emission</a></b> — The final stage
285 actually puts out the code for the current function, either in the target
286 assembler format or in machine code.</li>
289 <p>The code generator is based on the assumption that the instruction selector
290 will use an optimal pattern matching selector to create high-quality
291 sequences of native instructions. Alternative code generator designs based
292 on pattern expansion and aggressive iterative peephole optimization are much
293 slower. This design permits efficient compilation (important for JIT
294 environments) and aggressive optimization (used when generating code offline)
295 by allowing components of varying levels of sophistication to be used for any
296 step of compilation.</p>
298 <p>In addition to these stages, target implementations can insert arbitrary
299 target-specific passes into the flow. For example, the X86 target uses a
300 special pass to handle the 80x87 floating point stack architecture. Other
301 targets with unusual requirements can be supported with custom passes as
306 <!-- ======================================================================= -->
308 <a name="tablegen">Using TableGen for target description</a>
313 <p>The target description classes require a detailed description of the target
314 architecture. These target descriptions often have a large amount of common
315 information (e.g., an <tt>add</tt> instruction is almost identical to a
316 <tt>sub</tt> instruction). In order to allow the maximum amount of
317 commonality to be factored out, the LLVM code generator uses
318 the <a href="TableGenFundamentals.html">TableGen</a> tool to describe big
319 chunks of the target machine, which allows the use of domain-specific and
320 target-specific abstractions to reduce the amount of repetition.</p>
322 <p>As LLVM continues to be developed and refined, we plan to move more and more
323 of the target description to the <tt>.td</tt> form. Doing so gives us a
324 number of advantages. The most important is that it makes it easier to port
325 LLVM because it reduces the amount of C++ code that has to be written, and
326 the surface area of the code generator that needs to be understood before
327 someone can get something working. Second, it makes it easier to change
328 things. In particular, if tables and other things are all emitted
329 by <tt>tblgen</tt>, we only need a change in one place (<tt>tblgen</tt>) to
330 update all of the targets to a new interface.</p>
336 <!-- *********************************************************************** -->
338 <a name="targetdesc">Target description classes</a>
340 <!-- *********************************************************************** -->
344 <p>The LLVM target description classes (located in the
345 <tt>include/llvm/Target</tt> directory) provide an abstract description of
346 the target machine independent of any particular client. These classes are
347 designed to capture the <i>abstract</i> properties of the target (such as the
348 instructions and registers it has), and do not incorporate any particular
349 pieces of code generation algorithms.</p>
351 <p>All of the target description classes (except the
352 <tt><a href="#targetdata">TargetData</a></tt> class) are designed to be
353 subclassed by the concrete target implementation, and have virtual methods
354 implemented. To get to these implementations, the
355 <tt><a href="#targetmachine">TargetMachine</a></tt> class provides accessors
356 that should be implemented by the target.</p>
358 <!-- ======================================================================= -->
360 <a name="targetmachine">The <tt>TargetMachine</tt> class</a>
365 <p>The <tt>TargetMachine</tt> class provides virtual methods that are used to
366 access the target-specific implementations of the various target description
367 classes via the <tt>get*Info</tt> methods (<tt>getInstrInfo</tt>,
368 <tt>getRegisterInfo</tt>, <tt>getFrameInfo</tt>, etc.). This class is
369 designed to be specialized by a concrete target implementation
370 (e.g., <tt>X86TargetMachine</tt>) which implements the various virtual
371 methods. The only required target description class is
372 the <a href="#targetdata"><tt>TargetData</tt></a> class, but if the code
373 generator components are to be used, the other interfaces should be
374 implemented as well.</p>
378 <!-- ======================================================================= -->
380 <a name="targetdata">The <tt>TargetData</tt> class</a>
385 <p>The <tt>TargetData</tt> class is the only required target description class,
386 and it is the only class that is not extensible (you cannot derived a new
387 class from it). <tt>TargetData</tt> specifies information about how the
388 target lays out memory for structures, the alignment requirements for various
389 data types, the size of pointers in the target, and whether the target is
390 little-endian or big-endian.</p>
394 <!-- ======================================================================= -->
396 <a name="targetlowering">The <tt>TargetLowering</tt> class</a>
401 <p>The <tt>TargetLowering</tt> class is used by SelectionDAG based instruction
402 selectors primarily to describe how LLVM code should be lowered to
403 SelectionDAG operations. Among other things, this class indicates:</p>
406 <li>an initial register class to use for various <tt>ValueType</tt>s,</li>
408 <li>which operations are natively supported by the target machine,</li>
410 <li>the return type of <tt>setcc</tt> operations,</li>
412 <li>the type to use for shift amounts, and</li>
414 <li>various high-level characteristics, like whether it is profitable to turn
415 division by a constant into a multiplication sequence</li>
420 <!-- ======================================================================= -->
422 <a name="targetregisterinfo">The <tt>TargetRegisterInfo</tt> class</a>
427 <p>The <tt>TargetRegisterInfo</tt> class is used to describe the register file
428 of the target and any interactions between the registers.</p>
430 <p>Registers in the code generator are represented in the code generator by
431 unsigned integers. Physical registers (those that actually exist in the
432 target description) are unique small numbers, and virtual registers are
433 generally large. Note that register #0 is reserved as a flag value.</p>
435 <p>Each register in the processor description has an associated
436 <tt>TargetRegisterDesc</tt> entry, which provides a textual name for the
437 register (used for assembly output and debugging dumps) and a set of aliases
438 (used to indicate whether one register overlaps with another).</p>
440 <p>In addition to the per-register description, the <tt>TargetRegisterInfo</tt>
441 class exposes a set of processor specific register classes (instances of the
442 <tt>TargetRegisterClass</tt> class). Each register class contains sets of
443 registers that have the same properties (for example, they are all 32-bit
444 integer registers). Each SSA virtual register created by the instruction
445 selector has an associated register class. When the register allocator runs,
446 it replaces virtual registers with a physical register in the set.</p>
448 <p>The target-specific implementations of these classes is auto-generated from
449 a <a href="TableGenFundamentals.html">TableGen</a> description of the
454 <!-- ======================================================================= -->
456 <a name="targetinstrinfo">The <tt>TargetInstrInfo</tt> class</a>
461 <p>The <tt>TargetInstrInfo</tt> class is used to describe the machine
462 instructions supported by the target. It is essentially an array of
463 <tt>TargetInstrDescriptor</tt> objects, each of which describes one
464 instruction the target supports. Descriptors define things like the mnemonic
465 for the opcode, the number of operands, the list of implicit register uses
466 and defs, whether the instruction has certain target-independent properties
467 (accesses memory, is commutable, etc), and holds any target-specific
472 <!-- ======================================================================= -->
474 <a name="targetframeinfo">The <tt>TargetFrameInfo</tt> class</a>
479 <p>The <tt>TargetFrameInfo</tt> class is used to provide information about the
480 stack frame layout of the target. It holds the direction of stack growth, the
481 known stack alignment on entry to each function, and the offset to the local
482 area. The offset to the local area is the offset from the stack pointer on
483 function entry to the first location where function data (local variables,
484 spill locations) can be stored.</p>
488 <!-- ======================================================================= -->
490 <a name="targetsubtarget">The <tt>TargetSubtarget</tt> class</a>
495 <p>The <tt>TargetSubtarget</tt> class is used to provide information about the
496 specific chip set being targeted. A sub-target informs code generation of
497 which instructions are supported, instruction latencies and instruction
498 execution itinerary; i.e., which processing units are used, in what order,
499 and for how long.</p>
504 <!-- ======================================================================= -->
506 <a name="targetjitinfo">The <tt>TargetJITInfo</tt> class</a>
511 <p>The <tt>TargetJITInfo</tt> class exposes an abstract interface used by the
512 Just-In-Time code generator to perform target-specific activities, such as
513 emitting stubs. If a <tt>TargetMachine</tt> supports JIT code generation, it
514 should provide one of these objects through the <tt>getJITInfo</tt>
521 <!-- *********************************************************************** -->
523 <a name="codegendesc">Machine code description classes</a>
525 <!-- *********************************************************************** -->
529 <p>At the high-level, LLVM code is translated to a machine specific
530 representation formed out of
531 <a href="#machinefunction"><tt>MachineFunction</tt></a>,
532 <a href="#machinebasicblock"><tt>MachineBasicBlock</tt></a>,
533 and <a href="#machineinstr"><tt>MachineInstr</tt></a> instances (defined
534 in <tt>include/llvm/CodeGen</tt>). This representation is completely target
535 agnostic, representing instructions in their most abstract form: an opcode
536 and a series of operands. This representation is designed to support both an
537 SSA representation for machine code, as well as a register allocated, non-SSA
540 <!-- ======================================================================= -->
542 <a name="machineinstr">The <tt>MachineInstr</tt> class</a>
547 <p>Target machine instructions are represented as instances of the
548 <tt>MachineInstr</tt> class. This class is an extremely abstract way of
549 representing machine instructions. In particular, it only keeps track of an
550 opcode number and a set of operands.</p>
552 <p>The opcode number is a simple unsigned integer that only has meaning to a
553 specific backend. All of the instructions for a target should be defined in
554 the <tt>*InstrInfo.td</tt> file for the target. The opcode enum values are
555 auto-generated from this description. The <tt>MachineInstr</tt> class does
556 not have any information about how to interpret the instruction (i.e., what
557 the semantics of the instruction are); for that you must refer to the
558 <tt><a href="#targetinstrinfo">TargetInstrInfo</a></tt> class.</p>
560 <p>The operands of a machine instruction can be of several different types: a
561 register reference, a constant integer, a basic block reference, etc. In
562 addition, a machine operand should be marked as a def or a use of the value
563 (though only registers are allowed to be defs).</p>
565 <p>By convention, the LLVM code generator orders instruction operands so that
566 all register definitions come before the register uses, even on architectures
567 that are normally printed in other orders. For example, the SPARC add
568 instruction: "<tt>add %i1, %i2, %i3</tt>" adds the "%i1", and "%i2" registers
569 and stores the result into the "%i3" register. In the LLVM code generator,
570 the operands should be stored as "<tt>%i3, %i1, %i2</tt>": with the
571 destination first.</p>
573 <p>Keeping destination (definition) operands at the beginning of the operand
574 list has several advantages. In particular, the debugging printer will print
575 the instruction like this:</p>
577 <div class="doc_code">
583 <p>Also if the first operand is a def, it is easier to <a href="#buildmi">create
584 instructions</a> whose only def is the first operand.</p>
586 <!-- _______________________________________________________________________ -->
588 <a name="buildmi">Using the <tt>MachineInstrBuilder.h</tt> functions</a>
593 <p>Machine instructions are created by using the <tt>BuildMI</tt> functions,
594 located in the <tt>include/llvm/CodeGen/MachineInstrBuilder.h</tt> file. The
595 <tt>BuildMI</tt> functions make it easy to build arbitrary machine
596 instructions. Usage of the <tt>BuildMI</tt> functions look like this:</p>
598 <div class="doc_code">
600 // Create a 'DestReg = mov 42' (rendered in X86 assembly as 'mov DestReg, 42')
601 // instruction. The '1' specifies how many operands will be added.
602 MachineInstr *MI = BuildMI(X86::MOV32ri, 1, DestReg).addImm(42);
604 // Create the same instr, but insert it at the end of a basic block.
605 MachineBasicBlock &MBB = ...
606 BuildMI(MBB, X86::MOV32ri, 1, DestReg).addImm(42);
608 // Create the same instr, but insert it before a specified iterator point.
609 MachineBasicBlock::iterator MBBI = ...
610 BuildMI(MBB, MBBI, X86::MOV32ri, 1, DestReg).addImm(42);
612 // Create a 'cmp Reg, 0' instruction, no destination reg.
613 MI = BuildMI(X86::CMP32ri, 2).addReg(Reg).addImm(0);
614 // Create an 'sahf' instruction which takes no operands and stores nothing.
615 MI = BuildMI(X86::SAHF, 0);
617 // Create a self looping branch instruction.
618 BuildMI(MBB, X86::JNE, 1).addMBB(&MBB);
622 <p>The key thing to remember with the <tt>BuildMI</tt> functions is that you
623 have to specify the number of operands that the machine instruction will
624 take. This allows for efficient memory allocation. You also need to specify
625 if operands default to be uses of values, not definitions. If you need to
626 add a definition operand (other than the optional destination register), you
627 must explicitly mark it as such:</p>
629 <div class="doc_code">
631 MI.addReg(Reg, RegState::Define);
637 <!-- _______________________________________________________________________ -->
639 <a name="fixedregs">Fixed (preassigned) registers</a>
644 <p>One important issue that the code generator needs to be aware of is the
645 presence of fixed registers. In particular, there are often places in the
646 instruction stream where the register allocator <em>must</em> arrange for a
647 particular value to be in a particular register. This can occur due to
648 limitations of the instruction set (e.g., the X86 can only do a 32-bit divide
649 with the <tt>EAX</tt>/<tt>EDX</tt> registers), or external factors like
650 calling conventions. In any case, the instruction selector should emit code
651 that copies a virtual register into or out of a physical register when
654 <p>For example, consider this simple LLVM example:</p>
656 <div class="doc_code">
658 define i32 @test(i32 %X, i32 %Y) {
665 <p>The X86 instruction selector produces this machine code for the <tt>div</tt>
666 and <tt>ret</tt> (use "<tt>llc X.bc -march=x86 -print-machineinstrs</tt>" to
669 <div class="doc_code">
672 %EAX = mov %reg1024 ;; Copy X (in reg1024) into EAX
673 %reg1027 = sar %reg1024, 31
674 %EDX = mov %reg1027 ;; Sign extend X into EDX
675 idiv %reg1025 ;; Divide by Y (in reg1025)
676 %reg1026 = mov %EAX ;; Read the result (Z) out of EAX
679 %EAX = mov %reg1026 ;; 32-bit return value goes in EAX
684 <p>By the end of code generation, the register allocator has coalesced the
685 registers and deleted the resultant identity moves producing the following
688 <div class="doc_code">
690 ;; X is in EAX, Y is in ECX
698 <p>This approach is extremely general (if it can handle the X86 architecture, it
699 can handle anything!) and allows all of the target specific knowledge about
700 the instruction stream to be isolated in the instruction selector. Note that
701 physical registers should have a short lifetime for good code generation, and
702 all physical registers are assumed dead on entry to and exit from basic
703 blocks (before register allocation). Thus, if you need a value to be live
704 across basic block boundaries, it <em>must</em> live in a virtual
709 <!-- _______________________________________________________________________ -->
711 <a name="ssa">Machine code in SSA form</a>
716 <p><tt>MachineInstr</tt>'s are initially selected in SSA-form, and are
717 maintained in SSA-form until register allocation happens. For the most part,
718 this is trivially simple since LLVM is already in SSA form; LLVM PHI nodes
719 become machine code PHI nodes, and virtual registers are only allowed to have
720 a single definition.</p>
722 <p>After register allocation, machine code is no longer in SSA-form because
723 there are no virtual registers left in the code.</p>
729 <!-- ======================================================================= -->
731 <a name="machinebasicblock">The <tt>MachineBasicBlock</tt> class</a>
736 <p>The <tt>MachineBasicBlock</tt> class contains a list of machine instructions
737 (<tt><a href="#machineinstr">MachineInstr</a></tt> instances). It roughly
738 corresponds to the LLVM code input to the instruction selector, but there can
739 be a one-to-many mapping (i.e. one LLVM basic block can map to multiple
740 machine basic blocks). The <tt>MachineBasicBlock</tt> class has a
741 "<tt>getBasicBlock</tt>" method, which returns the LLVM basic block that it
746 <!-- ======================================================================= -->
748 <a name="machinefunction">The <tt>MachineFunction</tt> class</a>
753 <p>The <tt>MachineFunction</tt> class contains a list of machine basic blocks
754 (<tt><a href="#machinebasicblock">MachineBasicBlock</a></tt> instances). It
755 corresponds one-to-one with the LLVM function input to the instruction
756 selector. In addition to a list of basic blocks,
757 the <tt>MachineFunction</tt> contains a a <tt>MachineConstantPool</tt>,
758 a <tt>MachineFrameInfo</tt>, a <tt>MachineFunctionInfo</tt>, and a
759 <tt>MachineRegisterInfo</tt>. See
760 <tt>include/llvm/CodeGen/MachineFunction.h</tt> for more information.</p>
766 <!-- *********************************************************************** -->
768 <a name="mc">The "MC" Layer</a>
770 <!-- *********************************************************************** -->
775 The MC Layer is used to represent and process code at the raw machine code
776 level, devoid of "high level" information like "constant pools", "jump tables",
777 "global variables" or anything like that. At this level, LLVM handles things
778 like label names, machine instructions, and sections in the object file. The
779 code in this layer is used for a number of important purposes: the tail end of
780 the code generator uses it to write a .s or .o file, and it is also used by the
781 llvm-mc tool to implement standalone machine code assemblers and disassemblers.
785 This section describes some of the important classes. There are also a number
786 of important subsystems that interact at this layer, they are described later
790 <!-- ======================================================================= -->
792 <a name="mcstreamer">The <tt>MCStreamer</tt> API</a>
798 MCStreamer is best thought of as an assembler API. It is an abstract API which
799 is <em>implemented</em> in different ways (e.g. to output a .s file, output an
800 ELF .o file, etc) but whose API correspond directly to what you see in a .s
801 file. MCStreamer has one method per directive, such as EmitLabel,
802 EmitSymbolAttribute, SwitchSection, EmitValue (for .byte, .word), etc, which
803 directly correspond to assembly level directives. It also has an
804 EmitInstruction method, which is used to output an MCInst to the streamer.
808 This API is most important for two clients: the llvm-mc stand-alone assembler is
809 effectively a parser that parses a line, then invokes a method on MCStreamer. In
810 the code generator, the <a href="#codeemit">Code Emission</a> phase of the code
811 generator lowers higher level LLVM IR and Machine* constructs down to the MC
812 layer, emitting directives through MCStreamer.</p>
815 On the implementation side of MCStreamer, there are two major implementations:
816 one for writing out a .s file (MCAsmStreamer), and one for writing out a .o
817 file (MCObjectStreamer). MCAsmStreamer is a straight-forward implementation
818 that prints out a directive for each method (e.g. EmitValue -> .byte), but
819 MCObjectStreamer implements a full assembler.
824 <!-- ======================================================================= -->
826 <a name="mccontext">The <tt>MCContext</tt> class</a>
832 The MCContext class is the owner of a variety of uniqued data structures at the
833 MC layer, including symbols, sections, etc. As such, this is the class that you
834 interact with to create symbols and sections. This class can not be subclassed.
839 <!-- ======================================================================= -->
841 <a name="mcsymbol">The <tt>MCSymbol</tt> class</a>
847 The MCSymbol class represents a symbol (aka label) in the assembly file. There
848 are two interesting kinds of symbols: assembler temporary symbols, and normal
849 symbols. Assembler temporary symbols are used and processed by the assembler
850 but are discarded when the object file is produced. The distinction is usually
851 represented by adding a prefix to the label, for example "L" labels are
852 assembler temporary labels in MachO.
855 <p>MCSymbols are created by MCContext and uniqued there. This means that
856 MCSymbols can be compared for pointer equivalence to find out if they are the
857 same symbol. Note that pointer inequality does not guarantee the labels will
858 end up at different addresses though. It's perfectly legal to output something
859 like this to the .s file:<p>
867 <p>In this case, both the foo and bar symbols will have the same address.</p>
871 <!-- ======================================================================= -->
873 <a name="mcsection">The <tt>MCSection</tt> class</a>
879 The MCSection class represents an object-file specific section. It is subclassed
880 by object file specific implementations (e.g. <tt>MCSectionMachO</tt>,
881 <tt>MCSectionCOFF</tt>, <tt>MCSectionELF</tt>) and these are created and uniqued
882 by MCContext. The MCStreamer has a notion of the current section, which can be
883 changed with the SwitchToSection method (which corresponds to a ".section"
884 directive in a .s file).
889 <!-- ======================================================================= -->
891 <a name="mcinst">The <tt>MCInst</tt> class</a>
897 The MCInst class is a target-independent representation of an instruction. It
898 is a simple class (much more so than <a href="#machineinstr">MachineInstr</a>)
899 that holds a target-specific opcode and a vector of MCOperands. MCOperand, in
900 turn, is a simple discriminated union of three cases: 1) a simple immediate,
901 2) a target register ID, 3) a symbolic expression (e.g. "Lfoo-Lbar+42") as an
905 <p>MCInst is the common currency used to represent machine instructions at the
906 MC layer. It is the type used by the instruction encoder, the instruction
907 printer, and the type generated by the assembly parser and disassembler.
914 <!-- *********************************************************************** -->
916 <a name="codegenalgs">Target-independent code generation algorithms</a>
918 <!-- *********************************************************************** -->
922 <p>This section documents the phases described in the
923 <a href="#high-level-design">high-level design of the code generator</a>.
924 It explains how they work and some of the rationale behind their design.</p>
926 <!-- ======================================================================= -->
928 <a name="instselect">Instruction Selection</a>
933 <p>Instruction Selection is the process of translating LLVM code presented to
934 the code generator into target-specific machine instructions. There are
935 several well-known ways to do this in the literature. LLVM uses a
936 SelectionDAG based instruction selector.</p>
938 <p>Portions of the DAG instruction selector are generated from the target
939 description (<tt>*.td</tt>) files. Our goal is for the entire instruction
940 selector to be generated from these <tt>.td</tt> files, though currently
941 there are still things that require custom C++ code.</p>
943 <!-- _______________________________________________________________________ -->
945 <a name="selectiondag_intro">Introduction to SelectionDAGs</a>
950 <p>The SelectionDAG provides an abstraction for code representation in a way
951 that is amenable to instruction selection using automatic techniques
952 (e.g. dynamic-programming based optimal pattern matching selectors). It is
953 also well-suited to other phases of code generation; in particular,
954 instruction scheduling (SelectionDAG's are very close to scheduling DAGs
955 post-selection). Additionally, the SelectionDAG provides a host
956 representation where a large variety of very-low-level (but
957 target-independent) <a href="#selectiondag_optimize">optimizations</a> may be
958 performed; ones which require extensive information about the instructions
959 efficiently supported by the target.</p>
961 <p>The SelectionDAG is a Directed-Acyclic-Graph whose nodes are instances of the
962 <tt>SDNode</tt> class. The primary payload of the <tt>SDNode</tt> is its
963 operation code (Opcode) that indicates what operation the node performs and
964 the operands to the operation. The various operation node types are
965 described at the top of the <tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt>
968 <p>Although most operations define a single value, each node in the graph may
969 define multiple values. For example, a combined div/rem operation will
970 define both the dividend and the remainder. Many other situations require
971 multiple values as well. Each node also has some number of operands, which
972 are edges to the node defining the used value. Because nodes may define
973 multiple values, edges are represented by instances of the <tt>SDValue</tt>
974 class, which is a <tt><SDNode, unsigned></tt> pair, indicating the node
975 and result value being used, respectively. Each value produced by
976 an <tt>SDNode</tt> has an associated <tt>MVT</tt> (Machine Value Type)
977 indicating what the type of the value is.</p>
979 <p>SelectionDAGs contain two different kinds of values: those that represent
980 data flow and those that represent control flow dependencies. Data values
981 are simple edges with an integer or floating point value type. Control edges
982 are represented as "chain" edges which are of type <tt>MVT::Other</tt>.
983 These edges provide an ordering between nodes that have side effects (such as
984 loads, stores, calls, returns, etc). All nodes that have side effects should
985 take a token chain as input and produce a new one as output. By convention,
986 token chain inputs are always operand #0, and chain results are always the
987 last value produced by an operation.</p>
989 <p>A SelectionDAG has designated "Entry" and "Root" nodes. The Entry node is
990 always a marker node with an Opcode of <tt>ISD::EntryToken</tt>. The Root
991 node is the final side-effecting node in the token chain. For example, in a
992 single basic block function it would be the return node.</p>
994 <p>One important concept for SelectionDAGs is the notion of a "legal" vs.
995 "illegal" DAG. A legal DAG for a target is one that only uses supported
996 operations and supported types. On a 32-bit PowerPC, for example, a DAG with
997 a value of type i1, i8, i16, or i64 would be illegal, as would a DAG that
998 uses a SREM or UREM operation. The
999 <a href="#selectinodag_legalize_types">legalize types</a> and
1000 <a href="#selectiondag_legalize">legalize operations</a> phases are
1001 responsible for turning an illegal DAG into a legal DAG.</p>
1005 <!-- _______________________________________________________________________ -->
1007 <a name="selectiondag_process">SelectionDAG Instruction Selection Process</a>
1012 <p>SelectionDAG-based instruction selection consists of the following steps:</p>
1015 <li><a href="#selectiondag_build">Build initial DAG</a> — This stage
1016 performs a simple translation from the input LLVM code to an illegal
1019 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> — This
1020 stage performs simple optimizations on the SelectionDAG to simplify it,
1021 and recognize meta instructions (like rotates
1022 and <tt>div</tt>/<tt>rem</tt> pairs) for targets that support these meta
1023 operations. This makes the resultant code more efficient and
1024 the <a href="#selectiondag_select">select instructions from DAG</a> phase
1025 (below) simpler.</li>
1027 <li><a href="#selectiondag_legalize_types">Legalize SelectionDAG Types</a>
1028 — This stage transforms SelectionDAG nodes to eliminate any types
1029 that are unsupported on the target.</li>
1031 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> — The
1032 SelectionDAG optimizer is run to clean up redundancies exposed by type
1035 <li><a href="#selectiondag_legalize">Legalize SelectionDAG Ops</a> —
1036 This stage transforms SelectionDAG nodes to eliminate any operations
1037 that are unsupported on the target.</li>
1039 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> — The
1040 SelectionDAG optimizer is run to eliminate inefficiencies introduced by
1041 operation legalization.</li>
1043 <li><a href="#selectiondag_select">Select instructions from DAG</a> —
1044 Finally, the target instruction selector matches the DAG operations to
1045 target instructions. This process translates the target-independent input
1046 DAG into another DAG of target instructions.</li>
1048 <li><a href="#selectiondag_sched">SelectionDAG Scheduling and Formation</a>
1049 — The last phase assigns a linear order to the instructions in the
1050 target-instruction DAG and emits them into the MachineFunction being
1051 compiled. This step uses traditional prepass scheduling techniques.</li>
1054 <p>After all of these steps are complete, the SelectionDAG is destroyed and the
1055 rest of the code generation passes are run.</p>
1057 <p>One great way to visualize what is going on here is to take advantage of a
1058 few LLC command line options. The following options pop up a window
1059 displaying the SelectionDAG at specific times (if you only get errors printed
1060 to the console while using this, you probably
1061 <a href="ProgrammersManual.html#ViewGraph">need to configure your system</a>
1062 to add support for it).</p>
1065 <li><tt>-view-dag-combine1-dags</tt> displays the DAG after being built,
1066 before the first optimization pass.</li>
1068 <li><tt>-view-legalize-dags</tt> displays the DAG before Legalization.</li>
1070 <li><tt>-view-dag-combine2-dags</tt> displays the DAG before the second
1071 optimization pass.</li>
1073 <li><tt>-view-isel-dags</tt> displays the DAG before the Select phase.</li>
1075 <li><tt>-view-sched-dags</tt> displays the DAG before Scheduling.</li>
1078 <p>The <tt>-view-sunit-dags</tt> displays the Scheduler's dependency graph.
1079 This graph is based on the final SelectionDAG, with nodes that must be
1080 scheduled together bundled into a single scheduling-unit node, and with
1081 immediate operands and other nodes that aren't relevant for scheduling
1086 <!-- _______________________________________________________________________ -->
1088 <a name="selectiondag_build">Initial SelectionDAG Construction</a>
1093 <p>The initial SelectionDAG is naïvely peephole expanded from the LLVM
1094 input by the <tt>SelectionDAGLowering</tt> class in the
1095 <tt>lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp</tt> file. The intent of
1096 this pass is to expose as much low-level, target-specific details to the
1097 SelectionDAG as possible. This pass is mostly hard-coded (e.g. an
1098 LLVM <tt>add</tt> turns into an <tt>SDNode add</tt> while a
1099 <tt>getelementptr</tt> is expanded into the obvious arithmetic). This pass
1100 requires target-specific hooks to lower calls, returns, varargs, etc. For
1101 these features, the <tt><a href="#targetlowering">TargetLowering</a></tt>
1102 interface is used.</p>
1106 <!-- _______________________________________________________________________ -->
1108 <a name="selectiondag_legalize_types">SelectionDAG LegalizeTypes Phase</a>
1113 <p>The Legalize phase is in charge of converting a DAG to only use the types
1114 that are natively supported by the target.</p>
1116 <p>There are two main ways of converting values of unsupported scalar types to
1117 values of supported types: converting small types to larger types
1118 ("promoting"), and breaking up large integer types into smaller ones
1119 ("expanding"). For example, a target might require that all f32 values are
1120 promoted to f64 and that all i1/i8/i16 values are promoted to i32. The same
1121 target might require that all i64 values be expanded into pairs of i32
1122 values. These changes can insert sign and zero extensions as needed to make
1123 sure that the final code has the same behavior as the input.</p>
1125 <p>There are two main ways of converting values of unsupported vector types to
1126 value of supported types: splitting vector types, multiple times if
1127 necessary, until a legal type is found, and extending vector types by adding
1128 elements to the end to round them out to legal types ("widening"). If a
1129 vector gets split all the way down to single-element parts with no supported
1130 vector type being found, the elements are converted to scalars
1131 ("scalarizing").</p>
1133 <p>A target implementation tells the legalizer which types are supported (and
1134 which register class to use for them) by calling the
1135 <tt>addRegisterClass</tt> method in its TargetLowering constructor.</p>
1139 <!-- _______________________________________________________________________ -->
1141 <a name="selectiondag_legalize">SelectionDAG Legalize Phase</a>
1146 <p>The Legalize phase is in charge of converting a DAG to only use the
1147 operations that are natively supported by the target.</p>
1149 <p>Targets often have weird constraints, such as not supporting every operation
1150 on every supported datatype (e.g. X86 does not support byte conditional moves
1151 and PowerPC does not support sign-extending loads from a 16-bit memory
1152 location). Legalize takes care of this by open-coding another sequence of
1153 operations to emulate the operation ("expansion"), by promoting one type to a
1154 larger type that supports the operation ("promotion"), or by using a
1155 target-specific hook to implement the legalization ("custom").</p>
1157 <p>A target implementation tells the legalizer which operations are not
1158 supported (and which of the above three actions to take) by calling the
1159 <tt>setOperationAction</tt> method in its <tt>TargetLowering</tt>
1162 <p>Prior to the existence of the Legalize passes, we required that every target
1163 <a href="#selectiondag_optimize">selector</a> supported and handled every
1164 operator and type even if they are not natively supported. The introduction
1165 of the Legalize phases allows all of the canonicalization patterns to be
1166 shared across targets, and makes it very easy to optimize the canonicalized
1167 code because it is still in the form of a DAG.</p>
1171 <!-- _______________________________________________________________________ -->
1173 <a name="selectiondag_optimize">
1174 SelectionDAG Optimization Phase: the DAG Combiner
1180 <p>The SelectionDAG optimization phase is run multiple times for code
1181 generation, immediately after the DAG is built and once after each
1182 legalization. The first run of the pass allows the initial code to be
1183 cleaned up (e.g. performing optimizations that depend on knowing that the
1184 operators have restricted type inputs). Subsequent runs of the pass clean up
1185 the messy code generated by the Legalize passes, which allows Legalize to be
1186 very simple (it can focus on making code legal instead of focusing on
1187 generating <em>good</em> and legal code).</p>
1189 <p>One important class of optimizations performed is optimizing inserted sign
1190 and zero extension instructions. We currently use ad-hoc techniques, but
1191 could move to more rigorous techniques in the future. Here are some good
1192 papers on the subject:</p>
1194 <p>"<a href="http://www.eecs.harvard.edu/~nr/pubs/widen-abstract.html">Widening
1195 integer arithmetic</a>"<br>
1196 Kevin Redwine and Norman Ramsey<br>
1197 International Conference on Compiler Construction (CC) 2004</p>
1199 <p>"<a href="http://portal.acm.org/citation.cfm?doid=512529.512552">Effective
1200 sign extension elimination</a>"<br>
1201 Motohiro Kawahito, Hideaki Komatsu, and Toshio Nakatani<br>
1202 Proceedings of the ACM SIGPLAN 2002 Conference on Programming Language Design
1203 and Implementation.</p>
1207 <!-- _______________________________________________________________________ -->
1209 <a name="selectiondag_select">SelectionDAG Select Phase</a>
1214 <p>The Select phase is the bulk of the target-specific code for instruction
1215 selection. This phase takes a legal SelectionDAG as input, pattern matches
1216 the instructions supported by the target to this DAG, and produces a new DAG
1217 of target code. For example, consider the following LLVM fragment:</p>
1219 <div class="doc_code">
1221 %t1 = fadd float %W, %X
1222 %t2 = fmul float %t1, %Y
1223 %t3 = fadd float %t2, %Z
1227 <p>This LLVM code corresponds to a SelectionDAG that looks basically like
1230 <div class="doc_code">
1232 (fadd:f32 (fmul:f32 (fadd:f32 W, X), Y), Z)
1236 <p>If a target supports floating point multiply-and-add (FMA) operations, one of
1237 the adds can be merged with the multiply. On the PowerPC, for example, the
1238 output of the instruction selector might look like this DAG:</p>
1240 <div class="doc_code">
1242 (FMADDS (FADDS W, X), Y, Z)
1246 <p>The <tt>FMADDS</tt> instruction is a ternary instruction that multiplies its
1247 first two operands and adds the third (as single-precision floating-point
1248 numbers). The <tt>FADDS</tt> instruction is a simple binary single-precision
1249 add instruction. To perform this pattern match, the PowerPC backend includes
1250 the following instruction definitions:</p>
1252 <div class="doc_code">
1254 def FMADDS : AForm_1<59, 29,
1255 (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
1256 "fmadds $FRT, $FRA, $FRC, $FRB",
1257 [<b>(set F4RC:$FRT, (fadd (fmul F4RC:$FRA, F4RC:$FRC),
1258 F4RC:$FRB))</b>]>;
1259 def FADDS : AForm_2<59, 21,
1260 (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
1261 "fadds $FRT, $FRA, $FRB",
1262 [<b>(set F4RC:$FRT, (fadd F4RC:$FRA, F4RC:$FRB))</b>]>;
1266 <p>The portion of the instruction definition in bold indicates the pattern used
1267 to match the instruction. The DAG operators
1268 (like <tt>fmul</tt>/<tt>fadd</tt>) are defined in
1269 the <tt>include/llvm/Target/TargetSelectionDAG.td</tt> file. "
1270 <tt>F4RC</tt>" is the register class of the input and result values.</p>
1272 <p>The TableGen DAG instruction selector generator reads the instruction
1273 patterns in the <tt>.td</tt> file and automatically builds parts of the
1274 pattern matching code for your target. It has the following strengths:</p>
1277 <li>At compiler-compiler time, it analyzes your instruction patterns and tells
1278 you if your patterns make sense or not.</li>
1280 <li>It can handle arbitrary constraints on operands for the pattern match. In
1281 particular, it is straight-forward to say things like "match any immediate
1282 that is a 13-bit sign-extended value". For examples, see the
1283 <tt>immSExt16</tt> and related <tt>tblgen</tt> classes in the PowerPC
1286 <li>It knows several important identities for the patterns defined. For
1287 example, it knows that addition is commutative, so it allows the
1288 <tt>FMADDS</tt> pattern above to match "<tt>(fadd X, (fmul Y, Z))</tt>" as
1289 well as "<tt>(fadd (fmul X, Y), Z)</tt>", without the target author having
1290 to specially handle this case.</li>
1292 <li>It has a full-featured type-inferencing system. In particular, you should
1293 rarely have to explicitly tell the system what type parts of your patterns
1294 are. In the <tt>FMADDS</tt> case above, we didn't have to tell
1295 <tt>tblgen</tt> that all of the nodes in the pattern are of type 'f32'.
1296 It was able to infer and propagate this knowledge from the fact that
1297 <tt>F4RC</tt> has type 'f32'.</li>
1299 <li>Targets can define their own (and rely on built-in) "pattern fragments".
1300 Pattern fragments are chunks of reusable patterns that get inlined into
1301 your patterns during compiler-compiler time. For example, the integer
1302 "<tt>(not x)</tt>" operation is actually defined as a pattern fragment
1303 that expands as "<tt>(xor x, -1)</tt>", since the SelectionDAG does not
1304 have a native '<tt>not</tt>' operation. Targets can define their own
1305 short-hand fragments as they see fit. See the definition of
1306 '<tt>not</tt>' and '<tt>ineg</tt>' for examples.</li>
1308 <li>In addition to instructions, targets can specify arbitrary patterns that
1309 map to one or more instructions using the 'Pat' class. For example, the
1310 PowerPC has no way to load an arbitrary integer immediate into a register
1311 in one instruction. To tell tblgen how to do this, it defines:
1314 <div class="doc_code">
1316 // Arbitrary immediate support. Implement in terms of LIS/ORI.
1317 def : Pat<(i32 imm:$imm),
1318 (ORI (LIS (HI16 imm:$imm)), (LO16 imm:$imm))>;
1322 If none of the single-instruction patterns for loading an immediate into a
1323 register match, this will be used. This rule says "match an arbitrary i32
1324 immediate, turning it into an <tt>ORI</tt> ('or a 16-bit immediate') and
1325 an <tt>LIS</tt> ('load 16-bit immediate, where the immediate is shifted to
1326 the left 16 bits') instruction". To make this work, the
1327 <tt>LO16</tt>/<tt>HI16</tt> node transformations are used to manipulate
1328 the input immediate (in this case, take the high or low 16-bits of the
1331 <li>While the system does automate a lot, it still allows you to write custom
1332 C++ code to match special cases if there is something that is hard to
1336 <p>While it has many strengths, the system currently has some limitations,
1337 primarily because it is a work in progress and is not yet finished:</p>
1340 <li>Overall, there is no way to define or match SelectionDAG nodes that define
1341 multiple values (e.g. <tt>SMUL_LOHI</tt>, <tt>LOAD</tt>, <tt>CALL</tt>,
1342 etc). This is the biggest reason that you currently still <em>have
1343 to</em> write custom C++ code for your instruction selector.</li>
1345 <li>There is no great way to support matching complex addressing modes yet.
1346 In the future, we will extend pattern fragments to allow them to define
1347 multiple values (e.g. the four operands of the <a href="#x86_memory">X86
1348 addressing mode</a>, which are currently matched with custom C++ code).
1349 In addition, we'll extend fragments so that a fragment can match multiple
1350 different patterns.</li>
1352 <li>We don't automatically infer flags like isStore/isLoad yet.</li>
1354 <li>We don't automatically generate the set of supported registers and
1355 operations for the <a href="#selectiondag_legalize">Legalizer</a>
1358 <li>We don't have a way of tying in custom legalized nodes yet.</li>
1361 <p>Despite these limitations, the instruction selector generator is still quite
1362 useful for most of the binary and logical operations in typical instruction
1363 sets. If you run into any problems or can't figure out how to do something,
1364 please let Chris know!</p>
1368 <!-- _______________________________________________________________________ -->
1370 <a name="selectiondag_sched">SelectionDAG Scheduling and Formation Phase</a>
1375 <p>The scheduling phase takes the DAG of target instructions from the selection
1376 phase and assigns an order. The scheduler can pick an order depending on
1377 various constraints of the machines (i.e. order for minimal register pressure
1378 or try to cover instruction latencies). Once an order is established, the
1379 DAG is converted to a list
1380 of <tt><a href="#machineinstr">MachineInstr</a></tt>s and the SelectionDAG is
1383 <p>Note that this phase is logically separate from the instruction selection
1384 phase, but is tied to it closely in the code because it operates on
1389 <!-- _______________________________________________________________________ -->
1391 <a name="selectiondag_future">Future directions for the SelectionDAG</a>
1397 <li>Optional function-at-a-time selection.</li>
1399 <li>Auto-generate entire selector from <tt>.td</tt> file.</li>
1406 <!-- ======================================================================= -->
1408 <a name="ssamco">SSA-based Machine Code Optimizations</a>
1410 <div><p>To Be Written</p></div>
1412 <!-- ======================================================================= -->
1414 <a name="liveintervals">Live Intervals</a>
1419 <p>Live Intervals are the ranges (intervals) where a variable is <i>live</i>.
1420 They are used by some <a href="#regalloc">register allocator</a> passes to
1421 determine if two or more virtual registers which require the same physical
1422 register are live at the same point in the program (i.e., they conflict).
1423 When this situation occurs, one virtual register must be <i>spilled</i>.</p>
1425 <!-- _______________________________________________________________________ -->
1427 <a name="livevariable_analysis">Live Variable Analysis</a>
1432 <p>The first step in determining the live intervals of variables is to calculate
1433 the set of registers that are immediately dead after the instruction (i.e.,
1434 the instruction calculates the value, but it is never used) and the set of
1435 registers that are used by the instruction, but are never used after the
1436 instruction (i.e., they are killed). Live variable information is computed
1437 for each <i>virtual</i> register and <i>register allocatable</i> physical
1438 register in the function. This is done in a very efficient manner because it
1439 uses SSA to sparsely compute lifetime information for virtual registers
1440 (which are in SSA form) and only has to track physical registers within a
1441 block. Before register allocation, LLVM can assume that physical registers
1442 are only live within a single basic block. This allows it to do a single,
1443 local analysis to resolve physical register lifetimes within each basic
1444 block. If a physical register is not register allocatable (e.g., a stack
1445 pointer or condition codes), it is not tracked.</p>
1447 <p>Physical registers may be live in to or out of a function. Live in values are
1448 typically arguments in registers. Live out values are typically return values
1449 in registers. Live in values are marked as such, and are given a dummy
1450 "defining" instruction during live intervals analysis. If the last basic
1451 block of a function is a <tt>return</tt>, then it's marked as using all live
1452 out values in the function.</p>
1454 <p><tt>PHI</tt> nodes need to be handled specially, because the calculation of
1455 the live variable information from a depth first traversal of the CFG of the
1456 function won't guarantee that a virtual register used by the <tt>PHI</tt>
1457 node is defined before it's used. When a <tt>PHI</tt> node is encountered,
1458 only the definition is handled, because the uses will be handled in other
1461 <p>For each <tt>PHI</tt> node of the current basic block, we simulate an
1462 assignment at the end of the current basic block and traverse the successor
1463 basic blocks. If a successor basic block has a <tt>PHI</tt> node and one of
1464 the <tt>PHI</tt> node's operands is coming from the current basic block, then
1465 the variable is marked as <i>alive</i> within the current basic block and all
1466 of its predecessor basic blocks, until the basic block with the defining
1467 instruction is encountered.</p>
1471 <!-- _______________________________________________________________________ -->
1473 <a name="liveintervals_analysis">Live Intervals Analysis</a>
1478 <p>We now have the information available to perform the live intervals analysis
1479 and build the live intervals themselves. We start off by numbering the basic
1480 blocks and machine instructions. We then handle the "live-in" values. These
1481 are in physical registers, so the physical register is assumed to be killed
1482 by the end of the basic block. Live intervals for virtual registers are
1483 computed for some ordering of the machine instructions <tt>[1, N]</tt>. A
1484 live interval is an interval <tt>[i, j)</tt>, where <tt>1 <= i <= j
1485 < N</tt>, for which a variable is live.</p>
1487 <p><i><b>More to come...</b></i></p>
1493 <!-- ======================================================================= -->
1495 <a name="regalloc">Register Allocation</a>
1500 <p>The <i>Register Allocation problem</i> consists in mapping a program
1501 <i>P<sub>v</sub></i>, that can use an unbounded number of virtual registers,
1502 to a program <i>P<sub>p</sub></i> that contains a finite (possibly small)
1503 number of physical registers. Each target architecture has a different number
1504 of physical registers. If the number of physical registers is not enough to
1505 accommodate all the virtual registers, some of them will have to be mapped
1506 into memory. These virtuals are called <i>spilled virtuals</i>.</p>
1508 <!-- _______________________________________________________________________ -->
1511 <a name="regAlloc_represent">How registers are represented in LLVM</a>
1516 <p>In LLVM, physical registers are denoted by integer numbers that normally
1517 range from 1 to 1023. To see how this numbering is defined for a particular
1518 architecture, you can read the <tt>GenRegisterNames.inc</tt> file for that
1519 architecture. For instance, by
1520 inspecting <tt>lib/Target/X86/X86GenRegisterNames.inc</tt> we see that the
1521 32-bit register <tt>EAX</tt> is denoted by 15, and the MMX register
1522 <tt>MM0</tt> is mapped to 48.</p>
1524 <p>Some architectures contain registers that share the same physical location. A
1525 notable example is the X86 platform. For instance, in the X86 architecture,
1526 the registers <tt>EAX</tt>, <tt>AX</tt> and <tt>AL</tt> share the first eight
1527 bits. These physical registers are marked as <i>aliased</i> in LLVM. Given a
1528 particular architecture, you can check which registers are aliased by
1529 inspecting its <tt>RegisterInfo.td</tt> file. Moreover, the method
1530 <tt>TargetRegisterInfo::getAliasSet(p_reg)</tt> returns an array containing
1531 all the physical registers aliased to the register <tt>p_reg</tt>.</p>
1533 <p>Physical registers, in LLVM, are grouped in <i>Register Classes</i>.
1534 Elements in the same register class are functionally equivalent, and can be
1535 interchangeably used. Each virtual register can only be mapped to physical
1536 registers of a particular class. For instance, in the X86 architecture, some
1537 virtuals can only be allocated to 8 bit registers. A register class is
1538 described by <tt>TargetRegisterClass</tt> objects. To discover if a virtual
1539 register is compatible with a given physical, this code can be used:</p>
1541 <div class="doc_code">
1543 bool RegMapping_Fer::compatible_class(MachineFunction &mf,
1546 assert(TargetRegisterInfo::isPhysicalRegister(p_reg) &&
1547 "Target register must be physical");
1548 const TargetRegisterClass *trc = mf.getRegInfo().getRegClass(v_reg);
1549 return trc->contains(p_reg);
1554 <p>Sometimes, mostly for debugging purposes, it is useful to change the number
1555 of physical registers available in the target architecture. This must be done
1556 statically, inside the <tt>TargetRegsterInfo.td</tt> file. Just <tt>grep</tt>
1557 for <tt>RegisterClass</tt>, the last parameter of which is a list of
1558 registers. Just commenting some out is one simple way to avoid them being
1559 used. A more polite way is to explicitly exclude some registers from
1560 the <i>allocation order</i>. See the definition of the <tt>GR8</tt> register
1561 class in <tt>lib/Target/X86/X86RegisterInfo.td</tt> for an example of this.
1564 <p>Virtual registers are also denoted by integer numbers. Contrary to physical
1565 registers, different virtual registers never share the same number. Whereas
1566 physical registers are statically defined in a <tt>TargetRegisterInfo.td</tt>
1567 file and cannot be created by the application developer, that is not the case
1568 with virtual registers. In order to create new virtual registers, use the
1569 method <tt>MachineRegisterInfo::createVirtualRegister()</tt>. This method
1570 will return a new virtual register. Use an <tt>IndexedMap<Foo,
1571 VirtReg2IndexFunctor></tt> to hold information per virtual register. If you
1572 need to enumerate all virtual registers, use the function
1573 <tt>TargetRegisterInfo::index2VirtReg()</tt> to find the virtual register
1576 <div class="doc_code">
1578 for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
1579 unsigned VirtReg = TargetRegisterInfo::index2VirtReg(i);
1585 <p>Before register allocation, the operands of an instruction are mostly virtual
1586 registers, although physical registers may also be used. In order to check if
1587 a given machine operand is a register, use the boolean
1588 function <tt>MachineOperand::isRegister()</tt>. To obtain the integer code of
1589 a register, use <tt>MachineOperand::getReg()</tt>. An instruction may define
1590 or use a register. For instance, <tt>ADD reg:1026 := reg:1025 reg:1024</tt>
1591 defines the registers 1024, and uses registers 1025 and 1026. Given a
1592 register operand, the method <tt>MachineOperand::isUse()</tt> informs if that
1593 register is being used by the instruction. The
1594 method <tt>MachineOperand::isDef()</tt> informs if that registers is being
1597 <p>We will call physical registers present in the LLVM bitcode before register
1598 allocation <i>pre-colored registers</i>. Pre-colored registers are used in
1599 many different situations, for instance, to pass parameters of functions
1600 calls, and to store results of particular instructions. There are two types
1601 of pre-colored registers: the ones <i>implicitly</i> defined, and
1602 those <i>explicitly</i> defined. Explicitly defined registers are normal
1603 operands, and can be accessed
1604 with <tt>MachineInstr::getOperand(int)::getReg()</tt>. In order to check
1605 which registers are implicitly defined by an instruction, use
1606 the <tt>TargetInstrInfo::get(opcode)::ImplicitDefs</tt>,
1607 where <tt>opcode</tt> is the opcode of the target instruction. One important
1608 difference between explicit and implicit physical registers is that the
1609 latter are defined statically for each instruction, whereas the former may
1610 vary depending on the program being compiled. For example, an instruction
1611 that represents a function call will always implicitly define or use the same
1612 set of physical registers. To read the registers implicitly used by an
1614 use <tt>TargetInstrInfo::get(opcode)::ImplicitUses</tt>. Pre-colored
1615 registers impose constraints on any register allocation algorithm. The
1616 register allocator must make sure that none of them are overwritten by
1617 the values of virtual registers while still alive.</p>
1621 <!-- _______________________________________________________________________ -->
1624 <a name="regAlloc_howTo">Mapping virtual registers to physical registers</a>
1629 <p>There are two ways to map virtual registers to physical registers (or to
1630 memory slots). The first way, that we will call <i>direct mapping</i>, is
1631 based on the use of methods of the classes <tt>TargetRegisterInfo</tt>,
1632 and <tt>MachineOperand</tt>. The second way, that we will call <i>indirect
1633 mapping</i>, relies on the <tt>VirtRegMap</tt> class in order to insert loads
1634 and stores sending and getting values to and from memory.</p>
1636 <p>The direct mapping provides more flexibility to the developer of the register
1637 allocator; however, it is more error prone, and demands more implementation
1638 work. Basically, the programmer will have to specify where load and store
1639 instructions should be inserted in the target function being compiled in
1640 order to get and store values in memory. To assign a physical register to a
1641 virtual register present in a given operand,
1642 use <tt>MachineOperand::setReg(p_reg)</tt>. To insert a store instruction,
1643 use <tt>TargetInstrInfo::storeRegToStackSlot(...)</tt>, and to insert a
1644 load instruction, use <tt>TargetInstrInfo::loadRegFromStackSlot</tt>.</p>
1646 <p>The indirect mapping shields the application developer from the complexities
1647 of inserting load and store instructions. In order to map a virtual register
1648 to a physical one, use <tt>VirtRegMap::assignVirt2Phys(vreg, preg)</tt>. In
1649 order to map a certain virtual register to memory,
1650 use <tt>VirtRegMap::assignVirt2StackSlot(vreg)</tt>. This method will return
1651 the stack slot where <tt>vreg</tt>'s value will be located. If it is
1652 necessary to map another virtual register to the same stack slot,
1653 use <tt>VirtRegMap::assignVirt2StackSlot(vreg, stack_location)</tt>. One
1654 important point to consider when using the indirect mapping, is that even if
1655 a virtual register is mapped to memory, it still needs to be mapped to a
1656 physical register. This physical register is the location where the virtual
1657 register is supposed to be found before being stored or after being
1660 <p>If the indirect strategy is used, after all the virtual registers have been
1661 mapped to physical registers or stack slots, it is necessary to use a spiller
1662 object to place load and store instructions in the code. Every virtual that
1663 has been mapped to a stack slot will be stored to memory after been defined
1664 and will be loaded before being used. The implementation of the spiller tries
1665 to recycle load/store instructions, avoiding unnecessary instructions. For an
1666 example of how to invoke the spiller,
1667 see <tt>RegAllocLinearScan::runOnMachineFunction</tt>
1668 in <tt>lib/CodeGen/RegAllocLinearScan.cpp</tt>.</p>
1672 <!-- _______________________________________________________________________ -->
1674 <a name="regAlloc_twoAddr">Handling two address instructions</a>
1679 <p>With very rare exceptions (e.g., function calls), the LLVM machine code
1680 instructions are three address instructions. That is, each instruction is
1681 expected to define at most one register, and to use at most two registers.
1682 However, some architectures use two address instructions. In this case, the
1683 defined register is also one of the used register. For instance, an
1684 instruction such as <tt>ADD %EAX, %EBX</tt>, in X86 is actually equivalent
1685 to <tt>%EAX = %EAX + %EBX</tt>.</p>
1687 <p>In order to produce correct code, LLVM must convert three address
1688 instructions that represent two address instructions into true two address
1689 instructions. LLVM provides the pass <tt>TwoAddressInstructionPass</tt> for
1690 this specific purpose. It must be run before register allocation takes
1691 place. After its execution, the resulting code may no longer be in SSA
1692 form. This happens, for instance, in situations where an instruction such
1693 as <tt>%a = ADD %b %c</tt> is converted to two instructions such as:</p>
1695 <div class="doc_code">
1702 <p>Notice that, internally, the second instruction is represented as
1703 <tt>ADD %a[def/use] %c</tt>. I.e., the register operand <tt>%a</tt> is both
1704 used and defined by the instruction.</p>
1708 <!-- _______________________________________________________________________ -->
1710 <a name="regAlloc_ssaDecon">The SSA deconstruction phase</a>
1715 <p>An important transformation that happens during register allocation is called
1716 the <i>SSA Deconstruction Phase</i>. The SSA form simplifies many analyses
1717 that are performed on the control flow graph of programs. However,
1718 traditional instruction sets do not implement PHI instructions. Thus, in
1719 order to generate executable code, compilers must replace PHI instructions
1720 with other instructions that preserve their semantics.</p>
1722 <p>There are many ways in which PHI instructions can safely be removed from the
1723 target code. The most traditional PHI deconstruction algorithm replaces PHI
1724 instructions with copy instructions. That is the strategy adopted by
1725 LLVM. The SSA deconstruction algorithm is implemented
1726 in <tt>lib/CodeGen/PHIElimination.cpp</tt>. In order to invoke this pass, the
1727 identifier <tt>PHIEliminationID</tt> must be marked as required in the code
1728 of the register allocator.</p>
1732 <!-- _______________________________________________________________________ -->
1734 <a name="regAlloc_fold">Instruction folding</a>
1739 <p><i>Instruction folding</i> is an optimization performed during register
1740 allocation that removes unnecessary copy instructions. For instance, a
1741 sequence of instructions such as:</p>
1743 <div class="doc_code">
1745 %EBX = LOAD %mem_address
1750 <p>can be safely substituted by the single instruction:</p>
1752 <div class="doc_code">
1754 %EAX = LOAD %mem_address
1758 <p>Instructions can be folded with
1759 the <tt>TargetRegisterInfo::foldMemoryOperand(...)</tt> method. Care must be
1760 taken when folding instructions; a folded instruction can be quite different
1762 instruction. See <tt>LiveIntervals::addIntervalsForSpills</tt>
1763 in <tt>lib/CodeGen/LiveIntervalAnalysis.cpp</tt> for an example of its
1768 <!-- _______________________________________________________________________ -->
1771 <a name="regAlloc_builtIn">Built in register allocators</a>
1776 <p>The LLVM infrastructure provides the application developer with three
1777 different register allocators:</p>
1780 <li><i>Fast</i> — This register allocator is the default for debug
1781 builds. It allocates registers on a basic block level, attempting to keep
1782 values in registers and reusing registers as appropriate.</li>
1784 <li><i>Basic</i> — This is an incremental approach to register
1785 allocation. Live ranges are assigned to registers one at a time in
1786 an order that is driven by heuristics. Since code can be rewritten
1787 on-the-fly during allocation, this framework allows interesting
1788 allocators to be developed as extensions. It is not itself a
1789 production register allocator but is a potentially useful
1790 stand-alone mode for triaging bugs and as a performance baseline.
1792 <li><i>Greedy</i> — <i>The default allocator</i>. This is a
1793 highly tuned implementation of the <i>Basic</i> allocator that
1794 incorporates global live range splitting. This allocator works hard
1795 to minimize the cost of spill code.
1797 <li><i>PBQP</i> — A Partitioned Boolean Quadratic Programming (PBQP)
1798 based register allocator. This allocator works by constructing a PBQP
1799 problem representing the register allocation problem under consideration,
1800 solving this using a PBQP solver, and mapping the solution back to a
1801 register assignment.</li>
1804 <p>The type of register allocator used in <tt>llc</tt> can be chosen with the
1805 command line option <tt>-regalloc=...</tt>:</p>
1807 <div class="doc_code">
1809 $ llc -regalloc=linearscan file.bc -o ln.s;
1810 $ llc -regalloc=fast file.bc -o fa.s;
1811 $ llc -regalloc=pbqp file.bc -o pbqp.s;
1819 <!-- ======================================================================= -->
1821 <a name="proepicode">Prolog/Epilog Code Insertion</a>
1826 <!-- _______________________________________________________________________ -->
1828 <a name="compact_unwind">Compact Unwind</a>
1833 <p>Throwing an exception requires <em>unwinding</em> out of a function. The
1834 information on how to unwind a given function is traditionally expressed in
1835 DWARF unwind (a.k.a. frame) info. But that format was originally developed
1836 for debuggers to backtrace, and each Frame Description Entry (FDE) requires
1837 ~20-30 bytes per function. There is also the cost of mapping from an address
1838 in a function to the corresponding FDE at runtime. An alternative unwind
1839 encoding is called <em>compact unwind</em> and requires just 4-bytes per
1842 <p>The compact unwind encoding is a 32-bit value, which is encoded in an
1843 architecture-specific way. It specifies which registers to restore and from
1844 where, and how to unwind out of the function. When the linker creates a final
1845 linked image, it will create a <code>__TEXT,__unwind_info</code>
1846 section. This section is a small and fast way for the runtime to access
1847 unwind info for any given function. If we emit compact unwind info for the
1848 function, that compact unwind info will be encoded in
1849 the <code>__TEXT,__unwind_info</code> section. If we emit DWARF unwind info,
1850 the <code>__TEXT,__unwind_info</code> section will contain the offset of the
1851 FDE in the <code>__TEXT,__eh_frame</code> section in the final linked
1854 <p>For X86, there are three modes for the compact unwind encoding:</p>
1857 <dt><i>Function with a Frame Pointer (<code>EBP</code> or <code>RBP</code>)</i></dt>
1858 <dd><p><code>EBP/RBP</code>-based frame, where <code>EBP/RBP</code> is pushed
1859 onto the stack immediately after the return address,
1860 then <code>ESP/RSP</code> is moved to <code>EBP/RBP</code>. Thus to
1861 unwind, <code>ESP/RSP</code> is restored with the
1862 current <code>EBP/RBP</code> value, then <code>EBP/RBP</code> is restored
1863 by popping the stack, and the return is done by popping the stack once
1864 more into the PC. All non-volatile registers that need to be restored must
1865 have been saved in a small range on the stack that
1866 starts <code>EBP-4</code> to <code>EBP-1020</code> (<code>RBP-8</code>
1867 to <code>RBP-1020</code>). The offset (divided by 4 in 32-bit mode and 8
1868 in 64-bit mode) is encoded in bits 16-23 (mask: <code>0x00FF0000</code>).
1869 The registers saved are encoded in bits 0-14
1870 (mask: <code>0x00007FFF</code>) as five 3-bit entries from the following
1872 <table border="1" cellspacing="0">
1874 <th>Compact Number</th>
1875 <th>i386 Register</th>
1876 <th>x86-64 Regiser</th>
1880 <td><code>EBX</code></td>
1881 <td><code>RBX</code></td>
1885 <td><code>ECX</code></td>
1886 <td><code>R12</code></td>
1890 <td><code>EDX</code></td>
1891 <td><code>R13</code></td>
1895 <td><code>EDI</code></td>
1896 <td><code>R14</code></td>
1900 <td><code>ESI</code></td>
1901 <td><code>R15</code></td>
1905 <td><code>EBP</code></td>
1906 <td><code>RBP</code></td>
1912 <dt><i>Frameless with a Small Constant Stack Size (<code>EBP</code>
1913 or <code>RBP</code> is not used as a frame pointer)</i></dt>
1914 <dd><p>To return, a constant (encoded in the compact unwind encoding) is added
1915 to the <code>ESP/RSP</code>. Then the return is done by popping the stack
1916 into the PC. All non-volatile registers that need to be restored must have
1917 been saved on the stack immediately after the return address. The stack
1918 size (divided by 4 in 32-bit mode and 8 in 64-bit mode) is encoded in bits
1919 16-23 (mask: <code>0x00FF0000</code>). There is a maximum stack size of
1920 1024 bytes in 32-bit mode and 2048 in 64-bit mode. The number of registers
1921 saved is encoded in bits 9-12 (mask: <code>0x00001C00</code>). Bits 0-9
1922 (mask: <code>0x000003FF</code>) contain which registers were saved and
1924 the <code>encodeCompactUnwindRegistersWithoutFrame()</code> function
1925 in <code>lib/Target/X86FrameLowering.cpp</code> for the encoding
1926 algorithm.)</p></dd>
1928 <dt><i>Frameless with a Large Constant Stack Size (<code>EBP</code>
1929 or <code>RBP</code> is not used as a frame pointer)</i></dt>
1930 <dd><p>This case is like the "Frameless with a Small Constant Stack Size"
1931 case, but the stack size is too large to encode in the compact unwind
1932 encoding. Instead it requires that the function contains "<code>subl
1933 $nnnnnn, %esp</code>" in its prolog. The compact encoding contains the
1934 offset to the <code>$nnnnnn</code> value in the function in bits 9-12
1935 (mask: <code>0x00001C00</code>).</p></dd>
1942 <!-- ======================================================================= -->
1944 <a name="latemco">Late Machine Code Optimizations</a>
1946 <div><p>To Be Written</p></div>
1948 <!-- ======================================================================= -->
1950 <a name="codeemit">Code Emission</a>
1955 <p>The code emission step of code generation is responsible for lowering from
1956 the code generator abstractions (like <a
1957 href="#machinefunction">MachineFunction</a>, <a
1958 href="#machineinstr">MachineInstr</a>, etc) down
1959 to the abstractions used by the MC layer (<a href="#mcinst">MCInst</a>,
1960 <a href="#mcstreamer">MCStreamer</a>, etc). This is
1961 done with a combination of several different classes: the (misnamed)
1962 target-independent AsmPrinter class, target-specific subclasses of AsmPrinter
1963 (such as SparcAsmPrinter), and the TargetLoweringObjectFile class.</p>
1965 <p>Since the MC layer works at the level of abstraction of object files, it
1966 doesn't have a notion of functions, global variables etc. Instead, it thinks
1967 about labels, directives, and instructions. A key class used at this time is
1968 the MCStreamer class. This is an abstract API that is implemented in different
1969 ways (e.g. to output a .s file, output an ELF .o file, etc) that is effectively
1970 an "assembler API". MCStreamer has one method per directive, such as EmitLabel,
1971 EmitSymbolAttribute, SwitchSection, etc, which directly correspond to assembly
1975 <p>If you are interested in implementing a code generator for a target, there
1976 are three important things that you have to implement for your target:</p>
1979 <li>First, you need a subclass of AsmPrinter for your target. This class
1980 implements the general lowering process converting MachineFunction's into MC
1981 label constructs. The AsmPrinter base class provides a number of useful methods
1982 and routines, and also allows you to override the lowering process in some
1983 important ways. You should get much of the lowering for free if you are
1984 implementing an ELF, COFF, or MachO target, because the TargetLoweringObjectFile
1985 class implements much of the common logic.</li>
1987 <li>Second, you need to implement an instruction printer for your target. The
1988 instruction printer takes an <a href="#mcinst">MCInst</a> and renders it to a
1989 raw_ostream as text. Most of this is automatically generated from the .td file
1990 (when you specify something like "<tt>add $dst, $src1, $src2</tt>" in the
1991 instructions), but you need to implement routines to print operands.</li>
1993 <li>Third, you need to implement code that lowers a <a
1994 href="#machineinstr">MachineInstr</a> to an MCInst, usually implemented in
1995 "<target>MCInstLower.cpp". This lowering process is often target
1996 specific, and is responsible for turning jump table entries, constant pool
1997 indices, global variable addresses, etc into MCLabels as appropriate. This
1998 translation layer is also responsible for expanding pseudo ops used by the code
1999 generator into the actual machine instructions they correspond to. The MCInsts
2000 that are generated by this are fed into the instruction printer or the encoder.
2005 <p>Finally, at your choosing, you can also implement an subclass of
2006 MCCodeEmitter which lowers MCInst's into machine code bytes and relocations.
2007 This is important if you want to support direct .o file emission, or would like
2008 to implement an assembler for your target.</p>
2012 <!-- ======================================================================= -->
2014 <a name="vliw_packetizer">VLIW Packetizer</a>
2019 <p>In a Very Long Instruction Word (VLIW) architecture, the compiler is
2020 responsible for mapping instructions to functional-units available on
2021 the architecture. To that end, the compiler creates groups of instructions
2022 called <i>packets</i> or <i>bundles</i>. The VLIW packetizer in LLVM is
2023 a target-independent mechanism to enable the packetization of machine
2026 <!-- _______________________________________________________________________ -->
2029 <a name="vliw_mapping">Mapping from instructions to functional units</a>
2034 <p>Instructions in a VLIW target can typically be mapped to multiple functional
2035 units. During the process of packetizing, the compiler must be able to reason
2036 about whether an instruction can be added to a packet. This decision can be
2037 complex since the compiler has to examine all possible mappings of instructions
2038 to functional units. Therefore to alleviate compilation-time complexity, the
2039 VLIW packetizer parses the instruction classes of a target and generates tables
2040 at compiler build time. These tables can then be queried by the provided
2041 machine-independent API to determine if an instruction can be accommodated in a
2045 <!-- ======================================================================= -->
2047 <a name="vliw_repr">
2048 How the packetization tables are generated and used
2054 <p>The packetizer reads instruction classes from a target's itineraries and
2055 creates a deterministic finite automaton (DFA) to represent the state of a
2056 packet. A DFA consists of three major elements: inputs, states, and
2057 transitions. The set of inputs for the generated DFA represents the instruction
2058 being added to a packet. The states represent the possible consumption
2059 of functional units by instructions in a packet. In the DFA, transitions from
2060 one state to another occur on the addition of an instruction to an existing
2061 packet. If there is a legal mapping of functional units to instructions, then
2062 the DFA contains a corresponding transition. The absence of a transition
2063 indicates that a legal mapping does not exist and that the instruction cannot
2064 be added to the packet.</p>
2066 <p>To generate tables for a VLIW target, add <i>Target</i>GenDFAPacketizer.inc
2067 as a target to the Makefile in the target directory. The exported API provides
2068 three functions: <tt>DFAPacketizer::clearResources()</tt>,
2069 <tt>DFAPacketizer::reserveResources(MachineInstr *MI)</tt>, and
2070 <tt>DFAPacketizer::canReserveResources(MachineInstr *MI)</tt>. These functions
2071 allow a target packetizer to add an instruction to an existing packet and to
2072 check whether an instruction can be added to a packet. See
2073 <tt>llvm/CodeGen/DFAPacketizer.h</tt> for more information.</p>
2081 <!-- *********************************************************************** -->
2083 <a name="nativeassembler">Implementing a Native Assembler</a>
2085 <!-- *********************************************************************** -->
2089 <p>Though you're probably reading this because you want to write or maintain a
2090 compiler backend, LLVM also fully supports building a native assemblers too.
2091 We've tried hard to automate the generation of the assembler from the .td files
2092 (in particular the instruction syntax and encodings), which means that a large
2093 part of the manual and repetitive data entry can be factored and shared with the
2096 <!-- ======================================================================= -->
2097 <h3 id="na_instparsing">Instruction Parsing</h3>
2099 <div><p>To Be Written</p></div>
2102 <!-- ======================================================================= -->
2103 <h3 id="na_instaliases">
2104 Instruction Alias Processing
2108 <p>Once the instruction is parsed, it enters the MatchInstructionImpl function.
2109 The MatchInstructionImpl function performs alias processing and then does
2110 actual matching.</p>
2112 <p>Alias processing is the phase that canonicalizes different lexical forms of
2113 the same instructions down to one representation. There are several different
2114 kinds of alias that are possible to implement and they are listed below in the
2115 order that they are processed (which is in order from simplest/weakest to most
2116 complex/powerful). Generally you want to use the first alias mechanism that
2117 meets the needs of your instruction, because it will allow a more concise
2120 <!-- _______________________________________________________________________ -->
2121 <h4>Mnemonic Aliases</h4>
2125 <p>The first phase of alias processing is simple instruction mnemonic
2126 remapping for classes of instructions which are allowed with two different
2127 mnemonics. This phase is a simple and unconditionally remapping from one input
2128 mnemonic to one output mnemonic. It isn't possible for this form of alias to
2129 look at the operands at all, so the remapping must apply for all forms of a
2130 given mnemonic. Mnemonic aliases are defined simply, for example X86 has:
2133 <div class="doc_code">
2135 def : MnemonicAlias<"cbw", "cbtw">;
2136 def : MnemonicAlias<"smovq", "movsq">;
2137 def : MnemonicAlias<"fldcww", "fldcw">;
2138 def : MnemonicAlias<"fucompi", "fucomip">;
2139 def : MnemonicAlias<"ud2a", "ud2">;
2143 <p>... and many others. With a MnemonicAlias definition, the mnemonic is
2144 remapped simply and directly. Though MnemonicAlias's can't look at any aspect
2145 of the instruction (such as the operands) they can depend on global modes (the
2146 same ones supported by the matcher), through a Requires clause:</p>
2148 <div class="doc_code">
2150 def : MnemonicAlias<"pushf", "pushfq">, Requires<[In64BitMode]>;
2151 def : MnemonicAlias<"pushf", "pushfl">, Requires<[In32BitMode]>;
2155 <p>In this example, the mnemonic gets mapped into different a new one depending
2156 on the current instruction set.</p>
2160 <!-- _______________________________________________________________________ -->
2161 <h4>Instruction Aliases</h4>
2165 <p>The most general phase of alias processing occurs while matching is
2166 happening: it provides new forms for the matcher to match along with a specific
2167 instruction to generate. An instruction alias has two parts: the string to
2168 match and the instruction to generate. For example:
2171 <div class="doc_code">
2173 def : InstAlias<"movsx $src, $dst", (MOVSX16rr8W GR16:$dst, GR8 :$src)>;
2174 def : InstAlias<"movsx $src, $dst", (MOVSX16rm8W GR16:$dst, i8mem:$src)>;
2175 def : InstAlias<"movsx $src, $dst", (MOVSX32rr8 GR32:$dst, GR8 :$src)>;
2176 def : InstAlias<"movsx $src, $dst", (MOVSX32rr16 GR32:$dst, GR16 :$src)>;
2177 def : InstAlias<"movsx $src, $dst", (MOVSX64rr8 GR64:$dst, GR8 :$src)>;
2178 def : InstAlias<"movsx $src, $dst", (MOVSX64rr16 GR64:$dst, GR16 :$src)>;
2179 def : InstAlias<"movsx $src, $dst", (MOVSX64rr32 GR64:$dst, GR32 :$src)>;
2183 <p>This shows a powerful example of the instruction aliases, matching the
2184 same mnemonic in multiple different ways depending on what operands are present
2185 in the assembly. The result of instruction aliases can include operands in a
2186 different order than the destination instruction, and can use an input
2187 multiple times, for example:</p>
2189 <div class="doc_code">
2191 def : InstAlias<"clrb $reg", (XOR8rr GR8 :$reg, GR8 :$reg)>;
2192 def : InstAlias<"clrw $reg", (XOR16rr GR16:$reg, GR16:$reg)>;
2193 def : InstAlias<"clrl $reg", (XOR32rr GR32:$reg, GR32:$reg)>;
2194 def : InstAlias<"clrq $reg", (XOR64rr GR64:$reg, GR64:$reg)>;
2198 <p>This example also shows that tied operands are only listed once. In the X86
2199 backend, XOR8rr has two input GR8's and one output GR8 (where an input is tied
2200 to the output). InstAliases take a flattened operand list without duplicates
2201 for tied operands. The result of an instruction alias can also use immediates
2202 and fixed physical registers which are added as simple immediate operands in the
2203 result, for example:</p>
2205 <div class="doc_code">
2207 // Fixed Immediate operand.
2208 def : InstAlias<"aad", (AAD8i8 10)>;
2210 // Fixed register operand.
2211 def : InstAlias<"fcomi", (COM_FIr ST1)>;
2214 def : InstAlias<"fcomi $reg", (COM_FIr RST:$reg)>;
2219 <p>Instruction aliases can also have a Requires clause to make them
2220 subtarget specific.</p>
2222 <p>If the back-end supports it, the instruction printer can automatically emit
2223 the alias rather than what's being aliased. It typically leads to better,
2224 more readable code. If it's better to print out what's being aliased, then
2225 pass a '0' as the third parameter to the InstAlias definition.</p>
2231 <!-- ======================================================================= -->
2232 <h3 id="na_matching">Instruction Matching</h3>
2234 <div><p>To Be Written</p></div>
2238 <!-- *********************************************************************** -->
2240 <a name="targetimpls">Target-specific Implementation Notes</a>
2242 <!-- *********************************************************************** -->
2246 <p>This section of the document explains features or design decisions that are
2247 specific to the code generator for a particular target. First we start
2248 with a table that summarizes what features are supported by each target.</p>
2250 <!-- ======================================================================= -->
2252 <a name="targetfeatures">Target Feature Matrix</a>
2257 <p>Note that this table does not include the C backend or Cpp backends, since
2258 they do not use the target independent code generator infrastructure. It also
2259 doesn't list features that are not supported fully by any target yet. It
2260 considers a feature to be supported if at least one subtarget supports it. A
2261 feature being supported means that it is useful and works for most cases, it
2262 does not indicate that there are zero known bugs in the implementation. Here
2266 <table border="1" cellspacing="0">
2270 <th>Partial Support</th>
2271 <th>Complete Support</th>
2274 <td class="unknown"></td>
2275 <td class="no"></td>
2276 <td class="partial"></td>
2277 <td class="yes"></td>
2281 <p>Here is the table:</p>
2283 <table width="689" border="1" cellspacing="0">
2285 <td colspan="13" align="center" style="background-color:#ffc">Target</td>
2302 <td><a href="#feat_reliable">is generally reliable</a></td>
2303 <td class="yes"></td> <!-- ARM -->
2304 <td class="no"></td> <!-- CellSPU -->
2305 <td class="no"></td> <!-- MBlaze -->
2306 <td class="unknown"></td> <!-- MSP430 -->
2307 <td class="yes"></td> <!-- Mips -->
2308 <td class="no"></td> <!-- PTX -->
2309 <td class="yes"></td> <!-- PowerPC -->
2310 <td class="yes"></td> <!-- Sparc -->
2311 <td class="yes"></td> <!-- X86 -->
2312 <td class="unknown"></td> <!-- XCore -->
2316 <td><a href="#feat_asmparser">assembly parser</a></td>
2317 <td class="no"></td> <!-- ARM -->
2318 <td class="no"></td> <!-- CellSPU -->
2319 <td class="yes"></td> <!-- MBlaze -->
2320 <td class="no"></td> <!-- MSP430 -->
2321 <td class="no"></td> <!-- Mips -->
2322 <td class="no"></td> <!-- PTX -->
2323 <td class="no"></td> <!-- PowerPC -->
2324 <td class="no"></td> <!-- Sparc -->
2325 <td class="yes"></td> <!-- X86 -->
2326 <td class="no"></td> <!-- XCore -->
2330 <td><a href="#feat_disassembler">disassembler</a></td>
2331 <td class="yes"></td> <!-- ARM -->
2332 <td class="no"></td> <!-- CellSPU -->
2333 <td class="yes"></td> <!-- MBlaze -->
2334 <td class="no"></td> <!-- MSP430 -->
2335 <td class="no"></td> <!-- Mips -->
2336 <td class="no"></td> <!-- PTX -->
2337 <td class="no"></td> <!-- PowerPC -->
2338 <td class="no"></td> <!-- Sparc -->
2339 <td class="yes"></td> <!-- X86 -->
2340 <td class="no"></td> <!-- XCore -->
2344 <td><a href="#feat_inlineasm">inline asm</a></td>
2345 <td class="yes"></td> <!-- ARM -->
2346 <td class="no"></td> <!-- CellSPU -->
2347 <td class="yes"></td> <!-- MBlaze -->
2348 <td class="unknown"></td> <!-- MSP430 -->
2349 <td class="no"></td> <!-- Mips -->
2350 <td class="unknown"></td> <!-- PTX -->
2351 <td class="yes"></td> <!-- PowerPC -->
2352 <td class="unknown"></td> <!-- Sparc -->
2353 <td class="yes"></td> <!-- X86 -->
2354 <td class="unknown"></td> <!-- XCore -->
2358 <td><a href="#feat_jit">jit</a></td>
2359 <td class="partial"><a href="#feat_jit_arm">*</a></td> <!-- ARM -->
2360 <td class="no"></td> <!-- CellSPU -->
2361 <td class="no"></td> <!-- MBlaze -->
2362 <td class="unknown"></td> <!-- MSP430 -->
2363 <td class="yes"></td> <!-- Mips -->
2364 <td class="unknown"></td> <!-- PTX -->
2365 <td class="yes"></td> <!-- PowerPC -->
2366 <td class="unknown"></td> <!-- Sparc -->
2367 <td class="yes"></td> <!-- X86 -->
2368 <td class="unknown"></td> <!-- XCore -->
2372 <td><a href="#feat_objectwrite">.o file writing</a></td>
2373 <td class="no"></td> <!-- ARM -->
2374 <td class="no"></td> <!-- CellSPU -->
2375 <td class="yes"></td> <!-- MBlaze -->
2376 <td class="no"></td> <!-- MSP430 -->
2377 <td class="no"></td> <!-- Mips -->
2378 <td class="no"></td> <!-- PTX -->
2379 <td class="no"></td> <!-- PowerPC -->
2380 <td class="no"></td> <!-- Sparc -->
2381 <td class="yes"></td> <!-- X86 -->
2382 <td class="no"></td> <!-- XCore -->
2386 <td><a href="#feat_tailcall">tail calls</a></td>
2387 <td class="yes"></td> <!-- ARM -->
2388 <td class="no"></td> <!-- CellSPU -->
2389 <td class="no"></td> <!-- MBlaze -->
2390 <td class="unknown"></td> <!-- MSP430 -->
2391 <td class="no"></td> <!-- Mips -->
2392 <td class="unknown"></td> <!-- PTX -->
2393 <td class="yes"></td> <!-- PowerPC -->
2394 <td class="unknown"></td> <!-- Sparc -->
2395 <td class="yes"></td> <!-- X86 -->
2396 <td class="unknown"></td> <!-- XCore -->
2400 <td><a href="#feat_segstacks">segmented stacks</a></td>
2401 <td class="no"></td> <!-- ARM -->
2402 <td class="no"></td> <!-- CellSPU -->
2403 <td class="no"></td> <!-- MBlaze -->
2404 <td class="no"></td> <!-- MSP430 -->
2405 <td class="no"></td> <!-- Mips -->
2406 <td class="no"></td> <!-- PTX -->
2407 <td class="no"></td> <!-- PowerPC -->
2408 <td class="no"></td> <!-- Sparc -->
2409 <td class="partial"><a href="#feat_segstacks_x86">*</a></td> <!-- X86 -->
2410 <td class="no"></td> <!-- XCore -->
2416 <!-- _______________________________________________________________________ -->
2417 <h4 id="feat_reliable">Is Generally Reliable</h4>
2420 <p>This box indicates whether the target is considered to be production quality.
2421 This indicates that the target has been used as a static compiler to
2422 compile large amounts of code by a variety of different people and is in
2426 <!-- _______________________________________________________________________ -->
2427 <h4 id="feat_asmparser">Assembly Parser</h4>
2430 <p>This box indicates whether the target supports parsing target specific .s
2431 files by implementing the MCAsmParser interface. This is required for llvm-mc
2432 to be able to act as a native assembler and is required for inline assembly
2433 support in the native .o file writer.</p>
2438 <!-- _______________________________________________________________________ -->
2439 <h4 id="feat_disassembler">Disassembler</h4>
2442 <p>This box indicates whether the target supports the MCDisassembler API for
2443 disassembling machine opcode bytes into MCInst's.</p>
2447 <!-- _______________________________________________________________________ -->
2448 <h4 id="feat_inlineasm">Inline Asm</h4>
2451 <p>This box indicates whether the target supports most popular inline assembly
2452 constraints and modifiers.</p>
2456 <!-- _______________________________________________________________________ -->
2457 <h4 id="feat_jit">JIT Support</h4>
2460 <p>This box indicates whether the target supports the JIT compiler through
2461 the ExecutionEngine interface.</p>
2463 <p id="feat_jit_arm">The ARM backend has basic support for integer code
2464 in ARM codegen mode, but lacks NEON and full Thumb support.</p>
2468 <!-- _______________________________________________________________________ -->
2469 <h4 id="feat_objectwrite">.o File Writing</h4>
2473 <p>This box indicates whether the target supports writing .o files (e.g. MachO,
2474 ELF, and/or COFF) files directly from the target. Note that the target also
2475 must include an assembly parser and general inline assembly support for full
2476 inline assembly support in the .o writer.</p>
2478 <p>Targets that don't support this feature can obviously still write out .o
2479 files, they just rely on having an external assembler to translate from a .s
2480 file to a .o file (as is the case for many C compilers).</p>
2484 <!-- _______________________________________________________________________ -->
2485 <h4 id="feat_tailcall">Tail Calls</h4>
2489 <p>This box indicates whether the target supports guaranteed tail calls. These
2490 are calls marked "<a href="LangRef.html#i_call">tail</a>" and use the fastcc
2491 calling convention. Please see the <a href="#tailcallopt">tail call section
2492 more more details</a>.</p>
2496 <!-- _______________________________________________________________________ -->
2497 <h4 id="feat_segstacks">Segmented Stacks</h4>
2501 <p>This box indicates whether the target supports segmented stacks. This
2502 replaces the traditional large C stack with many linked segments. It
2503 is compatible with the <a href="http://gcc.gnu.org/wiki/SplitStacks">gcc
2504 implementation</a> used by the Go front end.</p>
2506 <p id="feat_segstacks_x86">Basic support exists on the X86 backend. Currently
2507 vararg doesn't work and the object files are not marked the way the gold
2508 linker expects, but simple Go programs can be built by dragonegg.</p>
2514 <!-- ======================================================================= -->
2516 <a name="tailcallopt">Tail call optimization</a>
2521 <p>Tail call optimization, callee reusing the stack of the caller, is currently
2522 supported on x86/x86-64 and PowerPC. It is performed if:</p>
2525 <li>Caller and callee have the calling convention <tt>fastcc</tt> or
2526 <tt>cc 10</tt> (GHC call convention).</li>
2528 <li>The call is a tail call - in tail position (ret immediately follows call
2529 and ret uses value of call or is void).</li>
2531 <li>Option <tt>-tailcallopt</tt> is enabled.</li>
2533 <li>Platform specific constraints are met.</li>
2536 <p>x86/x86-64 constraints:</p>
2539 <li>No variable argument lists are used.</li>
2541 <li>On x86-64 when generating GOT/PIC code only module-local calls (visibility
2542 = hidden or protected) are supported.</li>
2545 <p>PowerPC constraints:</p>
2548 <li>No variable argument lists are used.</li>
2550 <li>No byval parameters are used.</li>
2552 <li>On ppc32/64 GOT/PIC only module-local calls (visibility = hidden or protected) are supported.</li>
2557 <p>Call as <tt>llc -tailcallopt test.ll</tt>.</p>
2559 <div class="doc_code">
2561 declare fastcc i32 @tailcallee(i32 inreg %a1, i32 inreg %a2, i32 %a3, i32 %a4)
2563 define fastcc i32 @tailcaller(i32 %in1, i32 %in2) {
2564 %l1 = add i32 %in1, %in2
2565 %tmp = tail call fastcc i32 @tailcallee(i32 %in1 inreg, i32 %in2 inreg, i32 %in1, i32 %l1)
2571 <p>Implications of <tt>-tailcallopt</tt>:</p>
2573 <p>To support tail call optimization in situations where the callee has more
2574 arguments than the caller a 'callee pops arguments' convention is used. This
2575 currently causes each <tt>fastcc</tt> call that is not tail call optimized
2576 (because one or more of above constraints are not met) to be followed by a
2577 readjustment of the stack. So performance might be worse in such cases.</p>
2580 <!-- ======================================================================= -->
2582 <a name="sibcallopt">Sibling call optimization</a>
2587 <p>Sibling call optimization is a restricted form of tail call optimization.
2588 Unlike tail call optimization described in the previous section, it can be
2589 performed automatically on any tail calls when <tt>-tailcallopt</tt> option
2590 is not specified.</p>
2592 <p>Sibling call optimization is currently performed on x86/x86-64 when the
2593 following constraints are met:</p>
2596 <li>Caller and callee have the same calling convention. It can be either
2597 <tt>c</tt> or <tt>fastcc</tt>.
2599 <li>The call is a tail call - in tail position (ret immediately follows call
2600 and ret uses value of call or is void).</li>
2602 <li>Caller and callee have matching return type or the callee result is not
2605 <li>If any of the callee arguments are being passed in stack, they must be
2606 available in caller's own incoming argument stack and the frame offsets
2611 <div class="doc_code">
2613 declare i32 @bar(i32, i32)
2615 define i32 @foo(i32 %a, i32 %b, i32 %c) {
2617 %0 = tail call i32 @bar(i32 %a, i32 %b)
2624 <!-- ======================================================================= -->
2626 <a name="x86">The X86 backend</a>
2631 <p>The X86 code generator lives in the <tt>lib/Target/X86</tt> directory. This
2632 code generator is capable of targeting a variety of x86-32 and x86-64
2633 processors, and includes support for ISA extensions such as MMX and SSE.</p>
2635 <!-- _______________________________________________________________________ -->
2637 <a name="x86_tt">X86 Target Triples supported</a>
2642 <p>The following are the known target triples that are supported by the X86
2643 backend. This is not an exhaustive list, and it would be useful to add those
2644 that people test.</p>
2647 <li><b>i686-pc-linux-gnu</b> — Linux</li>
2649 <li><b>i386-unknown-freebsd5.3</b> — FreeBSD 5.3</li>
2651 <li><b>i686-pc-cygwin</b> — Cygwin on Win32</li>
2653 <li><b>i686-pc-mingw32</b> — MingW on Win32</li>
2655 <li><b>i386-pc-mingw32msvc</b> — MingW crosscompiler on Linux</li>
2657 <li><b>i686-apple-darwin*</b> — Apple Darwin on X86</li>
2659 <li><b>x86_64-unknown-linux-gnu</b> — Linux</li>
2664 <!-- _______________________________________________________________________ -->
2666 <a name="x86_cc">X86 Calling Conventions supported</a>
2672 <p>The following target-specific calling conventions are known to backend:</p>
2675 <li><b>x86_StdCall</b> — stdcall calling convention seen on Microsoft
2676 Windows platform (CC ID = 64).</li>
2677 <li><b>x86_FastCall</b> — fastcall calling convention seen on Microsoft
2678 Windows platform (CC ID = 65).</li>
2679 <li><b>x86_ThisCall</b> — Similar to X86_StdCall. Passes first argument
2680 in ECX, others via stack. Callee is responsible for stack cleaning. This
2681 convention is used by MSVC by default for methods in its ABI
2687 <!-- _______________________________________________________________________ -->
2689 <a name="x86_memory">Representing X86 addressing modes in MachineInstrs</a>
2694 <p>The x86 has a very flexible way of accessing memory. It is capable of
2695 forming memory addresses of the following expression directly in integer
2696 instructions (which use ModR/M addressing):</p>
2698 <div class="doc_code">
2700 SegmentReg: Base + [1,2,4,8] * IndexReg + Disp32
2704 <p>In order to represent this, LLVM tracks no less than 5 operands for each
2705 memory operand of this form. This means that the "load" form of
2706 '<tt>mov</tt>' has the following <tt>MachineOperand</tt>s in this order:</p>
2708 <div class="doc_code">
2710 Index: 0 | 1 2 3 4 5
2711 Meaning: DestReg, | BaseReg, Scale, IndexReg, Displacement Segment
2712 OperandTy: VirtReg, | VirtReg, UnsImm, VirtReg, SignExtImm PhysReg
2716 <p>Stores, and all other instructions, treat the four memory operands in the
2717 same way and in the same order. If the segment register is unspecified
2718 (regno = 0), then no segment override is generated. "Lea" operations do not
2719 have a segment register specified, so they only have 4 operands for their
2720 memory reference.</p>
2724 <!-- _______________________________________________________________________ -->
2726 <a name="x86_memory">X86 address spaces supported</a>
2731 <p>x86 has a feature which provides
2732 the ability to perform loads and stores to different address spaces
2733 via the x86 segment registers. A segment override prefix byte on an
2734 instruction causes the instruction's memory access to go to the specified
2735 segment. LLVM address space 0 is the default address space, which includes
2736 the stack, and any unqualified memory accesses in a program. Address spaces
2737 1-255 are currently reserved for user-defined code. The GS-segment is
2738 represented by address space 256, while the FS-segment is represented by
2739 address space 257. Other x86 segments have yet to be allocated address space
2742 <p>While these address spaces may seem similar to TLS via the
2743 <tt>thread_local</tt> keyword, and often use the same underlying hardware,
2744 there are some fundamental differences.</p>
2746 <p>The <tt>thread_local</tt> keyword applies to global variables and
2747 specifies that they are to be allocated in thread-local memory. There are
2748 no type qualifiers involved, and these variables can be pointed to with
2749 normal pointers and accessed with normal loads and stores.
2750 The <tt>thread_local</tt> keyword is target-independent at the LLVM IR
2751 level (though LLVM doesn't yet have implementations of it for some
2754 <p>Special address spaces, in contrast, apply to static types. Every
2755 load and store has a particular address space in its address operand type,
2756 and this is what determines which address space is accessed.
2757 LLVM ignores these special address space qualifiers on global variables,
2758 and does not provide a way to directly allocate storage in them.
2759 At the LLVM IR level, the behavior of these special address spaces depends
2760 in part on the underlying OS or runtime environment, and they are specific
2761 to x86 (and LLVM doesn't yet handle them correctly in some cases).</p>
2763 <p>Some operating systems and runtime environments use (or may in the future
2764 use) the FS/GS-segment registers for various low-level purposes, so care
2765 should be taken when considering them.</p>
2769 <!-- _______________________________________________________________________ -->
2771 <a name="x86_names">Instruction naming</a>
2776 <p>An instruction name consists of the base name, a default operand size, and a
2777 a character per operand with an optional special size. For example:</p>
2779 <div class="doc_code">
2781 ADD8rr -> add, 8-bit register, 8-bit register
2782 IMUL16rmi -> imul, 16-bit register, 16-bit memory, 16-bit immediate
2783 IMUL16rmi8 -> imul, 16-bit register, 16-bit memory, 8-bit immediate
2784 MOVSX32rm16 -> movsx, 32-bit register, 16-bit memory
2792 <!-- ======================================================================= -->
2794 <a name="ppc">The PowerPC backend</a>
2799 <p>The PowerPC code generator lives in the lib/Target/PowerPC directory. The
2800 code generation is retargetable to several variations or <i>subtargets</i> of
2801 the PowerPC ISA; including ppc32, ppc64 and altivec.</p>
2803 <!-- _______________________________________________________________________ -->
2805 <a name="ppc_abi">LLVM PowerPC ABI</a>
2810 <p>LLVM follows the AIX PowerPC ABI, with two deviations. LLVM uses a PC
2811 relative (PIC) or static addressing for accessing global values, so no TOC
2812 (r2) is used. Second, r31 is used as a frame pointer to allow dynamic growth
2813 of a stack frame. LLVM takes advantage of having no TOC to provide space to
2814 save the frame pointer in the PowerPC linkage area of the caller frame.
2815 Other details of PowerPC ABI can be found at <a href=
2816 "http://developer.apple.com/documentation/DeveloperTools/Conceptual/LowLevelABI/Articles/32bitPowerPC.html"
2817 >PowerPC ABI.</a> Note: This link describes the 32 bit ABI. The 64 bit ABI
2818 is similar except space for GPRs are 8 bytes wide (not 4) and r13 is reserved
2823 <!-- _______________________________________________________________________ -->
2825 <a name="ppc_frame">Frame Layout</a>
2830 <p>The size of a PowerPC frame is usually fixed for the duration of a
2831 function's invocation. Since the frame is fixed size, all references
2832 into the frame can be accessed via fixed offsets from the stack pointer. The
2833 exception to this is when dynamic alloca or variable sized arrays are
2834 present, then a base pointer (r31) is used as a proxy for the stack pointer
2835 and stack pointer is free to grow or shrink. A base pointer is also used if
2836 llvm-gcc is not passed the -fomit-frame-pointer flag. The stack pointer is
2837 always aligned to 16 bytes, so that space allocated for altivec vectors will
2838 be properly aligned.</p>
2840 <p>An invocation frame is laid out as follows (low memory at top);</p>
2842 <table class="layout">
2844 <td>Linkage<br><br></td>
2847 <td>Parameter area<br><br></td>
2850 <td>Dynamic area<br><br></td>
2853 <td>Locals area<br><br></td>
2856 <td>Saved registers area<br><br></td>
2858 <tr style="border-style: none hidden none hidden;">
2862 <td>Previous Frame<br><br></td>
2866 <p>The <i>linkage</i> area is used by a callee to save special registers prior
2867 to allocating its own frame. Only three entries are relevant to LLVM. The
2868 first entry is the previous stack pointer (sp), aka link. This allows
2869 probing tools like gdb or exception handlers to quickly scan the frames in
2870 the stack. A function epilog can also use the link to pop the frame from the
2871 stack. The third entry in the linkage area is used to save the return
2872 address from the lr register. Finally, as mentioned above, the last entry is
2873 used to save the previous frame pointer (r31.) The entries in the linkage
2874 area are the size of a GPR, thus the linkage area is 24 bytes long in 32 bit
2875 mode and 48 bytes in 64 bit mode.</p>
2877 <p>32 bit linkage area</p>
2879 <table class="layout">
2882 <td>Saved SP (r1)</td>
2902 <td>Saved FP (r31)</td>
2906 <p>64 bit linkage area</p>
2908 <table class="layout">
2911 <td>Saved SP (r1)</td>
2931 <td>Saved FP (r31)</td>
2935 <p>The <i>parameter area</i> is used to store arguments being passed to a callee
2936 function. Following the PowerPC ABI, the first few arguments are actually
2937 passed in registers, with the space in the parameter area unused. However,
2938 if there are not enough registers or the callee is a thunk or vararg
2939 function, these register arguments can be spilled into the parameter area.
2940 Thus, the parameter area must be large enough to store all the parameters for
2941 the largest call sequence made by the caller. The size must also be
2942 minimally large enough to spill registers r3-r10. This allows callees blind
2943 to the call signature, such as thunks and vararg functions, enough space to
2944 cache the argument registers. Therefore, the parameter area is minimally 32
2945 bytes (64 bytes in 64 bit mode.) Also note that since the parameter area is
2946 a fixed offset from the top of the frame, that a callee can access its spilt
2947 arguments using fixed offsets from the stack pointer (or base pointer.)</p>
2949 <p>Combining the information about the linkage, parameter areas and alignment. A
2950 stack frame is minimally 64 bytes in 32 bit mode and 128 bytes in 64 bit
2953 <p>The <i>dynamic area</i> starts out as size zero. If a function uses dynamic
2954 alloca then space is added to the stack, the linkage and parameter areas are
2955 shifted to top of stack, and the new space is available immediately below the
2956 linkage and parameter areas. The cost of shifting the linkage and parameter
2957 areas is minor since only the link value needs to be copied. The link value
2958 can be easily fetched by adding the original frame size to the base pointer.
2959 Note that allocations in the dynamic space need to observe 16 byte
2962 <p>The <i>locals area</i> is where the llvm compiler reserves space for local
2965 <p>The <i>saved registers area</i> is where the llvm compiler spills callee
2966 saved registers on entry to the callee.</p>
2970 <!-- _______________________________________________________________________ -->
2972 <a name="ppc_prolog">Prolog/Epilog</a>
2977 <p>The llvm prolog and epilog are the same as described in the PowerPC ABI, with
2978 the following exceptions. Callee saved registers are spilled after the frame
2979 is created. This allows the llvm epilog/prolog support to be common with
2980 other targets. The base pointer callee saved register r31 is saved in the
2981 TOC slot of linkage area. This simplifies allocation of space for the base
2982 pointer and makes it convenient to locate programatically and during
2987 <!-- _______________________________________________________________________ -->
2989 <a name="ppc_dynamic">Dynamic Allocation</a>
2994 <p><i>TODO - More to come.</i></p>
3000 <!-- ======================================================================= -->
3002 <a name="ptx">The PTX backend</a>
3007 <p>The PTX code generator lives in the lib/Target/PTX directory. It is
3008 currently a work-in-progress, but already supports most of the code
3009 generation functionality needed to generate correct PTX kernels for
3012 <p>The code generator can target PTX 2.0+, and shader model 1.0+. The
3013 PTX ISA Reference Manual is used as the primary source of ISA
3014 information, though an effort is made to make the output of the code
3015 generator match the output of the NVidia nvcc compiler, whenever
3018 <p>Code Generator Options:</p>
3019 <table border="1" cellspacing="0">
3022 <th>Description</th>
3025 <td><code>double</code></td>
3026 <td align="left">If enabled, the map_f64_to_f32 directive is
3027 disabled in the PTX output, allowing native double-precision
3031 <td><code>no-fma</code></td>
3032 <td align="left">Disable generation of Fused-Multiply Add
3033 instructions, which may be beneficial for some devices</td>
3036 <td><code>smxy / computexy</code></td>
3037 <td align="left">Set shader model/compute capability to x.y,
3038 e.g. sm20 or compute13</td>
3044 <li>Arithmetic instruction selection (including combo FMA)</li>
3045 <li>Bitwise instruction selection</li>
3046 <li>Control-flow instruction selection</li>
3047 <li>Function calls (only on SM 2.0+ and no return arguments)</li>
3048 <li>Addresses spaces (0 = global, 1 = constant, 2 = local, 4 =
3050 <li>Thread synchronization (bar.sync)</li>
3051 <li>Special register reads ([N]TID, [N]CTAID, PMx, CLOCK, etc.)</li>
3056 <li>Robust call instruction selection</li>
3057 <li>Stack frame allocation</li>
3058 <li>Device-specific instruction scheduling optimizations</li>
3066 <!-- *********************************************************************** -->
3069 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
3070 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
3071 <a href="http://validator.w3.org/check/referer"><img
3072 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
3074 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
3075 <a href="http://llvm.org/">The LLVM Compiler Infrastructure</a><br>
3076 Last modified: $Date$