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">
11 <div class="doc_title">
12 The LLVM Target-Independent Code Generator
16 <li><a href="#introduction">Introduction</a>
18 <li><a href="#required">Required components in the code generator</a></li>
19 <li><a href="#high-level-design">The high-level design of the code
21 <li><a href="#tablegen">Using TableGen for target description</a></li>
24 <li><a href="#targetdesc">Target description classes</a>
26 <li><a href="#targetmachine">The <tt>TargetMachine</tt> class</a></li>
27 <li><a href="#targetdata">The <tt>TargetData</tt> class</a></li>
28 <li><a href="#targetlowering">The <tt>TargetLowering</tt> class</a></li>
29 <li><a href="#targetregisterinfo">The <tt>TargetRegisterInfo</tt> class</a></li>
30 <li><a href="#targetinstrinfo">The <tt>TargetInstrInfo</tt> class</a></li>
31 <li><a href="#targetframeinfo">The <tt>TargetFrameInfo</tt> class</a></li>
32 <li><a href="#targetsubtarget">The <tt>TargetSubtarget</tt> class</a></li>
33 <li><a href="#targetjitinfo">The <tt>TargetJITInfo</tt> class</a></li>
36 <li><a href="#codegendesc">Machine code description classes</a>
38 <li><a href="#machineinstr">The <tt>MachineInstr</tt> class</a></li>
39 <li><a href="#machinebasicblock">The <tt>MachineBasicBlock</tt>
41 <li><a href="#machinefunction">The <tt>MachineFunction</tt> class</a></li>
44 <li><a href="#codegenalgs">Target-independent code generation algorithms</a>
46 <li><a href="#instselect">Instruction Selection</a>
48 <li><a href="#selectiondag_intro">Introduction to SelectionDAGs</a></li>
49 <li><a href="#selectiondag_process">SelectionDAG Code Generation
51 <li><a href="#selectiondag_build">Initial SelectionDAG
53 <li><a href="#selectiondag_legalize_types">SelectionDAG LegalizeTypes Phase</a></li>
54 <li><a href="#selectiondag_legalize">SelectionDAG Legalize Phase</a></li>
55 <li><a href="#selectiondag_optimize">SelectionDAG Optimization
56 Phase: the DAG Combiner</a></li>
57 <li><a href="#selectiondag_select">SelectionDAG Select Phase</a></li>
58 <li><a href="#selectiondag_sched">SelectionDAG Scheduling and Formation
60 <li><a href="#selectiondag_future">Future directions for the
63 <li><a href="#liveintervals">Live Intervals</a>
65 <li><a href="#livevariable_analysis">Live Variable Analysis</a></li>
66 <li><a href="#liveintervals_analysis">Live Intervals Analysis</a></li>
68 <li><a href="#regalloc">Register Allocation</a>
70 <li><a href="#regAlloc_represent">How registers are represented in
72 <li><a href="#regAlloc_howTo">Mapping virtual registers to physical
74 <li><a href="#regAlloc_twoAddr">Handling two address instructions</a></li>
75 <li><a href="#regAlloc_ssaDecon">The SSA deconstruction phase</a></li>
76 <li><a href="#regAlloc_fold">Instruction folding</a></li>
77 <li><a href="#regAlloc_builtIn">Built in register allocators</a></li>
79 <li><a href="#codeemit">Code Emission</a>
81 <li><a href="#codeemit_asm">Generating Assembly Code</a></li>
82 <li><a href="#codeemit_bin">Generating Binary Machine Code</a></li>
86 <li><a href="#targetimpls">Target-specific Implementation Notes</a>
88 <li><a href="#tailcallopt">Tail call optimization</a></li>
89 <li><a href="#x86">The X86 backend</a></li>
90 <li><a href="#ppc">The PowerPC backend</a>
92 <li><a href="#ppc_abi">LLVM PowerPC ABI</a></li>
93 <li><a href="#ppc_frame">Frame Layout</a></li>
94 <li><a href="#ppc_prolog">Prolog/Epilog</a></li>
95 <li><a href="#ppc_dynamic">Dynamic Allocation</a></li>
101 <div class="doc_author">
102 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a>,
103 <a href="mailto:isanbard@gmail.com">Bill Wendling</a>,
104 <a href="mailto:pronesto@gmail.com">Fernando Magno Quintao
106 <a href="mailto:jlaskey@mac.com">Jim Laskey</a></p>
109 <div class="doc_warning">
110 <p>Warning: This is a work in progress.</p>
113 <!-- *********************************************************************** -->
114 <div class="doc_section">
115 <a name="introduction">Introduction</a>
117 <!-- *********************************************************************** -->
119 <div class="doc_text">
121 <p>The LLVM target-independent code generator is a framework that provides a
122 suite of reusable components for translating the LLVM internal representation to
123 the machine code for a specified target—either in assembly form (suitable
124 for a static compiler) or in binary machine code format (usable for a JIT
125 compiler). The LLVM target-independent code generator consists of five main
129 <li><a href="#targetdesc">Abstract target description</a> interfaces which
130 capture important properties about various aspects of the machine, independently
131 of how they will be used. These interfaces are defined in
132 <tt>include/llvm/Target/</tt>.</li>
134 <li>Classes used to represent the <a href="#codegendesc">machine code</a> being
135 generated for a target. These classes are intended to be abstract enough to
136 represent the machine code for <i>any</i> target machine. These classes are
137 defined in <tt>include/llvm/CodeGen/</tt>.</li>
139 <li><a href="#codegenalgs">Target-independent algorithms</a> used to implement
140 various phases of native code generation (register allocation, scheduling, stack
141 frame representation, etc). This code lives in <tt>lib/CodeGen/</tt>.</li>
143 <li><a href="#targetimpls">Implementations of the abstract target description
144 interfaces</a> for particular targets. These machine descriptions make use of
145 the components provided by LLVM, and can optionally provide custom
146 target-specific passes, to build complete code generators for a specific target.
147 Target descriptions live in <tt>lib/Target/</tt>.</li>
149 <li><a href="#jit">The target-independent JIT components</a>. The LLVM JIT is
150 completely target independent (it uses the <tt>TargetJITInfo</tt> structure to
151 interface for target-specific issues. The code for the target-independent
152 JIT lives in <tt>lib/ExecutionEngine/JIT</tt>.</li>
157 Depending on which part of the code generator you are interested in working on,
158 different pieces of this will be useful to you. In any case, you should be
159 familiar with the <a href="#targetdesc">target description</a> and <a
160 href="#codegendesc">machine code representation</a> classes. If you want to add
161 a backend for a new target, you will need to <a href="#targetimpls">implement the
162 target description</a> classes for your new target and understand the <a
163 href="LangRef.html">LLVM code representation</a>. If you are interested in
164 implementing a new <a href="#codegenalgs">code generation algorithm</a>, it
165 should only depend on the target-description and machine code representation
166 classes, ensuring that it is portable.
171 <!-- ======================================================================= -->
172 <div class="doc_subsection">
173 <a name="required">Required components in the code generator</a>
176 <div class="doc_text">
178 <p>The two pieces of the LLVM code generator are the high-level interface to the
179 code generator and the set of reusable components that can be used to build
180 target-specific backends. The two most important interfaces (<a
181 href="#targetmachine"><tt>TargetMachine</tt></a> and <a
182 href="#targetdata"><tt>TargetData</tt></a>) are the only ones that are
183 required to be defined for a backend to fit into the LLVM system, but the others
184 must be defined if the reusable code generator components are going to be
187 <p>This design has two important implications. The first is that LLVM can
188 support completely non-traditional code generation targets. For example, the C
189 backend does not require register allocation, instruction selection, or any of
190 the other standard components provided by the system. As such, it only
191 implements these two interfaces, and does its own thing. Another example of a
192 code generator like this is a (purely hypothetical) backend that converts LLVM
193 to the GCC RTL form and uses GCC to emit machine code for a target.</p>
195 <p>This design also implies that it is possible to design and
196 implement radically different code generators in the LLVM system that do not
197 make use of any of the built-in components. Doing so is not recommended at all,
198 but could be required for radically different targets that do not fit into the
199 LLVM machine description model: FPGAs for example.</p>
203 <!-- ======================================================================= -->
204 <div class="doc_subsection">
205 <a name="high-level-design">The high-level design of the code generator</a>
208 <div class="doc_text">
210 <p>The LLVM target-independent code generator is designed to support efficient and
211 quality code generation for standard register-based microprocessors. Code
212 generation in this model is divided into the following stages:</p>
215 <li><b><a href="#instselect">Instruction Selection</a></b> - This phase
216 determines an efficient way to express the input LLVM code in the target
218 This stage produces the initial code for the program in the target instruction
219 set, then makes use of virtual registers in SSA form and physical registers that
220 represent any required register assignments due to target constraints or calling
221 conventions. This step turns the LLVM code into a DAG of target
224 <li><b><a href="#selectiondag_sched">Scheduling and Formation</a></b> - This
225 phase takes the DAG of target instructions produced by the instruction selection
226 phase, determines an ordering of the instructions, then emits the instructions
227 as <tt><a href="#machineinstr">MachineInstr</a></tt>s with that ordering. Note
228 that we describe this in the <a href="#instselect">instruction selection
229 section</a> because it operates on a <a
230 href="#selectiondag_intro">SelectionDAG</a>.
233 <li><b><a href="#ssamco">SSA-based Machine Code Optimizations</a></b> - This
234 optional stage consists of a series of machine-code optimizations that
235 operate on the SSA-form produced by the instruction selector. Optimizations
236 like modulo-scheduling or peephole optimization work here.
239 <li><b><a href="#regalloc">Register Allocation</a></b> - The
240 target code is transformed from an infinite virtual register file in SSA form
241 to the concrete register file used by the target. This phase introduces spill
242 code and eliminates all virtual register references from the program.</li>
244 <li><b><a href="#proepicode">Prolog/Epilog Code Insertion</a></b> - Once the
245 machine code has been generated for the function and the amount of stack space
246 required is known (used for LLVM alloca's and spill slots), the prolog and
247 epilog code for the function can be inserted and "abstract stack location
248 references" can be eliminated. This stage is responsible for implementing
249 optimizations like frame-pointer elimination and stack packing.</li>
251 <li><b><a href="#latemco">Late Machine Code Optimizations</a></b> - Optimizations
252 that operate on "final" machine code can go here, such as spill code scheduling
253 and peephole optimizations.</li>
255 <li><b><a href="#codeemit">Code Emission</a></b> - The final stage actually
256 puts out the code for the current function, either in the target assembler
257 format or in machine code.</li>
261 <p>The code generator is based on the assumption that the instruction selector
262 will use an optimal pattern matching selector to create high-quality sequences of
263 native instructions. Alternative code generator designs based on pattern
264 expansion and aggressive iterative peephole optimization are much slower. This
265 design permits efficient compilation (important for JIT environments) and
266 aggressive optimization (used when generating code offline) by allowing
267 components of varying levels of sophistication to be used for any step of
270 <p>In addition to these stages, target implementations can insert arbitrary
271 target-specific passes into the flow. For example, the X86 target uses a
272 special pass to handle the 80x87 floating point stack architecture. Other
273 targets with unusual requirements can be supported with custom passes as
279 <!-- ======================================================================= -->
280 <div class="doc_subsection">
281 <a name="tablegen">Using TableGen for target description</a>
284 <div class="doc_text">
286 <p>The target description classes require a detailed description of the target
287 architecture. These target descriptions often have a large amount of common
288 information (e.g., an <tt>add</tt> instruction is almost identical to a
289 <tt>sub</tt> instruction).
290 In order to allow the maximum amount of commonality to be factored out, the LLVM
291 code generator uses the <a href="TableGenFundamentals.html">TableGen</a> tool to
292 describe big chunks of the target machine, which allows the use of
293 domain-specific and target-specific abstractions to reduce the amount of
296 <p>As LLVM continues to be developed and refined, we plan to move more and more
297 of the target description to the <tt>.td</tt> form. Doing so gives us a
298 number of advantages. The most important is that it makes it easier to port
299 LLVM because it reduces the amount of C++ code that has to be written, and the
300 surface area of the code generator that needs to be understood before someone
301 can get something working. Second, it makes it easier to change things. In
302 particular, if tables and other things are all emitted by <tt>tblgen</tt>, we
303 only need a change in one place (<tt>tblgen</tt>) to update all of the targets
304 to a new interface.</p>
308 <!-- *********************************************************************** -->
309 <div class="doc_section">
310 <a name="targetdesc">Target description classes</a>
312 <!-- *********************************************************************** -->
314 <div class="doc_text">
316 <p>The LLVM target description classes (located in the
317 <tt>include/llvm/Target</tt> directory) provide an abstract description of the
318 target machine independent of any particular client. These classes are
319 designed to capture the <i>abstract</i> properties of the target (such as the
320 instructions and registers it has), and do not incorporate any particular pieces
321 of code generation algorithms.</p>
323 <p>All of the target description classes (except the <tt><a
324 href="#targetdata">TargetData</a></tt> class) are designed to be subclassed by
325 the concrete target implementation, and have virtual methods implemented. To
326 get to these implementations, the <tt><a
327 href="#targetmachine">TargetMachine</a></tt> class provides accessors that
328 should be implemented by the target.</p>
332 <!-- ======================================================================= -->
333 <div class="doc_subsection">
334 <a name="targetmachine">The <tt>TargetMachine</tt> class</a>
337 <div class="doc_text">
339 <p>The <tt>TargetMachine</tt> class provides virtual methods that are used to
340 access the target-specific implementations of the various target description
341 classes via the <tt>get*Info</tt> methods (<tt>getInstrInfo</tt>,
342 <tt>getRegisterInfo</tt>, <tt>getFrameInfo</tt>, etc.). This class is
343 designed to be specialized by
344 a concrete target implementation (e.g., <tt>X86TargetMachine</tt>) which
345 implements the various virtual methods. The only required target description
346 class is the <a href="#targetdata"><tt>TargetData</tt></a> class, but if the
347 code generator components are to be used, the other interfaces should be
348 implemented as well.</p>
353 <!-- ======================================================================= -->
354 <div class="doc_subsection">
355 <a name="targetdata">The <tt>TargetData</tt> class</a>
358 <div class="doc_text">
360 <p>The <tt>TargetData</tt> class is the only required target description class,
361 and it is the only class that is not extensible (you cannot derived a new
362 class from it). <tt>TargetData</tt> specifies information about how the target
363 lays out memory for structures, the alignment requirements for various data
364 types, the size of pointers in the target, and whether the target is
365 little-endian or big-endian.</p>
369 <!-- ======================================================================= -->
370 <div class="doc_subsection">
371 <a name="targetlowering">The <tt>TargetLowering</tt> class</a>
374 <div class="doc_text">
376 <p>The <tt>TargetLowering</tt> class is used by SelectionDAG based instruction
377 selectors primarily to describe how LLVM code should be lowered to SelectionDAG
378 operations. Among other things, this class indicates:</p>
381 <li>an initial register class to use for various <tt>ValueType</tt>s</li>
382 <li>which operations are natively supported by the target machine</li>
383 <li>the return type of <tt>setcc</tt> operations</li>
384 <li>the type to use for shift amounts</li>
385 <li>various high-level characteristics, like whether it is profitable to turn
386 division by a constant into a multiplication sequence</li>
391 <!-- ======================================================================= -->
392 <div class="doc_subsection">
393 <a name="targetregisterinfo">The <tt>TargetRegisterInfo</tt> class</a>
396 <div class="doc_text">
398 <p>The <tt>TargetRegisterInfo</tt> class is used to describe the register
399 file of the target and any interactions between the registers.</p>
401 <p>Registers in the code generator are represented in the code generator by
402 unsigned integers. Physical registers (those that actually exist in the target
403 description) are unique small numbers, and virtual registers are generally
404 large. Note that register #0 is reserved as a flag value.</p>
406 <p>Each register in the processor description has an associated
407 <tt>TargetRegisterDesc</tt> entry, which provides a textual name for the
408 register (used for assembly output and debugging dumps) and a set of aliases
409 (used to indicate whether one register overlaps with another).
412 <p>In addition to the per-register description, the <tt>TargetRegisterInfo</tt>
413 class exposes a set of processor specific register classes (instances of the
414 <tt>TargetRegisterClass</tt> class). Each register class contains sets of
415 registers that have the same properties (for example, they are all 32-bit
416 integer registers). Each SSA virtual register created by the instruction
417 selector has an associated register class. When the register allocator runs, it
418 replaces virtual registers with a physical register in the set.</p>
421 The target-specific implementations of these classes is auto-generated from a <a
422 href="TableGenFundamentals.html">TableGen</a> description of the register file.
427 <!-- ======================================================================= -->
428 <div class="doc_subsection">
429 <a name="targetinstrinfo">The <tt>TargetInstrInfo</tt> class</a>
432 <div class="doc_text">
433 <p>The <tt>TargetInstrInfo</tt> class is used to describe the machine
434 instructions supported by the target. It is essentially an array of
435 <tt>TargetInstrDescriptor</tt> objects, each of which describes one
436 instruction the target supports. Descriptors define things like the mnemonic
437 for the opcode, the number of operands, the list of implicit register uses
438 and defs, whether the instruction has certain target-independent properties
439 (accesses memory, is commutable, etc), and holds any target-specific
443 <!-- ======================================================================= -->
444 <div class="doc_subsection">
445 <a name="targetframeinfo">The <tt>TargetFrameInfo</tt> class</a>
448 <div class="doc_text">
449 <p>The <tt>TargetFrameInfo</tt> class is used to provide information about the
450 stack frame layout of the target. It holds the direction of stack growth,
451 the known stack alignment on entry to each function, and the offset to the
452 local area. The offset to the local area is the offset from the stack
453 pointer on function entry to the first location where function data (local
454 variables, spill locations) can be stored.</p>
457 <!-- ======================================================================= -->
458 <div class="doc_subsection">
459 <a name="targetsubtarget">The <tt>TargetSubtarget</tt> class</a>
462 <div class="doc_text">
463 <p>The <tt>TargetSubtarget</tt> class is used to provide information about the
464 specific chip set being targeted. A sub-target informs code generation of
465 which instructions are supported, instruction latencies and instruction
466 execution itinerary; i.e., which processing units are used, in what order, and
471 <!-- ======================================================================= -->
472 <div class="doc_subsection">
473 <a name="targetjitinfo">The <tt>TargetJITInfo</tt> class</a>
476 <div class="doc_text">
477 <p>The <tt>TargetJITInfo</tt> class exposes an abstract interface used by the
478 Just-In-Time code generator to perform target-specific activities, such as
479 emitting stubs. If a <tt>TargetMachine</tt> supports JIT code generation, it
480 should provide one of these objects through the <tt>getJITInfo</tt>
484 <!-- *********************************************************************** -->
485 <div class="doc_section">
486 <a name="codegendesc">Machine code description classes</a>
488 <!-- *********************************************************************** -->
490 <div class="doc_text">
492 <p>At the high-level, LLVM code is translated to a machine specific
493 representation formed out of
494 <a href="#machinefunction"><tt>MachineFunction</tt></a>,
495 <a href="#machinebasicblock"><tt>MachineBasicBlock</tt></a>, and <a
496 href="#machineinstr"><tt>MachineInstr</tt></a> instances
497 (defined in <tt>include/llvm/CodeGen</tt>). This representation is completely
498 target agnostic, representing instructions in their most abstract form: an
499 opcode and a series of operands. This representation is designed to support
500 both an SSA representation for machine code, as well as a register allocated,
505 <!-- ======================================================================= -->
506 <div class="doc_subsection">
507 <a name="machineinstr">The <tt>MachineInstr</tt> class</a>
510 <div class="doc_text">
512 <p>Target machine instructions are represented as instances of the
513 <tt>MachineInstr</tt> class. This class is an extremely abstract way of
514 representing machine instructions. In particular, it only keeps track of
515 an opcode number and a set of operands.</p>
517 <p>The opcode number is a simple unsigned integer that only has meaning to a
518 specific backend. All of the instructions for a target should be defined in
519 the <tt>*InstrInfo.td</tt> file for the target. The opcode enum values
520 are auto-generated from this description. The <tt>MachineInstr</tt> class does
521 not have any information about how to interpret the instruction (i.e., what the
522 semantics of the instruction are); for that you must refer to the
523 <tt><a href="#targetinstrinfo">TargetInstrInfo</a></tt> class.</p>
525 <p>The operands of a machine instruction can be of several different types:
526 a register reference, a constant integer, a basic block reference, etc. In
527 addition, a machine operand should be marked as a def or a use of the value
528 (though only registers are allowed to be defs).</p>
530 <p>By convention, the LLVM code generator orders instruction operands so that
531 all register definitions come before the register uses, even on architectures
532 that are normally printed in other orders. For example, the SPARC add
533 instruction: "<tt>add %i1, %i2, %i3</tt>" adds the "%i1", and "%i2" registers
534 and stores the result into the "%i3" register. In the LLVM code generator,
535 the operands should be stored as "<tt>%i3, %i1, %i2</tt>": with the destination
538 <p>Keeping destination (definition) operands at the beginning of the operand
539 list has several advantages. In particular, the debugging printer will print
540 the instruction like this:</p>
542 <div class="doc_code">
548 <p>Also if the first operand is a def, it is easier to <a
549 href="#buildmi">create instructions</a> whose only def is the first
554 <!-- _______________________________________________________________________ -->
555 <div class="doc_subsubsection">
556 <a name="buildmi">Using the <tt>MachineInstrBuilder.h</tt> functions</a>
559 <div class="doc_text">
561 <p>Machine instructions are created by using the <tt>BuildMI</tt> functions,
562 located in the <tt>include/llvm/CodeGen/MachineInstrBuilder.h</tt> file. The
563 <tt>BuildMI</tt> functions make it easy to build arbitrary machine
564 instructions. Usage of the <tt>BuildMI</tt> functions look like this:</p>
566 <div class="doc_code">
568 // Create a 'DestReg = mov 42' (rendered in X86 assembly as 'mov DestReg, 42')
569 // instruction. The '1' specifies how many operands will be added.
570 MachineInstr *MI = BuildMI(X86::MOV32ri, 1, DestReg).addImm(42);
572 // Create the same instr, but insert it at the end of a basic block.
573 MachineBasicBlock &MBB = ...
574 BuildMI(MBB, X86::MOV32ri, 1, DestReg).addImm(42);
576 // Create the same instr, but insert it before a specified iterator point.
577 MachineBasicBlock::iterator MBBI = ...
578 BuildMI(MBB, MBBI, X86::MOV32ri, 1, DestReg).addImm(42);
580 // Create a 'cmp Reg, 0' instruction, no destination reg.
581 MI = BuildMI(X86::CMP32ri, 2).addReg(Reg).addImm(0);
582 // Create an 'sahf' instruction which takes no operands and stores nothing.
583 MI = BuildMI(X86::SAHF, 0);
585 // Create a self looping branch instruction.
586 BuildMI(MBB, X86::JNE, 1).addMBB(&MBB);
590 <p>The key thing to remember with the <tt>BuildMI</tt> functions is that you
591 have to specify the number of operands that the machine instruction will take.
592 This allows for efficient memory allocation. You also need to specify if
593 operands default to be uses of values, not definitions. If you need to add a
594 definition operand (other than the optional destination register), you must
595 explicitly mark it as such:</p>
597 <div class="doc_code">
599 MI.addReg(Reg, MachineOperand::Def);
605 <!-- _______________________________________________________________________ -->
606 <div class="doc_subsubsection">
607 <a name="fixedregs">Fixed (preassigned) registers</a>
610 <div class="doc_text">
612 <p>One important issue that the code generator needs to be aware of is the
613 presence of fixed registers. In particular, there are often places in the
614 instruction stream where the register allocator <em>must</em> arrange for a
615 particular value to be in a particular register. This can occur due to
616 limitations of the instruction set (e.g., the X86 can only do a 32-bit divide
617 with the <tt>EAX</tt>/<tt>EDX</tt> registers), or external factors like calling
618 conventions. In any case, the instruction selector should emit code that
619 copies a virtual register into or out of a physical register when needed.</p>
621 <p>For example, consider this simple LLVM example:</p>
623 <div class="doc_code">
625 define i32 @test(i32 %X, i32 %Y) {
632 <p>The X86 instruction selector produces this machine code for the <tt>div</tt>
633 and <tt>ret</tt> (use
634 "<tt>llc X.bc -march=x86 -print-machineinstrs</tt>" to get this):</p>
636 <div class="doc_code">
639 %EAX = mov %reg1024 ;; Copy X (in reg1024) into EAX
640 %reg1027 = sar %reg1024, 31
641 %EDX = mov %reg1027 ;; Sign extend X into EDX
642 idiv %reg1025 ;; Divide by Y (in reg1025)
643 %reg1026 = mov %EAX ;; Read the result (Z) out of EAX
646 %EAX = mov %reg1026 ;; 32-bit return value goes in EAX
651 <p>By the end of code generation, the register allocator has coalesced
652 the registers and deleted the resultant identity moves producing the
655 <div class="doc_code">
657 ;; X is in EAX, Y is in ECX
665 <p>This approach is extremely general (if it can handle the X86 architecture,
666 it can handle anything!) and allows all of the target specific
667 knowledge about the instruction stream to be isolated in the instruction
668 selector. Note that physical registers should have a short lifetime for good
669 code generation, and all physical registers are assumed dead on entry to and
670 exit from basic blocks (before register allocation). Thus, if you need a value
671 to be live across basic block boundaries, it <em>must</em> live in a virtual
676 <!-- _______________________________________________________________________ -->
677 <div class="doc_subsubsection">
678 <a name="ssa">Machine code in SSA form</a>
681 <div class="doc_text">
683 <p><tt>MachineInstr</tt>'s are initially selected in SSA-form, and
684 are maintained in SSA-form until register allocation happens. For the most
685 part, this is trivially simple since LLVM is already in SSA form; LLVM PHI nodes
686 become machine code PHI nodes, and virtual registers are only allowed to have a
687 single definition.</p>
689 <p>After register allocation, machine code is no longer in SSA-form because there
690 are no virtual registers left in the code.</p>
694 <!-- ======================================================================= -->
695 <div class="doc_subsection">
696 <a name="machinebasicblock">The <tt>MachineBasicBlock</tt> class</a>
699 <div class="doc_text">
701 <p>The <tt>MachineBasicBlock</tt> class contains a list of machine instructions
702 (<tt><a href="#machineinstr">MachineInstr</a></tt> instances). It roughly
703 corresponds to the LLVM code input to the instruction selector, but there can be
704 a one-to-many mapping (i.e. one LLVM basic block can map to multiple machine
705 basic blocks). The <tt>MachineBasicBlock</tt> class has a
706 "<tt>getBasicBlock</tt>" method, which returns the LLVM basic block that it
711 <!-- ======================================================================= -->
712 <div class="doc_subsection">
713 <a name="machinefunction">The <tt>MachineFunction</tt> class</a>
716 <div class="doc_text">
718 <p>The <tt>MachineFunction</tt> class contains a list of machine basic blocks
719 (<tt><a href="#machinebasicblock">MachineBasicBlock</a></tt> instances). It
720 corresponds one-to-one with the LLVM function input to the instruction selector.
721 In addition to a list of basic blocks, the <tt>MachineFunction</tt> contains a
722 a <tt>MachineConstantPool</tt>, a <tt>MachineFrameInfo</tt>, a
723 <tt>MachineFunctionInfo</tt>, and a <tt>MachineRegisterInfo</tt>. See
724 <tt>include/llvm/CodeGen/MachineFunction.h</tt> for more information.</p>
728 <!-- *********************************************************************** -->
729 <div class="doc_section">
730 <a name="codegenalgs">Target-independent code generation algorithms</a>
732 <!-- *********************************************************************** -->
734 <div class="doc_text">
736 <p>This section documents the phases described in the <a
737 href="#high-level-design">high-level design of the code generator</a>. It
738 explains how they work and some of the rationale behind their design.</p>
742 <!-- ======================================================================= -->
743 <div class="doc_subsection">
744 <a name="instselect">Instruction Selection</a>
747 <div class="doc_text">
749 Instruction Selection is the process of translating LLVM code presented to the
750 code generator into target-specific machine instructions. There are several
751 well-known ways to do this in the literature. LLVM uses a SelectionDAG based
752 instruction selector.
755 <p>Portions of the DAG instruction selector are generated from the target
756 description (<tt>*.td</tt>) files. Our goal is for the entire instruction
757 selector to be generated from these <tt>.td</tt> files, though currently
758 there are still things that require custom C++ code.</p>
761 <!-- _______________________________________________________________________ -->
762 <div class="doc_subsubsection">
763 <a name="selectiondag_intro">Introduction to SelectionDAGs</a>
766 <div class="doc_text">
768 <p>The SelectionDAG provides an abstraction for code representation in a way
769 that is amenable to instruction selection using automatic techniques
770 (e.g. dynamic-programming based optimal pattern matching selectors). It is also
771 well-suited to other phases of code generation; in particular,
772 instruction scheduling (SelectionDAG's are very close to scheduling DAGs
773 post-selection). Additionally, the SelectionDAG provides a host representation
774 where a large variety of very-low-level (but target-independent)
775 <a href="#selectiondag_optimize">optimizations</a> may be
776 performed; ones which require extensive information about the instructions
777 efficiently supported by the target.</p>
779 <p>The SelectionDAG is a Directed-Acyclic-Graph whose nodes are instances of the
780 <tt>SDNode</tt> class. The primary payload of the <tt>SDNode</tt> is its
781 operation code (Opcode) that indicates what operation the node performs and
782 the operands to the operation.
783 The various operation node types are described at the top of the
784 <tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt> file.</p>
786 <p>Although most operations define a single value, each node in the graph may
787 define multiple values. For example, a combined div/rem operation will define
788 both the dividend and the remainder. Many other situations require multiple
789 values as well. Each node also has some number of operands, which are edges
790 to the node defining the used value. Because nodes may define multiple values,
791 edges are represented by instances of the <tt>SDValue</tt> class, which is
792 a <tt><SDNode, unsigned></tt> pair, indicating the node and result
793 value being used, respectively. Each value produced by an <tt>SDNode</tt> has
794 an associated <tt>MVT</tt> (Machine Value Type) indicating what the type of the
797 <p>SelectionDAGs contain two different kinds of values: those that represent
798 data flow and those that represent control flow dependencies. Data values are
799 simple edges with an integer or floating point value type. Control edges are
800 represented as "chain" edges which are of type <tt>MVT::Other</tt>. These edges
801 provide an ordering between nodes that have side effects (such as
802 loads, stores, calls, returns, etc). All nodes that have side effects should
803 take a token chain as input and produce a new one as output. By convention,
804 token chain inputs are always operand #0, and chain results are always the last
805 value produced by an operation.</p>
807 <p>A SelectionDAG has designated "Entry" and "Root" nodes. The Entry node is
808 always a marker node with an Opcode of <tt>ISD::EntryToken</tt>. The Root node
809 is the final side-effecting node in the token chain. For example, in a single
810 basic block function it would be the return node.</p>
812 <p>One important concept for SelectionDAGs is the notion of a "legal" vs.
813 "illegal" DAG. A legal DAG for a target is one that only uses supported
814 operations and supported types. On a 32-bit PowerPC, for example, a DAG with
815 a value of type i1, i8, i16, or i64 would be illegal, as would a DAG that uses a
816 SREM or UREM operation. The
817 <a href="#selectinodag_legalize_types">legalize types</a> and
818 <a href="#selectiondag_legalize">legalize operations</a> phases are
819 responsible for turning an illegal DAG into a legal DAG.</p>
823 <!-- _______________________________________________________________________ -->
824 <div class="doc_subsubsection">
825 <a name="selectiondag_process">SelectionDAG Instruction Selection Process</a>
828 <div class="doc_text">
830 <p>SelectionDAG-based instruction selection consists of the following steps:</p>
833 <li><a href="#selectiondag_build">Build initial DAG</a> - This stage
834 performs a simple translation from the input LLVM code to an illegal
836 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> - This stage
837 performs simple optimizations on the SelectionDAG to simplify it, and
838 recognize meta instructions (like rotates and <tt>div</tt>/<tt>rem</tt>
839 pairs) for targets that support these meta operations. This makes the
840 resultant code more efficient and the <a href="#selectiondag_select">select
841 instructions from DAG</a> phase (below) simpler.</li>
842 <li><a href="#selectiondag_legalize_types">Legalize SelectionDAG Types</a> - This
843 stage transforms SelectionDAG nodes to eliminate any types that are
844 unsupported on the target.</li>
845 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> - The
846 SelectionDAG optimizer is run to clean up redundancies exposed
847 by type legalization.</li>
848 <li><a href="#selectiondag_legalize">Legalize SelectionDAG Types</a> - This
849 stage transforms SelectionDAG nodes to eliminate any types that are
850 unsupported on the target.</li>
851 <li><a href="#selectiondag_optimize">Optimize SelectionDAG</a> - The
852 SelectionDAG optimizer is run to eliminate inefficiencies introduced
853 by operation legalization.</li>
854 <li><a href="#selectiondag_select">Select instructions from DAG</a> - Finally,
855 the target instruction selector matches the DAG operations to target
856 instructions. This process translates the target-independent input DAG into
857 another DAG of target instructions.</li>
858 <li><a href="#selectiondag_sched">SelectionDAG Scheduling and Formation</a>
859 - The last phase assigns a linear order to the instructions in the
860 target-instruction DAG and emits them into the MachineFunction being
861 compiled. This step uses traditional prepass scheduling techniques.</li>
864 <p>After all of these steps are complete, the SelectionDAG is destroyed and the
865 rest of the code generation passes are run.</p>
867 <p>One great way to visualize what is going on here is to take advantage of a
868 few LLC command line options. The following options pop up a window displaying
869 the SelectionDAG at specific times (if you only get errors printed to the console
870 while using this, you probably
871 <a href="ProgrammersManual.html#ViewGraph">need to configure your system</a> to
872 add support for it).</p>
875 <li><tt>-view-dag-combine1-dags</tt> displays the DAG after being built, before
876 the first optimization pass.</li>
877 <li><tt>-view-legalize-dags</tt> displays the DAG before Legalization.</li>
878 <li><tt>-view-dag-combine2-dags</tt> displays the DAG before the second
879 optimization pass.</li>
880 <li><tt>-view-isel-dags</tt> displays the DAG before the Select phase.</li>
881 <li><tt>-view-sched-dags</tt> displays the DAG before Scheduling.</li>
884 <p>The <tt>-view-sunit-dags</tt> displays the Scheduler's dependency graph.
885 This graph is based on the final SelectionDAG, with nodes that must be
886 scheduled together bundled into a single scheduling-unit node, and with
887 immediate operands and other nodes that aren't relevant for scheduling
893 <!-- _______________________________________________________________________ -->
894 <div class="doc_subsubsection">
895 <a name="selectiondag_build">Initial SelectionDAG Construction</a>
898 <div class="doc_text">
900 <p>The initial SelectionDAG is naïvely peephole expanded from the LLVM
901 input by the <tt>SelectionDAGLowering</tt> class in the
902 <tt>lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp</tt> file. The intent of this
903 pass is to expose as much low-level, target-specific details to the SelectionDAG
904 as possible. This pass is mostly hard-coded (e.g. an LLVM <tt>add</tt> turns
905 into an <tt>SDNode add</tt> while a <tt>getelementptr</tt> is expanded into the
906 obvious arithmetic). This pass requires target-specific hooks to lower calls,
907 returns, varargs, etc. For these features, the
908 <tt><a href="#targetlowering">TargetLowering</a></tt> interface is used.</p>
912 <!-- _______________________________________________________________________ -->
913 <div class="doc_subsubsection">
914 <a name="selectiondag_legalize_types">SelectionDAG LegalizeTypes Phase</a>
917 <div class="doc_text">
919 <p>The Legalize phase is in charge of converting a DAG to only use the types
920 that are natively supported by the target.</p>
922 <p>There are two main ways of converting values of unsupported scalar types
923 to values of supported types: converting small types to
924 larger types ("promoting"), and breaking up large integer types
925 into smaller ones ("expanding"). For example, a target might require
926 that all f32 values are promoted to f64 and that all i1/i8/i16 values
927 are promoted to i32. The same target might require that all i64 values
928 be expanded into pairs of i32 values. These changes can insert sign and
929 zero extensions as needed to make sure that the final code has the same
930 behavior as the input.</p>
932 <p>There are two main ways of converting values of unsupported vector types
933 to value of supported types: splitting vector types, multiple times if
934 necessary, until a legal type is found, and extending vector types by
935 adding elements to the end to round them out to legal types ("widening").
936 If a vector gets split all the way down to single-element parts with
937 no supported vector type being found, the elements are converted to
938 scalars ("scalarizing").</p>
940 <p>A target implementation tells the legalizer which types are supported
941 (and which register class to use for them) by calling the
942 <tt>addRegisterClass</tt> method in its TargetLowering constructor.</p>
946 <!-- _______________________________________________________________________ -->
947 <div class="doc_subsubsection">
948 <a name="selectiondag_legalize">SelectionDAG Legalize Phase</a>
951 <div class="doc_text">
953 <p>The Legalize phase is in charge of converting a DAG to only use the
954 operations that are natively supported by the target.</p>
956 <p>Targets often have weird constraints, such as not supporting every
957 operation on every supported datatype (e.g. X86 does not support byte
958 conditional moves and PowerPC does not support sign-extending loads from
959 a 16-bit memory location). Legalize takes care of this by open-coding
960 another sequence of operations to emulate the operation ("expansion"), by
961 promoting one type to a larger type that supports the operation
962 ("promotion"), or by using a target-specific hook to implement the
963 legalization ("custom").</p>
965 <p>A target implementation tells the legalizer which operations are not
966 supported (and which of the above three actions to take) by calling the
967 <tt>setOperationAction</tt> method in its <tt>TargetLowering</tt>
970 <p>Prior to the existence of the Legalize passes, we required that every target
971 <a href="#selectiondag_optimize">selector</a> supported and handled every
972 operator and type even if they are not natively supported. The introduction of
973 the Legalize phases allows all of the canonicalization patterns to be shared
974 across targets, and makes it very easy to optimize the canonicalized code
975 because it is still in the form of a DAG.</p>
979 <!-- _______________________________________________________________________ -->
980 <div class="doc_subsubsection">
981 <a name="selectiondag_optimize">SelectionDAG Optimization Phase: the DAG
985 <div class="doc_text">
987 <p>The SelectionDAG optimization phase is run multiple times for code generation,
988 immediately after the DAG is built and once after each legalization. The first
989 run of the pass allows the initial code to be cleaned up (e.g. performing
990 optimizations that depend on knowing that the operators have restricted type
991 inputs). Subsequent runs of the pass clean up the messy code generated by the
992 Legalize passes, which allows Legalize to be very simple (it can focus on making
993 code legal instead of focusing on generating <em>good</em> and legal code).</p>
995 <p>One important class of optimizations performed is optimizing inserted sign
996 and zero extension instructions. We currently use ad-hoc techniques, but could
997 move to more rigorous techniques in the future. Here are some good papers on
1001 "<a href="http://www.eecs.harvard.edu/~nr/pubs/widen-abstract.html">Widening
1002 integer arithmetic</a>"<br>
1003 Kevin Redwine and Norman Ramsey<br>
1004 International Conference on Compiler Construction (CC) 2004
1009 "<a href="http://portal.acm.org/citation.cfm?doid=512529.512552">Effective
1010 sign extension elimination</a>"<br>
1011 Motohiro Kawahito, Hideaki Komatsu, and Toshio Nakatani<br>
1012 Proceedings of the ACM SIGPLAN 2002 Conference on Programming Language Design
1018 <!-- _______________________________________________________________________ -->
1019 <div class="doc_subsubsection">
1020 <a name="selectiondag_select">SelectionDAG Select Phase</a>
1023 <div class="doc_text">
1025 <p>The Select phase is the bulk of the target-specific code for instruction
1026 selection. This phase takes a legal SelectionDAG as input, pattern matches the
1027 instructions supported by the target to this DAG, and produces a new DAG of
1028 target code. For example, consider the following LLVM fragment:</p>
1030 <div class="doc_code">
1032 %t1 = add float %W, %X
1033 %t2 = mul float %t1, %Y
1034 %t3 = add float %t2, %Z
1038 <p>This LLVM code corresponds to a SelectionDAG that looks basically like
1041 <div class="doc_code">
1043 (fadd:f32 (fmul:f32 (fadd:f32 W, X), Y), Z)
1047 <p>If a target supports floating point multiply-and-add (FMA) operations, one
1048 of the adds can be merged with the multiply. On the PowerPC, for example, the
1049 output of the instruction selector might look like this DAG:</p>
1051 <div class="doc_code">
1053 (FMADDS (FADDS W, X), Y, Z)
1057 <p>The <tt>FMADDS</tt> instruction is a ternary instruction that multiplies its
1058 first two operands and adds the third (as single-precision floating-point
1059 numbers). The <tt>FADDS</tt> instruction is a simple binary single-precision
1060 add instruction. To perform this pattern match, the PowerPC backend includes
1061 the following instruction definitions:</p>
1063 <div class="doc_code">
1065 def FMADDS : AForm_1<59, 29,
1066 (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
1067 "fmadds $FRT, $FRA, $FRC, $FRB",
1068 [<b>(set F4RC:$FRT, (fadd (fmul F4RC:$FRA, F4RC:$FRC),
1069 F4RC:$FRB))</b>]>;
1070 def FADDS : AForm_2<59, 21,
1071 (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
1072 "fadds $FRT, $FRA, $FRB",
1073 [<b>(set F4RC:$FRT, (fadd F4RC:$FRA, F4RC:$FRB))</b>]>;
1077 <p>The portion of the instruction definition in bold indicates the pattern used
1078 to match the instruction. The DAG operators (like <tt>fmul</tt>/<tt>fadd</tt>)
1079 are defined in the <tt>lib/Target/TargetSelectionDAG.td</tt> file.
1080 "<tt>F4RC</tt>" is the register class of the input and result values.</p>
1082 <p>The TableGen DAG instruction selector generator reads the instruction
1083 patterns in the <tt>.td</tt> file and automatically builds parts of the pattern
1084 matching code for your target. It has the following strengths:</p>
1087 <li>At compiler-compiler time, it analyzes your instruction patterns and tells
1088 you if your patterns make sense or not.</li>
1089 <li>It can handle arbitrary constraints on operands for the pattern match. In
1090 particular, it is straight-forward to say things like "match any immediate
1091 that is a 13-bit sign-extended value". For examples, see the
1092 <tt>immSExt16</tt> and related <tt>tblgen</tt> classes in the PowerPC
1094 <li>It knows several important identities for the patterns defined. For
1095 example, it knows that addition is commutative, so it allows the
1096 <tt>FMADDS</tt> pattern above to match "<tt>(fadd X, (fmul Y, Z))</tt>" as
1097 well as "<tt>(fadd (fmul X, Y), Z)</tt>", without the target author having
1098 to specially handle this case.</li>
1099 <li>It has a full-featured type-inferencing system. In particular, you should
1100 rarely have to explicitly tell the system what type parts of your patterns
1101 are. In the <tt>FMADDS</tt> case above, we didn't have to tell
1102 <tt>tblgen</tt> that all of the nodes in the pattern are of type 'f32'. It
1103 was able to infer and propagate this knowledge from the fact that
1104 <tt>F4RC</tt> has type 'f32'.</li>
1105 <li>Targets can define their own (and rely on built-in) "pattern fragments".
1106 Pattern fragments are chunks of reusable patterns that get inlined into your
1107 patterns during compiler-compiler time. For example, the integer
1108 "<tt>(not x)</tt>" operation is actually defined as a pattern fragment that
1109 expands as "<tt>(xor x, -1)</tt>", since the SelectionDAG does not have a
1110 native '<tt>not</tt>' operation. Targets can define their own short-hand
1111 fragments as they see fit. See the definition of '<tt>not</tt>' and
1112 '<tt>ineg</tt>' for examples.</li>
1113 <li>In addition to instructions, targets can specify arbitrary patterns that
1114 map to one or more instructions using the 'Pat' class. For example,
1115 the PowerPC has no way to load an arbitrary integer immediate into a
1116 register in one instruction. To tell tblgen how to do this, it defines:
1119 <div class="doc_code">
1121 // Arbitrary immediate support. Implement in terms of LIS/ORI.
1122 def : Pat<(i32 imm:$imm),
1123 (ORI (LIS (HI16 imm:$imm)), (LO16 imm:$imm))>;
1127 If none of the single-instruction patterns for loading an immediate into a
1128 register match, this will be used. This rule says "match an arbitrary i32
1129 immediate, turning it into an <tt>ORI</tt> ('or a 16-bit immediate') and an
1130 <tt>LIS</tt> ('load 16-bit immediate, where the immediate is shifted to the
1131 left 16 bits') instruction". To make this work, the
1132 <tt>LO16</tt>/<tt>HI16</tt> node transformations are used to manipulate the
1133 input immediate (in this case, take the high or low 16-bits of the
1135 <li>While the system does automate a lot, it still allows you to write custom
1136 C++ code to match special cases if there is something that is hard to
1140 <p>While it has many strengths, the system currently has some limitations,
1141 primarily because it is a work in progress and is not yet finished:</p>
1144 <li>Overall, there is no way to define or match SelectionDAG nodes that define
1145 multiple values (e.g. <tt>ADD_PARTS</tt>, <tt>LOAD</tt>, <tt>CALL</tt>,
1146 etc). This is the biggest reason that you currently still <em>have to</em>
1147 write custom C++ code for your instruction selector.</li>
1148 <li>There is no great way to support matching complex addressing modes yet. In
1149 the future, we will extend pattern fragments to allow them to define
1150 multiple values (e.g. the four operands of the <a href="#x86_memory">X86
1151 addressing mode</a>, which are currently matched with custom C++ code).
1152 In addition, we'll extend fragments so that a
1153 fragment can match multiple different patterns.</li>
1154 <li>We don't automatically infer flags like isStore/isLoad yet.</li>
1155 <li>We don't automatically generate the set of supported registers and
1156 operations for the <a href="#selectiondag_legalize">Legalizer</a> yet.</li>
1157 <li>We don't have a way of tying in custom legalized nodes yet.</li>
1160 <p>Despite these limitations, the instruction selector generator is still quite
1161 useful for most of the binary and logical operations in typical instruction
1162 sets. If you run into any problems or can't figure out how to do something,
1163 please let Chris know!</p>
1167 <!-- _______________________________________________________________________ -->
1168 <div class="doc_subsubsection">
1169 <a name="selectiondag_sched">SelectionDAG Scheduling and Formation Phase</a>
1172 <div class="doc_text">
1174 <p>The scheduling phase takes the DAG of target instructions from the selection
1175 phase and assigns an order. The scheduler can pick an order depending on
1176 various constraints of the machines (i.e. order for minimal register pressure or
1177 try to cover instruction latencies). Once an order is established, the DAG is
1178 converted to a list of <tt><a href="#machineinstr">MachineInstr</a></tt>s and
1179 the SelectionDAG is destroyed.</p>
1181 <p>Note that this phase is logically separate from the instruction selection
1182 phase, but is tied to it closely in the code because it operates on
1187 <!-- _______________________________________________________________________ -->
1188 <div class="doc_subsubsection">
1189 <a name="selectiondag_future">Future directions for the SelectionDAG</a>
1192 <div class="doc_text">
1195 <li>Optional function-at-a-time selection.</li>
1196 <li>Auto-generate entire selector from <tt>.td</tt> file.</li>
1201 <!-- ======================================================================= -->
1202 <div class="doc_subsection">
1203 <a name="ssamco">SSA-based Machine Code Optimizations</a>
1205 <div class="doc_text"><p>To Be Written</p></div>
1207 <!-- ======================================================================= -->
1208 <div class="doc_subsection">
1209 <a name="liveintervals">Live Intervals</a>
1212 <div class="doc_text">
1214 <p>Live Intervals are the ranges (intervals) where a variable is <i>live</i>.
1215 They are used by some <a href="#regalloc">register allocator</a> passes to
1216 determine if two or more virtual registers which require the same physical
1217 register are live at the same point in the program (i.e., they conflict). When
1218 this situation occurs, one virtual register must be <i>spilled</i>.</p>
1222 <!-- _______________________________________________________________________ -->
1223 <div class="doc_subsubsection">
1224 <a name="livevariable_analysis">Live Variable Analysis</a>
1227 <div class="doc_text">
1229 <p>The first step in determining the live intervals of variables is to
1230 calculate the set of registers that are immediately dead after the
1231 instruction (i.e., the instruction calculates the value, but it is
1232 never used) and the set of registers that are used by the instruction,
1233 but are never used after the instruction (i.e., they are killed). Live
1234 variable information is computed for each <i>virtual</i> register and
1235 <i>register allocatable</i> physical register in the function. This
1236 is done in a very efficient manner because it uses SSA to sparsely
1237 compute lifetime information for virtual registers (which are in SSA
1238 form) and only has to track physical registers within a block. Before
1239 register allocation, LLVM can assume that physical registers are only
1240 live within a single basic block. This allows it to do a single,
1241 local analysis to resolve physical register lifetimes within each
1242 basic block. If a physical register is not register allocatable (e.g.,
1243 a stack pointer or condition codes), it is not tracked.</p>
1245 <p>Physical registers may be live in to or out of a function. Live in values
1246 are typically arguments in registers. Live out values are typically return
1247 values in registers. Live in values are marked as such, and are given a dummy
1248 "defining" instruction during live intervals analysis. If the last basic block
1249 of a function is a <tt>return</tt>, then it's marked as using all live out
1250 values in the function.</p>
1252 <p><tt>PHI</tt> nodes need to be handled specially, because the calculation
1253 of the live variable information from a depth first traversal of the CFG of
1254 the function won't guarantee that a virtual register used by the <tt>PHI</tt>
1255 node is defined before it's used. When a <tt>PHI</tt> node is encountered, only
1256 the definition is handled, because the uses will be handled in other basic
1259 <p>For each <tt>PHI</tt> node of the current basic block, we simulate an
1260 assignment at the end of the current basic block and traverse the successor
1261 basic blocks. If a successor basic block has a <tt>PHI</tt> node and one of
1262 the <tt>PHI</tt> node's operands is coming from the current basic block,
1263 then the variable is marked as <i>alive</i> within the current basic block
1264 and all of its predecessor basic blocks, until the basic block with the
1265 defining instruction is encountered.</p>
1269 <!-- _______________________________________________________________________ -->
1270 <div class="doc_subsubsection">
1271 <a name="liveintervals_analysis">Live Intervals Analysis</a>
1274 <div class="doc_text">
1276 <p>We now have the information available to perform the live intervals analysis
1277 and build the live intervals themselves. We start off by numbering the basic
1278 blocks and machine instructions. We then handle the "live-in" values. These
1279 are in physical registers, so the physical register is assumed to be killed by
1280 the end of the basic block. Live intervals for virtual registers are computed
1281 for some ordering of the machine instructions <tt>[1, N]</tt>. A live interval
1282 is an interval <tt>[i, j)</tt>, where <tt>1 <= i <= j < N</tt>, for which a
1283 variable is live.</p>
1285 <p><i><b>More to come...</b></i></p>
1289 <!-- ======================================================================= -->
1290 <div class="doc_subsection">
1291 <a name="regalloc">Register Allocation</a>
1294 <div class="doc_text">
1296 <p>The <i>Register Allocation problem</i> consists in mapping a program
1297 <i>P<sub>v</sub></i>, that can use an unbounded number of virtual
1298 registers, to a program <i>P<sub>p</sub></i> that contains a finite
1299 (possibly small) number of physical registers. Each target architecture has
1300 a different number of physical registers. If the number of physical
1301 registers is not enough to accommodate all the virtual registers, some of
1302 them will have to be mapped into memory. These virtuals are called
1303 <i>spilled virtuals</i>.</p>
1307 <!-- _______________________________________________________________________ -->
1309 <div class="doc_subsubsection">
1310 <a name="regAlloc_represent">How registers are represented in LLVM</a>
1313 <div class="doc_text">
1315 <p>In LLVM, physical registers are denoted by integer numbers that
1316 normally range from 1 to 1023. To see how this numbering is defined
1317 for a particular architecture, you can read the
1318 <tt>GenRegisterNames.inc</tt> file for that architecture. For
1319 instance, by inspecting
1320 <tt>lib/Target/X86/X86GenRegisterNames.inc</tt> we see that the 32-bit
1321 register <tt>EAX</tt> is denoted by 15, and the MMX register
1322 <tt>MM0</tt> is mapped to 48.</p>
1324 <p>Some architectures contain registers that share the same physical
1325 location. A notable example is the X86 platform. For instance, in the
1326 X86 architecture, the registers <tt>EAX</tt>, <tt>AX</tt> and
1327 <tt>AL</tt> share the first eight bits. These physical registers are
1328 marked as <i>aliased</i> in LLVM. Given a particular architecture, you
1329 can check which registers are aliased by inspecting its
1330 <tt>RegisterInfo.td</tt> file. Moreover, the method
1331 <tt>TargetRegisterInfo::getAliasSet(p_reg)</tt> returns an array containing
1332 all the physical registers aliased to the register <tt>p_reg</tt>.</p>
1334 <p>Physical registers, in LLVM, are grouped in <i>Register Classes</i>.
1335 Elements in the same register class are functionally equivalent, and can
1336 be interchangeably used. Each virtual register can only be mapped to
1337 physical registers of a particular class. For instance, in the X86
1338 architecture, some virtuals can only be allocated to 8 bit registers.
1339 A register class is described by <tt>TargetRegisterClass</tt> objects.
1340 To discover if a virtual register is compatible with a given physical,
1341 this code can be used:
1344 <div class="doc_code">
1346 bool RegMapping_Fer::compatible_class(MachineFunction &mf,
1349 assert(TargetRegisterInfo::isPhysicalRegister(p_reg) &&
1350 "Target register must be physical");
1351 const TargetRegisterClass *trc = mf.getRegInfo().getRegClass(v_reg);
1352 return trc->contains(p_reg);
1357 <p>Sometimes, mostly for debugging purposes, it is useful to change
1358 the number of physical registers available in the target
1359 architecture. This must be done statically, inside the
1360 <tt>TargetRegsterInfo.td</tt> file. Just <tt>grep</tt> for
1361 <tt>RegisterClass</tt>, the last parameter of which is a list of
1362 registers. Just commenting some out is one simple way to avoid them
1363 being used. A more polite way is to explicitly exclude some registers
1364 from the <i>allocation order</i>. See the definition of the
1365 <tt>GR</tt> register class in
1366 <tt>lib/Target/IA64/IA64RegisterInfo.td</tt> for an example of this
1367 (e.g., <tt>numReservedRegs</tt> registers are hidden.)</p>
1369 <p>Virtual registers are also denoted by integer numbers. Contrary to
1370 physical registers, different virtual registers never share the same
1371 number. The smallest virtual register is normally assigned the number
1372 1024. This may change, so, in order to know which is the first virtual
1373 register, you should access
1374 <tt>TargetRegisterInfo::FirstVirtualRegister</tt>. Any register whose
1375 number is greater than or equal to
1376 <tt>TargetRegisterInfo::FirstVirtualRegister</tt> is considered a virtual
1377 register. Whereas physical registers are statically defined in a
1378 <tt>TargetRegisterInfo.td</tt> file and cannot be created by the
1379 application developer, that is not the case with virtual registers.
1380 In order to create new virtual registers, use the method
1381 <tt>MachineRegisterInfo::createVirtualRegister()</tt>. This method will return a
1382 virtual register with the highest code.
1385 <p>Before register allocation, the operands of an instruction are
1386 mostly virtual registers, although physical registers may also be
1387 used. In order to check if a given machine operand is a register, use
1388 the boolean function <tt>MachineOperand::isRegister()</tt>. To obtain
1389 the integer code of a register, use
1390 <tt>MachineOperand::getReg()</tt>. An instruction may define or use a
1391 register. For instance, <tt>ADD reg:1026 := reg:1025 reg:1024</tt>
1392 defines the registers 1024, and uses registers 1025 and 1026. Given a
1393 register operand, the method <tt>MachineOperand::isUse()</tt> informs
1394 if that register is being used by the instruction. The method
1395 <tt>MachineOperand::isDef()</tt> informs if that registers is being
1398 <p>We will call physical registers present in the LLVM bitcode before
1399 register allocation <i>pre-colored registers</i>. Pre-colored
1400 registers are used in many different situations, for instance, to pass
1401 parameters of functions calls, and to store results of particular
1402 instructions. There are two types of pre-colored registers: the ones
1403 <i>implicitly</i> defined, and those <i>explicitly</i>
1404 defined. Explicitly defined registers are normal operands, and can be
1405 accessed with <tt>MachineInstr::getOperand(int)::getReg()</tt>. In
1406 order to check which registers are implicitly defined by an
1407 instruction, use the
1408 <tt>TargetInstrInfo::get(opcode)::ImplicitDefs</tt>, where
1409 <tt>opcode</tt> is the opcode of the target instruction. One important
1410 difference between explicit and implicit physical registers is that
1411 the latter are defined statically for each instruction, whereas the
1412 former may vary depending on the program being compiled. For example,
1413 an instruction that represents a function call will always implicitly
1414 define or use the same set of physical registers. To read the
1415 registers implicitly used by an instruction, use
1416 <tt>TargetInstrInfo::get(opcode)::ImplicitUses</tt>. Pre-colored
1417 registers impose constraints on any register allocation algorithm. The
1418 register allocator must make sure that none of them is been
1419 overwritten by the values of virtual registers while still alive.</p>
1423 <!-- _______________________________________________________________________ -->
1425 <div class="doc_subsubsection">
1426 <a name="regAlloc_howTo">Mapping virtual registers to physical registers</a>
1429 <div class="doc_text">
1431 <p>There are two ways to map virtual registers to physical registers (or to
1432 memory slots). The first way, that we will call <i>direct mapping</i>,
1433 is based on the use of methods of the classes <tt>TargetRegisterInfo</tt>,
1434 and <tt>MachineOperand</tt>. The second way, that we will call
1435 <i>indirect mapping</i>, relies on the <tt>VirtRegMap</tt> class in
1436 order to insert loads and stores sending and getting values to and from
1439 <p>The direct mapping provides more flexibility to the developer of
1440 the register allocator; however, it is more error prone, and demands
1441 more implementation work. Basically, the programmer will have to
1442 specify where load and store instructions should be inserted in the
1443 target function being compiled in order to get and store values in
1444 memory. To assign a physical register to a virtual register present in
1445 a given operand, use <tt>MachineOperand::setReg(p_reg)</tt>. To insert
1446 a store instruction, use
1447 <tt>TargetRegisterInfo::storeRegToStackSlot(...)</tt>, and to insert a load
1448 instruction, use <tt>TargetRegisterInfo::loadRegFromStackSlot</tt>.</p>
1450 <p>The indirect mapping shields the application developer from the
1451 complexities of inserting load and store instructions. In order to map
1452 a virtual register to a physical one, use
1453 <tt>VirtRegMap::assignVirt2Phys(vreg, preg)</tt>. In order to map a
1454 certain virtual register to memory, use
1455 <tt>VirtRegMap::assignVirt2StackSlot(vreg)</tt>. This method will
1456 return the stack slot where <tt>vreg</tt>'s value will be located. If
1457 it is necessary to map another virtual register to the same stack
1458 slot, use <tt>VirtRegMap::assignVirt2StackSlot(vreg,
1459 stack_location)</tt>. One important point to consider when using the
1460 indirect mapping, is that even if a virtual register is mapped to
1461 memory, it still needs to be mapped to a physical register. This
1462 physical register is the location where the virtual register is
1463 supposed to be found before being stored or after being reloaded.</p>
1465 <p>If the indirect strategy is used, after all the virtual registers
1466 have been mapped to physical registers or stack slots, it is necessary
1467 to use a spiller object to place load and store instructions in the
1468 code. Every virtual that has been mapped to a stack slot will be
1469 stored to memory after been defined and will be loaded before being
1470 used. The implementation of the spiller tries to recycle load/store
1471 instructions, avoiding unnecessary instructions. For an example of how
1472 to invoke the spiller, see
1473 <tt>RegAllocLinearScan::runOnMachineFunction</tt> in
1474 <tt>lib/CodeGen/RegAllocLinearScan.cpp</tt>.</p>
1478 <!-- _______________________________________________________________________ -->
1479 <div class="doc_subsubsection">
1480 <a name="regAlloc_twoAddr">Handling two address instructions</a>
1483 <div class="doc_text">
1485 <p>With very rare exceptions (e.g., function calls), the LLVM machine
1486 code instructions are three address instructions. That is, each
1487 instruction is expected to define at most one register, and to use at
1488 most two registers. However, some architectures use two address
1489 instructions. In this case, the defined register is also one of the
1490 used register. For instance, an instruction such as <tt>ADD %EAX,
1491 %EBX</tt>, in X86 is actually equivalent to <tt>%EAX = %EAX +
1494 <p>In order to produce correct code, LLVM must convert three address
1495 instructions that represent two address instructions into true two
1496 address instructions. LLVM provides the pass
1497 <tt>TwoAddressInstructionPass</tt> for this specific purpose. It must
1498 be run before register allocation takes place. After its execution,
1499 the resulting code may no longer be in SSA form. This happens, for
1500 instance, in situations where an instruction such as <tt>%a = ADD %b
1501 %c</tt> is converted to two instructions such as:</p>
1503 <div class="doc_code">
1510 <p>Notice that, internally, the second instruction is represented as
1511 <tt>ADD %a[def/use] %c</tt>. I.e., the register operand <tt>%a</tt> is
1512 both used and defined by the instruction.</p>
1516 <!-- _______________________________________________________________________ -->
1517 <div class="doc_subsubsection">
1518 <a name="regAlloc_ssaDecon">The SSA deconstruction phase</a>
1521 <div class="doc_text">
1523 <p>An important transformation that happens during register allocation is called
1524 the <i>SSA Deconstruction Phase</i>. The SSA form simplifies many
1525 analyses that are performed on the control flow graph of
1526 programs. However, traditional instruction sets do not implement
1527 PHI instructions. Thus, in order to generate executable code, compilers
1528 must replace PHI instructions with other instructions that preserve their
1531 <p>There are many ways in which PHI instructions can safely be removed
1532 from the target code. The most traditional PHI deconstruction
1533 algorithm replaces PHI instructions with copy instructions. That is
1534 the strategy adopted by LLVM. The SSA deconstruction algorithm is
1535 implemented in <tt>lib/CodeGen/PHIElimination.cpp</tt>. In order to
1536 invoke this pass, the identifier <tt>PHIEliminationID</tt> must be
1537 marked as required in the code of the register allocator.</p>
1541 <!-- _______________________________________________________________________ -->
1542 <div class="doc_subsubsection">
1543 <a name="regAlloc_fold">Instruction folding</a>
1546 <div class="doc_text">
1548 <p><i>Instruction folding</i> is an optimization performed during
1549 register allocation that removes unnecessary copy instructions. For
1550 instance, a sequence of instructions such as:</p>
1552 <div class="doc_code">
1554 %EBX = LOAD %mem_address
1559 <p>can be safely substituted by the single instruction:</p>
1561 <div class="doc_code">
1563 %EAX = LOAD %mem_address
1567 <p>Instructions can be folded with the
1568 <tt>TargetRegisterInfo::foldMemoryOperand(...)</tt> method. Care must be
1569 taken when folding instructions; a folded instruction can be quite
1570 different from the original instruction. See
1571 <tt>LiveIntervals::addIntervalsForSpills</tt> in
1572 <tt>lib/CodeGen/LiveIntervalAnalysis.cpp</tt> for an example of its use.</p>
1576 <!-- _______________________________________________________________________ -->
1578 <div class="doc_subsubsection">
1579 <a name="regAlloc_builtIn">Built in register allocators</a>
1582 <div class="doc_text">
1584 <p>The LLVM infrastructure provides the application developer with
1585 three different register allocators:</p>
1588 <li><i>Simple</i> - This is a very simple implementation that does
1589 not keep values in registers across instructions. This register
1590 allocator immediately spills every value right after it is
1591 computed, and reloads all used operands from memory to temporary
1592 registers before each instruction.</li>
1593 <li><i>Local</i> - This register allocator is an improvement on the
1594 <i>Simple</i> implementation. It allocates registers on a basic
1595 block level, attempting to keep values in registers and reusing
1596 registers as appropriate.</li>
1597 <li><i>Linear Scan</i> - <i>The default allocator</i>. This is the
1598 well-know linear scan register allocator. Whereas the
1599 <i>Simple</i> and <i>Local</i> algorithms use a direct mapping
1600 implementation technique, the <i>Linear Scan</i> implementation
1601 uses a spiller in order to place load and stores.</li>
1604 <p>The type of register allocator used in <tt>llc</tt> can be chosen with the
1605 command line option <tt>-regalloc=...</tt>:</p>
1607 <div class="doc_code">
1609 $ llc -f -regalloc=simple file.bc -o sp.s;
1610 $ llc -f -regalloc=local file.bc -o lc.s;
1611 $ llc -f -regalloc=linearscan file.bc -o ln.s;
1617 <!-- ======================================================================= -->
1618 <div class="doc_subsection">
1619 <a name="proepicode">Prolog/Epilog Code Insertion</a>
1621 <div class="doc_text"><p>To Be Written</p></div>
1622 <!-- ======================================================================= -->
1623 <div class="doc_subsection">
1624 <a name="latemco">Late Machine Code Optimizations</a>
1626 <div class="doc_text"><p>To Be Written</p></div>
1627 <!-- ======================================================================= -->
1628 <div class="doc_subsection">
1629 <a name="codeemit">Code Emission</a>
1631 <div class="doc_text"><p>To Be Written</p></div>
1632 <!-- _______________________________________________________________________ -->
1633 <div class="doc_subsubsection">
1634 <a name="codeemit_asm">Generating Assembly Code</a>
1636 <div class="doc_text"><p>To Be Written</p></div>
1637 <!-- _______________________________________________________________________ -->
1638 <div class="doc_subsubsection">
1639 <a name="codeemit_bin">Generating Binary Machine Code</a>
1642 <div class="doc_text">
1643 <p>For the JIT or <tt>.o</tt> file writer</p>
1647 <!-- *********************************************************************** -->
1648 <div class="doc_section">
1649 <a name="targetimpls">Target-specific Implementation Notes</a>
1651 <!-- *********************************************************************** -->
1653 <div class="doc_text">
1655 <p>This section of the document explains features or design decisions that
1656 are specific to the code generator for a particular target.</p>
1660 <!-- ======================================================================= -->
1661 <div class="doc_subsection">
1662 <a name="tailcallopt">Tail call optimization</a>
1665 <div class="doc_text">
1666 <p>Tail call optimization, callee reusing the stack of the caller, is currently supported on x86/x86-64 and PowerPC. It is performed if:
1668 <li>Caller and callee have the calling convention <tt>fastcc</tt>.</li>
1669 <li>The call is a tail call - in tail position (ret immediately follows call and ret uses value of call or is void).</li>
1670 <li>Option <tt>-tailcallopt</tt> is enabled.</li>
1671 <li>Platform specific constraints are met.</li>
1675 <p>x86/x86-64 constraints:
1677 <li>No variable argument lists are used.</li>
1678 <li>On x86-64 when generating GOT/PIC code only module-local calls (visibility = hidden or protected) are supported.</li>
1681 <p>PowerPC constraints:
1683 <li>No variable argument lists are used.</li>
1684 <li>No byval parameters are used.</li>
1685 <li>On ppc32/64 GOT/PIC only module-local calls (visibility = hidden or protected) are supported.</li>
1689 <p>Call as <tt>llc -tailcallopt test.ll</tt>.
1690 <div class="doc_code">
1692 declare fastcc i32 @tailcallee(i32 inreg %a1, i32 inreg %a2, i32 %a3, i32 %a4)
1694 define fastcc i32 @tailcaller(i32 %in1, i32 %in2) {
1695 %l1 = add i32 %in1, %in2
1696 %tmp = tail call fastcc i32 @tailcallee(i32 %in1 inreg, i32 %in2 inreg, i32 %in1, i32 %l1)
1701 <p>Implications of <tt>-tailcallopt</tt>:</p>
1702 <p>To support tail call optimization in situations where the callee has more arguments than the caller a 'callee pops arguments' convention is used. This currently causes each <tt>fastcc</tt> call that is not tail call optimized (because one or more of above constraints are not met) to be followed by a readjustment of the stack. So performance might be worse in such cases.</p>
1703 <p>On x86 and x86-64 one register is reserved for indirect tail calls (e.g via a function pointer). So there is one less register for integer argument passing. For x86 this means 2 registers (if <tt>inreg</tt> parameter attribute is used) and for x86-64 this means 5 register are used.</p>
1705 <!-- ======================================================================= -->
1706 <div class="doc_subsection">
1707 <a name="x86">The X86 backend</a>
1710 <div class="doc_text">
1712 <p>The X86 code generator lives in the <tt>lib/Target/X86</tt> directory. This
1713 code generator is capable of targeting a variety of x86-32 and x86-64
1714 processors, and includes support for ISA extensions such as MMX and SSE.
1719 <!-- _______________________________________________________________________ -->
1720 <div class="doc_subsubsection">
1721 <a name="x86_tt">X86 Target Triples supported</a>
1724 <div class="doc_text">
1726 <p>The following are the known target triples that are supported by the X86
1727 backend. This is not an exhaustive list, and it would be useful to add those
1728 that people test.</p>
1731 <li><b>i686-pc-linux-gnu</b> - Linux</li>
1732 <li><b>i386-unknown-freebsd5.3</b> - FreeBSD 5.3</li>
1733 <li><b>i686-pc-cygwin</b> - Cygwin on Win32</li>
1734 <li><b>i686-pc-mingw32</b> - MingW on Win32</li>
1735 <li><b>i386-pc-mingw32msvc</b> - MingW crosscompiler on Linux</li>
1736 <li><b>i686-apple-darwin*</b> - Apple Darwin on X86</li>
1741 <!-- _______________________________________________________________________ -->
1742 <div class="doc_subsubsection">
1743 <a name="x86_cc">X86 Calling Conventions supported</a>
1747 <div class="doc_text">
1749 <p>The following target-specific calling conventions are known to backend:</p>
1752 <li><b>x86_StdCall</b> - stdcall calling convention seen on Microsoft Windows
1753 platform (CC ID = 64).</li>
1754 <li><b>x86_FastCall</b> - fastcall calling convention seen on Microsoft Windows
1755 platform (CC ID = 65).</li>
1760 <!-- _______________________________________________________________________ -->
1761 <div class="doc_subsubsection">
1762 <a name="x86_memory">Representing X86 addressing modes in MachineInstrs</a>
1765 <div class="doc_text">
1767 <p>The x86 has a very flexible way of accessing memory. It is capable of
1768 forming memory addresses of the following expression directly in integer
1769 instructions (which use ModR/M addressing):</p>
1771 <div class="doc_code">
1773 Base + [1,2,4,8] * IndexReg + Disp32
1777 <p>In order to represent this, LLVM tracks no less than 4 operands for each
1778 memory operand of this form. This means that the "load" form of '<tt>mov</tt>'
1779 has the following <tt>MachineOperand</tt>s in this order:</p>
1783 Meaning: DestReg, | BaseReg, Scale, IndexReg, Displacement
1784 OperandTy: VirtReg, | VirtReg, UnsImm, VirtReg, SignExtImm
1787 <p>Stores, and all other instructions, treat the four memory operands in the
1788 same way and in the same order.</p>
1792 <!-- _______________________________________________________________________ -->
1793 <div class="doc_subsubsection">
1794 <a name="x86_memory">X86 address spaces supported</a>
1797 <div class="doc_text">
1799 <p>x86 has the ability to perform loads and stores to different address spaces
1800 via the x86 segment registers. A segment override prefix byte on an instruction
1801 causes the instruction's memory access to go to the specified segment. LLVM
1802 address space 0 is the default address space, which includes the stack, and
1803 any unqualified memory accesses in a program. Address spaces 1-255 are
1804 currently reserved for user-defined code. The GS-segment is represented by
1805 address space 256. Other x86 segments have yet to be allocated address space
1808 <p>Some operating systems use the GS-segment to implement TLS, so care should be
1809 taken when reading and writing to address space 256 on these platforms.
1813 <!-- _______________________________________________________________________ -->
1814 <div class="doc_subsubsection">
1815 <a name="x86_names">Instruction naming</a>
1818 <div class="doc_text">
1820 <p>An instruction name consists of the base name, a default operand size, and a
1821 a character per operand with an optional special size. For example:</p>
1824 <tt>ADD8rr</tt> -> add, 8-bit register, 8-bit register<br>
1825 <tt>IMUL16rmi</tt> -> imul, 16-bit register, 16-bit memory, 16-bit immediate<br>
1826 <tt>IMUL16rmi8</tt> -> imul, 16-bit register, 16-bit memory, 8-bit immediate<br>
1827 <tt>MOVSX32rm16</tt> -> movsx, 32-bit register, 16-bit memory
1832 <!-- ======================================================================= -->
1833 <div class="doc_subsection">
1834 <a name="ppc">The PowerPC backend</a>
1837 <div class="doc_text">
1838 <p>The PowerPC code generator lives in the lib/Target/PowerPC directory. The
1839 code generation is retargetable to several variations or <i>subtargets</i> of
1840 the PowerPC ISA; including ppc32, ppc64 and altivec.
1844 <!-- _______________________________________________________________________ -->
1845 <div class="doc_subsubsection">
1846 <a name="ppc_abi">LLVM PowerPC ABI</a>
1849 <div class="doc_text">
1850 <p>LLVM follows the AIX PowerPC ABI, with two deviations. LLVM uses a PC
1851 relative (PIC) or static addressing for accessing global values, so no TOC (r2)
1852 is used. Second, r31 is used as a frame pointer to allow dynamic growth of a
1853 stack frame. LLVM takes advantage of having no TOC to provide space to save
1854 the frame pointer in the PowerPC linkage area of the caller frame. Other
1855 details of PowerPC ABI can be found at <a href=
1856 "http://developer.apple.com/documentation/DeveloperTools/Conceptual/LowLevelABI/Articles/32bitPowerPC.html"
1857 >PowerPC ABI.</a> Note: This link describes the 32 bit ABI. The
1858 64 bit ABI is similar except space for GPRs are 8 bytes wide (not 4) and r13 is
1859 reserved for system use.</p>
1862 <!-- _______________________________________________________________________ -->
1863 <div class="doc_subsubsection">
1864 <a name="ppc_frame">Frame Layout</a>
1867 <div class="doc_text">
1868 <p>The size of a PowerPC frame is usually fixed for the duration of a
1869 function’s invocation. Since the frame is fixed size, all references into
1870 the frame can be accessed via fixed offsets from the stack pointer. The
1871 exception to this is when dynamic alloca or variable sized arrays are present,
1872 then a base pointer (r31) is used as a proxy for the stack pointer and stack
1873 pointer is free to grow or shrink. A base pointer is also used if llvm-gcc is
1874 not passed the -fomit-frame-pointer flag. The stack pointer is always aligned to
1875 16 bytes, so that space allocated for altivec vectors will be properly
1877 <p>An invocation frame is laid out as follows (low memory at top);</p>
1880 <div class="doc_text">
1881 <table class="layout">
1883 <td>Linkage<br><br></td>
1886 <td>Parameter area<br><br></td>
1889 <td>Dynamic area<br><br></td>
1892 <td>Locals area<br><br></td>
1895 <td>Saved registers area<br><br></td>
1897 <tr style="border-style: none hidden none hidden;">
1901 <td>Previous Frame<br><br></td>
1906 <div class="doc_text">
1907 <p>The <i>linkage</i> area is used by a callee to save special registers prior
1908 to allocating its own frame. Only three entries are relevant to LLVM. The
1909 first entry is the previous stack pointer (sp), aka link. This allows probing
1910 tools like gdb or exception handlers to quickly scan the frames in the stack. A
1911 function epilog can also use the link to pop the frame from the stack. The
1912 third entry in the linkage area is used to save the return address from the lr
1913 register. Finally, as mentioned above, the last entry is used to save the
1914 previous frame pointer (r31.) The entries in the linkage area are the size of a
1915 GPR, thus the linkage area is 24 bytes long in 32 bit mode and 48 bytes in 64
1919 <div class="doc_text">
1920 <p>32 bit linkage area</p>
1921 <table class="layout">
1924 <td>Saved SP (r1)</td>
1944 <td>Saved FP (r31)</td>
1949 <div class="doc_text">
1950 <p>64 bit linkage area</p>
1951 <table class="layout">
1954 <td>Saved SP (r1)</td>
1974 <td>Saved FP (r31)</td>
1979 <div class="doc_text">
1980 <p>The <i>parameter area</i> is used to store arguments being passed to a callee
1981 function. Following the PowerPC ABI, the first few arguments are actually
1982 passed in registers, with the space in the parameter area unused. However, if
1983 there are not enough registers or the callee is a thunk or vararg function,
1984 these register arguments can be spilled into the parameter area. Thus, the
1985 parameter area must be large enough to store all the parameters for the largest
1986 call sequence made by the caller. The size must also be minimally large enough
1987 to spill registers r3-r10. This allows callees blind to the call signature,
1988 such as thunks and vararg functions, enough space to cache the argument
1989 registers. Therefore, the parameter area is minimally 32 bytes (64 bytes in 64
1990 bit mode.) Also note that since the parameter area is a fixed offset from the
1991 top of the frame, that a callee can access its spilt arguments using fixed
1992 offsets from the stack pointer (or base pointer.)</p>
1995 <div class="doc_text">
1996 <p>Combining the information about the linkage, parameter areas and alignment. A
1997 stack frame is minimally 64 bytes in 32 bit mode and 128 bytes in 64 bit
2001 <div class="doc_text">
2002 <p>The <i>dynamic area</i> starts out as size zero. If a function uses dynamic
2003 alloca then space is added to the stack, the linkage and parameter areas are
2004 shifted to top of stack, and the new space is available immediately below the
2005 linkage and parameter areas. The cost of shifting the linkage and parameter
2006 areas is minor since only the link value needs to be copied. The link value can
2007 be easily fetched by adding the original frame size to the base pointer. Note
2008 that allocations in the dynamic space need to observe 16 byte alignment.</p>
2011 <div class="doc_text">
2012 <p>The <i>locals area</i> is where the llvm compiler reserves space for local
2016 <div class="doc_text">
2017 <p>The <i>saved registers area</i> is where the llvm compiler spills callee saved
2018 registers on entry to the callee.</p>
2021 <!-- _______________________________________________________________________ -->
2022 <div class="doc_subsubsection">
2023 <a name="ppc_prolog">Prolog/Epilog</a>
2026 <div class="doc_text">
2027 <p>The llvm prolog and epilog are the same as described in the PowerPC ABI, with
2028 the following exceptions. Callee saved registers are spilled after the frame is
2029 created. This allows the llvm epilog/prolog support to be common with other
2030 targets. The base pointer callee saved register r31 is saved in the TOC slot of
2031 linkage area. This simplifies allocation of space for the base pointer and
2032 makes it convenient to locate programatically and during debugging.</p>
2035 <!-- _______________________________________________________________________ -->
2036 <div class="doc_subsubsection">
2037 <a name="ppc_dynamic">Dynamic Allocation</a>
2040 <div class="doc_text">
2044 <div class="doc_text">
2045 <p><i>TODO - More to come.</i></p>
2049 <!-- *********************************************************************** -->
2052 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
2053 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
2054 <a href="http://validator.w3.org/check/referer"><img
2055 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
2057 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
2058 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a><br>
2059 Last modified: $Date$