342 lines · plain
1================================2Frequently Asked Questions (FAQ)3================================4 5.. contents::6 :local:7 8 9License10=======11 12Can I modify LLVM source code and redistribute the modified source?13-------------------------------------------------------------------14Yes. The modified source distribution must retain the copyright notice and15follow the conditions listed in the `Apache License v2.0 with LLVM Exceptions16<https://github.com/llvm/llvm-project/blob/main/llvm/LICENSE.TXT>`_.17 18 19Can I modify the LLVM source code and redistribute binaries or other tools based on it, without redistributing the source?20--------------------------------------------------------------------------------------------------------------------------21Yes. This is why we distribute LLVM under a less restrictive license than GPL,22as explained in the first question above.23 24 25Can I use AI coding tools, such as GitHub co-pilot, to write LLVM patches?26--------------------------------------------------------------------------27Yes, as long as the resulting work can be licensed under the project license, as28covered in the :doc:`DeveloperPolicy`. Using an AI tool to reproduce copyrighted29work does not rinse it of copyright and grant you the right to relicense it.30 31 32Source Code33===========34 35In what language is LLVM written?36---------------------------------37All of the LLVM tools and libraries are written in C++ with extensive use of38the STL.39 40 41How portable is the LLVM source code?42-------------------------------------43The LLVM source code should be portable to most modern Unix-like operating44systems. LLVM also has excellent support on Windows systems.45Most of the code is written in standard C++ with operating system46services abstracted to a support library. The tools required to build and47test LLVM have been ported to a plethora of platforms.48 49 50What API do I use to store a value to one of the virtual registers in LLVM IR's SSA representation?51---------------------------------------------------------------------------------------------------52 53In short: you can't. It's actually kind of a silly question once you understand54what's going on. Basically, in code like:55 56.. code-block:: llvm57 58 %result = add i32 %foo, %bar59 60, ``%result`` is just a name given to the ``Value`` of the ``add``61instruction. In other words, ``%result`` *is* the add instruction. The62"assignment" doesn't explicitly "store" anything to any "virtual register";63the "``=``" is more like the mathematical sense of equality.64 65Longer explanation: In order to generate a textual representation of the66IR, some kind of name has to be given to each instruction so that other67instructions can textually reference it. However, the isomorphic in-memory68representation that you manipulate from C++ has no such restriction since69instructions can simply keep pointers to any other ``Value``'s that they70reference. In fact, the names of dummy numbered temporaries like ``%1`` are71not explicitly represented in the in-memory representation at all (see72``Value::getName()``).73 74 75Source Languages76================77 78What source languages are supported?79------------------------------------80 81LLVM currently has full support for C and C++ source languages through82`Clang <https://clang.llvm.org/>`_. Many other language frontends have83been written using LLVM; an incomplete list is available at84`projects with LLVM <https://llvm.org/ProjectsWithLLVM/>`_.85 86 87I'd like to write a self-hosting LLVM compiler. How should I interface with the LLVM middle-end optimizers and back-end code generators?88----------------------------------------------------------------------------------------------------------------------------------------89Your compiler front-end will communicate with LLVM by creating a module in the90LLVM intermediate representation (IR) format. Assuming you want to write your91language's compiler in the language itself (rather than C++), there are 392major ways to tackle generating LLVM IR from a front-end:93 941. **Call into the LLVM libraries using your language's FFI (foreign95 function interface).**96 97 * *for:* best tracks changes to the LLVM IR, .ll syntax, and .bc format98 99 * *for:* enables running LLVM optimization passes without an emit/parse100 overhead101 102 * *for:* adapts well to a JIT context103 104 * *against:* lots of ugly glue code to write105 1062. **Emit LLVM assembly from your compiler's native language.**107 108 * *for:* very straightforward to get started109 110 * *against:* the .ll parser is slower than the bitcode reader when111 interfacing to the middle end112 113 * *against:* it may be harder to track changes to the IR114 1153. **Emit LLVM bitcode from your compiler's native language.**116 117 * *for:* can use the more-efficient bitcode reader when interfacing to the118 middle end119 120 * *against:* you'll have to re-engineer the LLVM IR object model and bitcode121 writer in your language122 123 * *against:* it may be harder to track changes to the IR124 125If you go with the first option, the C bindings in include/llvm-c should help126a lot, since most languages have strong support for interfacing with C. The127most common hurdle with calling C from managed code is interfacing with the128garbage collector. The C interface was designed to require very little memory129management, and so is straightforward in this regard.130 131What support is there for a higher-level source language constructs for building a compiler?132--------------------------------------------------------------------------------------------133Currently, there isn't much. LLVM supports an intermediate representation134which is useful for code representation but will not support the high-level135(abstract syntax tree) representation needed by most compilers. There are no136facilities for lexical nor semantic analysis.137 138 139I don't understand the ``GetElementPtr`` instruction. Help!140-----------------------------------------------------------141See `The Often Misunderstood GEP Instruction <GetElementPtr.html>`_.142 143 144Using the C and C++ Front Ends145==============================146 147Can I compile C or C++ code to platform-independent LLVM bitcode?148-----------------------------------------------------------------149No. C and C++ are inherently platform-dependent languages. The most obvious150example of this is the preprocessor. A very common way that C code is made151portable is by using the preprocessor to include platform-specific code. In152practice, information about other platforms is lost after preprocessing, so153the result is inherently dependent on the platform that the preprocessing was154targeting.155 156Another example is ``sizeof``. It's common for ``sizeof(long)`` to vary157between platforms. In most C front-ends, ``sizeof`` is expanded to a158constant immediately, thus hard-wiring a platform-specific detail.159 160Also, since many platforms define their ABIs in terms of C, and since LLVM is161lower-level than C, front-ends currently must emit platform-specific IR in162order to have the result conform to the platform ABI.163 164 165Questions about code generated by the demo page166===============================================167 168What is this ``llvm.global_ctors`` and ``_GLOBAL__I_a...`` stuff that happens when I ``#include <iostream>``?169-------------------------------------------------------------------------------------------------------------170If you ``#include`` the ``<iostream>`` header into a C++ translation unit,171the file will probably use the ``std::cin``/``std::cout``/... global objects.172However, C++ does not guarantee an order of initialization between static173objects in different translation units, so if a static ctor/dtor in your .cpp174file used ``std::cout``, for example, the object would not necessarily be175automatically initialized before your use.176 177To make ``std::cout`` and friends work correctly in these scenarios, the STL178that we use declares a static object that gets created in every translation179unit that includes ``<iostream>``. This object has a static constructor180and destructor that initializes and destroys the global iostream objects181before they could possibly be used in the file. The code that you see in the182``.ll`` file corresponds to the constructor and destructor registration code.183 184If you would like to make it easier to *understand* the LLVM code generated185by the compiler in the demo page, consider using ``printf()`` instead of186``iostream``\s to print values.187 188 189Where did all of my code go??190-----------------------------191If you are using the LLVM demo page, you may often wonder what happened to192all of the code that you typed in. Remember that the demo script is running193the code through the LLVM optimizers, so if your code doesn't actually do194anything useful, it might all be deleted.195 196To prevent this, make sure that the code is actually needed. For example, if197you are computing some expression, return the value from the function instead198of leaving it in a local variable. If you really want to constrain the199optimizer, you can read from and assign to ``volatile`` global variables.200 201 202What is this "``undef``" thing that shows up in my code?203--------------------------------------------------------204``undef`` is the LLVM way of representing a value that is not defined. You205can get these if you do not initialize a variable before you use it. For206example, the C function:207 208.. code-block:: c209 210 int X() { int i; return i; }211 212Is compiled to "``ret i32 undef``" because "``i``" never has a value specified213for it.214 215 216Why does instcombine + simplifycfg turn a call to a function with a mismatched calling convention into "unreachable"? Why not make the verifier reject it?217----------------------------------------------------------------------------------------------------------------------------------------------------------218This is a common problem encountered by authors of front-ends that are using219custom calling conventions: you need to make sure to set the right calling220convention on both the function and on each call to the function. For221example, this code:222 223.. code-block:: llvm224 225 define fastcc void @foo() {226 ret void227 }228 define void @bar() {229 call void @foo()230 ret void231 }232 233Is optimized to:234 235.. code-block:: llvm236 237 define fastcc void @foo() {238 ret void239 }240 define void @bar() {241 unreachable242 }243 244... with "``opt -instcombine -simplifycfg``". This often bites people because245"all their code disappears". Setting the calling convention on the caller and246callee is required for indirect calls to work, so people often ask why not247make the verifier reject this sort of thing.248 249The answer is that this code has undefined behavior, but it is not illegal.250If we made it illegal, then every transformation that could potentially create251this would have to ensure that it doesn't, and there is valid code that can252create this sort of construct (in dead code). The sorts of things that can253cause this to happen are fairly contrived, but we still need to accept them.254Here's an example:255 256.. code-block:: llvm257 258 define fastcc void @foo() {259 ret void260 }261 define internal void @bar(void()* %FP, i1 %cond) {262 br i1 %cond, label %T, label %F263 T:264 call void %FP()265 ret void266 F:267 call fastcc void %FP()268 ret void269 }270 define void @test() {271 %X = or i1 false, false272 call void @bar(void()* @foo, i1 %X)273 ret void274 }275 276In this example, "test" always passes ``@foo``/``false`` into ``bar``, which277ensures that it is dynamically called with the right calling convention (thus, the278code is perfectly well defined). If you run this through the inliner, you279get this (the explicit "or" is there so that the inliner doesn't dead code280eliminate a bunch of stuff):281 282.. code-block:: llvm283 284 define fastcc void @foo() {285 ret void286 }287 define void @test() {288 %X = or i1 false, false289 br i1 %X, label %T.i, label %F.i290 T.i:291 call void @foo()292 br label %bar.exit293 F.i:294 call fastcc void @foo()295 br label %bar.exit296 bar.exit:297 ret void298 }299 300Here you can see that the inlining pass made an undefined call to ``@foo``301with the wrong calling convention. We really don't want to make the inliner302have to know about this sort of thing, so it needs to be valid code. In this303case, dead code elimination can trivially remove the undefined code. However,304if ``%X`` was an input argument to ``@test``, the inliner would produce this:305 306.. code-block:: llvm307 308 define fastcc void @foo() {309 ret void310 }311 312 define void @test(i1 %X) {313 br i1 %X, label %T.i, label %F.i314 T.i:315 call void @foo()316 br label %bar.exit317 F.i:318 call fastcc void @foo()319 br label %bar.exit320 bar.exit:321 ret void322 }323 324The interesting thing about this is that ``%X`` *must* be false for the325code to be well-defined, but no amount of dead code elimination will be able326to delete the broken call as unreachable. However, since327``instcombine``/``simplifycfg`` turns the undefined call into unreachable, we328end up with a branch on a condition that goes to unreachable: a branch to329unreachable can never happen, so "``-inline -instcombine -simplifycfg``" is330able to produce:331 332.. code-block:: llvm333 334 define fastcc void @foo() {335 ret void336 }337 define void @test(i1 %X) {338 F.i:339 call fastcc void @foo()340 ret void341 }342