LLV8 is an experimental top-tier compiler for V8. It leverages the power of LLVM MCJIT to produce highly optimized code. It is supposed to be used as a third tier for cases where it makes sense to spend more time compiling to achieve higher throughput.
LLV8 (backend) is implemented as a patch to V8 and it cannot function without the virtual machine. Although LLV8 is only a fraction of the entire patched VM, we also refer to the whole thing (our fork of V8) as LLV8.
LLV8 codebase effectively consists of two repositories, both of which are hosted at github:
Building instructions can be found in the project's wiki. They are duplicated in this readme for convenience.
We are going to check out sources of LLV8 and the modified LLVM. V8 (and thus LLV8) comes with binaries of a clang compiler, so we are going to use it to build both LLVM and LLV8 to avoid linking problems.
The easiest way is to follow the process of building regular V8, adding the LLV8 branch as a remote.
cd $LLV8_ROOT # Choose any directory you want. # E.g. cd /some/dir && export LLV8_ROOT=`pwd` Install depot_tools:
git clone https://chromium.googlesource.com/chromium/tools/depot_tools.git export PATH=`pwd`/depot_tools:"$PATH" Fetch all the code:
fetch v8 cd v8 git remote add llv8 https://github.com/ispras/llv8.git git fetch llv8 git checkout -b llv8 llv8/llv8 gclient sync Note that we don't run make yet, since first we need to build LLVM libraries to link against. We had to check out V8 first to obtain the clang compiler though.
Check out and build our version of LLVM:
cd $LLV8_ROOT git clone https://github.com/ispras/llvm-for-v8.git mkdir build-llvm cd build-llvm export CC=$LLV8_ROOT/v8/third_party/llvm-build/Release+Asserts/bin/clang export CXX=$LLV8_ROOT/v8/third_party/llvm-build/Release+Asserts/bin/clang++ ../llvm-for-v8/configure --enable-assertions --enable-optimized --disable-zlib make -j9 sudo make install # Note: this installs the built version of llvm system-wide. You can in theory pass --prefix to configure or not install llvm at all and use it from the build directory, because all you need to build llv8 is the llvm-configure of this freshly built llvm in your $PATH. But this makes the subsequent compilation of llv8 a bit more involved (the C++ compiler spews warnings-errors as it compiles LLVM headers).
Finally, run make (substitute "release" for "debug" if you'd like to test performance):
cd $LLV8_ROOT/v8 export LINK=$CXX make x64.debug -j9 i18nsupport=off gdbjit=off Design documentation, building and runnig insructions can be found on LLV8 wiki.
Let's compile a simple piece of JavaScript code with LLV8.
cat > a-plus-b.js var N = 10000; // Should be big enough to pass optimization threshold. function foo(a, b) { return a + b; } var k = 1; for (var i = 0; i < N; i++) { k += foo(i, i % k); } print(k); Now run it through LLV8:
$LLV8_ROOT/v8/out/x64.debug/d8 a-plus-b.js --llvm-filter=* It should spew a lot of debug information to stderr, mostly LLVM IR before and after various passes and disassembly. Here is an abridged stderr output:
... ====================vvv Module AFTER optimization vvv==================== ; ModuleID = '0' target triple = "x86_64-unknown-linux-gnu" define x86_64_v8cc i8* @"0"(i8* %pointer_0, i8* nocapture readnone, i8* nocapture readnone, i8* %pointer_1, i8* %pointer_2, i8* %pointer_3) #0 gc "v8-gc" { BlockEntry0: tail call void (i64, i32, ...) @llvm.experimental.stackmap(i64 0, i32 0) %2 = ptrtoint i8* %pointer_2 to i64 %3 = and i64 %2, 1 %4 = icmp eq i64 %3, 0 br i1 %4, label %BlockCont, label %DeoptBlock BlockCont: ; preds = %BlockEntry0 %5 = ptrtoint i8* %pointer_1 to i64 %6 = and i64 %5, 1 %7 = icmp eq i64 %6, 0 br i1 %7, label %BlockCont1, label %DeoptBlock2 DeoptBlock: ; preds = %BlockEntry0 %8 = tail call i64 (i64, i32, i8*, i32, ...) @llvm.experimental.patchpoint.i64(i64 1, i32 5, i8* null, i32 0, i8* %pointer_3, i8* %pointer_2, i8* %pointer_1, i8* %pointer_0) unreachable BlockCont1: ; preds = %BlockCont %9 = tail call { i64, i1 } @llvm.sadd.with.overflow.i64(i64 %2, i64 %5) %10 = extractvalue { i64, i1 } %9, 1 br i1 %10, label %DeoptBlock4, label %BlockCont3 DeoptBlock2: ; preds = %BlockCont %11 = tail call i64 (i64, i32, i8*, i32, ...) @llvm.experimental.patchpoint.i64(i64 2, i32 5, i8* null, i32 0, i8* %pointer_3, i8* %pointer_2, i8* %pointer_1, i8* %pointer_0) unreachable BlockCont3: ; preds = %BlockCont1 %12 = extractvalue { i64, i1 } %9, 0 %13 = inttoptr i64 %12 to i8* ret i8* %13 DeoptBlock4: ; preds = %BlockCont1 %14 = tail call i64 (i64, i32, i8*, i32, ...) @llvm.experimental.patchpoint.i64(i64 3, i32 5, i8* null, i32 0, i8* %pointer_3, i8* %pointer_2, i8* %pointer_1, i8* %pointer_0) unreachable } declare void @llvm.experimental.stackmap(i64, i32, ...) declare i64 @llvm.experimental.patchpoint.i64(i64, i32, i8*, i32, ...) ; Function Attrs: nounwind readnone declare { i64, i1 } @llvm.sadd.with.overflow.i64(i64, i64) #1 attributes #0 = { "no-frame-pointer-elim"="true" "put-constantpool-in-fn-section"="true" "put-jumptable-in-fn-section"="true" } attributes #1 = { nounwind readnone } ====================^^^ Module AFTER optimization ^^^==================== ... Version: 1 StackSizes: (off:48232192, size:24) Constants: Records: (#0, offset = 18, flags = 0, locations = [] live_outs = []) (#1, offset = 42, flags = 0, locations = [(Register, rcx, off:0, size:8), (Register, rdx, off:0, size:8), (Register, rbx, off:0, size:8), (Register, rsi, off:0, size:8), ] live_outs = []) (#2, offset = 47, flags = 0, locations = [(Register, rcx, off:0, size:8), (Register, rdx, off:0, size:8), (Register, rbx, off:0, size:8), (Register, rsi, off:0, size:8), ] live_outs = []) (#3, offset = 52, flags = 0, locations = [(Register, rcx, off:0, size:8), (Register, rdx, off:0, size:8), (Register, rbx, off:0, size:8), (Register, rsi, off:0, size:8), ] live_outs = []) Instruction start: 0xa31d33a040 0xa31d33a040 push rbp 0xa31d33a041 mov rbp, rsp 0xa31d33a044 push rsi 0xa31d33a045 push rdi 0xa31d33a046 mov rcx, qword ptr [rbp + 0x20] 0xa31d33a04a mov rdx, qword ptr [rbp + 0x18] 0xa31d33a04e mov rbx, qword ptr [rbp + 0x10] 0xa31d33a052 test dl, 0x1 0xa31d33a055 jne 0x13 0xa31d33a057 test bl, 0x1 0xa31d33a05a jne 0x13 0xa31d33a05c mov rax, rdx 0xa31d33a05f add rax, rbx 0xa31d33a062 jo 0x10 0xa31d33a064 pop rdi 0xa31d33a065 pop rsi 0xa31d33a066 pop rbp 0xa31d33a067 ret 0x18 0xa31d33a06a call -0x33406f 0xa31d33a06f call -0x33406a 0xa31d33a074 call -0x334065 ... RelocInfo (size = 6) 0xa31d33a06b runtime entry (deoptimization bailout 0) 0xa31d33a070 runtime entry (deoptimization bailout 1) 0xa31d33a075 runtime entry (deoptimization bailout 2) Safepoints (size = 8) And of course the answer is printed to stdout.
99989998