IR - Lightweight JIT Compilation Framework

IR Framework is a practical solution for implementing JIT in medium-size projects. It defines Intermediate Representation (IR), provides a simple API for IR construction and a set of algorithms for optimization, scheduling, register allocation and code generation. The resulting generated in-memory code, may be directly executed.

A presentation about IR framework design and implementation details is available at researchgate.

IR is used as a base for PHP JIT since PHP-8.4. It’s also used as a back-end for experimental Rational C Compiler.

Anyway, this is not a stable finished product yet. It’s still under development.

IR - Intermediate Representation

The Framework uses single Medium level Intermediate Representation during all phases of optimization, register allocation and code generation. It is inspired by Sea-Of-Nodes introduced by Cliff Click [1]. Sea-Of-Nodes is used in Java HotSpot Server Compiler, V8 TurboFan JavaScript Compiler, Java Graal Compiler…

This representation unifies data and control dependencies into a single graph, where each instruction represented as a Node and each dependency as an Edge between Nodes. There are no classical CFG (Control Flow Graph) with Basic Blocks. Instead, IR uses special Control Nodes that start and finish some code Regions. The data part of the IR is very similar to SSA (Static Single Assignment) form. Each variable may be assigned only once, but except to SSA, in our IR, we don’t have any variables, their versions and name. Everything is represented by computation Nodes and Edges between them. Of course, we have special PHI Node that represents the Phi() function.

Internally, our graph-based IR is represented as a plain two-way grow-able array of Nodes. Dependency Edge are represented as indexes of the other Node. This physical representation is almost completely repeats the LuaJIT IR designed by Mike Pall [3].

IR Generation

IR Framework provides a simple IR Builder API (ir_builder.h). It’s implemented as a number of C preprocessor macros, where each macro just creates a corresponding IR node and ties it with the sources.

IR Optimization

In comparison to classical optimizing compilers (like GCC and LLVM), IR Framework uses very short optimization pipeline. Together with compact IR representation, this makes it extremely fast and allows to generate quite good machine code in reasonable time.

Folding

Folding is done on the fly, during IR generation. It performs a set of local transformations, but because of the graph nature of the IR where most data operations (like ADD) are “floating” (not “pinned” to Basic Block), the scope of transformations is not limited by Basic Block. It’s important to generate IR in a proper (Reverse Post) order, that would emit all Nodes before their first usage. (In case of different order the scope of the folding should be limited).

All the folding rules are written in a declarative style in the single file - ir_fold.h

Folding Engine performs Constants Folding, Copy Propagation, Algebraic Simplifications, Algebraic Re-Association and Common Sub-Expression Elimination. The simple and fast declarative implementation is borrowed from LuaJIT [3].

Sparse Conditional Constant Propagation

This pass implements a classical algorithm originally designed by M. N. Wegman and F. K. Zadeck [4] for SSA form. Unification of data and control dependencies made its implementation even simpler. Despite of constant propagation itself this pass also performs global Copy Propagation and re-applies the folding rules. At the end all the “dead” instructions (those whose results go unused) are replaced with NOPs.

Global Code Motion

Now we have to “fix” places of “floating” instructions. This pass builds CFG (Control Flow Graph) skeleton and then ”pin” each “floating” instruction to the best Basic Block. The algorithm is developed by Cliff Click [2].

Local Scheduling

As the final IR transformation pass, we reorder instructions inside each Basic Block to satisfy the dependencies. Currently this is done by a simple topological sorting.

Target Instruction Selection

This is the first target dependent step of compilation. It aims to combine instruction Nodes into tiles that allows better instruction fusion. For example 10 + a + b * 4 may be calculated by a single x86 instruction lea 10(%eax, %ebx, 4), %ecx. The selection is done by a constrained tree pattern matching. The current implementation uses simple Max-Munch approach. (This may be replaced by a smarter BURS method).

Register Allocation

CPU independent implementation of Linear Scan Register Allocation for SSA form with second chance bin-packing. [5] [6]

Machine Code Generations

IR Framework implements X86_64, x86 and AArch64 back-ends. The current implementation uses DynAsm [?]. (In the future, this should be replaced with a faster “binary” encoder). Code generator walks throw all instructions of each basic blocks and emits some code according to “rules” selected during instruction selection pass. It uses registers, selected by register allocator and inserts the necessary spill load/store and SSA deconstruction code.

Tooling

• Ability to load and save IR in a textual form
• Ability to visualize IR graph through graphviz dot.
• Target CPU disassembler for generated code (uses libcapstone [?])
• GDB/JIT interface to allow debugging of JIT-ed code
• Linux perf interface to analyze the code performance

Building and Playing with IR

IR Framework is under active development and doesn’t provide any stable libraries yet. In case you like to use IR, it’s better to embed the necessary sources into your project (like PHP does].

However, we provide a simple driver that may be built to run tests and play with IR.

git clone https://github.com/dstogov/ir.git
cd ir
make
make test
./ir bench/mandelbrit.ir --run
./ir bench/mandelbrit.ir -S
./ir --help

IR Example

The complete described example may be found in examples/mandelbrot.c and built using make examples.

It generates the code for the following C function:

int32_t mandelbrot(double x, double y)
{
    double cr = y - 0.5;
    double ci = x;
    double zi = 0.0;
    double zr = 0.0;
    int i = 0;

    while(1) {
        i++;
        double temp = zr * zi;
        double zr2 = zr * zr;
        double zi2 = zi * zi;
        zr = zr2 - zi2 + cr;
        zi = temp + temp + ci;
        if (zi2 + zr2 > 16)
            return i;
        if (i > 1000)
            return 0;
    }
    
}

This is done through IR builder API by the following code:

void gen_mandelbrot(ir_ctx *ctx)
{
    ir_START();
    ir_ref x = ir_PARAM(IR_DOUBLE, "x", 1);
    ir_ref y = ir_PARAM(IR_DOUBLE, "y", 2);
    ir_ref cr = ir_SUB_D(y, ir_CONST_DOUBLE(0.5));
    ir_ref ci = ir_COPY_D(x);
    ir_ref zi = ir_COPY_D(ir_CONST_DOUBLE(0.0));
    ir_ref zr = ir_COPY_D(ir_CONST_DOUBLE(0.0));
    ir_ref i = ir_COPY_I32(ir_CONST_I32(0));

    ir_ref loop = ir_LOOP_BEGIN(ir_END());
        ir_ref zi_1 = ir_PHI_2(IR_DOUBLE, zi, IR_UNUSED);
        ir_ref zr_1 = ir_PHI_2(IR_DOUBLE, zr, IR_UNUSED);
        ir_ref i_1 = ir_PHI_2(IR_I32, i, IR_UNUSED);

        ir_ref i_2 = ir_ADD_I32(i_1, ir_CONST_I32(1));
        ir_ref temp = ir_MUL_D(zr_1, zi_1);
        ir_ref zr2 = ir_MUL_D(zr_1, zr_1);
        ir_ref zi2 = ir_MUL_D(zi_1, zi_1);
        ir_ref zr_2 = ir_ADD_D(ir_SUB_D(zr2, zi2), cr);
        ir_ref zi_2 = ir_ADD_D(ir_ADD_D(temp, temp), ci);
        ir_ref if_1 = ir_IF(ir_GT(ir_ADD_D(zi2, zr2), ir_CONST_DOUBLE(16.0)));
            ir_IF_TRUE(if_1);
                ir_RETURN(i_2);
            ir_IF_FALSE(if_1);
                ir_ref if_2 = ir_IF(ir_GT(i_2, ir_CONST_I32(1000)));
                ir_IF_TRUE(if_2);
                    ir_RETURN(ir_CONST_I32(0));
                ir_IF_FALSE(if_2);
                    ir_ref loop_end = ir_LOOP_END();

    /* close loop */
    ir_MERGE_SET_OP(loop, 2, loop_end);
    ir_PHI_SET_OP(zi_1, 2, zi_2);
    ir_PHI_SET_OP(zr_1, 2, zr_2);
    ir_PHI_SET_OP(i_1, 2, i_2);
}

The textual representation of the IR after system independent optimizations:

{
    uintptr_t c_1 = 0;
    bool c_2 = 0;
    bool c_3 = 1;
    double c_4 = 0.5;
    double c_5 = 0;
    int32_t c_6 = 0;
    int32_t c_7 = 1;
    double c_8 = 16;
    int32_t c_9 = 1000;
    l_1 = START(l_22);
    double d_2 = PARAM(l_1, "x", 1);
    double d_3 = PARAM(l_1, "y", 2);
    double d_4 = SUB(d_3, c_4);
    l_5 = END(l_1);
    l_6 = LOOP_BEGIN(l_5, l_29);
    double d_7 = PHI(l_6, c_5, d_28);
    double d_8 = PHI(l_6, c_5, d_26);
    int32_t d_9 = PHI(l_6, c_6, d_10);
    int32_t d_10 = ADD(d_9, c_7);
    double d_11 = MUL(d_8, d_8);
    double d_12 = MUL(d_7, d_7);
    double d_13 = ADD(d_12, d_11);
    bool d_14 = GT(d_13, c_8);
    l_15 = IF(l_6, d_14);
    l_16 = IF_TRUE(l_15);
    l_17 = RETURN(l_16, d_10);
    l_18 = IF_FALSE(l_15);
    bool d_19 = GT(d_10, c_9);
    l_20 = IF(l_18, d_19);
    l_21 = IF_TRUE(l_20);
    l_22 = RETURN(l_21, c_6, l_17);
    l_23 = IF_FALSE(l_20);
    double d_24 = MUL(d_7, d_8);
    double d_25 = SUB(d_11, d_12);
    double d_26 = ADD(d_25, d_4);
    double d_27 = ADD(d_24, d_24);
    double d_28 = ADD(d_27, d_2);
    l_29 = LOOP_END(l_23);
}

The visualized graph:

The graph was generated by the commands:

./ir bench/mandelbrot.ir --dot mandelbrot.dot
dot -Tsvg mandelbrot.dot -o mandelbrot.svg

The final generated assembler code:

test:
    subsd .L4(%rip), %xmm1
    xorpd %xmm3, %xmm3
    xorpd %xmm2, %xmm2
    xorl %eax, %eax
.L1:
    leal 1(%rax), %eax
    movapd %xmm2, %xmm4
    mulsd %xmm2, %xmm4
    movapd %xmm3, %xmm5
    mulsd %xmm3, %xmm5
    movapd %xmm5, %xmm6
    addsd %xmm4, %xmm6
    ucomisd .L5(%rip), %xmm6
    ja .L2
    cmpl $0x3e8, %eax
    jg .L3
    mulsd %xmm2, %xmm3
    subsd %xmm5, %xmm4
    movapd %xmm4, %xmm2
    addsd %xmm1, %xmm2
    addsd %xmm3, %xmm3
    addsd %xmm0, %xmm3
    jmp .L1
.L2:
    retq
.L3:
    xorl %eax, %eax
    retq
.rodata
.L4:
    .db 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xe0, 0x3f
.L5:
    .db 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x30, 0x40

LLVM interoperability

IR is partially compatible with LLVM. It’s possible to convert IR into LLVM and then compile it.

./ir bench/mandelbrot.ir --emit-llvm mandelbrot.ll
llc mandelbrot.ll
gcc mandelbrot.s
./a.out

It’s also possible to read an LLVM file and convert it to IR, but this requires LLVM loader and therefore IR should be rebuilt with LLVM support.

make clean
make HAVE_LLVM=yes

Now you may compile some C file into LLVM, and then convert it to IR.

clang -O2 -fno-vectorize -fno-slp-vectorize -S -emit-llvm -o minilua.ll ./dynasm/minilua.c
./ir --llvm-asm minilua.ll --save 2>minilua.ir
./ir minilua.ir --run bench/mandelbrot.lua

Note that the last command above compiles and runs the Lua interpreter.

Also note that the LLVM code produced by clang is already optimized. In case of benchmarking, it may be more honest to avoid LLVM optimizations.

clang -O -Xclang -disable-llvm-passes -c -emit-llvm -S -o minilua.ll ./dynasm/minilua.c

Performance

The following table shows the benchmarks execution time in comparison to the same benchmarks compiled by gcc -O2 (the more the better). The C benchmarks were compiled by CLAG into LLVM code (without any LLVM optimizations, only SSA construction is necessary now) and then loaded, compiled and executed by IR framework.

Benchmark	Execution time (relative to GCC -02)
array	0.93
binary-trees	0.99
funnkuch-reduce	1.01
hash	1.13
hash2	0.87
heapsort	1.03
lists	0.98
matrix	1.12
method-call	1.00
mandelbrot	0.95
nbody	0.98
sieve	0.93
spectral-norm	0.94
strcat	0.97
oggenc	0.88
minilua	0.99
gzip	0.80
bzip2	0.81
AVERAGE	0.96
GEOMEAN	0.96

Most of the benchmarks are very simple (few screens of code), but oggenc, minilua gzip and bzip2 are real applications. As you can see, IR produces code that is in average 5% slower than GCC -O2, ~20% slower in the worst case, and on some benchmarks it’s even faster. The chart shows the same data graphically.

The next table shows the compilation time relative to gcc -O2 (the more the better)

Benchmark	Compilation time (relative to GCC -02)
oggenc	35.22
minilua	40.50
gzip	37.50
bzip2	49.67
AVERAGE	40.69
GEOMEAN	40.34

This comparison is not completely fair, because GCC compiles C source, but IR takes precompiled LLVM asm

Anyway, IR framework provides code that is in average 5% slower, but does this up to ~40 times faster.

PHP JIT based on IR

IR-based JIT integrated into PHP-8.4.0 and above. See JIT in action

php -d opcache.jit=tracing -d opcache.jit_debug=1 Zend/bench.php
php -d opcache.jit=function -d opcache.jit_debug=1 Zend/bench.php

References

1. C. Click, M. Paleczny. “A Simple Graph-Based Intermediate Representation” In ACM SIGPLAN Workshop on Intermediate Representations (IR ‘95), pages 35-49, Jan. 1995.
2. C. Click. “Global Code Motion Global Value Numbering” In ACM SIGPLAN Notices, Volume 30, Issue 6, pp 246–257, June 1995
3. M. Pall. “LuaJIT 2.0 intellectual property disclosure and research opportunities” November 2009 http://lua-users.org/lists/lua-l/2009-11/msg00089.html
4. M. N. Wegman and F. K. Zadeck. “Constant propagation with conditional branches” ACM Transactions on Programming Languages and Systems, 13(2):181-210, April 1991
5. C. Wimmer. “Optimized Interval Splitting in a Linear Scan Register Allocator” In VEE ‘05: Proceedings of the 1st ACM/USENIX international conference on Virtual execution environments, pages 132–141, June 2005
6. C. Wimmer and M. Franz. “Linear Scan Register Allocation on SSA Form” In CGO ‘10: Proceedings of the 8-th annual IEEE/ACM international symposium on Code generation and optimization, pages 170–179, April 2010