rvemu
rvemu is a fast Linux emulator, which can run statically linked RV64 programs.
Yeah, it's a toy, but a fairly complete and fast one, which is perfect for learning how a JIT emulator works.
Features
- Faster than QEMU in most cases.
- (Almost*) architecture independent, we've tested it under x86_64.
- Tiny, and easy to understand.
- Targeting RV64IMFDC w/ Newlib (only a small subset of syscalls is implemented, adding more).
*Support for new architecture requires handling relocations in src/compile.c, but it's relatively easy. A few lines of code would do.
Usage
make -j
riscv64-unknown-elf-gcc hello.c
./rvemu a.out
rvemu can only run under Linux at present, and clang needs to be installed to run, as rvemu uses clang to generate jit code.
TODOs
- Add more syscalls.
- Consider using a mmaped MMU.
Notes
-
rvemuusesclang -O3to generate highly optimized target code. -
rvemuuses hardfloat technique to gain more performance, just like NEMU, this actually violates the RISC-V standard, but it produces correct results in most cases, and it's way faster than softfloat.
Benchmark
All the tests is compiled with
riscv64-unknown-elf-gcc -O3and runs on Intel Xeon Platinum 8269CY
Recursive Fibonacci
#include <stdio.h>
int fib(int n) {
if (n == 0 || n == 1) return n;
return fib(n-1) + fib(n-2);
}
int main(void) {
printf("%d\n", fib(42));
return 0;
}rvemu
real 0m2.518s
user 0m2.483s
sys 0m0.035s
QEMU
real 0m2.800s
user 0m2.795s
sys 0m0.006s
Prime Numbers
In this very case, rvemu is 3x faster than QEMU!
rvemu
real 0m54.067s
user 0m54.019s
sys 0m0.047s
QEMU
real 2m58.996s
user 2m58.916s
sys 0m0.050s
nbench
rvemu
NUMERIC SORT : Iterations/sec.: 1413.29 Index: 36.24
STRING SORT : Iterations/sec.: 71.17 Index: 31.80
BITFIELD : Iterations/sec.: 653766423.76 Index: 112.14
FP EMULATION : Iterations/sec.: 527.33 Index: 253.04
FOURIER : Iterations/sec.: 6345.69 Index: 7.22
ASSIGNMENT : Iterations/sec.: 75.19 Index: 286.12
IDEA : Iterations/sec.: 10489.30 Index: 160.43
HUFFMAN : Iterations/sec.: 4384.40 Index: 121.58
LU DECOMPOSITION : Iterations/sec.: 614.64 Index: 31.84
QEMU
BYTEmark (tm) Native Mode Benchmark ver. 2 (10/95)
NUMERIC SORT : Iterations/sec.: 821.90 Index: 21.08
STRING SORT : Iterations/sec.: 25.24 Index: 11.28
BITFIELD : Iterations/sec.: 406495858.33 Index: 69.73
FP EMULATION : Iterations/sec.: 162.26 Index: 77.86
FOURIER : Iterations/sec.: 5867.78 Index: 6.67
ASSIGNMENT : Iterations/sec.: 31.98 Index: 121.69
IDEA : Iterations/sec.: 5234.18 Index: 80.06
HUFFMAN : Iterations/sec.: 2205.89 Index: 61.17
LU DECOMPOSITION : Iterations/sec.: 241.70 Index: 12.52
Refs
Inspired by