Skip to content
A brainfuck interpreter written on a cloudy afternoon
Branch: master
Clone or download


This is an optimising interpreter for the language Brainfuck with a small twist: brainstorm's output operation can be configured to print only numbers - the numeric value of the memory cell. In other words, brainstorm is oriented toward complex computational problems.

Other than that brainstorm does two very basic optimisations on the code: homogeneous data- and instruction-pointer-move instruction sequences are collapsed to one instruction. Instructions themselves are stored as 16-bit words, containing an optional immediate-operand field.

How to Build

Running the script in the source directiory builds the executable, if you happen to have clang++-3.6 (hardcoded in the script -- substitute for your preferred compiler).

There are three versions of the interpreter -- vanilla, alt and alt-alt. They differ by minor tweaks to the interpeter loop, which can benefit some combinations of microarchitectures and compilers. The vanilla version is built by default; to build the alt version, pass _alt as an arg to the build script, and to build the alt-alt version, pass _alt_alt, respectively.


Erik Bosman's mandelbrot generator (times include printout; 'alt' = alt version, 'alt^2' = alt-alt version):

CPU compiler time (real)
AMD C-60 (Bobcat) @ 1.33GHz clang++-3.5.2 0m36.497s [^1]
AMD C-60 (Bobcat) @ 1.33GHz g++-4.8.4 0m34.668s [^1]
AMD A8-7600 (Steamroller) @ 2.4GHz clang++-3.5.2 0m16.720s
AMD Ryzen 1700 (Zen) @ 3.0GHz g++-6.3.0 0m9.54s
AMD Ryzen 1700X (Zen) @ 2.2GHz g++-4.9.2 0m11.81s [^2]
Intel Core2 Quad Q6600 (Kentsfield) @ 3.2GHz clang++-3.7.x 0m12.650s
Intel Core2 Quad Q6600 (Kentsfield) @ 3.2GHz g++-4.9.2 0m16.430s
Intel Core2 Duo P8600 (Penryn) @ 2.4GHz (alt^2) apple clang++-8.1.0 0m13.383s
Intel i7-2600 (Sandy Bridge) @ 3.8GHz clang++-3.7.1 0m9.200s
Intel i7-2600 (Sandy Bridge) @ 3.8GHz g++-5.3.0 0m8.350s
Intel E5-2687W (Sandy Bridge) @ 3.1GHz clang++-3.6.2 0m9.390s
Intel E5-2687W (Sandy Bridge) @ 3.1GHz g++-4.7.2 0m9.946s
Intel i5-3470S (Ivy Bridge) @ 1.6GHz g++-4.8.2 0m17.087s
Intel i5-3570 (Ivy Bridge) @ 1.6GHz g++-4.8.4 0m17.086s
Intel i5-3570 (Ivy Bridge) @ 3.8GHz g++-4.8.4 0m7.200s
Intel E3-1270v2 (Ivy Bridge) @ 1.6GHz clang++-3.5.0 0m17.189s
Intel E3-1270v2 (Ivy Bridge) @ 1.6GHz g++-4.8.2 0m17.088s
Intel E3-1270v2 (Ivy Bridge) @ 3.9GHz g++-4.8.2 0m7.033s
Intel i7-5820K (Haswell) @ 3.6GHz clang++-3.6.x 0m7.276s
Intel i7-5820K (Haswell) @ 3.6GHz g++-4.8.x 0m6.905s
Freescale iMX53 (Cortex-A8 r2p5) @ 1.0GHz clang++-3.6.2 1m0.941s
Freescale iMX53 (Cortex-A8 r2p5) @ 1.0GHz g++-4.9.2 0m55.673s
Samsung Exynos 5422 (Cortex-A15 r2p3) @ 2.0GHz clang++-3.6.2 0m28.754s
Samsung Exynos 5422 (Cortex-A15 r2p3) @ 2.0GHz g++-6.3.0 0m20.007s
Allwinner A64 (Cortex-A53 r0p4) @ 1.152GHz (alt) clang++-3.6.2 0m40.832s
AppliedMicro X-Gene 1 @ 2.4GHz (alt) clang++-3.5.0 0m19.623s
AppliedMicro X-Gene 1 @ 2.4GHz (alt) g++-4.9.1 0m19.608s
Rockchip RK3368 (Cortex-A53 r0p3) A32 @ 1.51GHz (alt^2) g++-4.9.3 0m32.623s
Rockchip RK3368 (Cortex-A53 r0p3) A64 @ 1.51GHz (alt^2) clang++-3.6.0 0m30.224s
MediaTek MT8163A (Cortex-A53 r0p3) A32 @ 1.5GHz (alt^2) g++-4.9.2 0m33.430s [^4]
MediaTek MT8163A (Cortex-A53 r0p3) A64 @ 1.5GHz (alt^2) clang++-3.6.2 0m30.587s [^4]
MediaTek MT8173C (Cortex-A72 r0p0) A64 @ 2.1GHz clang++-5.0.1 0m19.083s [^5]
MediaTek MT8173C (Cortex-A72 r0p0) A64 @ 2.1GHz (alt^2) g++-7.2.1 0m14.011s [^5]
Marvell ARMADA 8040 (Cortex-A72 r0p1) A64 @ 1.3GHz (alt^2) g++-7.2.1 0m22.703s [^5]
Marvell ARMADA 8040 (Cortex-A72 r0p1) A64 @ 1.3GHz clang++-3.5.2 0m30.836s [^5]
Marvell ARMADA 8040 (Cortex-A72 r0p1) A64 @ 2.0GHz clang++-3.5.2 0m20.035s [^5]
AWS Graviton (Cortex-A72 r0p3) A64 @ 2.28GHz (alt^2) g++-7.3.0 0m12.881s
AWS Graviton (Cortex-A72 r0p3) A64 @ 2.28GHz clang++-6.0.0 0m16.924s [^5]
Baikal-T1 (MIPS P5600) @ 1.2GHz (alt) g++-6.3.0 0m37.212s
Baikal-T1 (MIPS P5600) @ 1.2GHz (alt^2) g++-7.3.0 0m31.822s
Amlogic S922X (Cortex-A73 r0p2) @ 1.8GHz (alt^2) g++-7.3.0 0m15.459s

[^1]: Generic compiler tuning; native tuning does not pose any speed advantage
[^2]: Non-native compiler tuning -march=corei7
[^3]: Non-native compiler tuning -mcpu=cortex-a7
[^4]: This MT8163A resides in a BQ M10 Ubuntu tablet, and as such is subject to the following performance detriments:

(1) Power management causes cores to pop in and out of existence, rather than just scaling them by frequency.  
(2) There is an entire (albeit minimal) Android running in a lxc container on that tablet.

[^5]: Non-native compiler tuning -mcpu=cortex-a57

Note: There are two compiler snafus in all A64 alt-alt entries built by clang. First, the interpereter loop does not get aligned to a multiple-of-16 address, so one has to inject nops before the loop to get optimal loop alignment. Second, the code generated for the loop could be better:

400f00:       787a7aa8        ldrh    w8, [x21,x26,lsl #1]
400f04:       12000909        and     w9, w8, #0x7
400f08:       71001d3f        cmp     w9, #0x7            // since we just AND'd its source register with 7, w9 could not be larger than 7,
400f0c:       54000468        b.hi    400f98              // so this pair of ops is essentially nops in the innermost loop
400f98:       f94007e8        ldr     x8, [sp,#8]
400f9c:       9100075a        add     x26, x26, #0x1
400fa0:       eb08035f        cmp     x26, x8
400fa4:       54fffae3    400f00

Normalized performance from the above as ticks = duration x CPU_GHz (lower is better):

CPU compiler ticks
Freescale iMX53 (Cortex-A8 r2p5) g++-4.9.2 55.67
AppliedMicro X-Gene 1 (alt) g++-4.9.1 47.06
Allwinner A64 (Cortex-A53 r0p4) (alt) clang++-3.6.2 47.04
AMD C-60 (Bobcat) g++-4.8.4 46.11
MediaTek MT8163A (Cortex-A53 r0p3) (alt^2) clang++-3.6.2 45.88
Rockchip RK3368 (Cortex-A53 r0p3) (alt^2) clang++-3.6.0 45.64
Intel Core2 Quad Q6600 (Kentsfield) clang++-3.7.x 40.48
AMD A8-7600 (Steamroller) clang++-3.5.2 40.13
Samsung Exynos 5422 (Cortex-A15 r2p3) g++-6.3.0 40.01
Baikal-T1 (MIPS P5600) (alt^2) g++-7.3.0 38.19
Intel Core2 Duo P8600 (Penryn) (alt^2) apple clang++-8.1.0 32.12
Intel i7-2600 (Sandy Bridge) g++-5.3.0 31.73
Marvell ARMADA 8040 (Cortex-A72 r0p1) (alt^2) g++-7.2.1 29.51
MediaTek MT8173C (Cortex-A72 r0p0) (alt^2) g++-7.2.1 29.42
AWS Graviton (Cortex-A72 r0p3) (alt^2) g++-7.3.0 29.37
Intel E5-2687W (Sandy Bridge) clang++-3.6.2 29.11
AMD Ryzen 1700 (Zen) g++-6.3.0 28.62
Amlogic S922X (Cortex-A73 r0p2) (alt^2) g++-7.3.0 27.83
Intel E3-1270v2 (Ivy Bridge) g++-4.8.2 27.34
AMD Ryzen 1700X (Zen) g++-4.9.2 25.98
Intel i7-5820K (Haswell) g++-4.8.x 24.86


Since brainstorm is a classic example of 100%-integer, branchy code, it can be indicative of some fundamental microarchitecture characteristics. For instance, in the mandelbrot test above:

  • Bobcat (a dual-decode/dual-issue/six-dispatch, out-of-order microarchitecture) is ~1.2x faster per-clock than Cortex-A8 (a dual-decode/dual-issue, in-order microarchitecture). Unfortunately, we cannot tell what part of Bobcat's advantage comes from the out-of-orderness, and what - from a potentially better branch predictor (currently perf does not support very well Cortex-A8's performance counters).
  • Ivy Bridge (a quad-decode/quad-issue/six-dispatch, aggressively out-of-order microarchitecture) is ~1.7x faster per-clock than Bobcat. IVB does not quite outperform Bobcat twice in this test, even though IVB should have a clear branch-predictor advantage, on top of its twice-wider decoder. Indeed, running the test under perf stat shows that IVB achieves an average IPC of 2.46 and no branch mispredictions to speak of (0.34%), versus IPC of 1.42 for Bobcat and slightly worse but still negligible branch mispredictions (1.36%). So we have an apparent case of diminishing returns from the decode width, even when branch mispredictions are not an issue. Actually this is true not only for this test as 2.35-2.5 is normal IPC for the late 4-decode Intel microarchitectures across many loads.
You can’t perform that action at this time.