x86 32-bit machine code function, 21 bytes
x86-64 machine code function, 22 bytes
The 1B saving in 32-bit mode requires using separator = filler-1, e.g. fill=0
and sep=/
. The 22-byte version can use an arbitrary choice of separator and filler.
This is the 21-byte version, with input-separator = \n
(0xa), output-filler = 0
, output-separator = /
= filler-1. These constants can be easily changed.
; see the source for more comments
; RDI points to the output buffer, RSI points to the src string
; EDX holds the base
; This is the 32-bit version.
; The 64-bit version is the same, but the DEC is one byte longer (or we can just mov al,output_separator)
08048080 <str_exp>:
8048080: 6a 01 push 0x1
8048082: 59 pop ecx ; ecx = 1 = base**0
8048083: ac lods al,BYTE PTR ds:[esi] ; skip the first char so we don't do too many multiplies
; read an input row and accumulate base**n as we go.
08048084 <str_exp.read_bar>:
8048084: 0f af ca imul ecx,edx ; accumulate the exponential
8048087: ac lods al,BYTE PTR ds:[esi]
8048088: 3c 0a cmp al,0xa ; input_separator = newline
804808a: 77 f8 ja 8048084 <str_exp.read_bar>
; AL = separator or terminator
; flags = below (CF=1) or equal (ZF=1). Equal also implies CF=0 in this case.
; store the output row
804808c: b0 30 mov al,0x30 ; output_filler
804808e: f3 aa rep stos BYTE PTR es:[edi],al ; ecx bytes of filler
8048090: 48 dec eax ; mov al,output_separator
8048091: aa stos BYTE PTR es:[edi],al ;append delim
; CF still set from the inner loop, even after DEC clobbers the other flags
8048092: 73 ec jnc 8048080 <str_exp> ; new row if this is a separator, not terminator
8048094: c3 ret
08048095 <end_of_function>
; 0x95 - 0x80 = 0x15 = 21 bytes
The 64-bit version is 1 byte longer, using a 2-byte DEC or a mov al, output_separator
. Other than that, the machine-code is the same for both versions, but some register names change (e.g. rcx
instead of ecx
in the pop
).
Sample output from running the test program (base 3):
$ ./string-exponential $'.\n..\n...\n....' $(seq 3);echo
000/000000000/000000000000000000000000000/000000000000000000000000000000000000000000000000000000000000000000000000000000000/
Algorithm:
Loop over the input, doing exp *= base
for every filler char. On delimiters and the terminating zero byte, append exp
bytes of filler and then a separator to the output string and reset to exp=1
. It's very convenient that the input is guaranteed not to end with both a newline and a terminator.
On input, any byte value above the separator (unsigned compare) is treated as filler, and any byte value below the separator is treated as an end-of-string marker. (Checking explicitly for a zero-byte would take an extra test al,al
vs. branching on flags set by the inner loop).
The rules only allow a trailing separator when it's a trailing newline. My implementation always appends the separator. To get the 1B saving in 32-bit mode, that rule requires separator = 0xa ('\n'
ASCII LF = linefeed), filler = 0xb ('\v'
ASCII VT = vertical tab). That's not very human-friendly, but satisfies the letter of the law. (You can hexdump or
tr $'\v' x
the output to verify that it works, or change the constant so the output separator and filler are printable. I also noticed that the rules seem to require that it can accept input with the same fill/sep it uses for output, but I don't see anything to gain from breaking that rule.).
NASM/YASM source. Build as 32 or 64-bit code, using the %if
stuff included with the test program or just change rcx to ecx.
input_separator equ 0xa ; `\n` in NASM syntax, but YASM doesn't do C-style escapes
output_filler equ '0' ; For strict rules-compliance, needs to be input_separator+1
output_separator equ output_filler-1 ; saves 1B in 32-bit vs. an arbitrary choice
;; Using output_filler+1 is also possible, but isn't compatible with using the same filler and separator for input and output.
global str_exp
str_exp: ; void str_exp(char *out /*rdi*/, const char *src /*rsi*/,
; unsigned base /*edx*/);
.new_row:
push 1
pop rcx ; ecx=1 = base**0
lodsb ; Skip the first char, since we multiply for the separator
.read_bar:
imul ecx, edx ; accumulate the exponential
lodsb
cmp al, input_separator
ja .read_bar ; anything > separator is treated as filler
; AL = separator or terminator
; flags = below (CF=1) or equal (ZF=1). Equal also implies CF=0, since x-x doesn't produce carry.
mov al, output_filler
rep stosb ; append ecx bytes of filler to the output string
%if output_separator == output_filler-1
dec eax ; saves 1B in the 32-bit version. Use dec even in 64-bit for easier testing
%else
mov al, output_separator
%endif
stosb ; append the delimiter
; CF is still set from the .read_bar loop, even if DEC clobbered the other flags
; JNC/JNB here is equivalent to JE on the original flags, because we can only be here if the char was below-or-equal the separator
jnc .new_row ; separator means more rows, else it's a terminator
; (f+s)+f+ full-match guarantees that the input doesn't end with separator + terminator
ret
The function follows the x86-64 SystemV ABI, with signature
void str_exp(char *out /*rdi*/, const char *src /*rsi*/, unsigned base /*edx*/);
It only informs the caller of the length of the output string by leaving a one-past-the-end pointer to it in rdi
, so you could consider this the return value in a non-standard calling convention.
It would cost 1 or 2 bytes (xchg eax,edi
) to return the end-pointer in eax or rax. (If using the x32 ABI, pointers are guaranteed to be only 32 bits, otherwise we have to use xchg rax,rdi
in case the caller passes a pointer to a buffer outside the low 32 bits.) I didn't include this in the version I'm posting because there are workarounds the caller can use without getting the value from rdi
, so you could call this from C with no wrapper.
We don't even null-terminate the output string or anything, so it's only newline-terminated. It would take 2 bytes to fix that: xchg eax,ecx / stosb
(rcx is zero from rep stosb
.)
The ways to find out the output-string length are:
- rdi points to one-past-the-end of the string on return (so the caller can do len=end-start)
- the caller can just know how many rows were in the input and count newlines
- the caller can use a large zeroed buffer and
strlen()
afterwards.
They're not pretty or efficient (except for using the RDI return value from an asm caller), but if you want that then don't call golfed asm functions from C. :P
Size/range limitations
The max output string size is only limited by virtual memory address-space limitations. (Mainly that current x86-64 hardware only supports 48 significant bits in virtual addresses, split in half because they sign-extend instead of zero-extend. See the diagram in the linked answer.)
Each row can only have a maximum of 2**32 - 1 filler bytes, since I accumulate the exponential in a 32-bit register.
The function works correctly for bases from 0 to 2**32 - 1. (Correct for base 0 is 0^x = 0, i.e. just blank lines with no filler bytes. Correct for base 1 is 1^x = 1, so always 1 filler per line.)
It's also blazingly fast on Intel IvyBridge and later, especially for large rows being written to aligned memory. rep stosb
is an optimal implementation of memset()
for large counts with aligned pointers on CPUs with the ERMSB feature. e.g. 180**4 is 0.97GB, and takes 0.27 seconds on my i7-6700k Skylake (with ~256k soft page-faults) to write to /dev/null. (On Linux the device-driver for /dev/null doesn't copy the data anywhere, it just returns. So all of the time is in the rep stosb
and the soft page-faults that triggers when touching the memory for the first time. It's unfortunately not using transparent hugepages for the array in the BSS. Probably an madvise()
system call would speed it up.)
Test program:
Build a static binary and run as ./string-exponential $'#\n##\n###' $(seq 2)
for base 2. To avoid implementing an atoi
, it uses base = argc-2
. (Command-line length limits prevent testing ridiculously large bases.)
This wrapper works for output strings up to 1 GB. (It only makes a single write() system call even for gigantic strings, but Linux supports this even for writing to pipes). For counting characters, either pipe into wc -c
or use strace ./foo ... > /dev/null
to see the arg to the write syscall.
This takes advantage of the RDI return-value to calculate the string length as an arg for write()
.
;;; Test program that calls it
;;; Assembles correctly for either x86-64 or i386, using the following %if stuff.
;;; This block of macro-stuff also lets us build the function itself as 32 or 64-bit with no source changes.
%ifidn __OUTPUT_FORMAT__, elf64
%define CPUMODE 64
%define STACKWIDTH 8 ; push / pop 8 bytes
%define PTRWIDTH 8
%elifidn __OUTPUT_FORMAT__, elfx32
%define CPUMODE 64
%define STACKWIDTH 8 ; push / pop 8 bytes
%define PTRWIDTH 4
%else
%define CPUMODE 32
%define STACKWIDTH 4 ; push / pop 4 bytes
%define PTRWIDTH 4
%define rcx ecx ; Use the 32-bit names everywhere, even in addressing modes and push/pop, for 32-bit code
%define rsi esi
%define rdi edi
%define rsp esp
%endif
global _start
_start:
mov rsi, [rsp+PTRWIDTH + PTRWIDTH*1] ; rsi = argv[1]
mov edx, [rsp] ; base = argc
sub edx, 2 ; base = argc-2 (so it's possible to test base=0 and base=1, and so ./foo $'xxx\nxx\nx' $(seq 2) has the actual base in the arg to seq)
mov edi, outbuf ; output buffer. static data is in the low 2G of address space, so 32-bit mov is fine. This part isn't golfed, though
call str_exp ; str_exp(outbuf, argv[1], argc-2)
; leaves RDI pointing to one-past-the-end of the string
mov esi, outbuf
mov edx, edi
sub edx, esi ; length = end - start
%if CPUMODE == 64 ; use the x86-64 ABI
mov edi, 1 ; fd=1 (stdout)
mov eax, 1 ; SYS_write (Linux x86-64 ABI, from /usr/include/asm/unistd_64.h)
syscall ; write(1, outbuf, length);
xor edi,edi
mov eax,231 ; exit_group(0)
syscall
%else ; Use the i386 32-bit ABI (with legacy int 0x80 instead of sysenter for convenience)
mov ebx, 1
mov eax, 4 ; SYS_write (Linux i386 ABI, from /usr/include/asm/unistd_32.h)
mov ecx, esi ; outbuf
; 3rd arg goes in edx for both ABIs, conveniently enough
int 0x80 ; write(1, outbuf, length)
xor ebx,ebx
mov eax, 1
int 0x80 ; 32-bit ABI _exit(0)
%endif
section .bss
align 2*1024*1024 ; hugepage alignment (32-bit uses 4M hugepages, but whatever)
outbuf: resb 1024*1024*1024 * 1
; 2GB of code+data is the limit for the default 64-bit code model.
; But with -m32, a 2GB bss doesn't get mapped, so we segfault. 1GB is plenty anyway.
This was a fun challenge that lent itself very well to asm, especially x86 string ops. The rules are nicely designed to avoid having to handle a newline and then a terminator at the end of the input string.
An exponential with repeated multiplication is just like multiplying with repeated addition, and I needed to loop to count chars in each input row anyway.
I considered using one-operand mul
or imul
instead of the longer imul r,r
, but its implicit use of EAX would conflict with LODSB.
I also tried SCASB instead of load and compare, but I needed xchg esi,edi
before and after the inner loop, because SCASB and STOSB both use EDI. (So the 64-bit version has to use the x32 ABI to avoid truncating 64-bit pointers).
Avoiding STOSB is not an option; nothing else is anywhere near as short. And half the benefit of using SCASB is that AL=filler after leaving the inner loop, so we don't need any setup for REP STOSB.
SCASB compares in the other direction from what I had been doing, so I needed to reverse the comparisons.
My best attempt with xchg and scasb. Works, but isn't shorter. (32-bit code, using the inc
/dec
trick to change filler into separator).
; SCASB version, 24 bytes. Also experimenting with a different loop structure for the inner loop, but all these ideas are break-even at best
; Using separator = filler+1 instead of filler-1 was necessary to distinguish separator from terminator from just CF.
input_filler equ '.' ; bytes below this -> terminator. Bytes above this -> separator
output_filler equ input_filler ; implicit
output_separator equ input_filler+1 ; ('/') implicit
8048080: 89 d1 mov ecx,edx ; ecx=base**1
8048082: b0 2e mov al,0x2e ; input_filler= .
8048084: 87 fe xchg esi,edi
8048086: ae scas al,BYTE PTR es:[edi]
08048087 <str_exp.read_bar>:
8048087: ae scas al,BYTE PTR es:[edi]
8048088: 75 05 jne 804808f <str_exp.bar_end>
804808a: 0f af ca imul ecx,edx ; exit the loop before multiplying for non-filler
804808d: eb f8 jmp 8048087 <str_exp.read_bar> ; The other loop structure (ending with the conditional) would work with SCASB, too. Just showing this for variety.
0804808f <str_exp.bar_end>:
; flags = below if CF=1 (filler<separator), above if CF=0 (filler<terminator)
; (CF=0 is the AE condition, but we can't be here on equal)
; So CF is enough info to distinguish separator from terminator if we clobber ZF with INC
; AL = input_filler = output_filler
804808f: 87 fe xchg esi,edi
8048091: f3 aa rep stos BYTE PTR es:[edi],al
8048093: 40 inc eax ; output_separator
8048094: aa stos BYTE PTR es:[edi],al
8048095: 72 e9 jc 8048080 <str_exp> ; CF is still set from the inner loop
8048097: c3 ret
For an input of ../.../.
, produces ..../......../../
. I'm not going to bother showing a hexdump of the version with separator=newline.