h-y-b / riscvv-test Goto Github PK

This repository contains RISC-V assembly examples.

These example use RISC-V "V" vector instructions (version 0.8).

• bcd2ascii.s - convert BCD strings into ASCII strings. See also https://gms.tf/riscv-vector.html for details.
• ascii2bcd.s - convert ASCII strings into BCD strings.
• memchr.s - vector version of the well-known memchr() libc function
• mempchr.s - similar to memchr(), but mempchr() returns the one-past-the-end pointer instead of NULL. This is more useful e.g. in text scanning code.
• memrchr.s - vector version of the well-known memrchr() libc function
• memcmp.s - vector version of the well-known memcmp() libc function
• rawmemchr.s - vector version of the well-known rawmemchr() glibc function
• strlen.s, strlen_small.s - vector versions of the well-known strlen() libc function

The RISC-V "V" extension specification contains several vector assembly code example, e.g. vector versions of memcpy(), strcpy(), strncpy() and strlen().

• rotate.s - the general purpose RISC-V instruction sets don't include rotate instructions, instead they are part of the Bitmanip "B" extension. This example shows how to bit-rotate without the "B" extension.

To stress the point that a length agnostic vector instruction set allows to write very compact (and arguably elegant) code (cf. https://gms.tf/riscv-vector.html), this repository also contains an x86 SIMD version of bcd2ascii.s: x86_64/bcd2ascii.S It requires many more instructions because of extra code that deals with input that isn't dividable by the SIMD register size.

It also demonstrates how many x86 SIMD instructions encode into 4 or more bytes due to the bloated x86 opcode space:

$ objdump -M intel -d bcd2ascii.o | grep '^.*\<v[^ ]*' -o \
     | awk '{ print NF-2, $NF }'

(whereas RISC-V vector instructions uniformly encode into 4 bytes)

2020, Georg Sauthoff [email protected]