I am confused about the operation MSL, which is used in a variant of the MOVI and MVNI instructions.
There's not a lot of information out there, but I have seen it referred to as "Masking Shift Left". Could anyone give an example of what a masking shift is? The instruction variant is also described as "shifting ones variant". I am not sure what this means either.
https://developer.arm.com/documentation/dui0801/c/A64-SIMD-Vector-Instructions/MOVI--vector-
CodePudding user response:
“Shifting ones” means that the bit positions where nothing is shifted to are set to 1. In a “logical shift”, or generally in a “shift” without explicit qualification, those bit positions are set to 0.
Demo (with 64-bit code):
#include <stdio.h>
int main() {
unsigned long v0u, v0l, v1u, v1l, v2u, v2l;
asm (
"movi v0.4s, #42\n"
"mov %[v0u], v0.d[0]\n"
"mov %[v0l], v0.d[1]\n"
"movi v1.4s, #42, lsl #8\n"
"mov %[v1u], v1.d[0]\n"
"mov %[v1l], v1.d[1]\n"
"movi v2.4s, #42, msl #8\n"
"mov %[v2u], v2.d[0]\n"
"mov %[v2l], v2.d[1]\n"
: [v0u] "=r" (v0u), [v0l] "=r" (v0l),
[v1u] "=r" (v1u), [v1l] "=r" (v1l),
[v2u] "=r" (v2u), [v2l] "=r" (v2l)
:
: "v0");
printf("plain : 6lx6lx\n", v0u, v0l);
printf("lsl #8 : 6lx6lx\n", v1u, v1l);
printf("msl #8 : 6lx6lx\n", v2u, v2l);
}
Output:
plain : 0000002a0000002a0000002a0000002a
lsl #8 : 00002a0000002a0000002a0000002a00
msl #8 : 00002aff00002aff00002aff00002aff
I can't find this documented in the assembler manual. It is documented in the Architecture Reference Manual if you care to dig through it (references for version DDI0487H.a):
- C7.2.204
MOVI
p.2490: “shifting ones” iscmode == 110x
. AdvSIMDExpandImm
p.11185:Zeros(X):imm8:Ones(8)
whereX
is either 8 or 16.
CodePudding user response:
Running movi v0.4s, #32, msl #8
through gdb shows v0
containing {0x20FF, 0x20FF, 0x20FF, 0x20FF}
, implying that MSL shifts to the left but pads in with 1's rather than 0's
In binary:
32 would be 00100000
shifting left 8 places padding in 1's gives 0010000011111111
which is 20FF
in hex
The same is true with a negative immediate, with movi v1.4s #-32, msl #8
giving {0xE0FF, 0xE0FF, 0xE0FF, 0xE0FF}