Files
libsst/libsst-atomic/SST_Atomic_mips64.asm
2026-04-03 00:22:39 -05:00

278 lines
7.9 KiB
NASM

# SST_Atomic_mips64.asm
# Author: Patrick Baggett <ptbaggett@762studios.com>
# Created: 7/1/2013
#
# Purpose:
#
# 64-bit assembly for atomic operations for MIPS CPUs. Targets MIPS64R2 ISA, but should run on
# older MIPS machines (e.g. MIPS-III / Loongson).
# Assembles with GNU as (doubtfully IRIX)
#
# License:
#
# This program is free software. It comes without any warranty, to
# the extent permitted by applicable law. You can redistribute it
# and/or modify it under the terms of the Do What The Fuck You Want
# To Public License, Version 2, as published by Sam Hocevar. See
# http://sam.zoy.org/wtfpl/COPYING for more details.
.text
#Don't reorder instructions
.set noreorder
# ELF symbol names
.global SST_Atomic_Add
.global SST_Atomic_AddPtr
.global SST_Atomic_And
.global SST_Atomic_Or
.global SST_Atomic_Xor
.global SST_Atomic_Not
.global SST_Atomic_AddReturn
.global SST_Atomic_AddPtrReturn
.global SST_Atomic_AndReturn
.global SST_Atomic_OrReturn
.global SST_Atomic_XorReturn
.global SST_Atomic_NotReturn
.global SST_Atomic_ExchangeAdd
.global SST_Atomic_ExchangeAddPtr
.global SST_Atomic_Exchange
.global SST_Atomic_ExchangePtr
.global SST_Atomic_CAS
.global SST_Atomic_CASPtr
.global SST_Atomic_LoadAcquire
.global SST_Atomic_LoadAcquirePtr
.global SST_Atomic_StoreRelease
.global SST_Atomic_StoreReleasePtr
# Unbelievable. binutils-2.22 doesn't support register names in n64 (-mabi=64) mode, but it does in o32 (-mabi=32)
# Here are the symbolic names for registers used
# $v0-$v1 = $2-$3
# $a0-$a4 = $4-$7
# $t4-$t5 = $8-$9
# $ra = $31
# MIPS has a "sync" instruction with a bitfield describing "what" to sync, sort of like SPARC's "membar" instruction.
# By default, "sync" means "sync 0" -- a full memory barrier including I/O devices/instructions. This is a very heavy-weight operation and used for things like ordering MMIO accesses
# MIPS64R2 defines "sync 0x10" as a full memory barrier that doesn't synchronize I/O, and "sync 0x04" as a "write/write" memory barrier. However,
# we can't use MIPS64R2 code only. But as it would turn out, it is OK because "sync <k>" is treated like "sync 0" in older MIPS CPUs, meaning full barrier,
# but it will automatically be faster on MIPS32R2 CPUs. Neat.
# void SST_Atomic_Add(volatile int* x, int value)
# int SST_Atomic_AddReturn(volatile int* x, int value)
SST_Atomic_Add:
SST_Atomic_AddReturn:
sync 0x10 # sync_mb
1:
ll $12, 0($4) # t4 <- *x
addu $13, $12, $5 # t5 <- t4 + value
sc $13, 0($4) # try { *x <- t5 }, set t5 = 1/0 depending
beqz $13,1b # if failure, retry
addu $2, $12, $5 # (delay slot) $2 <- *x + value (return value)
jr $31 # return
sync 0x04 # (delay slot) sync_wmb
# void SST_Atomic_AddPtr(volatile void** x, int value)
# void* SST_Atomic_AddPtrReturn(volatile void** x, int value)
SST_Atomic_AddPtr:
SST_Atomic_AddPtrReturn:
sync 0x10 # sync_mb
1:
lld $12, 0($4) # t4 <- *x
daddu $13, $12, $5 # t5 <- t4 + value
scd $13, 0($4) # try { *x <- t5 }, set t5 = 1/0 depending
beqz $13,1b # if failure, retry
daddu $2, $12, $5 # (delay slot) $2 <- *x + value (return value)
jr $31 # return
sync 0x04 # (delay slot) sync_wmb
# void SST_Atomic_And(volatile int* x, int value)
# int SST_Atomic_AndReturn(volatile int* x, int value)
SST_Atomic_And:
SST_Atomic_AndReturn:
sync 0x10 # sync_mb
1:
ll $12, 0($4) # t4 <- *x
and $13, $12, $5 # t5 <- t4 & value
sc $13, 0($4) # try { *x <- t5 }, set t5 = 1/0 depending
beqz $13,1b # if failure, retry
and $2, $12, $5 # (delay slot) compute return value
jr $31 # return
sync 0x04 # (delay slot) sync_wmb
# void SST_Atomic_Or(volatile int* x, int value)
# int SST_Atomic_OrReturn(volatile int* x, int value)
SST_Atomic_Or:
SST_Atomic_OrReturn:
sync 0x10 # sync_mb
1:
ll $12, 0($4) # t4 <- *x
or $13, $12, $5 # t5 <- t4 | value
sc $13, 0($4) # try { *x <- t5 }, set t5 = 1/0 depending
beqz $13,1b # if failure, retry
or $2, $12, $5 # (delay slot) compute return value
jr $31 # return
sync 0x04 # (delay slot) sync_wmb
# void SST_Atomic_Xor(volatile int* x, int value)
# int SST_Atomic_XorReturn(volatile int* x, int value)
SST_Atomic_XorReturn:
SST_Atomic_Xor:
sync 0x10 # sync_mb
1:
ll $12, 0($4) # t4 <- *x
xor $13, $12, $5 # t5 <- t4 ^ value
sc $13, 0($4) # try { *x <- t5 }, set t5 = 1/0 depending
beqz $13,1b # if failure, retry
xor $2, $12, $5 # (delay slot) Nothing to do
jr $31 # return
sync 0x04 # (delay slot) sync_wmb
# void SST_Atomic_Not(volatile int* x)
SST_Atomic_Not:
SST_Atomic_NotReturn:
sync 0x10 # sync_mb
1:
ll $12, 0($4) # t4 <- *x
nor $13, $12, $0 # t5 <- ~t4
sc $13, 0($4) # try { *x <- t5 }, set t5 = 1/0 depending
beqz $13,1b # if failure, retry
nor $2, $12, $0 # (delay slot) compute return value
jr $31 # return
sync 0x04 # (delay slot) sync_wmb
# int SST_Atomic_Exchange(volatile int* x, int value)
SST_Atomic_Exchange:
sync 0x10 # sync_mb
move $12, $5 # t4 <- value
1:
ll $2, 0($4) # v0 <- *x
sc $12, 0($4) # try { *x <- t4 }, set t4 = 1/0 depending
beqz $12,1b # if failure, retry
move $12, $5 # (delay slot) reload t4 <- value for retry
jr $31 # return
sync 0x04 # (delay slot) sync_wmb
# void* SST_Atomic_ExchangePtr(volatile void** x, void* value)
SST_Atomic_ExchangePtr:
sync 0x10 # sync_mb
move $12, $5 # value for exchange
1:
lld $2, 0($4) # v0 <- *x
scd $12, 0($4) # try { *x <- t4 }, set t4 = 1/0 depending
beqz $12,1b # if failure, retry
move $12, $5 # (delay slot) t4 <- value for retry
jr $31 # return
sync 0x04 # (delay slot) sync_wmb
# int SST_Atomic_ExchangeAdd(volatile int* x, int value);
SST_Atomic_ExchangeAdd:
sync 0x10 # sync_mb
1:
ll $2, 0($4) # v0 <- *x
addu $12, $2, $5 # t4 <- v0 + value
sc $12, 0($4) # try { *x <- t4 }, set t4 = 1/0 depending
beqz $12,1b # if failure, retry
nop # (delay slot) do nothing
jr $31 # return
sync 0x04 # (delay slot) sync_wmb
# void* SST_Atomic_ExchangeAddPtr(volatile void** x, int value);
SST_Atomic_ExchangeAddPtr:
sync 0x10 # sync_mb
1:
lld $2, 0($4) # v0 <- *x
daddu $12, $2, $5 # t4 <- v0 + value
scd $12, 0($4) # try { *x <- t4 }, set t4 = 1/0 depending
beqz $12,1b # if failure, retry
nop # (delay slot) do nothing
jr $31 # return
sync 0x04 # (delay slot) sync_wmb
# int SST_Atomic_CAS(int* dest, int compare, int newValue);
# a0 = dest, a1 = compare, a2 = newValue
SST_Atomic_CAS:
sync 0x10 # sync_mb
1:
move $12, $6 # t4 <- newValue
ll $2, 0($4) # v0 <- *dest (captures "old" value of *dest, also is the return value)
bne $2, $5, 3f # compare v0 with a1 ("compare"), and if not equal, exit loop
nop # (delay slot) do nothing
sc $12, 0($4) # try { *x <- t4 }, set t4 = 1/0 depending
beqz $12,1b # if failure, retry
nop # (delay slot) do nothing
sync 0x04 # sync_wmb
# Return
3:
jr $31 # return
nop # (delay slot) do nothing
# void* SST_Atomic_CASPtr(void** dest, void* compare, void* newValue);
# a0 = dest, a1 = compare, a2 = newValue
SST_Atomic_CASPtr:
sync 0x10 # sync_mb
1:
move $12, $6 # t4 <- newValue
lld $2, 0($4) # v0 <- *dest (captures "old" value of *dest, also is the return value)
bne $2, $5, 3f # compare v0 with a1 ("compare"), and if not equal, exit loop
nop # (delay slot) do nothing
scd $12, 0($4) # try { *x <- t4 }, set t4 = 1/0 depending
beqz $12,1b # if failure, retry
nop # (delay slot) do nothing
sync 0x04 # sync_wmb
# Return
3:
jr $31 # return
nop # (delay slot) do nothing
SST_Atomic_LoadAcquire:
lw $2, 0($4)
sync 0x11 # sync_acquire: LoadLoad | LoadStore
jr $31
nop
SST_Atomic_StoreRelease:
sync 0x12 # sync_release: LoadStore | StoreStore
jr $31
sw $5, 0($4)
SST_Atomic_LoadAcquirePtr:
ld $2, 0($4)
sync 0x11 # sync_acquire: LoadLoad | LoadStore
jr $31
nop
SST_Atomic_StoreReleasePtr:
sync 0x12 # sync_release: LoadStore | StoreStore
jr $31
sd $5, 0($4)