When working on LLVM, I found this problem
https://github.com/llvm/llvm-project/issues/64974.
Maybe it's time for us to reconsider the way of getting GOT address
for PIC code.
1.Background[1]:
All of the accessing of global variables and normal function calls
needs help from GOT.
So normally, the first 3 instructions of a function are to compute the
address of GOT.
Normally like:
lui $gp,%hi(_gp_disp)
addiu $gp,$gp,%lo(_gp_disp)
addu $gp,$gp,$t9
These 3 instructions load the value of _gp_disp, which is a link-time
constant value,
and add them with $t9, which normally contains the address of the
current function.
So, why $t9? The reason is that the pre-R6 MIPS lacks instructions to
get the PC value.
Thus, the ABI defines that, the caller should call functions with
jr/jarl $t9. So the callee
can get the address of itself by $t9.
2. What's my proposal?
2.1 For MIPSr6, its has instructions to get the current PC value, so
we can use them,
which can gain performance improvement.
2.2 For pre-R6, in fact, it can get the value of PC with NAL
instruction [2], while the performance will be some worse. For the
worst case, the decreasing of performance will be 20%-40%
My plan for pre-R6, is to add a non-default option to use NAL to get
the address of GOT.
Some software, like Linux kernel (for Kalsr support) will need it.
Any suggestions?
[1] https://refspecs.linuxfoundation.org/elf/mipsabi.pdf
[2]
https://s3-eu-west-1.amazonaws.com/downloads-mips/documents/MD00087-2B-MIPS64BIS-AFP-6.06.pdf
Page 404
--
YunQiang Su
