I am trying to alter a little bit the LLVM Backend for X86 target, to produce some desired behaviour.
More specifically, I would like to emulate a flag like gcc's fcall-used-reg option, which instructs the compiler to convert a callee-saved register into a clobbered register (meaning that it may be altered during a function call).
Let's focus on r14. I manually clobber the register, like in this answer:
#include <inttypes.h>
uint64_t inc(uint64_t i) {
__asm__ __volatile__(
""
: "+m" (i)
:
: "r14"
);
return i + 1;
}
int main(int argc, char **argv) {
(void)argv;
return inc(argc);
}
Compile and disassemble:
gcc -std=gnu99 -O3 -ggdb3 -Wall -Wextra -pedantic -o main.out main.c
objdump -d main.out
Disassembly contains:
0000000000001150 <inc>:
1150: 41 56 push %r14
1152: 48 89 7c 24 f8 mov %rdi,-0x8(%rsp)
1157: 48 8b 44 24 f8 mov -0x8(%rsp),%rax
115c: 41 5e pop %r14
115e: 48 83 c0 01 add $0x1,%rax
1162: c3 retq
1163: 66 2e 0f 1f 84 00 00 nopw %cs:0x0(%rax,%rax,1)
116a: 00 00 00
116d: 0f 1f 00 nopl (%rax)
where we can see that r14, because it is tampered with, is pushed to the stack, and then popped to regain its original value. Now, repeat with the -fcall-used-r14 flag:
gcc -std=gnu99 -O3 -ggdb3 -fcall-used-r14 -Wall -Wextra -pedantic -o main.out main.c
objdump -d main.out
Disassembly contains:
0000000000001150 <inc>:
1150: 48 89 7c 24 f8 mov %rdi,-0x8(%rsp)
1155: 48 8b 44 24 f8 mov -0x8(%rsp),%rax
115a: 48 83 c0 01 add $0x1,%rax
115e: c3 retq
115f: 90 nop
where no push/pop happens.
Now, I have modified some LLVM Target files, compiled the source, and added(?) this functionality to the llc tool:
clang-11 -emit-llvm -S -c main.c -o main.ll
llc-11 main.ll -o main.s
Now, main.s contains:
# %bb.0:
pushq %rbp
.cfi_def_cfa_offset 16
.cfi_offset %rbp, -16
movq %rsp, %rbp
.cfi_def_cfa_register %rbp
pushq %r14
.cfi_offset %r14, -24
movq %rdi, -16(%rbp)
#APP
#NO_APP
movq -16(%rbp), %rax
addq $1, %rax
popq %r14
popq %rbp
.cfi_def_cfa %rsp, 8
retq
Apparently, r14 is still callee-saved.
Inside llvm/lib/Target/X86/X86CallingConv.td I have modified the following lines (removing R14), because they seemed the only relevant to the System V ABI for Linux and C calling conventions that I was interested in:
def CSR_64 : CalleeSavedRegs<(add R12, R13, R15, RBP)>;
...
def CSR_64_MostRegs : CalleeSavedRegs<(add RBX, RCX, RDX, RSI, RDI, R8, R9, R10,
R11, R12, R13, R15, RBP,
...
def CSR_64_AllRegs_NoSSE : CalleeSavedRegs<(add RAX, RBX, RCX, RDX, RSI, RDI, R8, R9,
R10, R11, R12, R13, R15, RBP)>;
My questions are:
X86CallingConv.td the only file I should modify? I think yes, but maybe I'm wrong.I am running LLVM 11 inside Debian 10.5.
EDIT:
Changing the line, removing R14 from "hidden" definition:
def CSR_SysV64_RegCall_NoSSE : CalleeSavedRegs<(add RBX, RBP, RSP,
(sequence "R%u", 12, 13), R15)>;
as Margaret correctly pointed out did not help either.
Turns out, the minimum modification was the line:
def CSR_64 : CalleeSavedRegs<(add RBX, R12, R13, R15, RBP)>;
The problem was with how I built the source.
By running cmake --build . again after the original installation, the llc tool was not modified globally (I thought it would have because I was building the default architecture - X86 - but that was irrelevant). So, I was calling an unmodified llc-11 tool. Thus, when I ran:
/path/to/llvm-project/build/bin/lcc main.ll -o main.s
main.s contained:
# %bb.0:
movq %rdi, -8(%rsp)
#APP
#NO_APP
movq -8(%rsp), %rax
addq $1, %rax
retq
which is what I wanted in the first place.