serge-sans-paille added a subscriber: eli.friedman.
serge-sans-paille added a comment.
@efriedma the free probe algorithm requires more testing, and I'd like to take
into account memset and memcpy as free probes too. To showcase this algorithm,
consider the following LLVM bitcode:
define i32 @foo() local_unnamed_addr {
%a = alloca i32, i64 2000, align 16
%b = getelementptr inbounds i32, i32* %a, i64 1198
store volatile i32 1, i32* %b
%c = load volatile i32, i32* %a
ret i32 %c
}
when compiled with `llc` it outputs the following assembly:
foo: # @foo
subq $7880, %rsp # imm = 0x1EC8
movl $1, 4664(%rsp)
movl -128(%rsp), %eax
addq $7880, %rsp # imm = 0x1EC8
retq
When `probe-stack` is set to `inline-asm` it outputs
foo: # @foo
subq $4096, %rsp # imm = 0x1000
movl $1, 880(%rsp)
subq $3784, %rsp # imm = 0xEC8
movq $0, (%rsp)
movl -128(%rsp), %eax
addq $7880, %rsp # imm = 0x1EC8
retq
The stack allocation is split in two, but only one MOV is added, the first one
is what I call a free probe. Turns out we could only use natural probes here, I
need to implement that :-)
As a comparison, setting `probe-stack` to a random function name like
`__probe_stack` outputs the following:
foo: # @foo
movl $8008, %eax # imm = 0x1F48
callq __probe_stack
subq %rax, %rsp
movl $1, 4792(%rsp)
movl (%rsp), %eax
addq $8008, %rsp # imm = 0x1F48
retq
which requires runtime support (to provide `__stack_probe`), and a function
call overhead, while ideally just an extra `sub %rsp` would be needed.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D68720/new/
https://reviews.llvm.org/D68720
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
