Skip to content

Commit 826e713

Browse files
tmandrytmandry
committed
Use global_asm! instead of inline asm!
1 parent c7b64af commit 826e713

File tree

2 files changed

+92
-120
lines changed

2 files changed

+92
-120
lines changed

‎src/lib.rs

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,7 @@
11
#![cfg_attr(feature = "compiler-builtins", compiler_builtins)]
22
#![feature(abi_unadjusted)]
33
#![feature(asm)]
4+
#![feature(global_asm)]
45
#![feature(cfg_target_has_atomic)]
56
#![feature(compiler_builtins)]
67
#![feature(core_intrinsics)]

‎src/probestack.rs

Lines changed: 91 additions & 120 deletions
Original file line numberDiff line numberDiff line change
@@ -61,30 +61,15 @@ macro_rules! define_rust_probestack {
6161
($body: expr) => {
6262
concat!(
6363
"
64-
// We are about to define a 'function within a function.' Because the
65-
// compiler will have emitted a .cfi_startproc at the beginning of
66-
// __rust_probestack_wrapper, we need .cfi_endproc before we can define
67-
// the contents of __rust_probestack.
68-
.cfi_endproc
69-
7064
.pushsection .text.__rust_probestack
7165
.globl __rust_probestack
7266
.type __rust_probestack, @function
7367
__rust_probestack:
74-
.cfi_startproc
75-
7668
",
7769
$body,
7870
"
79-
80-
.cfi_endproc
81-
8271
.size __rust_probestack, . - __rust_probestack
8372
.popsection
84-
85-
// Similar to above, we add .cfi_startproc here to match the
86-
// .cfi_endproc emitted at the end of __rust_probestack_wrapper.
87-
.cfi_startproc
8873
"
8974
)
9075
};
@@ -100,119 +85,105 @@ macro_rules! define_rust_probestack {
10085
($body: expr) => {
10186
concat!(
10287
"
103-
.cfi_endproc
10488
.globl __rust_probestack
10589
__rust_probestack:
106-
.cfi_startproc
107-
10890
",
109-
$body,
110-
"
111-
112-
.cfi_endproc
113-
.cfi_startproc
114-
"
91+
$body
11592
)
11693
};
11794
}
11895

119-
#[naked]
120-
#[no_mangle]
96+
// Our goal here is to touch each page between %rsp+8 and %rsp+8-%rax,
97+
// ensuring that if any pages are unmapped we'll make a page fault.
98+
//
99+
// The ABI here is that the stack frame size is located in `%rax`. Upon
100+
// return we're not supposed to modify `%rsp` or `%rax`.
121101
#[cfg(all(target_arch = "x86_64", not(feature = "mangled-names")))]
122-
pub unsafe extern "C" fn __rust_probestack_wrapper() {
123-
// Our goal here is to touch each page between %rsp+8 and %rsp+8-%rax,
124-
// ensuring that if any pages are unmapped we'll make a page fault.
102+
global_asm!(define_rust_probestack!("
103+
.cfi_startproc
104+
pushq %rbp
105+
.cfi_adjust_cfa_offset 8
106+
.cfi_offset %rbp, -16
107+
movq %rsp, %rbp
108+
.cfi_def_cfa_register %rbp
109+
110+
mov %rax,%r11 // duplicate %rax as we're clobbering %r11
111+
112+
// Main loop, taken in one page increments. We're decrementing rsp by
113+
// a page each time until there's less than a page remaining. We're
114+
// guaranteed that this function isn't called unless there's more than a
115+
// page needed.
125116
//
126-
// The ABI here is that the stack frame size is located in `%rax`. Upon
127-
// return we're not supposed to modify `%rsp` or `%rax`.
128-
asm!(define_rust_probestack!("
129-
pushq %rbp
130-
.cfi_adjust_cfa_offset 8
131-
.cfi_offset %rbp, -16
132-
movq %rsp, %rbp
133-
.cfi_def_cfa_register %rbp
134-
135-
mov %rax,%r11 // duplicate %rax as we're clobbering %r11
136-
137-
// Main loop, taken in one page increments. We're decrementing rsp by
138-
// a page each time until there's less than a page remaining. We're
139-
// guaranteed that this function isn't called unless there's more than a
140-
// page needed.
141-
//
142-
// Note that we're also testing against `8(%rsp)` to account for the 8
143-
// bytes pushed on the stack orginally with our return address. Using
144-
// `8(%rsp)` simulates us testing the stack pointer in the caller's
145-
// context.
146-
147-
// It's usually called when %rax >= 0x1000, but that's not always true.
148-
// Dynamic stack allocation, which is needed to implement unsized
149-
// rvalues, triggers stackprobe even if %rax < 0x1000.
150-
// Thus we have to check %r11 first to avoid segfault.
151-
cmp $$0x1000,%r11
152-
jna 3f
153-
2:
154-
sub $$0x1000,%rsp
155-
test %rsp,8(%rsp)
156-
sub $$0x1000,%r11
157-
cmp $$0x1000,%r11
158-
ja 2b
159-
160-
3:
161-
// Finish up the last remaining stack space requested, getting the last
162-
// bits out of r11
163-
sub %r11,%rsp
164-
test %rsp,8(%rsp)
165-
166-
// Restore the stack pointer to what it previously was when entering
167-
// this function. The caller will readjust the stack pointer after we
168-
// return.
169-
add %rax,%rsp
170-
171-
leave
172-
.cfi_def_cfa_register %rsp
173-
.cfi_adjust_cfa_offset -8
174-
ret
175-
") ::: "memory" : "volatile");
176-
::core::intrinsics::unreachable();
177-
}
117+
// Note that we're also testing against `8(%rsp)` to account for the 8
118+
// bytes pushed on the stack orginally with our return address. Using
119+
// `8(%rsp)` simulates us testing the stack pointer in the caller's
120+
// context.
121+
122+
// It's usually called when %rax >= 0x1000, but that's not always true.
123+
// Dynamic stack allocation, which is needed to implement unsized
124+
// rvalues, triggers stackprobe even if %rax < 0x1000.
125+
// Thus we have to check %r11 first to avoid segfault.
126+
cmp $0x1000,%r11
127+
jna 3f
128+
2:
129+
sub $0x1000,%rsp
130+
test %rsp,8(%rsp)
131+
sub $0x1000,%r11
132+
cmp $0x1000,%r11
133+
ja 2b
134+
135+
3:
136+
// Finish up the last remaining stack space requested, getting the last
137+
// bits out of r11
138+
sub %r11,%rsp
139+
test %rsp,8(%rsp)
140+
141+
// Restore the stack pointer to what it previously was when entering
142+
// this function. The caller will readjust the stack pointer after we
143+
// return.
144+
add %rax,%rsp
145+
146+
leave
147+
.cfi_def_cfa_register %rsp
148+
.cfi_adjust_cfa_offset -8
149+
ret
150+
.cfi_endproc
151+
"));
178152

179-
#[naked]
180-
#[no_mangle]
181153
#[cfg(all(target_arch = "x86", not(feature = "mangled-names")))]
182-
pub unsafe extern "C" fn __rust_probestack_wrapper() {
183-
// This is the same as x86_64 above, only translated for 32-bit sizes. Note
184-
// that on Unix we're expected to restore everything as it was, this
185-
// function basically can't tamper with anything.
186-
//
187-
// The ABI here is the same as x86_64, except everything is 32-bits large.
188-
asm!(define_rust_probestack!("
189-
push %ebp
190-
.cfi_adjust_cfa_offset 4
191-
.cfi_offset %ebp, -8
192-
mov %esp, %ebp
193-
.cfi_def_cfa_register %ebp
194-
push %ecx
195-
mov %eax,%ecx
196-
197-
cmp $$0x1000,%ecx
198-
jna 3f
199-
2:
200-
sub $$0x1000,%esp
201-
test %esp,8(%esp)
202-
sub $$0x1000,%ecx
203-
cmp $$0x1000,%ecx
204-
ja 2b
205-
206-
3:
207-
sub %ecx,%esp
208-
test %esp,8(%esp)
209-
210-
add %eax,%esp
211-
pop %ecx
212-
leave
213-
.cfi_def_cfa_register %esp
214-
.cfi_adjust_cfa_offset -4
215-
ret
216-
") ::: "memory" : "volatile");
217-
::core::intrinsics::unreachable();
218-
}
154+
// This is the same as x86_64 above, only translated for 32-bit sizes. Note
155+
// that on Unix we're expected to restore everything as it was, this
156+
// function basically can't tamper with anything.
157+
//
158+
// The ABI here is the same as x86_64, except everything is 32-bits large.
159+
global_asm!(define_rust_probestack!("
160+
.cfi_startproc
161+
push %ebp
162+
.cfi_adjust_cfa_offset 4
163+
.cfi_offset %ebp, -8
164+
mov %esp, %ebp
165+
.cfi_def_cfa_register %ebp
166+
push %ecx
167+
mov %eax,%ecx
168+
169+
cmp $0x1000,%ecx
170+
jna 3f
171+
2:
172+
sub $0x1000,%esp
173+
test %esp,8(%esp)
174+
sub $0x1000,%ecx
175+
cmp $0x1000,%ecx
176+
ja 2b
177+
178+
3:
179+
sub %ecx,%esp
180+
test %esp,8(%esp)
181+
182+
add %eax,%esp
183+
pop %ecx
184+
leave
185+
.cfi_def_cfa_register %esp
186+
.cfi_adjust_cfa_offset -4
187+
ret
188+
.cfi_endproc
189+
"));

0 commit comments

Comments
 (0)