| Issue |
87836
|
| Summary |
Over-aggressively optimization on infinite loops affects RISC-V platforms as well
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
leonardopsantos
|
Not exactly a duplicate of https://github.com/llvm/llvm-project/issues/60622 because that issue specifically mentions x86.
I just hit this issue in RISC-V bare-metal target with `clang 17.0.0`, so most definitively not restricted to x86. My code is fairly similar to the test case above:
```c
#define DO_WAIT 0
#define DO_ACTION 1
int g_lock = DO_WAIT;
void __attribute__((noinline)) worker() {
for (;;) {
volatile int i = 0;
i++;
g_lock = DO_WAIT;
while (g_lock != DO_ACTION) {
}
i = 0xCAFECAFE;
}
return;
}
void dontcallme() {
volatile int i = 0;
i++;
}
int main() {
worker();
return 0;
}
```
The disassembly for that is:
```asm
worker: # @worker
# %bb.0: # %for.cond
addi sp, sp, -16
sw zero, 12(sp)
lw a0, 12(sp)
addi a0, a0, 1
sw a0, 12(sp)
.Lfunc_end0:
.size worker, .Lfunc_end0-worker
# -- End function
.globl dontcallme # -- Begin function dontcallme
.p2align 2
.type dontcallme,@function
dontcallme: # @dontcallme
# %bb.0: # %entry
addi sp, sp, -16
sw zero, 12(sp)
lw a0, 12(sp)
addi a0, a0, 1
sw a0, 12(sp)
addi sp, sp, 16
ret
.Lfunc_end1:
.size dontcallme, .Lfunc_end1-dontcallme
# -- End function
.globl main # -- Begin function main
```
There's no control instruction at the end of `worker`, so it just falls through to `dontcallme`.
I'm baffled as to how can anyone think that's reasonable.
GCC for RISC-V _does the right thing_ and adds a `j`: https://godbolt.org/z/h55bYsdo9
I've tried the `-fsanitize=undefined` flag as suggested [here](https://github.com/llvm/llvm-project/issues/48943#issuecomment-981039980), but that doesn't work: https://godbolt.org/z/aYrqfsca7
The workaround is very counter-intuitive:
```c
while (g_lock != DO_ACTION) {
asm volatile("");
}
```
Yields:
```asm
.LBB0_2: # %while.body
# Parent Loop BB0_1 Depth=1
# => This Inner Loop Header: Depth=2
#APP
#NO_APP
lw a3, %pcrel_lo(.Lpcrel_hi0)(a0)
bne a3, a1, .LBB0_2
# %bb.3: # %while.end
# in Loop: Header=BB0_1 Depth=1
sw a2, 12(sp)
j .LBB0_1
.Lfunc_end0:
.size worker, .Lfunc_end0-worker
# -- End function
```
Also here: https://godbolt.org/z/f6Y6qdrMT
I also disagree that https://github.com/llvm/llvm-project/issues/60622, https://github.com/llvm/llvm-project/issues/60419, and https://github.com/llvm/llvm-project/issues/60588 should be closed. Especially because there are no proposed workarounds.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
