On Thu, Jun 5, 2025 at 3:37 PM Thomas Weißschuh thomas.weissschuh@linutronix.de wrote:
On Wed, Jun 04, 2025 at 10:30:55PM +0800, Xi Ruoyao wrote:
On Wed, 2025-06-04 at 22:05 +0800, Huacai Chen wrote:
On Tue, Jun 3, 2025 at 7:49 PM Thomas Weißschuh thomas.weissschuh@linutronix.de wrote:
The syscall wrappers use the "a0" register for two different register variables, both the first argument and the return value. The "ret" variable is used as both input and output while the argument register is only used as input. Clang treats the conflicting input parameters as undefined behaviour and optimizes away the argument assignment.
The code seems to work by chance for the most part today but that may change in the future. Specifically clock_gettime_fallback() fails with clockids from 16 to 23, as implemented by the upcoming auxiliary clocks.
Switch the "ret" register variable to a pure output, similar to the other architectures' vDSO code. This works in both clang and GCC.
Hmmm, at first the constraint is "=r", during the progress of upstream, Xuerui suggested me to use "+r" instead [1]. [1] https://lore.kernel.org/linux-arch/5b14144a-9725-41db-7179-c059c41814cf@xen0...
Based on the example at https://gcc.gnu.org/onlinedocs/gcc/Local-Register-Variables.html:
To force an operand into a register, create a local variable and specify the register name after the variable’s declaration. Then use the local variable for the asm operand and specify any constraint letter that matches the register:
register int *p1 asm ("r0") = …; register int *p2 asm ("r1") = …; register int *result asm ("r0"); asm ("sysint" : "=r" (result) : "0" (p1), "r" (p2));
I think this should actually be written
asm volatile( " syscall 0\n" : "=r" (ret) : "r" (nr), "0" (buffer), "r" (len), "r" (flags) : "$t0", "$t1", "$t2", "$t3", "$t4", "$t5", "$t6", "$t7","$t8", "memory");
i.e. "=" should be used for the output operand 0, and "0" should be used for the input operand 2 (buffer) to emphasis the same register as operand 0 is used.
I would have expected that matching constraints ("0") would only really make sense if the compiler selects the specific register to use. When the register is already selected manually it seems redundant. But my inline ASM knowledge is limited and this is a real example from the GCC docs, so it is probably more correct. On the other hand all the other vDSO implementations use "r" over "0" for the input operand 2 and I'd like to keep them consistent.
OK, if there are no objections, I will take this patch.
Huacai
Thomas