From 59f2954fcaacd9426827c69a729e2647cb9977e5 Mon Sep 17 00:00:00 2001 From: Palmer Dabbelt Date: Tue, 24 Sep 2019 20:30:15 -0700 Subject: [PATCH] correct the operand specifiers in the riscv64 CAS routines The operand sepcifiers in a_cas and a_casp for riscv64 were incorrect: there's a backwards branch in the routine, so despite tmp being written at the end of the assembly fragment it cannot be allocated in one of the input registers because the input values may be needed for another trip around the loop. For code that follows the guarnteed forward progress requirements, he backwards branch is rarely taken: SiFive's hardware only fails a store conditional on execptional cases (ie, instruction cache misses inside the loop), and until recently a bug in QEMU allowed back-to-back store conditionals to succeed. The bug has been fixed in the latest QEMU release, but it turns out that the fix caused this latent bug in musl to manifest. Full disclosure: I haven't actually even compiled musl. I just guessed this would fix a bug introducted by the new QEMU behavior, Alistair (CC'd) actually checked it fixes the problem. The rest is just conjecture. Upstream-Status: Submitted Signed-off-by: Khem Raj --- arch/riscv64/atomic_arch.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/arch/riscv64/atomic_arch.h b/arch/riscv64/atomic_arch.h index c9765342..41ad4d04 100644 --- a/arch/riscv64/atomic_arch.h +++ b/arch/riscv64/atomic_arch.h @@ -14,7 +14,7 @@ static inline int a_cas(volatile int *p, int t, int s) " sc.w.aqrl %1, %4, (%2)\n" " bnez %1, 1b\n" "1:" - : "=&r"(old), "=r"(tmp) + : "=&r"(old), "=&r"(tmp) : "r"(p), "r"(t), "r"(s) : "memory"); return old; @@ -31,7 +31,7 @@ static inline void *a_cas_p(volatile void *p, void *t, void *s) " sc.d.aqrl %1, %4, (%2)\n" " bnez %1, 1b\n" "1:" - : "=&r"(old), "=r"(tmp) + : "=&r"(old), "=&r"(tmp) : "r"(p), "r"(t), "r"(s) : "memory"); return old; -- 2.23.0