When we do a context switch, always set the FPU to "off" for the old proc.

Disable the optimization to mark the new proc "clean" since it causes
random failures in regress.  Hopefully we can revisit this soon.

ok drahn@

diff --git a/sys/arch/riscv64/riscv64/machdep.c b/sys/arch/riscv64/riscv64/machdep.c
index 43cdba3..71fa65a 100644 (file)
--- a/sys/arch/riscv64/riscv64/machdep.c
+++ b/sys/arch/riscv64/riscv64/machdep.c
@@ -1,4 +1,4 @@
-/*     $OpenBSD: machdep.c,v 1.20 2021/06/13 16:27:15 kettenis Exp $   */
+/*     $OpenBSD: machdep.c,v 1.21 2021/06/18 21:52:47 kettenis Exp $   */
 
 /*
  * Copyright (c) 2014 Patrick Wildt <patrick@blueri.se>
@@ -304,16 +304,20 @@ cpu_switchto(struct proc *old, struct proc *new)
                if ((tf->tf_sstatus & SSTATUS_FS_MASK) == SSTATUS_FS_DIRTY) {
                        fpu_save(old, tf);
                }
+               tf->tf_sstatus &= ~SSTATUS_FS_MASK;
        }
 
        cpu_switchto_asm(old, new);
 
        pcb = ci->ci_curpcb;
        tf = new->p_addr->u_pcb.pcb_tf;
+#if 0
+       /* XXX this optimization is subtly broken */    
        if (pcb->pcb_fpcpu == ci && ci->ci_fpuproc == new) {
                /* If fpu state is already loaded, allow it to be used */
                tf->tf_sstatus |= SSTATUS_FS_CLEAN;
        }
+#endif
 }
 
 /*