Index: sparc/genassym.cf
===================================================================
RCS file: /cvsroot/src/sys/arch/sparc/sparc/genassym.cf,v
retrieving revision 1.60
diff -p -r1.60 genassym.cf
*** sparc/genassym.cf	3 Jan 2010 11:44:58 -0000	1.60
--- sparc/genassym.cf	6 Jan 2010 01:30:14 -0000
*************** define	L_STAT		offsetof(struct lwp, l_st
*** 112,117 ****
--- 112,118 ----
  define	L_WCHAN		offsetof(struct lwp, l_wchan)
  define	L_CPU		offsetof(struct lwp, l_cpu)
  define	L_PRIORITY	offsetof(struct lwp, l_priority)
+ define  L_FPSTATE	offsetof(struct lwp, l_md.md_fpstate)
  define	P_VMSPACE	offsetof(struct proc, p_vmspace)
  define	P_MD_SYSCALL	offsetof(struct proc, p_md.md_syscall)
  define	LSRUN		LSRUN
*************** define	CPUINFO_IDLESPIN	offsetof(struct 
*** 166,171 ****
--- 167,173 ----
  define	CPUINFO_MTX_COUNT	offsetof(struct cpu_info, ci_mtx_count)
  define	CPUINFO_MTX_OLDSPL	offsetof(struct cpu_info, ci_mtx_oldspl)
  define	CPUINFO_IDEPTH		offsetof(struct cpu_info, ci_idepth)
+ define	CPUINFO_FPLWP		offsetof(struct cpu_info, fplwp)
  
  # PTE bits and related information
  define	PG_W		PG_W
Index: sparc/locore.s
===================================================================
RCS file: /cvsroot/src/sys/arch/sparc/sparc/locore.s,v
retrieving revision 1.254
diff -p -r1.254 locore.s
*** sparc/locore.s	4 Jan 2010 04:06:57 -0000	1.254
--- sparc/locore.s	6 Jan 2010 01:30:14 -0000
*************** Lfp_storeq:
*** 5921,5926 ****
--- 5921,5954 ----
  	 srl	%o3, 3, %o3		! (but first fix qsize)
  
  /*
+  * Entry point for savefpstate IPI's.  We skip the IPI if the fplwp is not
+  * the same as this LWP.  Since we zero cpuinfo.fplwp here, we call
+  * savefpstate() instead of inlining savefpstate() here, since we'd need to
+  * replicate the "special_fp_store" handling.
+  *
+  * %o0 = struct lwp *
+  */
+ #ifdef MULTIPROCESSOR
+ ENTRY(ipi_savefpstate)
+ 	sethi	%hi(CPUINFO_VA), %o1
+ 	ld	[%o1 + CPUINFO_FPLWP], %o2
+ 	cmp	%o2, %o0
+ 	bne	savefpstate
+ 	 nop
+ 
+ 	call	savefpstate
+ 	 ld	[%o2 + L_FPSTATE], %i0
+ 
+ 	! now zero cpuinfo.fplwp
+ 	sethi	%hi(CPUINFO_VA), %o1
+ 	st	%g0, [%o1 + CPUINFO_FPLWP]
+ 
+ 1:
+ 	retl
+ 	 nop
+ #endif
+ 
+ /*
   * The fsr store trapped.  Do it again; this time it will not trap.
   * We could just have the trap handler return to the `st %fsr', but
   * if for some reason it *does* trap, that would lock us into a tight
Index: sparc/machdep.c
===================================================================
RCS file: /cvsroot/src/sys/arch/sparc/sparc/machdep.c,v
retrieving revision 1.300
diff -p -r1.300 machdep.c
*** sparc/machdep.c	3 Jan 2010 23:03:21 -0000	1.300
--- sparc/machdep.c	6 Jan 2010 01:30:14 -0000
*************** setregs(struct lwp *l, struct exec_packa
*** 381,389 ****
  				savefpstate(fs);
  #if defined(MULTIPROCESSOR)
  			else
! 				XCALL1(savefpstate, fs, 1 << cpi->ci_cpuid);
  #endif
- 			cpi->fplwp = NULL;
  		}
  		l->l_md.md_fpu = NULL;
  		FPU_UNLOCK(s);
--- 381,388 ----
  				savefpstate(fs);
  #if defined(MULTIPROCESSOR)
  			else
! 				XCALL1(ipi_savefpstate, l, 1 << cpi->ci_cpuid);
  #endif
  		}
  		l->l_md.md_fpu = NULL;
  		FPU_UNLOCK(s);
Index: sparc/trap.c
===================================================================
RCS file: /cvsroot/src/sys/arch/sparc/sparc/trap.c,v
retrieving revision 1.179
diff -p -r1.179 trap.c
*** sparc/trap.c	21 Nov 2009 04:16:52 -0000	1.179
--- sparc/trap.c	6 Jan 2010 01:30:14 -0000
*************** badtrap:
*** 478,486 ****
  					panic("FPU(%d): state for %p",
  							cpi->ci_cpuid, l);
  #if defined(MULTIPROCESSOR)
! 				XCALL1(savefpstate, fs, 1 << cpi->ci_cpuid);
  #endif
- 				cpi->fplwp = NULL;
  			}
  			loadfpstate(fs);
  
--- 478,485 ----
  					panic("FPU(%d): state for %p",
  							cpi->ci_cpuid, l);
  #if defined(MULTIPROCESSOR)
! 				XCALL1(ipi_savefpstate, l, 1 << cpi->ci_cpuid);
  #endif
  			}
  			loadfpstate(fs);
  
Index: sparc/vm_machdep.c
===================================================================
RCS file: /cvsroot/src/sys/arch/sparc/sparc/vm_machdep.c,v
retrieving revision 1.101
diff -p -r1.101 vm_machdep.c
*** sparc/vm_machdep.c	21 Nov 2009 04:16:52 -0000	1.101
--- sparc/vm_machdep.c	6 Jan 2010 01:30:14 -0000
*************** cpu_lwp_fork(struct lwp *l1, struct lwp 
*** 228,235 ****
  				savefpstate(l1->l_md.md_fpstate);
  #if defined(MULTIPROCESSOR)
  			else
! 				XCALL1(savefpstate, l1->l_md.md_fpstate,
! 					1 << cpi->ci_cpuid);
  #endif
  		}
  		memcpy(l2->l_md.md_fpstate, l1->l_md.md_fpstate,
--- 228,234 ----
  				savefpstate(l1->l_md.md_fpstate);
  #if defined(MULTIPROCESSOR)
  			else
! 				XCALL1(ipi_savefpstate, l1, 1 << cpi->ci_cpuid);
  #endif
  		}
  		memcpy(l2->l_md.md_fpstate, l1->l_md.md_fpstate,
*************** cpu_lwp_free(struct lwp *l, int proc)
*** 306,312 ****
  				savefpstate(fs);
  #if defined(MULTIPROCESSOR)
  			else
! 				XCALL1(savefpstate, fs, 1 << cpi->ci_cpuid);
  #endif
  			cpi->fplwp = NULL;
  		}
--- 305,311 ----
  				savefpstate(fs);
  #if defined(MULTIPROCESSOR)
  			else
! 				XCALL1(ipi_savefpstate, l, 1 << cpi->ci_cpuid);
  #endif
  			cpi->fplwp = NULL;
  		}