[IA64] Reschedule break_fault() for better performance.

This patch reorganizes break_fault() to optimistically assume that a
system-call is being performed from user-space (which is almost always
the case).  If it turns out that (a) we're not being called due to a
system call or (b) we're being called from within the kernel, we fixup
the no-longer-valid assumptions in non_syscall() and .break_fixup(),
respectively.

With this approach, there are 3 major phases:

 - Phase 1: Read various control & application registers, in
	    particular the current task pointer from AR.K6.
 - Phase 2: Do all memory loads (load system-call entry,
	    load current_thread_info()->flags, prefetch
	    kernel register-backing store) and switch
	    to kernel register-stack.
 - Phase 3: Call ia64_syscall_setup() and invoke
	    syscall-handler.

Good for 26-30 cycles of improvement on break-based syscall-path.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>

arch/ia64/kernel/ivt.S

@@ -687,82 +687,118 @@ ENTRY(break_fault)
 	 * to prevent leaking bits from kernel to user level.
 	 */
 	DBG_FAULT(11)
-	mov r16=IA64_KR(CURRENT)		// r16 = current task; 12 cycle read lat.
-	mov r17=cr.iim
-	mov r18=__IA64_BREAK_SYSCALL
-	mov r21=ar.fpsr
-	mov r29=cr.ipsr
-	mov r19=b6
-	mov r25=ar.unat
-	mov r27=ar.rsc
-	mov r26=ar.pfs
-	mov r28=cr.iip
-	mov r31=pr				// prepare to save predicates
-	mov r20=r1
-	;;
+	mov.m r16=IA64_KR(CURRENT)		// M2 r16 <- current task (12 cyc)
+	mov r29=cr.ipsr				// M2 (12 cyc)
+	mov r31=pr				// I0 (2 cyc)
+
+	mov r17=cr.iim				// M2 (2 cyc)
+	mov.m r27=ar.rsc			// M2 (12 cyc)
+	mov r18=__IA64_BREAK_SYSCALL		// A
+
+	mov.m ar.rsc=0				// M2
+	mov.m r21=ar.fpsr			// M2 (12 cyc)
+	mov r19=b6				// I0 (2 cyc)
+	;;
+	mov.m r23=ar.bspstore			// M2 (12 cyc)
+	mov.m r24=ar.rnat			// M2 (5 cyc)
+	mov.i r26=ar.pfs			// I0 (2 cyc)
+
+	invala					// M0|1
+	nop.m 0					// M
+	mov r20=r1				// A			save r1
+
+	nop.m 0
+	movl r30=sys_call_table			// X
+
+	mov r28=cr.iip				// M2 (2 cyc)
+	cmp.eq p0,p7=r18,r17			// I0 is this a system call?
+(p7)	br.cond.spnt non_syscall		// B  no ->
+	//
+	// From this point on, we are definitely on the syscall-path
+	// and we can use (non-banked) scratch registers.
+	//
+///////////////////////////////////////////////////////////////////////
+	mov r1=r16				// A    move task-pointer to "addl"-addressable reg
+	mov r2=r16				// A    setup r2 for ia64_syscall_setup
+	add r9=TI_FLAGS+IA64_TASK_SIZE,r16	// A	r9 = &current_thread_info()->flags
+
 	adds r16=IA64_TASK_THREAD_ON_USTACK_OFFSET,r16
-	cmp.eq p0,p7=r18,r17			// is this a system call? (p7 <- false, if so)
-(p7)	br.cond.spnt non_syscall
+	adds r15=-1024,r15			// A    subtract 1024 from syscall number
+	mov r3=NR_syscalls - 1
 	;;
-	ld1 r17=[r16]				// load current->thread.on_ustack flag
-	st1 [r16]=r0				// clear current->thread.on_ustack flag
-	add r1=-IA64_TASK_THREAD_ON_USTACK_OFFSET,r16	// set r1 for MINSTATE_START_SAVE_MIN_VIRT
+	ld1.bias r17=[r16]			// M0|1 r17 = current->thread.on_ustack flag
+	ld4 r9=[r9]				// M0|1 r9 = current_thread_info()->flags
+	extr.u r8=r29,41,2			// I0   extract ei field from cr.ipsr
+
+	shladd r30=r15,3,r30			// A    r30 = sys_call_table + 8*(syscall-1024)
+	addl r22=IA64_RBS_OFFSET,r1		// A    compute base of RBS
+	cmp.leu p6,p7=r15,r3			// A    syscall number in range?
 	;;
-	invala
 
-	/* adjust return address so we skip over the break instruction: */
+	lfetch.fault.excl.nt1 [r22]		// M0|1 prefetch RBS
+(p6)	ld8 r30=[r30]				// M0|1 load address of syscall entry point
+	tnat.nz.or p7,p0=r15			// I0	is syscall nr a NaT?
 
-	extr.u r8=r29,41,2			// extract ei field from cr.ipsr
-	;;
-	cmp.eq p6,p7=2,r8			// isr.ei==2?
-	mov r2=r1				// setup r2 for ia64_syscall_setup
-	;;
-(p6)	mov r8=0				// clear ei to 0
-(p6)	adds r28=16,r28				// switch cr.iip to next bundle cr.ipsr.ei wrapped
-(p7)	adds r8=1,r8				// increment ei to next slot
-	;;
-	cmp.eq pKStk,pUStk=r0,r17		// are we in kernel mode already?
-	dep r29=r8,r29,41,2			// insert new ei into cr.ipsr
+	mov.m ar.bspstore=r22			// M2   switch to kernel RBS
+	cmp.eq p8,p9=2,r8			// A    isr.ei==2?
 	;;
 
-	// switch from user to kernel RBS:
-	MINSTATE_START_SAVE_MIN_VIRT
-	br.call.sptk.many b7=ia64_syscall_setup
-	;;
-	MINSTATE_END_SAVE_MIN_VIRT		// switch to bank 1
-	ssm psr.ic | PSR_DEFAULT_BITS
-	;;
-	srlz.i					// guarantee that interruption collection is on
-	mov r3=NR_syscalls - 1
-	;;
-(p15)	ssm psr.i				// restore psr.i
-	// p10==true means out registers are more than 8 or r15's Nat is true
-(p10)	br.cond.spnt.many ia64_ret_from_syscall
-	;;
-	movl r16=sys_call_table
+(p8)	mov r8=0				// A    clear ei to 0
+(p7)	movl r30=sys_ni_syscall			// X
 
-	adds r15=-1024,r15			// r15 contains the syscall number---subtract 1024
-	movl r2=ia64_ret_from_syscall
-	;;
-	shladd r20=r15,3,r16			// r20 = sys_call_table + 8*(syscall-1024)
-	cmp.leu p6,p7=r15,r3			// (syscall > 0 && syscall < 1024 + NR_syscalls) ?
-	mov rp=r2				// set the real return addr
+(p8)	adds r28=16,r28				// A    switch cr.iip to next bundle
+(p9)	adds r8=1,r8				// A    increment ei to next slot
+	nop.i 0
 	;;
-(p6)	ld8 r20=[r20]				// load address of syscall entry point
-(p7)	movl r20=sys_ni_syscall
 
-	add r2=TI_FLAGS+IA64_TASK_SIZE,r13
-	;;
-	ld4 r2=[r2]				// r2 = current_thread_info()->flags
-	;;
-	and r2=_TIF_SYSCALL_TRACEAUDIT,r2	// mask trace or audit
+	mov.m r25=ar.unat			// M2 (5 cyc)
+	dep r29=r8,r29,41,2			// I0   insert new ei into cr.ipsr
+	adds r15=1024,r15			// A    restore original syscall number
+	//
+	// If any of the above loads miss in L1D, we'll stall here until
+	// the data arrives.
+	//
+///////////////////////////////////////////////////////////////////////
+	st1 [r16]=r0				// M2|3 clear current->thread.on_ustack flag
+	mov b6=r30				// I0   setup syscall handler branch reg early
+	cmp.eq pKStk,pUStk=r0,r17		// A    were we on kernel stacks already?
+
+	and r9=_TIF_SYSCALL_TRACEAUDIT,r9	// A    mask trace or audit
+	mov r18=ar.bsp				// M2 (12 cyc)
+(pKStk)	br.cond.spnt .break_fixup		// B	we're already in kernel-mode -- fix up RBS
+	;;
+.back_from_break_fixup:
+(pUStk)	addl r1=IA64_STK_OFFSET-IA64_PT_REGS_SIZE,r1 // A    compute base of memory stack
+	cmp.eq p14,p0=r9,r0			// A    are syscalls being traced/audited?
+	br.call.sptk.many b7=ia64_syscall_setup	// B
+1:
+	mov ar.rsc=0x3				// M2   set eager mode, pl 0, LE, loadrs=0
+	nop 0
+	bsw.1					// B (6 cyc) regs are saved, switch to bank 1
 	;;
-	cmp.eq p8,p0=r2,r0
-	mov b6=r20
+
+	ssm psr.ic | PSR_DEFAULT_BITS		// M2	now it's safe to re-enable intr.-collection
+	movl r3=ia64_ret_from_syscall		// X
 	;;
-(p8)	br.call.sptk.many b6=b6			// ignore this return addr
-	br.cond.sptk ia64_trace_syscall
+
+	srlz.i					// M0   ensure interruption collection is on
+	mov rp=r3				// I0   set the real return addr
+(p10)	br.cond.spnt.many ia64_ret_from_syscall	// B    return if bad call-frame or r15 is a NaT
+
+(p15)	ssm psr.i				// M2   restore psr.i
+(p14)	br.call.sptk.many b6=b6			// B    invoke syscall-handker (ignore return addr)
+	br.cond.spnt.many ia64_trace_syscall	// B	do syscall-tracing thingamagic
 	// NOT REACHED
+///////////////////////////////////////////////////////////////////////
+	// On entry, we optimistically assumed that we're coming from user-space.
+	// For the rare cases where a system-call is done from within the kernel,
+	// we fix things up at this point:
+.break_fixup:
+	add r1=-IA64_PT_REGS_SIZE,sp		// A    allocate space for pt_regs structure
+	mov ar.rnat=r24				// M2	restore kernel's AR.RNAT
+	;;
+	mov ar.bspstore=r23			// M2	restore kernel's AR.BSPSTORE
+	br.cond.sptk .back_from_break_fixup
 END(break_fault)
 
 	.org ia64_ivt+0x3000
@@ -837,8 +873,6 @@ END(interrupt)
 	 *	- r31: saved pr
 	 *	-  b0: original contents (to be saved)
 	 * On exit:
-	 *	- executing on bank 1 registers
-	 *	- psr.ic enabled, interrupts restored
 	 *	-  p10: TRUE if syscall is invoked with more than 8 out
 	 *		registers or r15's Nat is true
 	 *	-  r1: kernel's gp
@@ -846,8 +880,11 @@ END(interrupt)
 	 *	-  r8: -EINVAL if p10 is true
 	 *	- r12: points to kernel stack
 	 *	- r13: points to current task
+	 *	- r14: preserved (same as on entry)
+	 *	- p13: preserved
 	 *	- p15: TRUE if interrupts need to be re-enabled
 	 *	- ar.fpsr: set to kernel settings
+	 *	-  b6: preserved (same as on entry)
 	 */
 GLOBAL_ENTRY(ia64_syscall_setup)
 #if PT(B6) != 0
@@ -915,7 +952,7 @@ GLOBAL_ENTRY(ia64_syscall_setup)
 (p13)	mov in5=-1
 	;;
 	st8 [r16]=r21,PT(R8)-PT(AR_FPSR)	// save ar.fpsr
-	tnat.nz p14,p0=in6
+	tnat.nz p13,p0=in6
 	cmp.lt p10,p9=r11,r8	// frame size can't be more than local+8
 	;;
 	mov r8=1
@@ -930,7 +967,7 @@ GLOBAL_ENTRY(ia64_syscall_setup)
 	movl r1=__gp				// establish kernel global pointer
 	;;
 	st8 [r16]=r8		// ensure pt_regs.r8 != 0 (see handle_syscall_error)
-(p14)	mov in6=-1
+(p13)	mov in6=-1
 (p8)	mov in7=-1
 
 	cmp.eq pSys,pNonSys=r0,r0		// set pSys=1, pNonSys=0
@@ -1002,6 +1039,8 @@ END(dispatch_illegal_op_fault)
 	FAULT(17)
 
 ENTRY(non_syscall)
+	mov ar.rsc=r27			// restore ar.rsc before SAVE_MIN_WITH_COVER
+	;;
 	SAVE_MIN_WITH_COVER
 
 	// There is no particular reason for this code to be here, other than that