This patch slightly optimizes hyperprivop emulation
especially hyper_rfi.
It shows about 2% faster in fstat system call on dom0.
Signed-off-by: Kouya Shimura <kouya@xxxxxxxxxxxxxx>
diff -r 71a8366fb212 xen/arch/ia64/xen/hyperprivop.S
--- a/xen/arch/ia64/xen/hyperprivop.S Fri Feb 29 09:19:58 2008 -0700
+++ b/xen/arch/ia64/xen/hyperprivop.S Fri Mar 07 17:18:44 2008 +0900
@@ -67,19 +67,18 @@
// r19 == ipsr.cpl
// r31 == pr
GLOBAL_ENTRY(fast_hyperprivop)
+ adds r20=XSI_PSR_I_ADDR_OFS-XSI_PSR_IC_OFS,r18
// HYPERPRIVOP_SSM_I?
// assumes domain interrupts pending, so just do it
cmp.eq p7,p6=HYPERPRIVOP_SSM_I,r17
(p7) br.sptk.many hyper_ssm_i;;
// Check pending event indication
- adds r20=XSI_PSR_I_ADDR_OFS-XSI_PSR_IC_OFS, r18;;
- ld8 r20=[r20]
+ ld8 r20=[r20] // interrupt_mask_addr
;;
ld1 r22=[r20],-1 // evtchn_upcall_mask
;;
ld1 r20=[r20] // evtchn_upcall_pending
- ;;
// HYPERPRIVOP_RFI?
cmp.eq p7,p6=HYPERPRIVOP_RFI,r17
@@ -210,9 +209,8 @@ ENTRY(hyper_ssm_i)
// give up for now if: ipsr.be==1, ipsr.pp==1
mov r30=cr.ipsr
mov r29=cr.iip;;
- extr.u r21=r30,IA64_PSR_PP_BIT,1 ;;
- cmp.ne p7,p0=r21,r0
-(p7) br.sptk.many dispatch_break_fault ;;
+ tbit.nz p7,p0=r30,IA64_PSR_PP_BIT
+(p7) br.spnt.many dispatch_break_fault ;;
#ifdef FAST_HYPERPRIVOP_CNT
movl r20=FAST_HYPERPRIVOP_PERFC(HYPERPRIVOP_SSM_I);;
ld4 r21=[r20];;
@@ -220,8 +218,7 @@ ENTRY(hyper_ssm_i)
st4 [r20]=r21;;
#endif
// set shared_mem iip to instruction after HYPER_SSM_I
- extr.u r20=r30,IA64_PSR_RI_BIT,2 ;;
- cmp.eq p6,p7=2,r20 ;;
+ tbit.nz p6,p7=r30,IA64_PSR_RI_BIT+1 ;; // cr.ipsr.ri >= 2 ?
(p6) mov r20=0
(p6) adds r29=16,r29
(p7) adds r20=1,r20 ;;
@@ -346,8 +343,7 @@ GLOBAL_ENTRY(fast_tick_reflect)
(p6) br.cond.spnt.few rp;;
mov r17=cr.ipsr;;
// slow path if: ipsr.pp==1
- extr.u r21=r17,IA64_PSR_PP_BIT,1 ;;
- cmp.ne p6,p0=r21,r0
+ tbit.nz p6,p0=r17,IA64_PSR_PP_BIT
(p6) br.cond.spnt.few rp;;
// definitely have a domain tick
mov cr.eoi=r0
@@ -537,8 +533,7 @@ GLOBAL_ENTRY(fast_break_reflect)
#endif
mov r30=cr.ipsr
mov r29=cr.iip;;
- extr.u r21=r30,IA64_PSR_PP_BIT,1 ;;
- cmp.ne p7,p0=r21,r0
+ tbit.nz p7,p0=r30,IA64_PSR_PP_BIT
(p7) br.spnt.few dispatch_break_fault ;;
movl r20=IA64_PSR_CPL ;;
and r22=r20,r30 ;;
@@ -722,8 +717,7 @@ GLOBAL_ENTRY(fast_access_reflect)
#endif
mov r30=cr.ipsr
mov r29=cr.iip;;
- extr.u r21=r30,IA64_PSR_PP_BIT,1 ;;
- cmp.ne p7,p0=r21,r0
+ tbit.nz p7,p0=r30,IA64_PSR_PP_BIT
(p7) br.spnt.few dispatch_reflection ;;
extr.u r21=r30,IA64_PSR_CPL0_BIT,2 ;;
cmp.eq p7,p0=r21,r0
@@ -769,8 +763,7 @@ GLOBAL_ENTRY(fast_tlb_miss_reflect)
cmp.eq p7,p0=r21,r0
(p7) br.spnt.few page_fault ;;
// slow path if strange ipsr or isr bits set
- extr.u r21=r30,IA64_PSR_PP_BIT,1 ;;
- cmp.ne p7,p0=r21,r0
+ tbit.nz p7,p0=r30,IA64_PSR_PP_BIT,1
(p7) br.spnt.few page_fault ;;
movl r21=IA64_ISR_IR|IA64_ISR_SP|IA64_ISR_NA ;;
and r21=r16,r21;;
@@ -1023,45 +1016,27 @@ ENTRY(hyper_rfi)
#ifndef FAST_RFI
br.spnt.few slow_vcpu_rfi ;;
#endif
- // if no interrupts pending, proceed
- mov r30=r0
- cmp.eq p7,p0=r20,r0
-(p7) br.sptk.many 1f
- ;;
- adds r20=XSI_IPSR_OFS-XSI_PSR_IC_OFS,r18 ;;
- ld8 r21=[r20];; // r21 = vcr.ipsr
- extr.u r22=r21,IA64_PSR_I_BIT,1 ;;
- mov r30=r22;;
- // r30 determines whether we might deliver an immediate extint
-#ifndef RFI_TO_INTERRUPT // see beginning of file
- cmp.ne p6,p0=r30,r0
-(p6) br.cond.spnt.few slow_vcpu_rfi ;;
-#endif
-1:
- adds r20=XSI_IPSR_OFS-XSI_PSR_IC_OFS,r18 ;;
- ld8 r21=[r20];; // r21 = vcr.ipsr
+ // if interrupts pending and vcr.ipsr.i=1, do it the slow way
+ adds r19=XSI_IPSR_OFS-XSI_PSR_IC_OFS,r18
+ adds r23=XSI_METAPHYS_OFS-XSI_PSR_IC_OFS,r18
+ cmp.ne p8,p0=r20,r0;; // evtchn_upcall_pending != 0
// if (!(vpsr.dt && vpsr.rt && vpsr.it)), do it the slow way
- movl r20=(IA64_PSR_DT|IA64_PSR_RT|IA64_PSR_IT);;
- and r22=r20,r21
- ;;
- cmp.ne p7,p0=r22,r20
-(p7) br.spnt.few slow_vcpu_rfi ;;
+ ld8 r21=[r19],XSI_IIP_OFS-XSI_IPSR_OFS // r21=vcr.ipsr
+ movl r20=~(IA64_PSR_DT|IA64_PSR_RT|IA64_PSR_IT);;
+ or r20=r20,r21
+ // p8 determines whether we might deliver an immediate extint
+(p8) tbit.nz p8,p0=r21,IA64_PSR_I_BIT;;
+ cmp.ne p7,p0=-1,r20
+ ld4 r23=[r23] // r23=metaphysical_mode
+#ifndef RFI_TO_INTERRUPT // see beginning of file
+(p8) br.cond.spnt.few slow_vcpu_rfi
+#endif
+(p7) br.spnt.few slow_vcpu_rfi;;
// if was in metaphys mode, do it the slow way (FIXME later?)
- adds r20=XSI_METAPHYS_OFS-XSI_PSR_IC_OFS,r18 ;;
- ld4 r20=[r20];;
- cmp.ne p7,p0=r20,r0
-(p7) br.spnt.few slow_vcpu_rfi ;;
-#if 0
- // if domain hasn't already done virtual bank switch
- // do it the slow way (FIXME later?)
- adds r20=XSI_BANKNUM_OFS-XSI_PSR_IC_OFS,r18 ;;
- ld4 r20=[r20];;
- cmp.eq p7,p0=r20,r0
-(p7) br.spnt.few slow_vcpu_rfi ;;
-#endif
- adds r20=XSI_IIP_OFS-XSI_PSR_IC_OFS,r18 ;;
- ld8 r22=[r20];;
-1: // OK now, let's do an rfi.
+ cmp.ne p7,p0=r23,r0
+ ld8 r22=[r19] // r22=vcr.iip
+(p7) br.spnt.few slow_vcpu_rfi;;
+ // OK now, let's do an rfi.
#ifdef FAST_HYPERPRIVOP_CNT
movl r20=FAST_HYPERPRIVOP_PERFC(HYPERPRIVOP_RFI);;
ld4 r23=[r20];;
@@ -1070,8 +1045,7 @@ 1: // OK now, let's do an rfi.
#endif
#ifdef RFI_TO_INTERRUPT
// maybe do an immediate interrupt delivery?
- cmp.ne p6,p0=r30,r0
-(p6) br.cond.spnt.few rfi_check_extint;;
+(p8) br.cond.spnt.few rfi_check_extint;;
#endif
just_do_rfi:
@@ -1090,15 +1064,13 @@ just_do_rfi:
// vpsr.i = vcr.ipsr.i; vpsr.ic = vcr.ipsr.ic
ld8 r20=[r20]
mov r19=1
- extr.u r23=r21,IA64_PSR_I_BIT,1 ;;
- cmp.ne p7,p6=r23,r0 ;;
+ tbit.nz p7,p6=r21,IA64_PSR_I_BIT
+ tbit.nz p9,p8=r21,IA64_PSR_IC_BIT;;
// not done yet
(p7) st1 [r20]=r0
-(p6) st1 [r20]=r19;;
- extr.u r23=r21,IA64_PSR_IC_BIT,1 ;;
- cmp.ne p7,p6=r23,r0 ;;
-(p7) st4 [r18]=r19;;
-(p6) st4 [r18]=r0;;
+(p6) st1 [r20]=r19
+(p9) st4 [r18]=r19
+(p8) st4 [r18]=r0
// force on psr.ic, i, dt, rt, it, bn
movl r20=(IA64_PSR_I|IA64_PSR_IC|IA64_PSR_DT|IA64_PSR_RT| \
IA64_PSR_IT|IA64_PSR_BN)
@@ -1228,9 +1200,8 @@ ENTRY(rfi_check_extint)
// r26 now contains the vector [0..255]
adds r20=XSI_TPR_OFS-XSI_PSR_IC_OFS,r18 ;;
ld8 r20=[r20] ;;
- extr.u r28=r20,16,1
- extr.u r29=r20,4,4 ;;
- cmp.ne p6,p0=r28,r0 // if tpr.mmi is set, just rfi
+ extr.u r29=r20,4,4
+ tbit.nz p6,p0=r20,16 // if tpr.mmi is set, just rfi
(p6) br.cond.spnt.few just_do_rfi;;
shl r29=r29,4;;
adds r29=15,r29;;
_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel
|
Home