ARM1136JF-S™ and ARM1136J-S™ Technical Reference Manual 3.3.39. c15, Performance Monitor Control Register (PMNC)

3.3.39. c15, Performance Monitor Control Register (PMNC)

The purpose of the Performance Monitor Control Register is to control the operation of:

the Cycle Counter Register (CCNT)
the Count Register 0 (PMN0)
the Count Register 1 (PMN1).

The Performance Monitor Control Register is:

in CP15 c15
a 32-bit read/write register
accessible in User and privileged modes.

Figure 3.62 shows the arrangement of bits in the register.

Figure 3.62. Performance Monitor Control Register format

Table 3.138 shows the bit functions of the Performance Monitor Control Register.

Table 3.138. Performance Monitor Control Register bit functions

Bit range	Field name	Function
[31:28]	-	UNP/SBZ.
[27:20]	EvtCount0	Identifies the source of events for Count Register 0, as defined in Table 3.141.
[19:12]	EvtCount1	Identifies the source of events for Count Register 1, as defined in Table 3.141.
[11]	X	Enable Export of the events to the event bus. This enables an external monitoring block, such as the ETM to trace events: 0 = Export disabled, EVNTBUS held at `0x0`. This is the reset value. 1 = Export enabled, EVNTBUS driven by the events.
[10]	CCR	Cycle Count Register overflow flag. See Table 3.139 for the meaning of the flag values.
[9]	CR1	Count Register 1 overflow flag. See Table 3.139 for the meaning of the flag values.
[8]	CR0	Count Register 0 overflow flag. See Table 3.139 for the meaning of the flag values.
[7]	-	UNP/SBZ.
[6]	ECC	Enable Cycle Counter interrupt. 0 = Disable interrupt. This is the reset value. 1 = Enable interrupt.
[5]	EC1	Enable Counter Register 1 interrupt. 0 = Disable interrupt. This is the reset value. 1 = Enable interrupt.
[4]	EC0	Enable Counter Register 0 interrupt. 0 = Disable interrupt. This is the reset value. 1 = Enable interrupt.
[3]	D	Cycle count divider: 0 = Cycle Counter Register counts every processor clock cycle. 1 = Cycle Counter Register counts every 64th processor clock cycle.
[2]	C	Cycle Counter Register Reset on Write, UNP on Read: Write 0 = no action Write 1 = reset the Cycle Counter Register to `0x0`.
[1]	P	Count Register Reset on Write, UNP on Read: Write 0 = no action Write 1 = reset both Count Registers to `0x0`.
[0]	E	Enable: 0 = all three counters disabled 1 = all three counters enabled. The PMUIRQ signal can only be cleared when this bit is set to 1. This signal is the mechanism by which a Performance Monitor Unit interrupt is signaled to the core, see Using the Performance Monitor Control Register for more information.

The meaning of the flags in the Performance Monitor Control Register, for read and write operations, are shown in Table 3.139:

Table 3.139. PMNC flag values

Flag value	On reads	On write
0	No overflow has occurred. This is the reset value.	No effect.
1	An overflow has occurred.	Clear this bit.

Accessing the Performance Monitor Control Register

Table 3.140 shows the results of attempted accesses to the Performance Monitor Control Register for each mode.

Table 3.140. Results of accesses to the Performance Monitor Control Register

Privileged read	Privileged write	User read or write
Data read	Data write	Undefined exception

To access the Performance Monitor Control Register you read or write CP15 with:

Opcode_1 set to 0
CRn set to c15
CRm set to c12
Opcode_2 set to 0.

For example:

MRC p15, 0, <Rd>, c15, c12, 0       ; Read Performance Monitor Control Register

MCR p15, 0, <Rd>, c15, c12, 0       ; Write Performance Monitor Control Register

Using the Performance Monitor Control Register

The Performance Monitor Control Register:

controls which events PMN0 and PMN1 monitor
detects which counter overflowed
enables and disables interrupt reporting
extends CCNT counting by six more bits (cycles between counter rollover = 2³⁸)
resets all counters to zero
enables the entire performance monitoring mechanism.

If an interrupt is generated by this unit, the ARM1136JF-S processor pin PMUIRQ is asserted. This output pin can then be routed to an external interrupt controller for prioritization and masking. This is the only mechanism by which the interrupt is signaled to the core. When asserted, the PMUIRQ signal can only be cleared if bit[0] of the register, the E bit, is HIGH.

There is a delay of three cycles between enabling the counter and the counter starting to count events. In addition, the information used to count events is taken from various pipeline stages. This means that the absolute counts recorded might vary because of pipeline effects. This has a negligible effect except in case where the counters are enabled for a very short time.

Table 3.141 shows the events that can be monitored using the Performance Monitor Control Register.

Table 3.141. Performance monitoring events

Event number	EVNTBUS bit position	Event definition
`0x0`	[0]	Instruction cache miss to a cachable location requires fetch from external memory.
`0x1`	[1]	Stall because instruction buffer cannot deliver an instruction. This could indicate an Instruction Cache miss or an Instruction MicroTLB miss. This event occurs every cycle in which the condition is present.
`0x2`	[2]	Stall because of a data dependency. This event occurs every cycle in which the condition is present.
`0x3`	[3]	Instruction MicroTLB miss.
`0x4`	[4]	Data MicroTLB miss.
`0x5`	[6:5]	Branch instruction executed, branch might or might not have changed program flow.
`0x6`	[7]	Branch mis-predicted.
`0x7`	[9:8]	Instruction executed.
`0x9`	[10]	Data cache access, not including Cache operations. This event occurs for each nonsequential access to a cache line, for cachable locations.
`0xA`	[11]	Data cache access, not including Cache Operations. This event occurs for each nonsequential access to a cache line, regardless of whether or not the location is cachable.
`0xB`	[12]	Data cache miss, not including Cache Operations.
`0xC`	[13]	Data cache write-back. This event occurs once for each half line of four words that are written back from the cache.
`0xD`	[15:14]	Software changed the PC. This event occurs any time the PC is changed by software and there is not a mode change. For example, a MOV instruction with PC as the destination triggers this event. Executing a SWI from User mode does not trigger this event, because it incurs a mode change.
`0xF`	[16]	Main TLB miss.
`0x10`	[17]	Explicit external data access. This includes Cache Refill, Noncachable and Write-Through accesses. It does not include Write-Backs, instruction cache line fills, and page table walks.
`0x11`	[18]	Stall because of Load Store Unit request queue being full. This event occurs each clock cycle in which the condition is met. A high incidence of this event indicates the BCU is often waiting for transactions to complete on the external bus.
`0x12`	[19]	The number of times the Write Buffer was drained because of a Data Synchronization Barrier command or Strongly Ordered operation.
`0x20`	-	ETMEXTOUT[0] signal was asserted for a cycle.
`0x21`	-	ETMEXTOUT[1] signal was asserted for a cycle.
`0x22`	-	If both ETMEXTOUT[0] and ETMEXTOUT[1] signals are asserted then the count is incremented by two.
`0xFF`	-	An increment each cycle.
All other values	-	Reserved. Unpredictable behavior.

In addition to the two counters within ARM1136JF-S processors, each of the events shown in Table 3.141 is available on an external bus, EVNTBUS. You can connect this bus to the ETM unit or other external trace hardware to enable the events to be monitored. If this functionality is not required, you must set the X bit in the Performance Monitor Control Register to the 0.

Copyright © 2002-2005 ARM Limited. All rights reserved.	ARM DDI 0211H
Non-Confidential
PDF version