freebsd-skq/lib/libpmc/pmc.mips.3

.\" Copyright (c) 2010 George Neville-Neil.  All rights reserved.
.\"
.\" Redistribution and use in source and binary forms, with or without
.\" modification, are permitted provided that the following conditions
.\" are met:
.\" 1. Redistributions of source code must retain the above copyright
.\"    notice, this list of conditions and the following disclaimer.
.\" 2. Redistributions in binary form must reproduce the above copyright
.\"    notice, this list of conditions and the following disclaimer in the
.\"    documentation and/or other materials provided with the distribution.
.\"
.\" This software is provided by ``as is'' and
.\" any express or implied warranties, including, but not limited to, the
.\" implied warranties of merchantability and fitness for a particular purpose
.\" are disclaimed.  in no event shall George Neville-Neil be liable
.\" for any direct, indirect, incidental, special, exemplary, or consequential
.\" damages (including, but not limited to, procurement of substitute goods
.\" or services; loss of use, data, or profits; or business interruption)
.\" however caused and on any theory of liability, whether in contract, strict
.\" liability, or tort (including negligence or otherwise) arising in any way
.\" out of the use of this software, even if advised of the possibility of
.\" such damage.
.\"
.\" $FreeBSD$
.\"
.Dd February 11, 2010
.Os
.Dt PMC.MIPS 3
.Sh NAME
.Nm pmc.mips
.Nd measurement events for
.Tn MIPS
family CPUs
.Sh LIBRARY
.Lb libpmc
.Sh SYNOPSIS
.In pmc.h
.Sh DESCRIPTION
MIPS PMCs are present in MIPS
.Tn "24k"
and other processors in the MIPS family.
.Pp
There are two counters supported by the hardware and each is 32 bits
wide.
.Pp
MIPS PMCs are documented in
.Rs
.%B "MIPS32 24K Processor Core Family Software User's Manual"
.%D December 2008
.%Q "MIPS Technologies Inc."
.Re
.Ss Event Specifiers (Programmable PMCs)
MIPS programmable PMCs support the following events:
.Bl -tag -width indent
.It Li CYCLE
.Pq Event 0, Counter 0/1
Total number of cycles. 
The performance counters are clocked by the
top-level gated clock. 
If the core is built with that clock gater
present, none of the counters will increment while the clock is
stopped - due to a WAIT instruction.
.It Li INSTR_EXECUTED
.Pq Event 1, Counter 0/1
Total number of instructions completed.
.It Li BRANCH_COMPLETED 
.Pq Event 2, Counter 0
Total number of branch instructions completed.
.It Li BRANCH_MISPRED
.Pq Event 2, Counter 1
Counts all branch instructions which completed, but were mispredicted.
.It Li RETURN
.Pq Event 3, Counter 0
Counts all JR R31 instructions completed.
.It Li RETURN_MISPRED
.Pq Event 3, Counter 1
Counts all JR $31 instructions which completed, used the RPS for a prediction, but were mispredicted.
.It Li RETURN_NOT_31
.Pq Event 4, Counter 0
Counts all JR $xx (not $31) and JALR instructions (indirect jumps).
.It Li RETURN_NOTPRED
.Pq Event 4, Counter 1
If RPS use is disabled, JR $31 will not be predicted.
.It Li ITLB_ACCESS
.Pq Event 5, Counter 0
Counts ITLB accesses that are due to fetches showing up in the
instruction fetch stage of the pipeline and which do not use a fixed
mapping or are not in unmapped space. 
If an address is fetched twice from the pipe (as in the case of a
cache miss), that instruction willcount as 2 ITLB accesses. 
Since each fetch gets us 2 instructions,there is one access marked per double
word.
.It Li ITLB_MISS
.Pq Event 5, Counter 1
Counts all misses in the ITLB except ones that are on the back of another
miss.
We cannot process back to back misses and thus those are
ignored.
They are also ignored if there is some form of address error.
.It Li DTLB_ACCESS
.Pq Event 6, Counter 0
Counts DTLB access including those in unmapped address spaces.
.It Li DTLB_MISS
.Pq Event 6, Counter 1
Counts DTLB misses. Back to back misses that result in only one DTLB
entry getting refilled are counted as a single miss.
.It Li JTLB_IACCESS
.Pq Event 7, Counter 0
Instruction JTLB accesses are counted exactly the same as ITLB misses.
.It Li JTLB_IMISS
.Pq Event 7, Counter 1
Counts instruction JTLB accesses that result in no match or a match on
an invalid translation.
.It Li JTLB_DACCESS
.Pq Event 8, Counter 0
Data JTLB accesses.
.It Li JTLB_DMISS
.Pq Event 8, Counter 1
Counts data JTLB accesses that result in no match or a match on an invalid translation.
.It Li IC_FETCH
.Pq Event 9, Counter 0
Counts every time the instruction cache is accessed. All replays,
wasted fetches etc. are counted.
For example, following a branch, even though the prediction is taken,
the fall through access is counted.

.It Li IC_MISS
.Pq Event 9, Counter 1
Counts all instruction cache misses that result in a bus request.
.It Li DC_LOADSTORE
.Pq Event 10, Counter 0
Counts cached loads and stores.
.It Li DC_WRITEBACK
.Pq Event 10, Counter 1
Counts cache lines written back to memory due to replacement or cacheops.
.It Li DC_MISS
.Pq Event 11,   Counter 0/1
Counts loads and stores that miss in the cache
.It Li LOAD_MISS
.Pq Event 13, Counter 0
Counts number of cacheable loads that miss in the cache.
.It Li STORE_MISS
.Pq Event 13, Counter 1
Counts number of cacheable stores that miss in the cache.
.It Li INTEGER_COMPLETED
.Pq Event 14, Counter 0
Non-floating point, non-Coprocessor 2 instructions.
.It Li FP_COMPLETED
.Pq Event 14, Counter 1
Floating point instructions completed.
.It Li LOAD_COMPLETED
.Pq Event 15, Counter 0
Integer and co-processor loads completed.
.It Li STORE_COMPLETED
.Pq Event 15, Counter 1
Integer and co-processor stores completed.
.It Li BARRIER_COMPLETED
.Pq Event 16, Counter 0
Direct jump (and link) instructions completed.
.It Li MIPS16_COMPLETED
.Pq Event 16, Counter 1
MIPS16c instructions completed.
.It Li NOP_COMPLETED
.Pq Event 17, Counter 0
NOPs completed.
This includes all instructions that normally write to a general
purpose register, but where the destination register was set to r0.
.It Li INTEGER_MULDIV_COMPLETED
.Pq Event 17, Counter 1
Integer multipy and divide instructions completed.  (MULxx, DIVx, MADDx, MSUBx).
.It Li RF_STALL
.Pq Event 18, Counter 0
Counts the total number of cycles where no instructions are issued
from the IFU to ALU (the RF stage does not advance) which includes
both of the previous two events.
The RT_STALL is different than the sum of them though because cycles
when both stalls are active will only be counted once.
.It Li INSTR_REFETCH
.Pq Event 18, Counter 1
replay traps (other than uTLB)
.It Li STORE_COND_COMPLETED
.Pq Event 19, Counter 0
Conditional stores completed.  Counts all events, including failed stores.
.It Li STORE_COND_FAILED
.Pq Event 19, Counter 1
Conditional store instruction that did not update memory.
Note: While this event and the SC instruction count event can be configured to
count in specific operating modes, the timing of the events is much
different and the observed operating mode could change between them,
causing some inaccuracy in the measured ratio.
.It Li ICACHE_REQUESTS
.Pq Event 20, Counter 0
Note that this only counts PREFs that are actually attempted. 
PREFs to uncached addresses or ones with translation errors are not counted
.It Li ICACHE_HIT
.Pq Event 20, Counter 1
Counts PREF instructions that hit in the cache
.It Li L2_WRITEBACK
.Pq Event 21, Counter 0
Counts cache lines written back to memory due to replacement or cacheops.
.It Li L2_ACCESS
.Pq Event 21, Counter 1
Number of accesses to L2 Cache.
.It Li L2_MISS
.Pq Event 22, Counter 0
Number of accesses that missed in the L2 cache.
.It Li L2_ERR_CORRECTED
.Pq Event 22, Counter 1
Single bit errors in L2 Cache that were detected and corrected.
.It Li EXCEPTIONS
.Pq Event 23, Counter 0
Any type of exception taken.
.It Li RF_CYCLES_STALLED
.Pq Event 24, Counter 0
Counts cycles where the LSU is in fixup and cannot accept a new
instruction from the ALU.
Fixups are replays within the LSU that occur when an instruction needs
to re-access the cache or the DTLB. 
.It Li IFU_CYCLES_STALLED
.Pq Event 25, Counter 0
Counts the number of cycles where the fetch unit is not providing a
valid instruction to the ALU.
.It Li ALU_CYCLES_STALLED
.Pq Event 25, Counter 1
Counts the number of cycles where the ALU pipeline cannot advance.
.It Li UNCACHED_LOAD
.Pq Event 33, Counter 0
Counts uncached and uncached accelerated loads.
.It Li UNCACHED_STORE
.Pq Event 33, Counter 1
Counts uncached and uncached accelerated stores.
.It Li CP2_REG_TO_REG_COMPLETED
.Pq Event 35, Counter 0
Co-processor 2 register to register instructions completed.
.It Li MFTC_COMPLETED
.Pq Event 35, Counter 1
Co-processor 2 move to and from instructions as well as loads and stores.
.It Li IC_BLOCKED_CYCLES
.Pq Event 37, Counter 0
Cycles when IFU stalls because an instruction miss caused the IFU not
to have any runnable instructions.
Ignores the stalls due to ITLB misses as well as the 4 cycles
following a redirect.
.It Li DC_BLOCKED_CYCLES
.Pq Event 37, Counter 1
Counts all cycles where integer pipeline waits on Load return data due
to a D-cache miss.
The LSU can signal a "long stall" on a D-cache misses, in which case
the waiting TC might be rescheduled so other TCs can execute
instructions till the data returns.
.It Li L2_IMISS_STALL_CYCLES
.Pq Event 38, Counter 0
Cycles where the main pipeline is stalled waiting for a SYNC to complete.
.It Li L2_DMISS_STALL_CYCLES
.Pq Event 38, Counter 1
Cycles where the main pipeline is stalled because of an index conflict
in the Fill Store Buffer.
.It Li DMISS_CYCLES
.Pq Event 39, Counter 0
Data miss is outstanding, but not necessarily stalling the pipeline. 
The difference between this and D$ miss stall cycles can show the gain
from non-blocking cache misses.
.It Li L2_MISS_CYCLES
.Pq Event 39, Counter 1
L2 miss is outstanding, but not necessarily stalling the pipeline.
.It Li UNCACHED_BLOCK_CYCLES
.Pq Event 40, Counter 0
Cycles where the processor is stalled on an uncached fetch, load, or store.
.It Li MDU_STALL_CYCLES
.Pq Event 41, Counter 0
Cycles where the processor is stalled on an uncached fetch, load, or store.
.It Li FPU_STALL_CYCLES
.Pq Event 41, Counter 1
Counts all cycles where integer pipeline waits on FPU return data.
.It Li CP2_STALL_CYCLES
.Pq Event 42, Counter 0
Counts all cycles where integer pipeline waits on CP2 return data.
.It Li COREXTEND_STALL_CYCLES
.Pq Event 42, Counter 1
Counts all cycles where integer pipeline waits on CorExtend return data.
.It Li ISPRAM_STALL_CYCLES
.Pq Event 43, Counter 0
Count all pipeline bubbles that are a result of multicycle ISPRAM
access.
Pipeline bubbles are defined as all cycles that IFU doesn't present an
instruction to ALU. The four cycles after a redirect are not counted.
.It Li DSPRAM_STALL_CYCLES
.Pq Event 43, Counter 1
Counts stall cycles created by an instruction waiting for access to DSPRAM.
.It Li CACHE_STALL_CYCLES
.Pq Event 44, Counter 0
Counts all cycles the where pipeline is stalled due to CACHE
instructions.
Includes cycles where CACHE instructions themselves are
stalled in the ALU, and cycles where CACHE instructions cause
subsequent instructions to be stalled.
.It Li LOAD_TO_USE_STALLS
.Pq Event 45, Counter 0
Counts all cycles where integer pipeline waits on Load return data.
.It Li BASE_MISPRED_STALLS
.Pq Event 45, Counter 1
Counts stall cycles due to skewed ALU where the bypass to the address
generation takes an extra cycle.
.It Li CPO_READ_STALLS
.Pq Event 46, Counter 0
Counts all cycles where integer pipeline waits on return data from
MFC0, RDHWR instructions.
.It Li BRANCH_MISPRED_CYCLES
.Pq Event 46, Counter 1
This counts the number of cycles from a mispredicted branch until the
next non-delay slot instruction executes.
.It Li IFETCH_BUFFER_FULL
.Pq Event 48, Counter 0
Counts the number of times an instruction cache miss was detected, but
both fill buffers were already allocated.
.It Li FETCH_BUFFER_ALLOCATED
.Pq Event 48, Counter 1
Number of cycles where at least one of the IFU fill buffers is
allocated (miss pending).
.It Li EJTAG_ITRIGGER
.Pq Event 49, Counter 0
Number of times an EJTAG Instruction Trigger Point condition matched.
.It Li EJTAG_DTRIGGER
.Pq Event 49, Counter 1
Number of times an EJTAG Data Trigger Point condition matched.
.It Li FSB_LT_QUARTER
.Pq Event 50, Counter 0
Fill store buffer less than one quarter full.
.It Li FSB_QUARTER_TO_HALF
.Pq Event 50, Counter 1
Fill store buffer between one quarter and one half full.
.It Li FSB_GT_HALF
.Pq Event 51, Counter 0
Fill store buffer more than half full.
.It Li FSB_FULL_PIPELINE_STALLS
.Pq Event 51, Counter 1
Cycles where the pipeline is stalled because the Fill-Store Buffer in LSU is full.
.It Li LDQ_LT_QUARTER
.Pq Event 52, Counter 0
Load data queue less than one quarter full.
.It Li LDQ_QUARTER_TO_HALF
.Pq Event 52, Counter 1
Load data queue between one quarter and one half full.
.It Li LDQ_GT_HALF
.Pq Event 53, Counter 0
Load data queue more than one half full.
.It Li LDQ_FULL_PIPELINE_STALLS
.Pq Event 53, Counter 1
Cycles where the pipeline is stalled because the Load Data Queue in the LSU is full.
.It Li WBB_LT_QUARTER
.Pq Event 54, Counter 0
Write back buffer less than one quarter full.
.It Li WBB_QUARTER_TO_HALF
.Pq Event 54, Counter 1
Write back buffer between one quarter and one half full.
.It Li WBB_GT_HALF
.Pq Event 55, Counter 0
Write back buffer more than one half full.
.It Li WBB_FULL_PIPELINE_STALLS
.Pq Event 55 Counter 1
Cycles where the pipeline is stalled because the Load Data Queue in the LSU is full.
.It Li REQUEST_LATENCY
.Pq Event 61, Counter 0
Measures latency from miss detection until critical dword of response
is returned, Only counts for cacheable reads.
.It Li REQUEST_COUNT
.Pq Event 61, Counter 1
Counts number of cacheable read requests used for previous latency counter.
.El
.Ss Event Name Aliases
The following table shows the mapping between the PMC-independent
aliases supported by
.Lb libpmc
and the underlying hardware events used.
.Bl -column "branch-mispredicts" "cpu_clk_unhalted.core_p"
.It Em Alias Ta Em Event Ta 
.It Li instructions Ta Li INSTR_EXECUTED Ta
.It Li branches Ta Li BRANCH_COMPLETED Ta 
.It Li branch-mispredicts Ta Li BRANCH_MISPRED Ta 
.El
.Sh SEE ALSO
.Xr pmc 3 ,
.Xr pmc.atom 3 ,
.Xr pmc.core 3 ,
.Xr pmc.iaf 3 ,
.Xr pmc.k7 3 ,
.Xr pmc.k8 3 ,
.Xr pmc.p4 3 ,
.Xr pmc.p5 3 ,
.Xr pmc.p6 3 ,
.Xr pmc.tsc 3 ,
.Xr pmc_cpuinfo 3 ,
.Xr pmclog 3 ,
.Xr hwpmc 4
.Sh HISTORY
The
.Nm pmc
library first appeared in
.Fx 6.0 .
.Sh AUTHORS
The
.Lb libpmc
library was written by
.An "Joseph Koshy"
.Aq jkoshy@FreeBSD.org .
MIPS support was added by
.An "George Neville-Neil"
.Aq gnn@FreeBSD.org .
.Sh CAVEATS
The MIPS code does not yet support sampling.