How Cachegrind works

Detailed technical notes for hackers, maintainers and the overly-curious

njn25@cam.ac.uk
http://valgrind.kde.org

Copyright © 2001-2003 Nick Nethercote


Cache profiling

Valgrind is a very nice platform for doing cache profiling and other kinds of simulation, because it converts horrible x86 instructions into nice clean RISC-like UCode. For example, for cache profiling we are interested in instructions that read and write memory; in UCode there are only four instructions that do this: LOAD, STORE, FPU_R and FPU_W. By contrast, because of the x86 addressing modes, almost every instruction can read or write memory.

Most of the cache profiling machinery is in the file vg_cachesim.c.

These notes are a somewhat haphazard guide to how Valgrind's cache profiling works.

Cost centres

Valgrind gathers cache profiling about every instruction executed, individually. Each instruction has a cost centre associated with it. There are two kinds of cost centre: one for instructions that don't reference memory (iCC), and one for instructions that do (idCC):
typedef struct _CC {
   ULong a;
   ULong m1;
   ULong m2;
} CC;

typedef struct _iCC {
   /* word 1 */
   UChar tag;
   UChar instr_size;

   /* words 2+ */
   Addr instr_addr;
   CC I;
} iCC;
   
typedef struct _idCC {
   /* word 1 */
   UChar tag;
   UChar instr_size;
   UChar data_size;

   /* words 2+ */
   Addr instr_addr;
   CC I; 
   CC D; 
} idCC; 
Each CC has three fields a, m1, m2 for recording references, level 1 misses and level 2 misses. Each of these is a 64-bit ULong -- the numbers can get very large, ie. greater than 4.2 billion allowed by a 32-bit unsigned int.

A iCC has one CC for instruction cache accesses. A idCC has two, one for instruction cache accesses, and one for data cache accesses.

The iCC and dCC structs also store unchanging information about the instruction:

Note that data address is not one of the fields for idCC. This is because for many memory-referencing instructions the data address can change each time it's executed (eg. if it uses register-offset addressing). We have to give this item to the cache simulation in a different way (see Instrumentation section below). Some memory-referencing instructions do always reference the same address, but we don't try to treat them specialy in order to keep things simple.

Also note that there is only room for recording info about one data cache access in an idCC. So what about instructions that do a read then a write, such as:

inc %(esi)
In a write-allocate cache, as simulated by Valgrind, the write cannot miss, since it immediately follows the read which will drag the block into the cache if it's not already there. So the write access isn't really interesting, and Valgrind doesn't record it. This means that Valgrind doesn't measure memory references, but rather memory references that could miss in the cache. This behaviour is the same as that used by the AMD Athlon hardware counters. It also has the benefit of simplifying the implementation -- instructions that read and write memory can be treated like instructions that read memory.

Storing cost-centres

Cost centres are stored in a way that makes them very cheap to lookup, which is important since one is looked up for every original x86 instruction executed.

Valgrind does JIT translations at the basic block level, and cost centres are also setup and stored at the basic block level. By doing things carefully, we store all the cost centres for a basic block in a contiguous array, and lookup comes almost for free.

Consider this part of a basic block (for exposition purposes, pretend it's an entire basic block):

movl $0x0,%eax
movl $0x99, -4(%ebp)
The translation to UCode looks like this:
MOVL      $0x0, t20
PUTL      t20, %EAX
INCEIPo   $5

LEA1L     -4(t4), t14
MOVL      $0x99, t18
STL       t18, (t14)
INCEIPo   $7
The first step is to allocate the cost centres. This requires a preliminary pass to count how many x86 instructions were in the basic block, and their types (and thus sizes). UCode translations for single x86 instructions are delimited by the INCEIPo instruction, the argument of which gives the byte size of the instruction (note that lazy INCEIP updating is turned off to allow this).

We can tell if an x86 instruction references memory by looking for LDL and STL UCode instructions, and thus what kind of cost centre is required. From this we can determine how many cost centres we need for the basic block, and their sizes. We can then allocate them in a single array.

Consider the example code above. After the preliminary pass, we know we need two cost centres, one iCC and one dCC. So we allocate an array to store these which looks like this:

|(uninit)|      tag         (1 byte)
|(uninit)|      instr_size  (1 bytes)
|(uninit)|      (padding)   (2 bytes)
|(uninit)|      instr_addr  (4 bytes)
|(uninit)|      I.a         (8 bytes)
|(uninit)|      I.m1        (8 bytes)
|(uninit)|      I.m2        (8 bytes)

|(uninit)|      tag         (1 byte)
|(uninit)|      instr_size  (1 byte)
|(uninit)|      data_size   (1 byte)
|(uninit)|      (padding)   (1 byte)
|(uninit)|      instr_addr  (4 bytes)
|(uninit)|      I.a         (8 bytes)
|(uninit)|      I.m1        (8 bytes)
|(uninit)|      I.m2        (8 bytes)
|(uninit)|      D.a         (8 bytes)
|(uninit)|      D.m1        (8 bytes)
|(uninit)|      D.m2        (8 bytes)
(We can see now why we need tags to distinguish between the two types of cost centres.)

We also record the size of the array. We look up the debug info of the first instruction in the basic block, and then stick the array into a table indexed by filename and function name. This makes it easy to dump the information quickly to file at the end.

Instrumentation

The instrumentation pass has two main jobs:
  1. Fill in the gaps in the allocated cost centres.
  2. Add UCode to call the cache simulator for each instruction.
The instrumentation pass steps through the UCode and the cost centres in tandem. As each original x86 instruction's UCode is processed, the appropriate gaps in the instructions cost centre are filled in, for example:
|INSTR_CC|      tag         (1 byte)
|5       |      instr_size  (1 bytes)
|(uninit)|      (padding)   (2 bytes)
|i_addr1 |      instr_addr  (4 bytes)
|0       |      I.a         (8 bytes)
|0       |      I.m1        (8 bytes)
|0       |      I.m2        (8 bytes)

|WRITE_CC|      tag         (1 byte)
|7       |      instr_size  (1 byte)
|4       |      data_size   (1 byte)
|(uninit)|      (padding)   (1 byte)
|i_addr2 |      instr_addr  (4 bytes)
|0       |      I.a         (8 bytes)
|0       |      I.m1        (8 bytes)
|0       |      I.m2        (8 bytes)
|0       |      D.a         (8 bytes)
|0       |      D.m1        (8 bytes)
|0       |      D.m2        (8 bytes)
(Note that this step is not performed if a basic block is re-translated; see here for more information.)

GCC inserts padding before the instr_size field so that it is word aligned.

The instrumentation added to call the cache simulation function looks like this (instrumentation is indented to distinguish it from the original UCode):

MOVL      $0x0, t20
PUTL      t20, %EAX
  PUSHL     %eax
  PUSHL     %ecx
  PUSHL     %edx
  MOVL      $0x4091F8A4, t46  # address of 1st CC
  PUSHL     t46
  CALLMo    $0x12             # second cachesim function
  CLEARo    $0x4
  POPL      %edx
  POPL      %ecx
  POPL      %eax
INCEIPo   $5

LEA1L     -4(t4), t14
MOVL      $0x99, t18
  MOVL      t14, t42
STL       t18, (t14)
  PUSHL     %eax
  PUSHL     %ecx
  PUSHL     %edx
  PUSHL     t42
  MOVL      $0x4091F8C4, t44  # address of 2nd CC
  PUSHL     t44
  CALLMo    $0x13             # second cachesim function
  CLEARo    $0x8
  POPL      %edx
  POPL      %ecx
  POPL      %eax
INCEIPo   $7
Consider the first instruction's UCode. Each call is surrounded by three PUSHL and POPL instructions to save and restore the caller-save registers. Then the address of the instruction's cost centre is pushed onto the stack, to be the first argument to the cache simulation function. The address is known at this point because we are doing a simultaneous pass through the cost centre array. This means the cost centre lookup for each instruction is almost free (just the cost of pushing an argument for a function call). Then the call to the cache simulation function for non-memory-reference instructions is made (note that the CALLMo UInstruction takes an offset into a table of predefined functions; it is not an absolute address), and the single argument is CLEARed from the stack.

The second instruction's UCode is similar. The only difference is that, as mentioned before, we have to pass the address of the data item referenced to the cache simulation function too. This explains the MOVL t14, t42 and PUSHL t42 UInstructions. (Note that the seemingly redundant MOVing will probably be optimised away during register allocation.)

Note that instead of storing unchanging information about each instruction (instruction size, data size, etc) in its cost centre, we could have passed in these arguments to the simulation function. But this would slow the calls down (two or three extra arguments pushed onto the stack). Also it would bloat the UCode instrumentation by amounts similar to the space required for them in the cost centre; bloated UCode would also fill the translation cache more quickly, requiring more translations for large programs and slowing them down more.

Handling basic block retranslations

The above description ignores one complication. Valgrind has a limited size cache for basic block translations; if it fills up, old translations are discarded. If a discarded basic block is executed again, it must be re-translated.

However, we can't use this approach for profiling -- we can't throw away cost centres for instructions in the middle of execution! So when a basic block is translated, we first look for its cost centre array in the hash table. If there is no cost centre array, it must be the first translation, so we proceed as described above. But if there is a cost centre array already, it must be a retranslation. In this case, we skip the cost centre allocation and initialisation steps, but still do the UCode instrumentation step.

The cache simulation

The cache simulation is fairly straightforward. It just tracks which memory blocks are in the cache at the moment (it doesn't track the contents, since that is irrelevant).

The interface to the simulation is quite clean. The functions called from the UCode contain calls to the simulation functions in the files vg_cachesim_{I1,D1,L2}.c; these calls are inlined so that only one function call is done per simulated x86 instruction. The file vg_cachesim.c simply #includes the three files containing the simulation, which makes plugging in new cache simulations is very easy -- you just replace the three files and recompile.

Output

Output is fairly straightforward, basically printing the cost centre for every instruction, grouped by files and functions. Total counts (eg. total cache accesses, total L1 misses) are calculated when traversing this structure rather than during execution, to save time; the cache simulation functions are called so often that even one or two extra adds can make a sizeable difference.

Input file has the following format:

file         ::= desc_line* cmd_line events_line data_line+ summary_line
desc_line    ::= "desc:" ws? non_nl_string
cmd_line     ::= "cmd:" ws? cmd
events_line  ::= "events:" ws? (event ws)+
data_line    ::= file_line | fn_line | count_line
file_line    ::= ("fl=" | "fi=" | "fe=") filename
fn_line      ::= "fn=" fn_name
count_line   ::= line_num ws? (count ws)+
summary_line ::= "summary:" ws? (count ws)+
count        ::= num | "."
Where: The contents of the "desc:" lines is printed out at the top of the summary. This is a generic way of providing simulation specific information, eg. for giving the cache configuration for cache simulation.

Counts can be "." to represent "N/A", eg. the number of write misses for an instruction that doesn't write to memory.

The number of counts in each line and the summary_line should not exceed the number of events in the event_line. If the number in each line is less, cg_annotate treats those missing as though they were a "." entry.

A file_line changes the current file name. A fn_line changes the current function name. A count_line contains counts that pertain to the current filename/fn_name. A "fn=" file_line and a fn_line must appear before any count_lines to give the context of the first count_lines.

Each file_line should be immediately followed by a fn_line. "fi=" file_lines are used to switch filenames for inlined functions; "fe=" file_lines are similar, but are put at the end of a basic block in which the file name hasn't been switched back to the original file name. (fi and fe lines behave the same, they are only distinguished to help debugging.)

Summary of performance features

Quite a lot of work has gone into making the profiling as fast as possible. This is a summary of the important features:

Annotation

Annotation is done by cg_annotate. It is a fairly straightforward Perl script that slurps up all the cost centres, and then runs through all the chosen source files, printing out cost centres with them. It too has been carefully optimised.

Similar work, extensions

It would be relatively straightforward to do other simulations and obtain line-by-line information about interesting events. A good example would be branch prediction -- all branches could be instrumented to interact with a branch prediction simulator, using very similar techniques to those described above.

In particular, cg_annotate would not need to change -- the file format is such that it is not specific to the cache simulation, but could be used for any kind of line-by-line information. The only part of cg_annotate that is specific to the cache simulation is the name of the input file (cachegrind.out), although it would be very simple to add an option to control this.