Skip to content

AVRtest is an AVR core simulator used for regression testing of avr-gcc and parts of AVR-LibC.

License

Notifications You must be signed in to change notification settings

sprintersb/atest

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

What is AVRtest?

AVRtest is a free software simulator for the AVR family of 8-bit microcontrollers distributed under the GNU General Public License.

The main intention of AVRtest is to supply a fast, light-weight and easy-to-use simulator to run the GCC testsuite for avr-gcc and parts of the AVR-LibC testsuite.

Table of Contents

Selected Options

Features

Introduction

avrtest is an instruction set simulator for AVR core families
avr51: ATmega128*, AT90USB128*, ATtiny2313, ... with a 2-byte PC
avr6: ATmega256* with a 3-byte PC

avrtest-xmega is an instruction set simulator for AVR XMEGA core families
avrxmega6: ATxmega128*, ... with a 3-byte PC
avrxmega3: ATtiny212, ATtiny816, ... that see flash in RAM address space
avrxmega4: ATxmega16*, ATxmega32*, ATxmega64*, ... with a 2-byte PC
avrxmega7: ATxmega128A1*, ... with a 3-byte PC that use RAMPx in RAM accesses

avrtest-tiny is an instruction set simulator for the Reduced AVR TINY core family
avrtiny: ATtiny40, ... with only 16 GPRs

Also supported are other cores like avr25, avrxmega2 etc. They are basically aliases for one of the cores from above.

The executable that supports AVR XMEGA cores is named avrtest-xmega. In addition to the avrtest simulator, it supports the XMEGA instructions XCH, LAS, LAC and LAT. For avrxmega3, it also supports the command line option -pm 0x4000 in order to set the location of the flash image in the RAM address space for devices like ATmega4808. The default for this option and for avrxmega3 is 0x8000.

The executable that supports Reduced AVR TINY cores is named avrtest-tiny. It only supports general purpose registers (GPRs) R16...R31, and many instructions like ADIW are not supported. The LDS and STS instructions are 1-word instructions that can access SRAM in the range 0x40...0xbf.

In the remainder, avrtest is explained. avrtest-xmega and avrtext-tiny work similar. AVRtest does not simulate internal peripherals like timers, I/O ports, interrupts, etc.

Special Features

The simulated program can communicate with the simulator: It can write to the host's standard output, and it can read from the host's standard input. For example, an easy hello world program for AVR could look like this:

#include <stdio.h>

int main (void)
{
    printf ("Hallo World!\n");
    return 0;
}

Compile this program as usual but link against exit-atmega128.o. How to build that object file is explained in the next chapter. Store the C source as hello.c and compile it with, e.g.

avr-gcc hello.c -O -mmcu=atmega128 -o hello.elf /someplace/avrtest/exit-atmega128.o

The exit-*.o object implements stdout as a stream writing through to the host computer's stdout, similar for stderr. Running the program

avrtest hello.elf

will print something like

Hallo World!

 exit status: EXIT
      reason: exit 0 function called
     program: hello.elf
exit address: 0001a4
total cycles: 837
total instr.: 503

You can also use the respective primitive functions directly:

#include "avrtest.h"

int main (void)
{
    char c;

    avrtest_putchar ('Q');

    c = avrtest_getchar();
    avrtest_putchar (c + 1);

    return 0;
}

The avrtest_putchar and avrtest_getchar functions are defined in avrtest.h and act like putchar resp. getchar in an ordinary, hosted program: putchar writes a character to the host's standard output, and getchar reads a character from the host's standard input.

Compile the C source with

avr-gcc inout.c -O -mmcu=atmega128 -o inout.elf -I /someplace/avrtest

When running the program with

avrtest inout.elf

the simulator will print a Q on the console and wait for input. Type 1<Enter> and the output will be

Q1
2
 exit status: EXIT
      reason: infinite loop detected (normal exit)
     program: inout.elf
exit address: 0000be
total cycles: 27
total instr.: 18

There are more functions available for program ↔ simulator interaction, see the respective sections below.

Running the avr-gcc Testsuite using the AVRtest Simulator

http://lists.gnu.org/archive/html/avr-gcc-list/2009-09/msg00016.html

  • Get avr-gcc, Binutils and AVR-LibC built from sources and working. For build instructions, see for example the AVR-LibC documentation.

  • Install DejaGNU, expect and TCL

  • Unpack the AVRtest sources top /someplace. You find a link to the package at

  • Run

    make
    

    inside /someplace/avrtest which will build executables from the AVRtest C-sources and AVR object files exit-*.o from dejagnuboards/exit.c.

  • Adjust your $HOME/.dejagnurc file (or create one if it not already exists) as indicated in /someplace/avrtest/.dejagnurc. Change the avrtest_dir path to the directory where the avrtest executable is located:

    set avrtest_dir "/someplace/avrtest"
    
  • You can add /someplace/avrtest to your PATH variable so that it's more convenient to start avrtest. In order to run the GCC test suite this is not necessary, though.

  • The .dejagnurc file sets the avrlibc_include_dir variable that points to a directory where the compiler can find stdio.h and similar headers from AVR-LibC. Adjust this path to point to the incude/avr subdir of a valid AVR-LibC installation. You can also adjust that path in the *-sim.exp file.

  • *-sim.exp just sets some custom variables and then loads avrtest.exp from the same directory to run avrtest.

  • Run the tests from the ./gcc subdirectory of your avr-gcc build directory like:

    make -k check-gcc RUNTESTFLAGS="--target_board=atmega128-sim"
    
  • Look in the ./gcc/testsuite subdirectory of avr-gcc build directory for logged results gcc.sum, gcc.log, or add the -all switch to see it working one test at a time:

    make -k check RUNTESTFLAGS="--target_board=atmega128-sim -all"
    
  • You can run parts of the testsuite like this:

    make -k check RUNTESTFLAGS="--target_board=atmega128-sim avr.exp"
    make -k check RUNTESTFLAGS="--target_board=atmega128-sim avr-torture.exp=pr41885*.c"
    

Voilà!

Building the exit.o Modules

The exit-*.o Modules are build during make using avr-gcc. When you want to use a different avr-gcc not found in PATH, use:

make clean-exit all-avr CC_FOR_AVR=/some-compiler-path/avr-gcc

When you need an exit module for a device not already present in the Makefile like ATmega16, you can just run

make exit-atmega16.o

Notice that an exit module is not strictly required in a program that's simulated with AVRtest. But it defines streams like stdout so that functions like printf behave like in a hosted environment without any further ado. Moreover, the exit module implements exit() in such a way that it reports the program's exit code to the host, whereas without the module, the simulator will just print "infinite loop detected (normal exit)" as it hits the rjmp .-2 in the stock exit implementation

-h: Getting Help

Running

avrtest -h

will print a help screen with available options:

  usage: avrtest [-d] [-e ENTRY] [-m MAXCOUNT] [-mmcu=ARCH] [-s SIZE]
                 [-no-log] [-no-stdin] [-no-stdout] [-no-stderr]
                 [-q] [-flush] [-runtime] [-v]
                 [-graph[=FILE]] [-sbox=FOLDER]
                 program [-args [...]]
         avrtest --help
Options:
  -h            Show this help and exit.
  -args ...     Pass all following parameters as argc and argv to main.
  -d            Initialize SRAM from .data (for ELF program)
  -e ENTRY      Byte address of program entry.  Default for ENTRY is
                the entry point from the ELF program and 0 for non-ELF.
  -pm OFFSET    Set OFFSET where the program memory is seen in the
                LD's instruction address space (avrxmega3 only).
  -m MAXCOUNT   Execute at most MAXCOUNT instructions. Supported suffixes
                are k for 1000 and M for a million.
  -s SIZE       The size of the simulated flash.  For a program built
                for ATmega8, SIZE would be 8K or 8192 or 0x2000.
  -q            Quiet operation.  Only print messages explicitly
                requested.  Pass exit status from the program.
  -v            Verbose mode.  Print the loaded ELF program headers.
  -runtime      Print avrtest execution time.
  -no-log       Disable logging in avrtest_log.  Useful when capturing
                performance data.  Logging can still be controlled by
                the running program, cf. README.
  -no-stdin     Disable avrtest_getchar (syscall 28) from avrtest.h.
  -no-stdout    Disable avrtest_putchar (syscall 29) from avrtest.h.
  -no-stderr    Disable avrtest_putchar_stderr (syscall 24).
  -flush        Flush the host's stdout resp. stderr stream after each
                (direct or indirect) call of avrtest_putchar (syscall 29)
                resp. avrtest_putchar_stderr (syscall 24).
  -sbox SANDBOX Provide the path to SANDBOX, which is a folder that the
                target program can access via file I/O (syscall 26).
  -graph[=FILE] Write a .dot FILE representing the dynamic call graph.
                For the dot tool see  http://graphviz.org
  -graph-help   Show more options to control graph generation and exit.
  -mmcu=ARCH    Select instruction set for ARCH

-s SIZE: Specifying simulated Flash Size

The simulator must know the flash size of the simulated controller so it can correctly simulare instructions like RJMP. For example, an RJMP .-4 located at address 0x0 will jump to 0x1ffe on ATmega8 but to 0x7fe on ATtiny2313.

For programs that use the start-up code from AVR-LibC in crt<mcu>.o, AVRtest will read the flash size from a special note section named .note.gnu.avr.deviceinfo contained in that startup-code. AVRtest will print its contents with -v.

In the following situations, -s SIZE can be used to specify the flash size:

  • The program does not use startup-code from AVR-LibC. In that case, you can specify, say -s 8192 or -s 0x2000 or -s 8k for ATmega8.

  • You want to use a flash size other than shipped with crt<mcu>.o.

  • Use -s -1 if you don't want to specify a flash size and also don't want to use the flash size from .note.gnu.avr.deviceinfo. AVRtest will use a core-specific default value.

-q: Quiet Operation

With -q turned on, AVRtest will print no messages except the ones explicitly requested by:

  • -runtime
  • LOG_U8 (X), PERF_DUMP_ALL and similar logging commands, see below.

The following exit stati will be returned with -q:

  • 0 – Everything went fine.
  • 1 – Target program called abort()
  • x – Target program called exit(x)
  • 10 – Program timeout as set by -m MAXCOUNT
  • 11 – Something is wrong with the program file: Does not fit into memory, not an AVR executable, ELF headers broken, ...
  • 12 – The program goes haywire: Stack overflow, illegal instruction or PC.
  • 13 – Problem with symbol information like an odd function address.
  • 14 – Problem with using file I/O with the host's file system: Bad file handle or file name, illegal argument. This does not encompass when fopen cannot open the file; this will just return a 0-handle.
  • 20 – Out of memory.
  • 21 – Wrong avrtest usage: Unknown options, etc.
  • 22 – Program file could not be found / read.
  • 42 – Fatal error in avrtest.

Without -q, the exit status will be 1 (EXIT_FAILURE) for the cases ≥ 20 from above. For all the cases < 20 from above, 0 (EXIT_SUCCESS) will be returned.

Speed of Simulation

The main incentive behind AVRtest is speed of execution. For AVRtest compiled with a reasonably optimizing compiler (like with gcc -O3), the speed of a simulated program is around

45 MHz/GHz  for  avrtest              on x86_64
12 MHz/GHz  for  avrtest_log -no-log  on x86_64

measured on an Intel Core2 Duo (x86_64). On x86, the performance is ca 70% of the above, i.e. around

30 MHz/GHz  for  avrtest              on x86
 8 MHz/GHz  for  avrtest_log -no-log  on x86

The GHz-values refer to the execution speed of the host computer, and the MHz-values refer to an AVR microcontroller. For example, with a x86_64 host running at 2 GHz, AVRtest will perform as fast as an AVR running at around 90 MHz.

AVRtest is a single-core application. The speed values are under the assumption that the simulator performs no logging or printing. The timing will also depend slighly on which instructions are being simulated.

-m MAXCOUNT: Maximum Instruction Count to simulate

Per default, there is no limitation of the number of simulated instructions. This is a problem with malfunctioning programs. When such cases may occur, the maximal number of instructions that will be simulated can be set with -m MAXCOUNT.

Supported suffixes are k for 1000 and M for a million, for example -m 1M and -m 1e6 will simulate no more than 1000000 instructions and terminate with a TIMEOUT exit status when the program requires more instructions. -m 0 means no limitation (default).

-args ...: Passing Arguments to the Program

All arguments after -args will be passed to the target program as if the program was running from the command line. This is accomplished by startup code from exit.c located in section .init8 right before main is called.

If you have own startup code in .init8 make sure it will be located before the code from exit.c, e.g. by appropriate order of the objects passed to the linker or by means of a custom linker script.

The first argument (argv[0]) is the program name with directories stripped off to save space. The last argument argv[argc] is set to NULL.

The simulator will transfer the arguments to RAM and set R24 to argc, R22 to argv and R20 to env so that you can write

int main (int argc, char *argv[])
{
    ...
}

resp.

int main (int argc, char *argv[], char *env[])
{
    ...
}

and use argc and argv as in any hosted application: If the program under simulation is main.elf and you run

avrtest main.elf -args hallo "4 2"

then:

argc       is 3
argv[0]    is (the address of) string "main.elf" (i.e. argv[0][0] = 'm')
argv[1]    is (the address of) string "hallo"
argv[2]    is (the address of) string "4 2"
argv[3]    is the NULL pointer

env is set to the NULL pointer if the program is executed by avrtest and to a non-NULL pointer when executed by avrtest_log.

exit.c:avrtest_init_argc_argv() requests the command args to be put starting at a RAM address hard-coded in exit.c::avrtest_init_argc_argv(). If you prefer a different location then adjust exit.c according to your needs, following the source comments.

-no-args

will act as if no -args switch was supplied: It skips all arguments thereafter and sets argc=0, argv=NULL and env as described above.

-no-log and Logging Control

⚠️ Instruction logging is only supported by the avrtest_log family.

avrtest_log will log address and action of each executed instruction to standard output of the host. In cases where that is too much a flood of information, you can start avrtest_log with -no-log and turn on instruction logging by executing special commands defined in avrtest.h:

// Turn on instruction logging.
LOG_ON;

// Turn it off again.
LOG_OFF;

// Push the current state of the logging (on or off) to an
// internal simulator stack, and then turn logging on / off.
LOG_PUSH_ON;
LOG_PUSH_OFF;

// Pop the logging state from the top of that stack.
LOG_POP;

// Turn on logging for the next N instructions and then switch
// it off again.
LOG_SET (N);

// Turn on logging during performance measurement described below,
// i.e. if any timer is in a START / STOP round.
LOG_PERF;

Please notice that this method of switching logging on / off is (low) intrusive to the program. The commands listed above are translated by the compiler to machine instructions that might affect code generation for the surrounding code (register allocation, jump offsets, ...). The commands have low overhead; it's not more than an AVRtest syscall.

LOG_ON, LOG_OFF, LOG_PUSH_ON, LOG_PUSH_OFF and LOG_POP are also supplied as macros that can be used from assembly code. They have no effect on the internal state of the program (except for the program counter: the respective syscalls consume 4 bytes of code).

Logging Values to the Host Computer

In order to get values out of the running program, the following low-overhead, low-intrusive commands might be useful:

LOG_U8 (X);         print X as unsigned 8-bit decimal value
LOG_S8 (X);         print X as signed 8-bit decimal value
LOG_X8 (X);         print X as unsigned 8-bit hexadecimal value

Besides these you can use similar commands for wider values with 16, 24 and 32 bits, e.g.

LOG_U16 (X);
LOG_X32 (X);

etc. Moreover, there are the following commands:

LOG_STR (X);        log string X located in RAM
LOG_PSTR (X);       log string X located in Flash
LOG_ADDR (X);       log address X
LOG_FLOAT (X);      log float X
LOG_DOUBLE (X);     log double X
LOG_LDOUBLE (X);    log long double X
LOG_D64 (X);        log an uint64_t, the bit representation of which
                    is interpreted as an IEEE double number
                    (11-bit exponent, 52-bit encoded mantissa)
LOG_F7T (X);        log LibF7's f7_t*

All of these logging commands have variants that take a custom printf format string to print the supplied value:

LOG_FMT_XXX (F, X);    log X using format string F located in RAM
LOG_PFMT_XXX (F, X);   log X using format string F located in Flash

⚠️ Please use format strings with care! You can easily crash AVRtest and make it raise a Segmentation Fault or other undefined behaviour by specifying a format string that does not match the value! Notice that the %-formats below refer to the HOST machine, i.e. the machine that executes AVRtest!

AVRtest uses double to represent floating point values and unsigned int for all other values up to 4 bytes. The default format strings for the vanilla LOG_XXX are:

" %u "              for the "U" variants
" %d "              for the "S" variants
" 0x%02x "          for the 8-bit "X" variants
...                 ...
" 0x%08x "          for the 32-bit "X" variants
"%s"                for strings
" 0x%04x "          for addresses
" %.6f "            for float, double, long double and 64-bit unsigned
                    interpreted as a 64-bit IEEE floating point number
"%s"                for f7_t*

For 8-byte values, the internal representation is unsigned long long, and the default format strings are:

" %llu "              for the "U" variant
" %lld "              for the "S" variant
" 0x%016llx "         for the "X" variant

Support of FLMAP

Devices from the AVR64* and AVR128* families see a 32 KiB segment of their program memory in the RAM address space. Which segment is visible and accessible is determined by bit field NVMCTRL_CTRLB.FLMAP. AVRtest copies the segment as specified by FLMAP from program memory into the RAM address space in .init4 by means of code from exit-avr*.o.

This means setting of NVMCTRL_CTRLB.FLMAP has only an effect when it happens prior to that in .init1 ... .init3.

If you really want to test an application that switches FLMAP back and forth, you have to trigger a re-write of the 32 KiB segment by means of a call to avrtest_misc_flmap (uint8_t flmap); each time, where flmap ranges from 0 to 3.

File I/O with the File System of the Host Computer

AVRtest supports basic file I/O capabilities that enable the target program to communicate with the host's file system. In order to enable this, option

-sbox SANDBOX

has to be set on the command line. SANDBOX is the name of a folder of the host file system. SANDBOX will be prepended to the file name passed to fopen() without any further ado. AVRtest implements several functionalities found in stdio.h and provides them by means of a syscall that can be accessed by functions from avrtest.h:

unsigned long avrtest_fileio_p (char what, const void *pargs);
unsigned long avrtest_fileio_1 (char what, unsigned char args);
unsigned long avrtest_fileio_2 (char what, unsigned      args);
unsigned long avrtest_fileio_4 (char what, unsigned long args);

WHAT specifies the host operation to perform. For each supported function, there is a macro AVRTEST_F that can be used as WHAT where F specifies the function. For example, AVRTEST_fopen specifies to call the host's fopen(). In order to specify FILEs, handles are passed around instead or FILE* pointers: If the prototype of the stdio.h function specifies a FILE* argument, then the AVRtest interface uses a handle instead. A handle is a small 8-bit value.

ARGS resp. PARGS specifies the arguments to pass to the host. If all of them -- after replacing FILE* by its respective handle -- fit into 4 bytes, then these arguments are passed compressed as ARGS. If they do not fit into 4 bytes, then the PARGS variant has to be used where PARGS points to a memory location where to find the arguments.

To give an example, fopen whould be called as

uint8_t handle = (uint8_t) avrtest_fileio_4 (AVRTEST_fopen, args);

where the low-word of ARGS is a const char* that points to the file name (which will be appended to SANDBOX as-is), and the high-word is a char pointer to the mode to use to open the file. The return value ships the handle associated to the host's FILE* in the low-byte of the return value, the other bytes of the return value are unused. HANDLE would be used in subsequent calls like AVRTEST_fclose, that passes the file handle in the low-byte of ARGS.

File I/O: Using the fileio.c Module

For convenience, AVRtest comes with a target module that supplies wrappers for the available file I/O functions. The interface of these functions is as specified by the C standard, and fileio.c then tries to map the FILE* pointer to an associated handle that's needed for the respective AVRtest syscall. The following functions from stdio.h are implemented:

fopen, fclose, feof, fflush, clearerr, fread, fwrite.

⚠️ However, using fileio.c comes with some limitations, and it needs special options when linking. See the next section for details.

If a handle is not found, a default action might be performed. Suppose for example a call to fread. If fileio.c finds a handle associated to the passed FILE*, it will call the host's fread with the host's FILE* derived from the handle; otherwise it calls the AVR-LibC implementation of fread which uses FILE* and operates byte-by-byte.

Most functions like fputc, fgetc, fputs, fprintf, etc. that are not included in the above list will work out-of-the-box and without special treatment or wrappers from the fileio module.

You can use fileio.c/h as a module in your application, or you can link against one of the pre-compiled object files built by make. For a program that simulates ATmega128 for example, the link command would read

avr-gcc -mmcu=atmega128 fileio-atmega128.o -Wl,-wrap,fclose -Wl,-wrap,feof ...

and then simulate program.elf with avrtest or avrtest_log.

In order to circumvent AVR-LibC stdio.h pecularities and also for better performance, the fileio module provides functions that use handles directly instead of the need for an intermediate FILE*. These functions with their obvious semantics are:

fileio_handle_t host_fopen (const char *filename, const char *mode);
int host_fclose (fileio_handle_t);
int host_fflush (fileio_handle_t)
int host_feof (fileio_handle_t);
int host_fseek (fileio_handle_t, long pos, uint8_t whence);
int host_fgetc (fileio_handle_t);
int host_fputc (char, fileio_handle_t);
void host_clearerr (fileio_handle_t);
size_t host_fwrite (const void*, size_t, size_t, fileio_handle_t);
size_t host_fread (void*, size_t, size_t, fileio_handle_t);

File I/O: Caveats, Restrictions and Limitations

⚠️ Due to peculiarities of AVR-LibC's stdio.h implementation, the following rules have to be obeyed when using file I/O via fileio.c:

  • The linker must be called with -wrap for the following symbols:

    feof, fwrite, fread, fclose, cleaerr
    

    i.e. avr-gcc has to be called with

    avr-gcc ... -Wl,-wrap,feof -Wl,-wrap,fwrite -Wl,-wrap,fread ...
    

    when linking. For the documentation of -wrap, see http://sourceware.org/binutils/docs-2.32/ld/Options.html#index-_002d_002dwrap_003dsymbol

  • When using AVRtest to run the GCC test suite for AVR, the host interactions that exit-*.o supplies via stdout and stderr should be sufficient. The complexity of fileio is not needed for the GCC test suite.

File I/O: Special Streams for the Host's stdin, stdout, stderr

The fileio module provides

FILE *host_stdin;
FILE *host_stdout;
FILE *host_stderr;

that are associated to the host's stdin, stdout and stderr, respectively. The respective handles to use with functions like host_fflush() are:

HANDLE_stdin
HANDLE_stdout
HANDLE_stderr

Performance Measurement

⚠️ This feature is only supported by avrtest_log and avrtest-xmega_log.

The simulator supports 7 independently operating performance-meters 1...7:

PERF_START (N);

starts perf-meter N which will start capturing values of the running program like program counter, instruction count, stack pointer, call depth, etc.

PERF_STOP (N);

will halt the perf-meter. Upon encountering the next PERF_START(N) the meter will be re-enabled and proceed. The collected values can be dumped at any time to the standard output of the host computer by

PERF_DUMP (N);

or

PERF_DUMP_ALL;

This will show a summary of the extremal values for the perf-meter(s) and reset it completely so that you can use it for a different task afterwards. Displayed values are:

 - Stack (abs):  Values of the stack pointer (SP).
 - Stack (rel):  Stack usage relative to the starting point.
 - Calls (abs):  The absolute call depth (CALLs increment, RETs decrement).
 - Calls (rel):  Call depth relative to the starting point.

Tracking of the call depth won't work as expected for setjmp / longjmp and similar functions that set SP to an unknown call depth.

In order to give a specific perf-meter N a more descriptive name there are

PERF_LABEL (N, L);
PERF_PLABEL (N, L);

where label L is a C-string residing in RAM or Flash, respectively. avrtest_log will cache the label so you can override it after PERF_LABEL.

If you want to find out what value lead to the round with mimimal or maximal instruction cycles, the next START/STOP round can be tagged with a value:

PERF_TAG_STR (N, STR);      Tag perf-meter N with string STR
PERF_TAG_U16 (N, U);        ... with a 16-bit unsigned integer U
PERF_TAG_S16 (N, U);        ... with a 16-bit signed integer U
PERF_TAG_U32 (N, U);        ... with a 32-bit unsigned integer U
PERF_TAG_S32 (N, U);        ... with a 32-bit signed integer U
PERF_TAG_FLOAT (N, F);      ... with float F

Just like with LOG_XXX(), S16 and S32 are represented internally as "int", U16 and U32 as "unsigned" and float as "double". The default format strings to print the tags are "%s" for string, " %u " for the unsigned versions, " %d " for the signed ones and " %.6f " for float.

Custom format strings can be supplied by

PERF_TAG_FMT_U16 (N, FMT, U);
PERF_TAG_PFMT_S32 (N, FMT, U);

etc. with a format string FMT located in RAM resp. Flash.

Example:

#include <math.h>
#include "avrtest.h"

float x, y;

int main (void)
{
    PERF_LABEL (1, "resource consumption of sin()");

    for (x = 0.0; x <= 3.14159; x += 0.01)
    {
        PERF_TAG_FMT_FLOAT (1, " sin (%.5f) ", x);
        PERF_START (1);
        y = sin (x);
        PERF_STOP (1);

        PERF_TAG_FMT_FLOAT (2, " x=%.5f ", x);
        PERF_STAT_FLOAT (2, fabs (x * (1 - x*x / 6) - y));
    }

    PERF_DUMP_ALL;
    return 0;
}

Compiling as usual, linking against exit-*.o and running

avrtest_log ... -no-log -q

produces an output like

--- Dump # 1:
 Timer T1 "resource consumption of sin()" (315 rounds):  027c--029e
              Instructions        Ticks
    Total:       411402          541887
    Mean:          1306            1720
    Stand.Dev:     75.5            85.6
    Min:            686            1029
    Max:           1515            1952
    Calls (abs) in [   1,   5] was:   1 now:   1
    Calls (rel) in [   0,   4] was:   0 now:   0
    Stack (abs) in [04f8,04e5] was:04f8 now:04f8
    Stack (rel) in [   0,  19] was:   0 now:   0

           Min round Max round    Min tag           /   Max tag
    Calls       -all-same-                          /
    Stack       -all-same-                          /
    Instr.         1       315     sin (0.00000)    /   sin (3.14000)
    Ticks          1       315     sin (0.00000)    /   sin (3.14000)

 Stat  T2 "" (315 Values)
    Mean:       3.589886e-001     round    tag
    Stand.Dev:  5.247231e-001
    Min:        0.000000e+000         1     x=0.00000
    Max:        2.021442e+000       315     x=3.14000

The computation of sin(x) took 1029...1952 cycles (with an average of 1720 cycles) for the 315 values it had been computed. sin() needs 19 bytes stack space. The maximal difference between sin (x) and the 3rd-order approximation x − x3 / 6 is 2.02144 and occurred for x = 3.14000.

PERF_STAT_FLOAT (N, F);

in the example makes perf-meter N gather statistics of float value F: mean (expected value), standard deviation, minimum, maximum, and the START/STOP round for which a minimal / maximal value of F has been seen. If rounds have been tagged, the tags for the rounds that yielded a minimal and maximal value of F are also displayed. Besides PERF_STAT_FLOAT there is also PERF_STAT_U32 and PERF_STAT_S32 to get statistics for (un)signed 32-bit integer values.

In the sample code from above, one is interested in the resource consumption of the sin function. In order to supply that function with a value x and to store the result y, additional instructions are needed:

PERF_START (1);
y = sin (x);
PERF_STOP (1);

If these additional instructions are of no interest, use

PERF_START_CALL (N);

which only adds costs of instructions that are at a call depth of at least 1 (relative to the starting point). This includes costs of CALL and RET that start / finish the function.

Make sure that the function of interest is not inlined or optimized away, e.g. by making inputs and outputs volatile and attributing the function with attribute((noinline,noclone)). The additional overhead caused by the volatile accesses does not matter as it is ignored by PERF_START_CALL.

Timing Data and Random Values

Following functions return an unsigned 32-bit value:

avrtest_cycles();       Program cycles of simulated instructions
avrtest_insns();        Number of simulated instructions
avrtest_prand();        A 32-bit pseudo random number
avrtest_rand();         A 32-bit random number

The values are "owned" by the program and are distinct from the counters used by performance meters or that are displayed when avrtest terminates. Except rand, the values can be reset to their value at program start:

avrtest_reset_cycles();       Set cycles of simulated instructions to 0
avrtest_reset_insns();        Set number of simulated instructions to 0
avrtest_reset_prand();        Reset the pseudo random number generator
avrtest_reset_all();          Reset all of them

The simulator does not account cycles to syscalls. The values returned by avrtest_prand() do not depend on the host machine. They only depend on the number of calls since the last reset.

IEEE single Emulation

AVRtest supports syscalls like

float avrtest_mulf (float, float);
float avrtest_sinf (float);

in order to help with IEEE single implementation.

Supported functions with one float argument are: sin, asin, sinh, asinh, cos, acos, cosh, acosh, tan, atan, tanh, atanh, exp, log, log2, sqrt, cbrt, trunc, ceil, floor, round, fabs.

Supported functions with two float arguments are: mul, div, add, sub, ulp, pow, atan2, hypot, fmin, fmax, fmod.

⚠️ Don't forget to append f to the function name for the float versions.

AVRtest will terminate with an error when the host IEEE single cannot be used for emulation.

Here is an example for the usage of avrtest_sinf:

#include "avrtest.h"

float compute_sinf (float x)
{
    float y = avrtest_sinf (x);
    LOG_FMT_FLOAT ("sinf(%f) = ", x);
    LOG_FMT_FLOAT ("%f\n", y);
    return y;
}

IEEE double Emulation

AVRtest supports syscalls like

long double avrtest_mull (long double, long double);
long double avrtest_sinl (long double);

Plus, there are syscalls like

uint64_t avrtest_mul_d64 (uint64_t, uint64_t);
uint64_t double avrtest_sin_d64 (uint64_t);

with the same functionality, but they interpret uint64_t as IEEE double. The _d64 versions are available irrespective of the layout of long double.

Supported functions are the same like for IEEE single but with l instead of f at the end of the syscall name.

AVRtest will terminate with an error when the host IEEE double cannot be used for emulation. The functions are not available for Reduced Tiny.

Here is an example for the usage of avrtest_sqrtl:

#include "avrtest.h"

long double compute_sqrtl (long double)
{
    long double y = avrtest_sqrtl (x);
    // The host uses "double" for floating-point values, thus
    // use "%f" (or similar) to print and NOT "%Lf".
    LOG_FMT_LDOUBLE ("square-root(%f) = ", x);
    LOG_FMT_LDOUBLE ("%f\n", y);
    return y;
}

Assembler Support in avrtest.h

avrtest.h adds assembler support for a few AVRtest features:

avrtest_syscall <sysno>
LOG_OFF             ;; Same as "avrtest_syscall 0"
LOG_ON              ;; Same as "avrtest_syscall 1"
LOG_PUSH_OFF        ;; Same as "avrtest_syscall 9"
LOG_PUSH_ON         ;; Same as "avrtest_syscall 10"
LOG_POP             ;; Same as "avrtest_syscall 11"
AVRTEST_ABORT       ;; Same as "avrtest_syscall 31"
AVRTEST_EXIT        ;; Same as "avrtest_syscall 30", exit value = R25:R24.
AVRTEST_PUTCHAR     ;; Same as "avrtest_syscall 29", char = R24
AVRTEST_ABORT_2ND_HIT ;; Same as "avrtest_syscall 25"

avrtest_syscall <sysno> is an assembler macro which expands to

CPSE  R<sysno>, R<sysno>
.word 0xffff

i.e. it does not change any register and does not change the program status (SREG).

Example:

#include "avrtest.h"

.text
.global func

func:
    LOG_ON
    ret

And then call this function from assembler or from C code after declaring it with extern void func (void);.

Compiler Support

  • avr-gcc can be used with avrtest.h.

  • To date (2024), clang has bugs so that programs that include avrtest.h cannot be compiled and terminate with an error message (or internal error) for the used inline assembly.

  • avrtest.h supports avr-gcc -mint8 with an 8-bit int type.

    ⚠️ Notice that -mint8 is not a multilib option and therefore functions from libgcc and AVR-LibC will not work when they take an int, long or long long parameter. In particular, [__builtin_]exit() will take an 8-bit value, and AVRtest will likely get a trashed 16-bit return value. The same applies to returning from main. What works though is to use avrtest_exit() to terminate the program. Also notice that with -mint8 no 64-bit integer types like long long or uint64_t are available.

About

AVRtest is an AVR core simulator used for regression testing of avr-gcc and parts of AVR-LibC.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published