Skip to content

Latest commit

 

History

History
296 lines (231 loc) · 15.3 KB

printf.md

File metadata and controls

296 lines (231 loc) · 15.3 KB

Printf and Scanf in Picolibc

Embedded systems are often tightly constrained on code space, which makes it important to be able to only include code and data actually needed by the application. The format-string based interface to printf and scanf makes it very difficult to determine which conversion operations might be needed by the application.

Picolibc handles this by providing multiple printf and scanf implementations in the library with varying levels of conversion support. The application developer selects among these at compile and link time based on knowledge of the requirements of the application. The selection is done using a compiler command line definition of a preprocessor symbol which maps the public printf and scanf names to internal names in the linker.

The function name mapping happens in picolibc.specs file which scans the compiler command line looking for the preprocessor token and uses that to add --defsym options when linking. This means the preprocessor definition must be set on the command line and not in a file.

Using the defsym approach, rather than renaming the functions with the C preprocessor has a couple of benefits:

  • Printf uses within picolibc (which are all integer-only) now share whichever printf the application needs, avoiding including both integer-only and float versions

  • Printf optimizations in the compiler which map to puts or fputs now work even when using integer-only or float printf functions.

Because the linker gets --defsym flags for both vfprintf and vfscanf, those functions will always get included in the link process. To avoid linking them into the application when they aren't otherwise needed, picolibc.specs includes the --gc-sections linker flag. This causes those functions to be discarded if they aren't used in the application.

However, the defsym approach does not work with link-time optimization. In that case, applications can only use the default function. For that case, the library needs to allow a different default version to be selected while compiling the library.

Printf and Scanf levels in Picolibc

There are five levels of printf support provided by Picolibc that can be selected when building applications. One of these is the default used when no symbol definitions are applied; that is selected using the picolibc built-time option, -Dformat-default, which defaults to double, selecting PICOLIBC_DOUBLE_PRINTF_SCANF. These are listed in order of decreasing functionality and size.

  • PICOLIBC_DOUBLE_PRINTF_SCANF (default when -Dformat-default=double). This offers full printf functionality, including both float and double conversions along with the C99 extensions and POSIX positional parameters. There is optional support for the upcoming %b format specifier via the io-percent-b setting and optional support for long double values via the io-long-double setting. The picolibc.specs stanza that matches this option maps __d_vfprintf to vfprintf and __d_vfscanf to vfscanf. This is equivalent to adding this when linking your application:

    cc -Wl,--defsym=vfprintf=__d_vfprintf -Wl,--defsym=vfscanf=__d_vfscanf

    If you're using a linker that supports -alias instead of --defsym, you would use:

    cc -Wl,-alias,___d_vfprintf,_vfprintf -Wl,-alias,___d_vfscanf,_vfscanf

  • PICOLIBC_FLOAT_PRINTF_SCANF (default when -Dformat-default=float). This provides support for float, but not double or long double conversions. When picolibc.specs finds -DPICOLIBC_FLOAT_PRINTF_SCANF on the command line during linking, it maps __f_vfprintf to vfprintf and __f_vfscanf to vfscanf. This is equivalent to adding this when linking your application:

    cc -Wl,--defsym=vfprintf=__f_vfprintf -Wl,--defsym=vfscanf=__f_vfscanf

    If you're using a linker that supports -alias instead of --defsym, you would use:

    cc -Wl,-alias,___f_vfprintf,_vfprintf -Wl,-alias,___f_vfscanf,_vfscanf

  • PICOLIBC_LONG_LONG_PRINTF_SCANF (default when -Dformat-default=long-long). This removes support for all float and double conversions and makes support for C99 extensions and POSIX positional parameters optional via the io-c99-formats and io-pos-args settings. The picolibc.specs stanza that matches this option maps __l_vfprintf to vfprintf and __l_vfscanf to vfscanf. This is equivalent to adding this when linking your application:

    cc -Wl,--defsym=vfprintf=__l_vfprintf -Wl,--defsym=vfscanf=__l_vfscanf

    If you're using a linker that supports -alias instead of --defsym, you would use:

    cc -Wl,-alias,___l_vfprintf,_vfprintf -Wl,-alias,___l_vfscanf,_vfscanf

  • PICOLIBC_INTEGER_PRINTF_SCANF (default when -Dformat-default=integer). This removes support for long long conversions where those values are larger than long values. The picolibc.specs stanza that matches this option maps __i_vfprintf to vfprintf and __i_vfscanf to vfscanf. This is equivalent to adding this when linking your application:

    cc -Wl,--defsym=vfprintf=__i_vfprintf -Wl,--defsym=vfscanf=__i_vfscanf

    If you're using a linker that supports -alias instead of --defsym, you would use:

    cc -Wl,-alias,___i_vfprintf,_vfprintf -Wl,-alias,___i_vfscanf,_vfscanf

  • PICOLIBC_MINIMAL_PRINTF_SCANF (default when -Dformat-default=minimal). This removes support for width and precision, alternate presentation modes and alternate sign presentation. All of these are still parsed correctly, so applications needn't adjust format strings in most cases, but the output will not be the same. It also disables any support for positional arguments and %b formats that might be selected with io-pos-args or io-percent-b. The picolibc.specs stanza that matches this option maps __m_vfprintf to vfprintf and __m_vfscanf to vfscanf. This is equivalent to adding this when linking your application:

    cc -Wl,--defsym=vfprintf=__m_vfprintf -Wl,--defsym=vfscanf=__m_vfscanf

    If you're using a linker that supports -alias instead of --defsym, you would use:

    cc -Wl,-alias,___m_vfprintf,_vfprintf -Wl,-alias,___m_vfscanf,_vfscanf

PICOLIBC_FLOAT_PRINTF_SCANF requires a special macro for float values: printf_float. To make it easier to switch between that and the default level, that macro is also correctly defined for the other two levels.

Here's an example program to experiment with these options:

#include <stdio.h>

void main(void) {
	printf(" 2⁶¹ = %lld π ≃ %.17g\n", 1ll << 61, printf_float(3.141592653589793));
}

Now we can build and run it with the double options:

$ arm-none-eabi-gcc -Os -march=armv7-m --specs=picolibc.specs --oslib=semihost --crt0=hosted -Wl,--defsym=__flash=0 -Wl,--defsym=__flash_size=0x00200000 -Wl,--defsym=__ram=0x20000000 -Wl,--defsym=__ram_size=0x200000 -DPICOLIBC_DOUBLE_PRINTF_SCANF -o printf.elf printf.c
$ arm-none-eabi-size printf.elf
   text	   data	    bss	    dec	    hex	filename
   8088	     80	   4104	  12272	   2ff0	printf.elf
$ qemu-system-arm -chardev stdio,id=stdio0 -semihosting-config enable=on,chardev=stdio0 -monitor none -serial none -machine mps2-an385,accel=tcg -kernel printf.elf -nographic
 2⁶¹ = 2305843009213693952 π ≃ 3.141592653589793

Switching to float-only reduces the size but lets this still work, although the floating point value has reduced precision:

$ arm-none-eabi-gcc -DPICOLIBC_FLOAT_PRINTF_SCANF -Os -march=armv7-m --specs=picolibc.specs --oslib=semihost --crt0=hosted -Wl,--defsym=__flash=0 -Wl,--defsym=__flash_size=0x00200000 -Wl,--defsym=__ram=0x20000000 -Wl,--defsym=__ram_size=0x200000 -o printf-float.elf printf.c
$ arm-none-eabi-size printf-float.elf
   text	   data	    bss	    dec	    hex	filename
   6792	     80	   4104	  10976	   2ae0	printf-float.elf
$ qemu-system-arm -chardev stdio,id=stdio0 -semihosting-config enable=on,chardev=stdio0 -monitor none -serial none -machine mps2-an385,accel=tcg -kernel printf-float.elf -nographic
 2⁶¹ = 2305843009213693952 π ≃ 3.1415927

Selecting the long-long variant reduces the size further, but now the floating point value is not displayed correctly:

$ arm-none-eabi-gcc -DPICOLIBC_LONG_LONG_PRINTF_SCANF -Os -march=armv7-m --specs=picolibc.specs --oslib=semihost --crt0=hosted -Wl,--defsym=__flash=0 -Wl,--defsym=__flash_size=0x00200000 -Wl,--defsym=__ram=0x20000000 -Wl,--defsym=__ram_size=0x200000 -o printf-long-long.elf printf.c
$ arm-none-eabi-size printf-long-long.elf
   text	   data	    bss	    dec	    hex	filename
   2216	     80	   4104	   6400	   1900	printf-long-long.elf
$ qemu-system-arm -chardev stdio,id=stdio0 -semihosting-config enable=on,chardev=stdio0 -monitor none -serial none -machine mps2-an385,accel=tcg -kernel printf-long-long.elf -nographic
 2⁶¹ = 2305843009213693952 π ≃ *float*

Going to integer-only reduces the size even further, but now it doesn't output the long long value correctly:

$ arm-none-eabi-gcc -DPICOLIBC_INTEGER_PRINTF_SCANF -Os -march=armv7-m --specs=picolibc.specs --oslib=semihost --crt0=hosted -Wl,--defsym=__flash=0 -Wl,--defsym=__flash_size=0x00200000 -Wl,--defsym=__ram=0x20000000 -Wl,--defsym=__ram_size=0x200000 -o printf-int.elf printf.c
$ arm-none-eabi-size printf-int.elf
   text	   data	    bss	    dec	    hex	filename
   2056	     80	   4104	   6240	   1860	printf-int.elf
$ qemu-system-arm -chardev stdio,id=stdio0 -semihosting-config enable=on,chardev=stdio0 -monitor none -serial none -machine mps2-an385,accel=tcg -kernel printf-int.elf -nographic
 2⁶¹ = 0 π ≃ *float*

To shrink things still further, use the 'minimal' variant. This doesn't even look at the %g formatting instruction, so that value displays as '%g'.

$ arm-none-eabi-gcc -DPICOLIBC_MINIMAL_PRINTF_SCANF -Os -march=armv7-m --specs=picolibc.specs --oslib=semihost --crt0=hosted -Wl,--defsym=__flash=0 -Wl,--defsym=__flash_size=0x00200000 -Wl,--defsym=__ram=0x20000000 -Wl,--defsym=__ram_size=0x200000 -o printf-min.elf printf.c
$ arm-none-eabi-size printf-min.elf
   text	   data	    bss	    dec	    hex	filename
   1520	     80	   4104	   5704	   1648	printf-min.elf
$ qemu-system-arm -chardev stdio,id=stdio0 -semihosting-config enable=on,chardev=stdio0 -monitor none -serial none -machine mps2-an385,accel=tcg -kernel printf-min.elf -nographic
 2⁶¹ = 0 π ≃ %g

There's a build-time option available that enables long-long support in the minimal printf variant, -Dminimal-io-long-long=true. Building with that increases the size modestly while fixing the long long output:

$ arm-none-eabi-size printf-min.elf
   text	   data	    bss	    dec	    hex	filename
   1632	     80	   4104	   5816	   16b8	printf-min.elf
$ qemu-system-arm -chardev stdio,id=stdio0 -semihosting-config enable=on,chardev=stdio0 -monitor none -serial none -machine mps2-an385,accel=tcg -kernel printf-min.elf -nographic
 2⁶¹ = 2305843009213693952 π ≃ %g

Picolibc build options for printf and scanf options

In addition to the application build-time options, picolibc includes a number of picolibc build-time options to control the feature set (and hence the size) of the library:

  • -Dio-c99-formats=true This option controls whether support for the C99 type-specific format modifiers 'j', 'z' and 't' and the hex float format 'a' are included in the long-long, integer and minimal printf and scanf variants. Support for the C99 format specifiers like PRId8 is always provided. This option is enabled by default.

  • -Dio-pos-args=true This option add support for C99 positional arguments to the long long and integer printf and scanf variant (e.g. "%1$"). Positional arguments are always supported in the double and float variants and never supported in the minimal variant. This option is disabled by default.

  • -Dio-long-long=true This deprecated option controls whether support for long long types is included in the integer variant of printf and scanf. Instead of using this option, applications should select the long long printf and scanf variants. This option is disabled by default.

  • -Dminimal-io-long-long=true This option controls whether support for long long types is included in the minimal variant of printf and scanf. This option is disabled by default.

  • -Dio-float-exact=true This option, which is enabled by default, controls whether the tinystdio code uses exact algorithms for printf and scanf. When enabled, printing at least 9 digits (e.g. "%.9g") for 32-bit floats and 17 digits (e.g. "%.17g") for 64-bit floats ensures that passing the output back to scanf will exactly re-create the original value.

  • -Dio-long-double=true This option add support for long double parameters. That is limited to systems using 80- and 128- bit long doubles, or systems for which long double is the same as double. This option is disabled by default

  • -Datomic-ungetc=true This option, which is enabled by default, controls whether getc/ungetc use atomic instruction sequences to make them re-entrant. Without this option, multiple threads using getc and ungetc may corrupt the state of the input buffer.

  • -Dprintf-small-ultoa=true This option, which is enabled by default, switches printf's binary-to-decimal conversion code to a version which avoids soft division for values larger than the machine registers as those functions are often quite large and slow. Applications using soft division on large values elsewhere will save space by disabling this option as that avoids including custom divide-and-modulus-by-ten implementations.

For compatibility with newlib printf and scanf functionality, picolibc can be compiled with the original newlib stdio code. That greatly increases the code and data sizes of the library, including adding a requirement for heap support in the run time system. Here are the picolibc build options for that code:

  • -Dtinystdio=false This disables the tinystdio code and uses original newlib stdio code.

  • -Dnewlib-stdio64=true This option changes the newlib stdio code to use 64 bit values for file sizes and offsets. It also adds 64-bit versions of stdio interfaces which are defined with types which may be 32-bits (like 'long'). This option is enabled by default.

Newlib floating point printf

To build the printf sample program using the original newlib stdio code, the first step is to build picolibc with the right options. The 'nano' printf code doesn't support long-long integer output, so we can't use that, and we need to enable long-long and floating point support in the full newlib stdio code:

    $ mkdir build-arm; cd build-arm
    $ ../scripts/do-arm-configure -Dtinystdio=false -Dio-long-long=true -Dnewlib-io-float=true
    $ ninja install

Now we can build the example with the library:

    $ arm-none-eabi-gcc -Os -march=armv7-m --specs=picolibc.specs --oslib=semihost --crt0=hosted -Wl,--defsym=__flash=0 -Wl,--defsym=__flash_size=0x00200000 -Wl,--defsym=__ram=0x20000000 -Wl,--defsym=__ram_size=0x200000 -o printf.elf printf.c
$ arm-none-eabi-size printf.elf
   text	   data	    bss	    dec	    hex	filename
  16632	    460	   4944	  22036	   5614	printf.elf
    $ qemu-system-arm -chardev stdio,id=stdio0 -semihosting-config enable=on,chardev=stdio0 -monitor none -serial none -machine mps2-an385,accel=tcg -kernel printf.elf -nographic
 2⁶¹ = 2305843009213693952 π ≃ 3.1415926535897931

This also uses 1252 bytes of space from the heap at runtime. Tinystdio saves 8544 bytes of text space and a total of 2472 bytes of data space.