std::mt19937 with ARM Compiler 6 uses sys_open and breaks retarget.c

I suggest simplifying to a really simple test case, and working back up. I tried to build a simple test for Cortex-M4 based on your code fragment: no build errors. (Also MDK 5.30, ARMCLANG 6.14, configured to use STDIN with 'User' code) But my case was only built to test compilation, not to run anywhere.

#include <random>
#include <iostream>

extern "C"{
// Set RTE option Compiler::I/O::STDIN to 'User'
int stdin_getchar (void);
}

std::mt19937 mt;
std::uniform_int_distribution< uint32_t > dist(0);
auto a = dist( mt );

int main(void)
{
   std::mt19937 mt;
   std::uniform_int_distribution< uint32_t > dist(0);

   auto b = dist( mt );
   //return a + b;
   std::cout << "Hello, world!\n" << a << std::endl;
   return 0;
}

The random number classes do define stream operator templates, so shouldn't be an issue unless you use stream operators with the distributions. I had no issues building with iostream or not. 

I'm curious about your file retarget.c -- the RTE manager will add retarget_io.c to your project when you configure IO using the RTE manager. Why and how did your retarget.c need 'tuning' for ARMCLANG? 

There is helpful documentation for retargeting I/O on the Keil website, here. Per that documentation, uVision doesn't support semihosting. This support article might also help.

This reminded me of an issue that I stumbled in three years ago - community.arm.com/.../using-itm-retarget-with-keil-software-packs

Bloated code size is the reason why I still use a hand-written retarget.cpp (not .c, that was a misprint earlier, sorry) instead of RTE Managing it.

Here it is, more or less:

#if ( (__ARMCC_VERSION) && (__ARMCC_VERSION >= 6010050) )

    asm(".global __use_no_semihosting_swi\n");
    
#elif __CC_ARM

    #pragma import(__use_no_semihosting_swi)
    
    namespace std { struct __FILE { int handle;} ; }

#endif

#include <stdio.h>
#include <rt_sys.h>
#include <rt_misc.h>


std::FILE std::__stdout;
std::FILE std::__stdin;
std::FILE std::__stderr;

extern "C"
{ 
    int fputc(int c, FILE *f)
    {
        return ITM_SendChar(c);
    }

    int fgetc(FILE *f)
    {
        char ch = 0;

        return((int)ch);
    }

    int ferror(FILE *f)
    {
        /* Your implementation of ferror */
        return EOF;
    }

    void _ttywrch(int ch)
    {
        ITM_SendChar(ch);            
    }
    
    char *_sys_command_string(char *cmd, int len)
    {
        return NULL;
    }
    
    void _sys_exit(int return_code) 
    {
        while(1);
    }        
}

So 'tuning for armclang' was this line about semihosting - asm(".global __use_no_semihosting_swi\n");

On its own this works very nicely, at least in Simulator (which is all I need at the moment).

This explains the original linker error: _sys_open is needed by the iostream functions pulled in from <random>, your retarget.cpp doesn't provide a definition and you disabled semihosting, so the linker complained that it couldn't find a suitable _sys_open. Allowing semihosting pulls in version of _sys_open that uses semihosting, which is why you end up on a breakpoint before main. (If you want to decode the semihosting call, see Chapter 7 in the ARM Compiler Software Development Guide.)

The bulk of your 84KB binary is most likely due to locale and other I/O support. <random> includes <istream> and <ostream>, which is enough to pull in the IO libraries, even if your <random> objects don't use streams.  Large binary sizes aren't really the fault of retarget_io.c -- it's just a very small part of the overall puzzle.