Assembler errors in code built with lto

static inline void delay_4cy(unsigned int cy)
{
        __asm__ __volatile__
        (
            "loop%=:" "\n\t"
            " subs %[cnt],#1" "\n\t"
            " bne loop%=" "\n\t"
            : [cnt]"+r"(cy) // output: +r means input+output
            : // input:
            : "cc" // clobbers:
        );
}

When buld with -flto -mcpu=cortex-m1 -mthumb -msoft-float -Os

get error:

C:\...\ccsln0ZZ.s: Assembler messages:
C:\...\ccsln0ZZ.s:2701: Error: instruction not supported in Thumb16 mode -- `subs r3,#1'
lto-wrapper.exe: fatal error: arm-none-eabi-gcc returned 1 exit status

Strange case#2
When my previous function delay_4cycase compile with -mcpu=cortex-m3 error message is gone.

Strange case#3
When in my previous function delay_4cycase i change instruction SUBS to SUB error message is gone, but in LST and LSS files I see SUBS!!!

080034bc <loop1613>:
 80034bc: 3b01 subs r3, #1
 80034be: d1fd bne.n 80034bc <loop1613>

Hi demiurg_spb,

The reason for this error is that inline assembly is considered divided for Thumb-1 instead of unified. This has nothing to do with lto (I can reproduce with your example without -flto). What happens is that the compiler will prefix the inline assembly with .syntax divided and thus the assembler will expect sub instead of subs.

Case #2 is explained by the fact that Thumb-2 will use unified syntax by default for inline assembly. When compiling for Cortex-M3, you are targeting Thumb-2 so the compiler will prefix the inline assembly with .syntax unified.

I'm not sure what you mean by LST and LSS in case #3 but I guess it's some form of disassembly. The reason for this is that this difference of syntax is only relevant to the assembler. sub in divided syntax is the same instruction as subs in unified syntax. It's akin to the difference between x:=1 in one programming language when another uses x=1. Once assembled, it's the same underlying instruction. The disassembler will choose one syntax to display its work which is independent to the one you used for the assembler. In this example the assembler gets fed divided syntax (sub) and the disassembler prefer to display unified syntax (subs).

Hope this helps.

Best regards.

Thank you very much!
Allow me to share the results!
The function delay_cycles generates a delay with an accuracy of 1 clock cycle:

delay.h:
...
//=============================================================================
static __is_always_inline void delay_4cycles(uint32_t cy) // +1 cycle
{
 #if ARCH_PIPELINE_RELOAD_CYCLES<2
 # define EXTRA_NOP_CYCLES "nop"
 #else
 # define EXTRA_NOP_CYCLES ""
 #endif

 __asm__ __volatile__
 (
  ".syntax unified" "\n\t" // is to prevent CM0,CM1 non-unified sintax
  "loop%=:" "\n\t"
  " subs %[cnt],#1" "\n\t"
          EXTRA_NOP_CYCLES "\n\t"
  " bne loop%=" "\n\t"
  : [cnt]"+r"(cy) // output: +r means input+output http://www.nongnu.org/avr-libc/user-manual/inline_asm.html
  : // input:
  : "cc" // clobbers:
 );
}

//=============================================================================
static __is_always_inline void delay_cycles(uint32_t x)
{
 #define MAXNOPS 4 // delay_4cycles
 if (x<=MAXNOPS)
 {
  if (x==1) {nop();}
  else if (x==2) {nop(); nop();}
  else if (x==3) {nop(); nop(); nop();}
  else if (x==4) {nop(); nop(); nop(); nop();}
 }
 else // because of +1 cycle inside delay_4cycles
 {
   uint32_t rem = (x-1)%MAXNOPS;
   if (rem==1) {nop();}
   else if (rem==2) {nop(); nop();}
   else if (rem==3) {nop(); nop(); nop();}
   if ((x=(x-1)/MAXNOPS)) delay_4cycles(x); // if need more then 4 nop loop is more optimal
 }
}

BUT with gcc version 4.9.3 even witout -flto and -fomit-frame-pointer i get my project compilation fail

C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s: Assembler messages:
C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s:194: Error: cannot honor width suffix -- `lsl r0,r0,r4'
C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s:201: Error: cannot honor width suffix -- `orr r0,r1'
C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s:231: Error: lo register required -- `sub r2,r2,#1'
C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s:279: Error: cannot honor width suffix -- `add r6,r0,r1'
C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s:292: Error: lo register required -- `add r4,r4,#1'
C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s:328: Error: cannot honor width suffix -- `add r6,r0,r1'
C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s:336: Error: cannot honor width suffix -- `mov r0,#0'
C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s:340: Error: lo register required -- `add r4,r4,#1'
C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s:372: Error: cannot honor width suffix -- `add r5,r0,r1'
C:\Users\A7DDC~1.IVA\AppData\Local\Temp\ccxI4e8j.s:383: Error: lo register required -- `add r4,r4,#1'

I have no such problems with gcc-5.2.1

This happen only if CM1 target is selected (with CM3 is OK).