Cortex-M0: Wrong registers in inline asm

[42Bastian]

Version:gcc version 7.3.1 20180425 [linaro-7.3-2018.05 revision d29120a424ecfbc167ef90065c0eeb7f91977701]

The compiler uses illegal registers for inline assembly.
Example:

static int MULSHIFT32(int a, int b)
{
  short t0,t1,t2;
  __asm__ volatile (
    " .syntax unified\n"
          "uxth %2,%0 // b\n"
          "lsrs %0,%0,#16 // a\n"
          "lsrs %3,%1,#16 // c\n"
          "uxth %1,%1 // d\n"
          "movs %4,%1 // d\n"
          "mul %1,%2 // bd\n"
          "mul %4,%0 // ad\n"
          "mul %0,%3 // ac\n"
          "mul %3,%2 // bc\n"
          "lsls %2,%4,#16 // ad => d0\n"
          "lsrs %4,%4,#16 // ad => 0ay"
          "adds %1,%1,%2 // bd + d0\n"
          "adcs %0,%4 // ac + 0a + C\n"
          "lsls %2,%3,#16 // bc => c0\n"
          "lsrs %3,%3,#16 // bc => 0b\n"
          "adds %1,%1,%2 // bd + c0\n"
          "adcs %0,%3 // ac + 0b + C\n"

          :"+r"(a):"r"(b),"r"(t0),"r"(t1),"r"(t2));
  return a;
}

Generated code:
  .syntax unified
uxth r6,r4 // b
lsrs r4,r4,#16 // a
lsrs r7,r5,#16 // c
uxth r5,r5 // d
movs ip,r5 // d
mul r5,r6 // bd
mul ip,r4 // ad
mul r4,r7 // ac
mul r7,r6 // bc
lsls r6,ip,#16 // ad => d0
lsrs ip,ip,#16 // ad => 0ayadds r5,r5,r6 // bd + d0
adcs r4,ip // ac + 0a + C
lsls r6,r7,#16 // bc => c0
lsrs r7,r7,#16 // bc => 0b
adds r5,r5,r6 // bd + c0
adcs r4,r7 // ac + 0b + C

(IP is r12, but only registers r0..7 are allowed)

[42Bastian (bastian-schick)]

Test source attached.
Compile options: -S -mthumb -mcpu=cortex-m0 -O2

[Mihail-Calin Ionescu (mihail.ionescu)]

Hi,

This page is for the GNU Arm Embedded Toolchain. You seem to be using the Linaro toolchain, in which case, the issue should be reported here: https://bugs.linaro.org/ (Product: GCC).

I could not reproduce the issue on GNU Arm Embedded, but it looks like it is caused by the use of "r" constraints; using "l" (the r0-r7 constraint for Thumb)[1] should fix the problem .

[1] https://gcc.gnu.org/onlinedocs/gcc-7.3.0/gcc/Machine-Constraints.html#Machine-Constraints

Regards,
   Mihail

[42Bastian (bastian-schick)]

Hi

oh, thought those are the same tool chains. My bad.

Thanks for pointing to the right constraints.

I'll try it, and in case it still does not work report on the proper site.