Can RealView compiler generate SMLAL/UMLAL instructions?

Yes, the compiler does generate SMLAL/UMLAL instructions.

Simple test code (UMLAL):

unsigned long long mac_test (unsigned long *a, unsigned long *b, int cnt) {
  unsigned long long res = 0;

  while (cnt--) {
    res += (unsigned long long)*a++ * (unsigned long long)*b++;
  }
  return (res);
}

Compiler output:

                  mac_test PROC
;;;1      unsigned long long mac_test (unsigned long *a, unsigned long *b, int cnt) {
000000  b570              PUSH     {r4-r6,lr}
000002  4603              MOV      r3,r0
000004  460c              MOV      r4,r1
000006  2000              MOVS     r0,#0
000008  4601              MOV      r1,r0
;;;2        unsigned long long res = 0;
;;;3
;;;4        while (cnt--) {
00000a  e005              B        |L1.24|
                  |L1.12|
;;;5          res += (unsigned long long)*a++ * (unsigned long long)*b++;
00000c  cb20              LDM      r3!,{r5}
00000e  cc40              LDM      r4!,{r6}
000010  fba56506          UMULL    r6,r5,r5,r6
000014  1830              ADDS     r0,r6,r0
000016  4169              ADCS     r1,r1,r5
                  |L1.24|
000018  1e52              SUBS     r2,r2,#1              ;4
00001a  d2f7              BCS      |L1.12|
;;;6        }
;;;7        return (res);
;;;8      }
00001c  bd70              POP      {r4-r6,pc}
                          ENDP

Yes, the compiler does generate SMLAL/UMLAL instructions.

Simple test code (UMLAL):

unsigned long long mac_test (unsigned long *a, unsigned long *b, int cnt) {
  unsigned long long res = 0;

  while (cnt--) {
    res += (unsigned long long)*a++ * (unsigned long long)*b++;
  }
  return (res);
}

Compiler output:

                  mac_test PROC
;;;1      unsigned long long mac_test (unsigned long *a, unsigned long *b, int cnt) {
000000  b570              PUSH     {r4-r6,lr}
000002  4603              MOV      r3,r0
000004  460c              MOV      r4,r1
000006  2000              MOVS     r0,#0
000008  4601              MOV      r1,r0
;;;2        unsigned long long res = 0;
;;;3
;;;4        while (cnt--) {
00000a  e005              B        |L1.24|
                  |L1.12|
;;;5          res += (unsigned long long)*a++ * (unsigned long long)*b++;
00000c  cb20              LDM      r3!,{r5}
00000e  cc40              LDM      r4!,{r6}
000010  fba56506          UMULL    r6,r5,r5,r6
000014  1830              ADDS     r0,r6,r0
000016  4169              ADCS     r1,r1,r5
                  |L1.24|
000018  1e52              SUBS     r2,r2,#1              ;4
00001a  d2f7              BCS      |L1.12|
;;;6        }
;;;7        return (res);
;;;8      }
00001c  bd70              POP      {r4-r6,pc}
                          ENDP

000010  fba56506          UMULL    r6,r5,r5,r6
000014  1830              ADDS     r0,r6,r0
000016  4169              ADCS     r1,r1,r5

Ironically, this compiler output doesn't show UMLAL being used here at all. But with my local versions of armcc (RVCT4.0 [Build 677], RVCT4.0 [Build 821]), UMLAL is generated indeed.

Regards
Marcus
http://www.doulos.com/arm/

Hi Marcus,

Thanks for the tip - I'm using armcc 4.1 Build 561 from the Keil UV4 installation. What are the command line options you use to see the UMLAL instructions?

Thanks,
Andrew

You mean other than "--cpu=cortex-m3"? Nothing in case of my Q&D test. I think that by default, "-O2 -Ospace" are selected by the compiler.

Regarding your statement

> Once I've forced the compiler to generate close to what I want, in this case the MAC
> instructions, I throw away the C-code and tweak the assembly.

May I ask why? The RealView compiler is fairly good at generating very efficient code. Implementing things in C increases the chance that you will remember what you did two weeks from now. I don't find many places where I could have outsmarted the compiler.

--
Marcus

>> Once I've forced the compiler to generate close to what I want, in this case the MAC
>> instructions, I throw away the C-code and tweak the assembly.

> May I ask why? The RealView compiler is fairly good at generating very efficient code.
> Implementing things in C increases the chance that you will remember what you did two
> weeks from now. I don't find many places where I could have outsmarted the compiler.

Exactly. In this particular case, if the compiler uses the MAC instructions I'm happy. Otherwise I might grudgingly resort to assembly but only if our profiling shows that optimizing this particular operation is worthwhile. Since our application is very power sensitive we want to get the RMS calculation done as quickly as possible so we can put the CPU back to sleep.

Andrew