[size] Redundant uxth/sxth insn are generated

[Yao Qi]

Compile this case by FSF GCC trunk reveals redundant uxth/sxth instructions.

short unPack( unsigned char c )
{
    /* Only want lower four bit nibble */
    c = c & (unsigned char)0x0F ;

    if( c > 7 ) {
        /* Negative nibble */
        return( ( short )( c - 16 ) ) ;
    }
    else
    {
        /* positive nibble */
        return( ( short )c ) ;
    }
}

# arm-none-linux-gnueabi-gcc -march=armv7-a -mthumb -Os unpack.c -S
unPack:
        @ args = 0, pretend = 0, frame = 0
        @ frame_needed = 0, uses_anonymous_args = 0
        @ link register save eliminated.
        and r0, r0, #15
        cmp r0, #7
        bls .L3
        subs r0, r0, #16
        uxth r0, r0
.L3:
        sxth r0, r0
        bx lr

uxth/sxth are still there with option {thumb,arm} {O2,O3}

[Yao Qi (yao-codesourcery)]

It is found in eembc/automotive/idctrn01/bmark.c

[Andrew Stubbs (ams-codesourcery)]

This program has the same problem:

  unsigned char
  f (unsigned int c)
  {
    if (c < 50)
      return c;
    return 0;
  }

Upstream GCC (pre-4.6) with "-mcpu=cortex-a8 -O2" gives:

        cmp r0, #49
        uxtbls r0, r0
        movhi r0, #0
        bx lr

As far as I can see, the only compiler pass that could "know" that the zero-extend is redundant is the VRP pass (Value Range Propagation). It's not totally clear to me whether it's supposed to be able to remove/annotate the type conversion, and it's not working, or whether this is just not addressed anywhere in the compiler.

[Ramana Radhakrishnan (ramana)]

I don't see how this can be done without VRP at the RTL level to show that the value of c is bounded by 50 above.

Ramana

[Ramana Radhakrishnan (ramana)]

Unless we can get VRP to deal with this in RTL I'm not sure how we could deal with this .

Ramana

[Kugan Vivekanandarajah (kugan-vivekanandarajah)]

Linaro bug created at https://bugs.linaro.org/show_bug.cgi?id=411