long long multiply-and-add far from optimal

[Andrew Stubbs]

This testcase compiles correctly, but give suboptimal code on ARM:

long long foolong (long long x, short *a, short *b)
   {
       return x + (long long)*a * (long long)*b;
   }

Here's the output from upstream gcc:

        ldrsh r3, [r3, #0]
        ldrsh r2, [r2, #0]
        push {r4, r5}
        asrs r4, r3, #31
        asrs r5, r2, #31
        mul r4, r2, r4
        mla r4, r3, r5, r4
        umull r2, r3, r2, r3
        adds r3, r4, r3
        adds r0, r0, r2
        adc r1, r1, r3
        pop {r4, r5}
        bx lr

It really should look more like this:

           ldrh r2, [r2, #0]
           ldrh r3, [r3, #0]
           smlalbb r0, r1, r2, r3
           bx lr

[Michael Hope (michaelh1)]

It's the same in gcc-4.6.0-RC-20110314

[Ramana Radhakrishnan (ramana)]

A recent Linaro GCC 4.6 generates

foolong:
        @ args = 0, pretend = 0, frame = 0
        @ frame_needed = 0, uses_anonymous_args = 0
        @ link register save eliminated.
        ldrh r2, [r2, #0]
        ldrh r3, [r3, #0]
        smlalbb r0, r1, r2, r3
        bx lr

I think this should be marked fixed as a result of the multiplication improvements that Andrew did

Ramana