Is this an optimisation bug?

Here is some 开发者_StackOverflow中文版output from my compiler in assembler. It's the MPLAB C30 C compiler, based on GCC v3.23, for a dsPIC33FJ128GP802, a 16-bit moderately high speed DSP/MCU.

212:               inline uint16_t ror_16(uint16_t word, int num)
213:               {
 078C4  608270     and.w w1,#16,w4
 078C6  DE0204     lsr w0,w4,w4
 078C8  780101     mov.w w1,w2
 078CA  EA8102     com.w w2,w2
 078CC  EA8183     com.w w3,w3
 078CE  610170     and.w w2,#16,w2
 078D0  DD0002     sl w0,w2,w0
 078D2  700004     ior.w w0,w4,w0
214:                num &= 16; // limit to 16 shifts
215:                return (word >> num) | (word << (16 - num));
216:               }
 078D4  060000     return

In particular I'm interested in the following:

and.w w1,#16,w4         AND W1 with 16, storing result in W4
lsr w0,w4,w4            Logical shift right W0 by W4 times storing result in W4
mov.w w1,w2             Move W1 to W2
com.w w2,w2             Logical complement of W2 stored in W2
com.w w3,w3             Logical complement of W3 stored in W3   <-- This line is confusing me
and.w w2,#16,w2         AND W2 with 16, storing result in W2
sl w0,w2,w0             (Logical) shift left W0 left by W2 times storing result in W0
ior.w w0,w4,w0          Inclusive OR of W0 and W4 stored in W0
return                  Return from function

W0..W15 are an array of sixteen on chip 16-bit registers.

Effectively this simplifies to (in a primitive RTL):

W4 := W1 & 16
W4 := W0 LSR W4
W1 := W2
W2 := COM W2
W3 := COM W3
W2 := W2 & 16
W0 := W0 SL W2
W0 := W0 | W4
return

Now I'm confused at why it is computing the complement of W3 when there are only two passed arguments (W0 and W1 - it uses the W array for passing arguments to functions for functions with smaller arguments.) W3 is never used in the calculation, and is never returned. In fact it doesn't even appear to have data in it: nothing is stored in it by the function, and only the callee will have some data in it (although functions are not required to preserve W0..W7 so the callee should not be relying on it.) Why is it included in the code? Is it just a compiler glitch or error, or am I missing something?

And it's not just this code - I'm seeing the very same oddness in other parts of code. Even code designed to calculate things like complements of a 16-bit variable always seem to use two registers. It has me lost!

The function is not coded to limit the count to 16 (which I suspect you mean 0 to 16) but limits it to 0 or 16.

Instead of

num &= 16

you perhaps want

num > 16 ? (num & 15) : num

Re: the question, since the function is inlined, it can only be answered by looking at where it is used. Perhaps W3 is used for something in the surrounding code. Or it could be a "bug," but one that only has performance, not correctness, impact.

If num can be only 0 or 16 (as in your code) then (16 - num) can also only be 16 or 0, which is why C30 can do the "subtract" with a complement and mask.

FYI, when I don't inline, in C30 I get:

34:                uint16_t ror_16(uint16_t word, int num)
35:                {
 05AF4  608170     and.w 0x0002,#16,0x0004
 05AF6  DE0102     lsr 0x0000,0x0004,0x0004
 05AF8  EA8081     com.w 0x0002,0x0002
 05AFA  6080F0     and.w 0x0002,#16,0x0002
 05AFC  DD0001     sl 0x0000,0x0002,0x0000
 05AFE  700002     ior.w 0x0000,0x0004,0x0000
36:                    num &= 16; // limit to 16 shifts
37:                    return (word >> num) | (word << (16 - num));
38:                }
 05B00  060000     return

I might code this as

34:                uint16_t ror_16(uint16_t word, int num)
35:                {
 05AF4  780100     mov.w 0x0000,0x0004
36:                    num &= 15; // mod 16
 05AF6  60806F     and.w 0x0002,#15,0x0000
37:                    return (num == 0) ? word : ((word >> num) | (word << (16 - num)));
 05AF8  320004     bra z, 0x005b02
 05AFA  DE1080     lsr 0x0004,0x0000,0x0002
 05AFC  100070     subr.w 0x0000,#16,0x0000
 05AFE  DD1000     sl 0x0004,0x0000,0x0000
 05B00  708100     ior.w 0x0002,0x0000,0x0004
38:                }
 05B02  780002     mov.w 0x0004,0x0000
 05B04  060000     return