[MPlayer-dev-eng] [PATCH] Detect "restrict" support

[Falk Hueffner]

Arpi <arpi at thot.banki.hu> writes:

> what is that 'restrict' keyword good for?

It tells the compiler that the data a pointer points to cannot be
aliased by another pointer. Take for example this function:

void foo(short *block, const unsigned char *pixels, int line_size) {
    int i;

    for (i = 0; i < 8; i++) {
	block[0] = pixels[0];
	block[1] = pixels[1];
	block[2] = pixels[2];
	block[3] = pixels[3];
	pixels += line_size;
	block += 8;
    }
}

The C standard mandates that char pointers can alias anything, so it
must assume block and pixels may point to the same region. It will
generate code like this (Alpha assembly):

        ldbu    t4,0(a1)
        addl    t3,0x1,t3
        cmple   t3,0x7,v0
        stw     t4,0(a0)
        ldbu    t2,1(a1)
        stw     t2,2(a0)
        ldbu    t0,2(a1)
        stw     t0,4(a0)
        ldbu    t1,3(a1)
        addq    a1,a2,a1
        stw     t1,6(a0)
        lda     a0,16(a0)
        bne     v0,10

i. e. it won't change the order of the loads (ldbu) and stores
(stw).  Every store will then stall 2 cycles till the load is
finished. This loop takes 13 cycles (assuming data is in L1 cache).

However, we know that block and pixel actually can never alias, so we
mark them as restrict. The compiler can then reorder loads and stores:

        ldbu    v0,0(a1)
        ldbu    t3,1(a1)
        ldbu    t4,2(a1)
        ldbu    t0,3(a1)
        addl    t5,0x1,t5
        addq    a1,a2,a1
        stw     v0,0(a0)
        cmple   t5,0x7,v0
        stw     t3,2(a0)
        stw     t4,4(a0)
        stw     t0,6(a0)
        lda     a0,16(a0)
        bne     v0,10

The stores now no longer stall, the loads can double issue, and the
loop runs in 8 cycles.

gcc does not support this well for a long time, so you might need a
very recent gcc version; it seems 3.1 doesn't do this reordering, but
the CVS head does.

-- 
	Falk