Compile this with llvm-gcc:

unsigned int *a, *b;

void joy(void) {
   a[20] = ~(~(a[10]|b[10]) | (a[10]&b[10]));
}

and you get:

define void @joy() {
entry:
        %tmp = load i32** @a            ; <i32*> [#uses=2]
        %tmp2 = getelementptr i32* %tmp, i32 10         ; <i32*> [#uses=1]
        %tmp3 = load i32* %tmp2         ; <i32> [#uses=2]
        %tmp4 = load i32** @b           ; <i32*> [#uses=1]
        %tmp5 = getelementptr i32* %tmp4, i32 10                ; <i32*> [#uses=1]
        %tmp6 = load i32* %tmp5         ; <i32> [#uses=2]
        %tmp7 = or i32 %tmp6, %tmp3             ; <i32> [#uses=1]
        %tmp7not = xor i32 %tmp7, -1            ; <i32> [#uses=1]
        %tmp14 = and i32 %tmp6, %tmp3           ; <i32> [#uses=1]
        %tmp15 = or i32 %tmp14, %tmp7not                ; <i32> [#uses=1]
        %tmp15not = xor i32 %tmp15, -1          ; <i32> [#uses=1]
        %tmp16 = getelementptr i32* %tmp, i32 20                ; <i32*> [#uses=1]
        store i32 %tmp15not, i32* %tmp16
        ret void
}


it could just use xor?

Created attachment 1011 [details]
this module is really just doing an xor

Here's a C vector example:

typedef unsigned __attribute__((vector_size(16))) vu;


vu test1(vu a, vu b) {
   return ~(~(a|b) | (a&b));
}

vu test2(vu a, vu b) {
   return (a|b) & ~(a&b);
}

Implemented.  There are many patches to this, the primary ones being:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070611/050487.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070611/050489.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070611/050495.html

Testcase here: Transforms/InstCombine/and-or-not.ll

This shrunk Duraid's nasty example from 32K lines of x86-64 .s file to 23k lines.

-Chris