aarch64: sbc is not being used

[pinskia]

Take:

unsigned f1(unsigned a, unsigned b, unsigned c)
{
  unsigned t = a < b;
  return c - t;
}

LLVM produces:

f1:
        cmp     w0, w1
        cset    w8, lo
        sub     w0, w2, w8
        ret

The cset/sub combo should be replaced with:

        sbc     w0, w2, wzr

Which is what GCC produces:

f1:
        cmp     w0, w1
        sbc     w0, w2, wzr
        ret

https://godbolt.org/z/5f83qTK79

[pinskia]

This shows up if you are doing multi-precision math e.g.:

void f1(unsigned long a, unsigned long c, unsigned long b, unsigned  long *d)
{
  *d = a - b;
  unsigned t = a < b;
  d[1] = c - t;
}
void f2(unsigned __int128 a, unsigned long b, unsigned  __int128 *d)
{
  *d = a - b;
}

These 2 should produce the same code, LLVM is able to handle __int128 case just fine but not the original one.

[pinskia]

Another example:

void f2(unsigned long a, unsigned long b, unsigned long c, unsigned long d, unsigned  long *e)
{
  e[0] = a - c;
  unsigned t = a < c;
  e[1] = b - d - t;
}
void f3(unsigned __int128 a, unsigned __int128 b, unsigned  __int128 *d)
{
  *d = a - b;
}

In this case there are 3 instructions that should just be one:

        sub     x10, x1, x3
        cset    w9, lo
        sub     x9, x10, x9

vs

        sbc     x9, x1, x3

Again LLVM __int128 just fine ...

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Andrew Pinski (pinskia)

Take: ``` unsigned f1(unsigned a, unsigned b, unsigned c) { unsigned t = a < b; return c - t; } ```

LLVM produces:

f1:
        cmp     w0, w1
        cset    w8, lo
        sub     w0, w2, w8
        ret

The cset/sub combo should be replaced with:

        sbc     w0, w2, wzr

Which is what GCC produces:

f1:
        cmp     w0, w1
        sbc     w0, w2, wzr
        ret

https://godbolt.org/z/5f83qTK79