[InstCombine] Simplify chain of GEP with constant indices

[huangjd]

Summary

Simplify chain of GEP in the form of GEP (GEP ... (GEP (GEP p C1) X1) ... Xn) C2 where C1 and C2 are constants by splitting GEP p C1 and breaking its user depedency, allowing visitGEPofGEP to reassociate the cloned inner GEP to outer GEP and perform constant index merging.

Example code that is optimized

struct S {
    char padding[123];
    int x;
};

int* P(S* s) {
    return (int*)((char*)&s->x + s->x - offsetof(S, x));
}

originally generates

P(S*):                                # @P(S*)
        movslq  124(%rdi), %rax
        addq    %rdi, %rax
        addq    $124, %rax
        addq    $-124, %rax
        retq

Now generates

P(S*):                                # @P(S*)
	movslq	124(%rdi), %rax
	addq	%rdi, %rax
	retq

[davidxl]

This is essentially forward propagation (with multi-use) and is a generally a good way to reduce dependence height. For instance it can be used to enable reassociation and LICM:

for (k = .... ;k++) {

for (i = ....  ; i++) {

      b = (a+i);
      call(b);
      y = b + k + x;
    }

}

>

for (k = .... ;k++) {

for (i = ....  ; i++) {

      b = (a+i);
      call(b);
      y  = (a + i) + k + x;
    }

}

>

for (k = .... ;k++) {

for (i = ....  ; i++) {

      b = (a+i);
      call(b);
      y  = (a + k) + i + x;
    }

}

>

for (k = .... ;k++) {

tmp = a + k;
for (i = ....  ; i++) {
      b = (a+i);
      call(b);
      y  = tmp + k + x;
    }

}

[davidxl]

y  = tmp + k + x;

should be

y  = tmp + i + x;

}

}

[spatel]

Please pre-commit the baseline tests as an NFC patch, so we'll just show the test diffs here.

[huangjd]

See D137664 for baseline

[huangjd]

Merged after adding baseline