Fix stack promotion of moved registers

This PR consists of two commits:

[X86-64] Do not look up values for xor operations that set a register…
[X86-64] Only replace dominating uses of promoted register with load …

I had to include commit 1., as without it, commit 2. would expose a bug that previously existed, but did not fail mctoll's unit tests.

Do not look up register operand values for xor eax, eax instructions that just zero out a register. I've encountered issues where getRegOperandValue would generate a load instruction from a promoted stack slot. The loaded value was not used, but no issues would arise during tests.
```
main:
movabs rdi, offset .L.str
call strlen # rax defined here
cmp eax, 0
je .eq
```

.neq: mov esi, eax cmp eax, 0 jne .print

.eq: mov esi, 0 jmp .print

.print: movabs rdi, offset .L.str mov al, 0 xor esi, esi call printf

xor eax, eax
ret


The above, compiled and raised, would generate the following output (just relevant parts shown here):

```ll
define dso_local i32 @main() {
entry:
  %ESI-SKT-LOC = alloca i32, align 4
  %EAX = call i32 @strlen(i8* getelementptr inbounds ([30 x i8], [30 x i8]* @rodata_11, i32 0, i32 4))
  %ld-stk-prom8 = load i32, i32* %ESI-SKT-LOC, align 4
  %0 = sub i32 %ld-stk-prom8, 0 ; <-- (A) loading from stack slot, but stack slot has not been written yet!
  %ZF = icmp eq i32 %0, 0
  %CmpZF_JE = icmp eq i1 %ZF, true
  br i1 %CmpZF_JE, label %bb.2, label %bb.1

bb.1:                                             ; preds = %entry
  %6 = sub i32 %EAX, 0
  %7 = call { i32, i1 } @llvm.usub.with.overflow.i32(i32 %EAX, i32 0)
  %CF1 = extractvalue { i32, i1 } %7, 1
  %ZF2 = icmp eq i32 %6, 0
  store i32 %EAX, i32* %ESI-SKT-LOC, align 1
  %CmpZF_JNE = icmp eq i1 %ZF2, false
  br i1 %CmpZF_JNE, label %bb.3, label %bb.2

bb.2:                                             ; preds = %bb.1, %entry
  store i32 0, i32* %ESI-SKT-LOC, align 1
  br label %bb.3

bb.3:                                             ; preds = %bb.2, %bb.1
  %ESI = load i32, i32* %ESI-SKT-LOC, align 1 ; <-- (B) not needed!
  %EAX9 = call i32 (i8*, ...) @printf(i8* getelementptr inbounds ([30 x i8], [30 x i8]* @rodata_11, i32 0, i32 4), i32 0)
  ret i32 0
}

After commit 1, (B) won't be generated in the code anymore.

In the above code, you can also see the issue that commit 2 fixes. Since we are doing mov esi, eax in bb.1, mctoll replaces all uses of %EAX with a load from the stack slot. This is seen at (A) in the above code. ESI-SKT-LOC has not been written yet, but we're loading from the stack slot. This commit now just replaces uses if they dominate the store to the stack slot.

test/asm_test/X86/mov-reaching-def.s is raised to the following after commit 2.:

define dso_local i32 @main() {
entry:
  %RSI-SKT-LOC = alloca i32, align 4
  %EAX = call i32 @strlen(i8* getelementptr inbounds ([30 x i8], [30 x i8]* @rodata_11, i32 0, i32 4))
  %0 = sub i32 %EAX, 0 ; <-- Using %EAX here!
  %ZF = icmp eq i32 %0, 0
  %CmpZF_JE = icmp eq i1 %ZF, true
  br i1 %CmpZF_JE, label %bb.2, label %bb.1

bb.1:                                             ; preds = %entry
  %6 = sub i32 %EAX, 0
  %ZF2 = icmp eq i32 %6, 0
  store i32 %EAX, i32* %RSI-SKT-LOC, align 1 ; <-- store happening here, after this all uses will be loaded from stack slot
  %CmpZF_JNE = icmp eq i1 %ZF2, false
  br i1 %CmpZF_JNE, label %bb.3, label %bb.2

bb.2:                                             ; preds = %bb.1, %entry
  store i32 0, i32* %RSI-SKT-LOC, align 1
  br label %bb.3

bb.3:                                             ; preds = %bb.2, %bb.1
  %12 = load i32, i32* %RSI-SKT-LOC, align 1 ; <-- load from stack slot
  %RSI = zext i32 %12 to i64
  %EAX7 = call i32 (i8*, ...) @printf(i8* getelementptr inbounds ([30 x i8], [30 x i8]* @rodata_11, i32 0, i32 4), i64 %RSI)
  ret i32 0
}

microsoft / llvm-mctoll

Fix stack promotion of moved registers #122