Fixed allocas insertion point for LLVM backend

[georgemitenkov]

With this PR alloca instructions are always inserted in the beginning of the function entry block. This is done to avoid them in the while or for loops, where allocations per iteration cause stack overflow (if the IR is not optimized).

Example:

/// loop.mod
PROCEDURE loop(a, b) {
    WHILE (a > 0) {
        LOCAL x 
        x = b
    }
}

Before, this code would produce the following LLVM IR:

define i32 @loop(double %a1, double %b2) {
  %a = alloca double, align 8
  store double %a1, double* %a, align 8
  %b = alloca double, align 8
  store double %b2, double* %b, align 8
  %ret_loop = alloca i32, align 4
  store i32 0, i32* %ret_loop, align 4
  br label %1

1:                                                ; preds = %4, %0
  %2 = load double, double* %a, align 8
  %3 = fcmp ogt double %2, 0.000000e+00
  br i1 %3, label %4, label %6

4:                                                ; preds = %1
  ; x is allocated inside the loop body - this is bad!
  %x = alloca double, align 8
  %5 = load double, double* %b, align 8
  store double %5, double* %x, align 8
  br label %1

6:                                                ; preds = %1
  %7 = load i32, i32* %ret_loop, align 4
  ret i32 %7
}

Now, the result is:

define i32 @loop(double %a1, double %b2) {
  %a = alloca double, align 8
  %b = alloca double, align 8
  %ret_loop = alloca i32, align 4
  ; x is not allocated inside the loop body!
  %x = alloca double, align 8
  store double %a1, double* %a, align 8
  store double %b2, double* %b, align 8
  store i32 0, i32* %ret_loop, align 4
  br label %1

1:                                                ; preds = %4, %0
  %2 = load double, double* %a, align 8
  %3 = fcmp ogt double %2, 0.000000e+00
  br i1 %3, label %4, label %6

4:                                                ; preds = %1
  %5 = load double, double* %b, align 8
  store double %5, double* %x, align 8
  br label %1

6:                                                ; preds = %1
  %7 = load i32, i32* %ret_loop, align 4
  ret i32 %7
}

[bbpbuildbot]

Can one of the admins verify this patch?

[pramodk]

LGTM! 👌