Adding new bb and conditional branch for string comparison

[KavinduZoysa]

Hi all,

public function main() {
     string s = "str1";
     if (s == "str2") {
     }
}

To perform a string comparison as shown above, first, we have the compare the pointers(pointer to "str1" and pointer to "str2"), if it fails then we have to call the rust side to compare two strings.

Here we have to do few changes to the LLVM IR.

1. Should create a new bb which includes the runtime call and conditional branch to continue with the rest
2. Should point the existing conditional branch to new bb

I have a problem where we should do this. After reading the BIR(before codegen) we can iterate through the bir package and find the place where we compare two string. At this point, we can inject the relevant changes to LLVM IR.

Highly appreciate your input on this approach.

[KavinduZoysa]

@ayonam, please check this discussion.

[ayonam]

@KavinduZoysa

Your original program is:

public function main() {
string s = "str1";
if (s == "str2") {
blah blah
}
more blah blah
}

So this boils down to something like:

public function main() {
string s = "str1";
string r = "str2";
if (pointer_of(s) == pointer_of(r)) ||
! deep_pointer_comparison_in_rust(s, r)) {
blah blah
}
more blah blah
}

In terms of LLVM IR, this should be something like:

bb0:
%0 = getelementptr of s
%1 = getelementptr of r
%2 = compare i64 %0, %1
br eq bb2

bb1:
%3 = call deep_pointer_comparison_in_rust (%0, %1)
%4 = compare i64 %3, 0
br ne bb3

bb2:
code for blah blah
br bb3

bb3:
code for more blah blah

Now you need to take a look at what does the BIR say. I expect the BIR to generate something like:

bb0:
%0 = getelementptr of s
%1 = getelementptr of r
%2 = compare i64 %0, %1
br ne bb3

bb2:
code for blah blah
br bb3

bb3:
code for more blah blah

So as you see, all you need to do is to add the bb1 and its code, and then patch the existing branch bb0 -> bb3 to bb0 -> bb2 and reverse the condition on the branch.

Do let me know if you have any doubts about this.

[KavinduZoysa]

Yes @ayonam, that is correct. My question is at which place we have to do this change. This is my opinion.
After reading the BIR(before codegen) we can iterate through the bir package and find the place where we compare two string. At this point, we can inject the relevant changes to LLVM IR.

Could you please tell me whether this one is correct or do you have any other suggestion on this?

[ayonam]

@KavinduZoysa
Yes, that is the right place. You need to generate the comparison of the pointers first and then add the new basic block and patch the branch to that basic block while keeping the original target saved up for use in the new basic block.

[KavinduZoysa]

@ayonam, Thank you for your input, I will start to implement

[ruvi-d]

The idiomatic approach will be to instead BBs in the instruction codegen scope. eg #284