llvm dev - May 2010 - [LLVMdev] Manipulating basic blocks with the C bindings

On Thu, May 27, 2010 at 6:38 PM, Kenneth Uildriks <kennethuil at
gmail.com> wrote:
> On Thu, May 27, 2010 at 5:47 PM, Evan Shaw <chickencha at gmail.com>
wrote:
>> I'm writing a frontend with the LLVM C bindings for a language that
>> has a goto statement, similar to C's. I'm having some trouble
figuring
>> out what to do for the case where the label is declared after the
>> goto, like this:
>>
>> goto label;
>> ...
>> label:
>> ...
>>
>> When I generate the code for the goto, I'd like to create a basic
>> block that's not inserted anywhere in particular and then put it in
>> the right place when I hit the label (or error out appropriately if I
>> never find it). It looks like the C++ API allows one to do what I'm
>> describing, but I don't see a way to do it with the C bindings.\
>
> When you hit the "goto", create an empty basic block with
> LLVMAppendBasicBlockInContext and put in an unconditional branch to
> it.  Later, when you hit the label, put another unconditional branch
> to that same block, then call LLVMPositionBuilderAtEnd(labelBlock) and
> continue on your way.
That would work. It just seems like there ought to be a better way. Thanks.

- Evan

> That would work. It just seems like there ought to be a better way. Thanks.
Since there's no problem using the C++ API, maybe the C bindings should be
improved?

Ciao,

Duncan.

On Thu, May 27, 2010 at 8:46 PM, Evan Shaw <chickencha at gmail.com>
wrote:
> On Thu, May 27, 2010 at 6:38 PM, Kenneth Uildriks <kennethuil at
gmail.com> wrote:
>> On Thu, May 27, 2010 at 5:47 PM, Evan Shaw <chickencha at
gmail.com> wrote:
>>> I'm writing a frontend with the LLVM C bindings for a language
that
>>> has a goto statement, similar to C's. I'm having some
trouble figuring
>>> out what to do for the case where the label is declared after the
>>> goto, like this:
>>>
>>> goto label;
>>> ...
>>> label:
>>> ...
>>>
>>> When I generate the code for the goto, I'd like to create a
basic
>>> block that's not inserted anywhere in particular and then put
it in
>>> the right place when I hit the label (or error out appropriately if
I
>>> never find it). It looks like the C++ API allows one to do what
I'm
>>> describing, but I don't see a way to do it with the C
bindings.\
>>
>> When you hit the "goto", create an empty basic block with
>> LLVMAppendBasicBlockInContext and put in an unconditional branch to
>> it.  Later, when you hit the label, put another unconditional branch
>> to that same block, then call LLVMPositionBuilderAtEnd(labelBlock) and
>> continue on your way.
> That would work. It just seems like there ought to be a better way. Thanks.
>
> - Evan
>
If you keep in mind that blocks (other than the entry) aren't really
ordered by anything other than branches, I can't see anything
unintuitive about it.

---------- Forwarded message ----------
From: Kenneth Uildriks <kennethuil at gmail.com>
Date: Fri, May 28, 2010 at 7:39 AM
Subject: Re: [LLVMdev] Manipulating basic blocks with the C bindings
To: Evan Shaw <chickencha at gmail.com>
Cc: llvmdev at cs.uiuc.edu


On Thu, May 27, 2010 at 8:46 PM, Evan Shaw <chickencha at gmail.com>
wrote:
> On Thu, May 27, 2010 at 6:38 PM, Kenneth Uildriks <kennethuil at
gmail.com> wrote:
>> On Thu, May 27, 2010 at 5:47 PM, Evan Shaw <chickencha at
gmail.com> wrote:
>>> I'm writing a frontend with the LLVM C bindings for a language
that
>>> has a goto statement, similar to C's. I'm having some
trouble figuring
>>> out what to do for the case where the label is declared after the
>>> goto, like this:
>>>
>>> goto label;
>>> ...
>>> label:
>>> ...
>>>
>>> When I generate the code for the goto, I'd like to create a
basic
>>> block that's not inserted anywhere in particular and then put
it in
>>> the right place when I hit the label (or error out appropriately if
I
>>> never find it). It looks like the C++ API allows one to do what
I'm
>>> describing, but I don't see a way to do it with the C
bindings.\
>>
>> When you hit the "goto", create an empty basic block with
>> LLVMAppendBasicBlockInContext and put in an unconditional branch to
>> it.  Later, when you hit the label, put another unconditional branch
>> to that same block, then call LLVMPositionBuilderAtEnd(labelBlock) and
>> continue on your way.
> That would work. It just seems like there ought to be a better way. Thanks.
>
> - Evan
>
If you keep in mind that blocks (other than the entry) aren't really
ordered by anything other than branches, I can't see anything
unintuitive about it.

On Fri, May 28, 2010 at 7:39 AM, Kenneth Uildriks <kennethuil at
gmail.com> wrote:
> If you keep in mind that blocks (other than the entry) aren't really
> ordered by anything other than branches, I can't see anything
> unintuitive about it.
>
It's true that the ordering doesn't matter to the computer, but it's
definitely easier for me to read the IR if its structure is similar to
the original source's structure.

I may just put together a patch and see what happens since it
shouldn't be very much work.

- Evan