llvm dev - Dec 2012 - [LLVMdev] ARM assembly

On Sun, Dec 9, 2012 at 7:48 PM, Evan Cheng <evan.cheng at apple.com>
wrote:
> Sound like you are not enabling optimization. Try with -O3.
Ah, of course! I'd forgotten about the extra allocas produced by clang.

To expand a little, the spills you're seeing are (probably)
specifically created by clang (it creates a shadow variable for each
local with alloca). "llc" on its own can't get rid of these, so
you'll
either want to give clang "-O3" or run the LLVM bitcode through
"opt"
before "llc".

Tim.

I used your code and it works fine. Once I increase the number of variable
it uses the extra registers which I added. I have another problem.

To generate the assembly file for new ARM machine with extra registers,I
follow these steps:

1) clang -emit-llvm main.ll main.c
2) llc -march=arm -o main.s main.bc

In this case it doesn't use the registers which I added to the
architecture. Do you have any ideas?

As far as I know, the generated ll code doesn't depend on the target
(right?), but I see this line in the ll code : target triple
"x86_64-unknown-linux-gnu"
Could it be the problem?

I really appreciate if anybody can help.






Best Regards,
A. Yazdanbakhsh
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
PhD. Student
School of Electrical and Computer Engineering
University of Wisconsin-Madison
E-mail:     yazdanbakhsh at wisc.edu


<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<



On Sun, Dec 9, 2012 at 4:17 PM, Tim Northover <t.p.northover at
gmail.com>wrote:
> On Sun, Dec 9, 2012 at 7:48 PM, Evan Cheng <evan.cheng at apple.com>
wrote:
> > Sound like you are not enabling optimization. Try with -O3.
>
> Ah, of course! I'd forgotten about the extra allocas produced by clang.
>
> To expand a little, the spills you're seeing are (probably)
> specifically created by clang (it creates a shadow variable for each
> local with alloca). "llc" on its own can't get rid of these,
so you'll
> either want to give clang "-O3" or run the LLVM bitcode through
"opt"
> before "llc".
>
> Tim.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20121210/36214c8a/attachment.html>

Hi Amir,
> 1) clang -emit-llvm main.ll main.c
> 2) llc -march=arm -o main.s main.bc
>
> In this case it doesn't use the registers which I added to the
architecture.
> Do you have any ideas?
Evan's suggestion was probably a good one. If you replace the clang command
with
$ clang -O3 -emit-llvm main.c

then llc should be able to use more registers for the resulting main.bc.
> As far as I know, the generated ll code doesn't depend on the target
> (right?), but I see this line in the ll code : target triple >
"x86_64-unknown-linux-gnu"
> Could it be the problem?
It's not *this* problem, and as long as you don't try to do anything
too complicated (including trying to run the output) you'll probably
get away with it, but it's not ideal.

What you really want is a cross-toolchain and to point clang at this
toolchain. Most importantly with the option "-target
arm-none-linux-gnueabi" (or similar), but possibly also telling it
where the headers and libraries are with other options.

Tim.