llvm dev - Jan 2012 - [LLVMdev] Opt pass 'Canonicalize Induction Variables' not working

I am new to LLVM and I've a question about the optimization passes and in
particular about the loop transformation passes. I've setup the LLVM
tool-chain with Polly installed.

The following example is causing problems with me:
http://llvm.org/docs/Passes.html#indvars

The input is a .c file with the following loops:

# 01 Standard, N=100
for (i=0; i<N; i++)
	b[i] = a[i] + 5;

# 02 Example loop -indvars
for (i = 7; i*i < 1000; ++i)
	b[i] = 4;

# 03 Start at 1 instead of 0
for (i = 1; i<10; i++)
	b[i] = 3;

I have the latest LLVM build installed on Mac OSX 10.6.8.

A script with (among others) the following lines is used:

-----------------------------------------------------------------------------------------------------
POLLY_DLIB=/Developer/llvm_build/tools/polly/Debug+Asserts/lib/LLVMPolly.dylib
alias opt="/Developer/llvm_build/Debug+Asserts/bin/opt -load
${POLLY_DLIB}"
alias llvm-as="/Developer/llvm_build/Debug+Asserts/bin/llvm-as"

# Set input filename
FILE="indvar"

# Run clang to generate LLVM-IR
clang -S -emit-llvm ${FILE}.c -o ${FILE}.s

# Make the human readable format into bitcode format
llvm-as ${FILE}.s -o ${FILE}.bc

# Prepare the LLVM-Bytecode into canonical form
opt -S -mem2reg -loop-simplify -indvars ${FILE}.bc > ${FILE}.ll
-----------------------------------------------------------------------------------------------------


What I want (and what is not working) is that loop 2 and loop 3 are transformed
into a loop starting at 0 with stride 1 (natural loop). Subsequent passes need
this to output information to form a Polyhedral model (PHM).

Up to now, only the first loop is recognized as natural loop (as it already is
one) and information about the loop is outputted as PHM. The other two loops
aren't changed.

Can somebody give me some advise? Maybe I'm using the wrong passes or in the
wrong order?

Many thanks in advance,

Pieter


-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120116/94ca2b40/attachment.html>

On 01/16/2012 01:19 PM, Pieter Custers wrote:
> I am new to LLVM and I've a question about the optimization passes and
> in particular about the loop transformation passes. I've setup the LLVM
> tool-chain with Polly installed.
>
> The following example is causing problems with me:
> http://llvm.org/docs/Passes.html#indvars
>
> The input is a .c file with the following loops:
>
> # 01 Standard, N=100
> for (i=0; i<N; i++)
> b[i] = a[i] + 5;
>
> # 02 Example loop -indvars
> for (i = 7; i*i < 1000; ++i)
> b[i] = 4;
>
> # 03 Start at 1 instead of 0
> for (i = 1; i<10; i++)
> b[i] = 3;
>
> I have the latest LLVM build installed on Mac OSX 10.6.8.
>
> A script with (among others) the following lines is used:
>
>
-----------------------------------------------------------------------------------------------------
>
POLLY_DLIB=/Developer/llvm_build/tools/polly/Debug+Asserts/lib/LLVMPolly.dylib
> alias opt="/Developer/llvm_build/Debug+Asserts/bin/opt -load
${POLLY_DLIB}"
> alias llvm-as="/Developer/llvm_build/Debug+Asserts/bin/llvm-as"
>
> # Set input filename
> FILE="indvar"
>
> # Run clang to generate LLVM-IR
> clang -S -emit-llvm ${FILE}.c -o ${FILE}.s
>
> # Make the human readable format into bitcode format
> llvm-as ${FILE}.s -o ${FILE}.bc
>
> # Prepare the LLVM-Bytecode into canonical form
> opt -S -mem2reg -loop-simplify -indvars ${FILE}.bc > ${FILE}.ll
>
-----------------------------------------------------------------------------------------------------
>
>
> What I want (and what is not working) is that loop 2 and loop 3 are
> transformed into a loop starting at 0 with stride 1 (natural loop).
> Subsequent passes need this to output information to form a Polyhedral
> model (PHM).
>
> Up to now, only the first loop is recognized as natural loop (as it
> already is one) and information about the loop is outputted as PHM. The
> other two loops aren't changed.
>
> Can somebody give me some advise? Maybe I'm using the wrong passes or
in
> the wrong order?
Hi Pieter,

the -indvars pass does by default not generate canonical induction 
variables. Most LLVM passes where converted to work without canonical 
induction variables. For Polly those induction variables are (still) 
required. You can enable the relevant conversion by adding
-enable-iv-rewrite to the opt flags.

In general, the to code for Polly a larger set of canonicalization 
passes is required. They can be run from clang with

clang -Xclang -load ${POLLY_DLIB} ${FILE}.c -o ${FILE}.ll -S -emit-llvm 
-mllvm -polly -O0

You need to make sure that Polly, LLVM and clang are compiled from the 
same source version.

Cheers
Tobi

Hello Tobi,

Thank you for the quick reply.

I updated my complete build suite with your script
(http://polly.grosser.es/polly.sh). The make-test fails with some errors, the
rest went smoothly;

--
Exit Code: 1
Command Output (stderr):
--
LLVM ERROR: Could not resolve external global address: stdout
--

********************
Testing Time: 10.01s
********************
Failing Tests (3):
    Polly :: CodeGen/do_pluto_matmult.ll
    Polly :: CodeGen/single_do_loop_int_max_iterations.ll
    Polly :: CodeGen/single_do_loop_scev_replace.ll

  Expected Passes    : 104
  Expected Failures  : 29
  Unexpected Failures: 3
make: *** [polly-test] Error 1

When I try your command 'clang -Xclang -load ${POLLY_DLIB} ${FILE}.c -o
${FILE}.ll -S -emit-llvm -mllvm -polly -O0' if fails with the following
errors:

--------------------------------------
clang: warning:
/Developer/llvm_build/tools/polly/Debug+Asserts/lib/LLVMPolly.dylib:
'linker' input unused when '-S' is present
clang: warning: argument unused during compilation: '-mllvm'
clang: warning: argument unused during compilation: '-polly'
Error opening '-o': dlopen(-o, 9): image not found
  -load request ignored.
error: error reading 'indvar.ll'
--------------------------------------

I tried a different order of arguments to no avail. My CLang version is 1.0.2

Also the '-enable-iv-rewrite' has no effect.

Any ideas?

Thanks in advance,

Pieter

On 16 jan 2012, at 15:01, Tobias Grosser wrote:
> On 01/16/2012 01:19 PM, Pieter Custers wrote:
>> I am new to LLVM and I've a question about the optimization passes
and
>> in particular about the loop transformation passes. I've setup the
LLVM
>> tool-chain with Polly installed.
>> 
>> The following example is causing problems with me:
>> http://llvm.org/docs/Passes.html#indvars
>> 
>> The input is a .c file with the following loops:
>> 
>> # 01 Standard, N=100
>> for (i=0; i<N; i++)
>> b[i] = a[i] + 5;
>> 
>> # 02 Example loop -indvars
>> for (i = 7; i*i < 1000; ++i)
>> b[i] = 4;
>> 
>> # 03 Start at 1 instead of 0
>> for (i = 1; i<10; i++)
>> b[i] = 3;
>> 
>> I have the latest LLVM build installed on Mac OSX 10.6.8.
>> 
>> A script with (among others) the following lines is used:
>> 
>>
-----------------------------------------------------------------------------------------------------
>>
POLLY_DLIB=/Developer/llvm_build/tools/polly/Debug+Asserts/lib/LLVMPolly.dylib
>> alias opt="/Developer/llvm_build/Debug+Asserts/bin/opt -load
${POLLY_DLIB}"
>> alias
llvm-as="/Developer/llvm_build/Debug+Asserts/bin/llvm-as"
>> 
>> # Set input filename
>> FILE="indvar"
>> 
>> # Run clang to generate LLVM-IR
>> clang -S -emit-llvm ${FILE}.c -o ${FILE}.s
>> 
>> # Make the human readable format into bitcode format
>> llvm-as ${FILE}.s -o ${FILE}.bc
>> 
>> # Prepare the LLVM-Bytecode into canonical form
>> opt -S -mem2reg -loop-simplify -indvars ${FILE}.bc > ${FILE}.ll
>>
-----------------------------------------------------------------------------------------------------
>> 
>> 
>> What I want (and what is not working) is that loop 2 and loop 3 are
>> transformed into a loop starting at 0 with stride 1 (natural loop).
>> Subsequent passes need this to output information to form a Polyhedral
>> model (PHM).
>> 
>> Up to now, only the first loop is recognized as natural loop (as it
>> already is one) and information about the loop is outputted as PHM. The
>> other two loops aren't changed.
>> 
>> Can somebody give me some advise? Maybe I'm using the wrong passes
or in
>> the wrong order?
> 
> Hi Pieter,
> 
> the -indvars pass does by default not generate canonical induction
variables. Most LLVM passes where converted to work without canonical induction
variables. For Polly those induction variables are (still) required. You can
enable the relevant conversion by adding
> -enable-iv-rewrite to the opt flags.
> 
> In general, the to code for Polly a larger set of canonicalization passes
is required. They can be run from clang with
> 
> clang -Xclang -load ${POLLY_DLIB} ${FILE}.c -o ${FILE}.ll -S -emit-llvm
-mllvm -polly -O0
> 
> You need to make sure that Polly, LLVM and clang are compiled from the same
source version.
> 
> Cheers
> Tobi