llvm dev - Sep 2012 - [LLVMdev] FileCheck for instructions of indeterminate order?

To test some recent changes, I need to verify that seven instructions
are generated.  However, the order of those instructions doesn't matter
(they are all independent loads from memory).  Is there a way to tell
FileCheck to reset its scan position rather than assuming all CHECK:
instructions must be in the given order?

My initial version of the test was to use -O0, attempting to ensure that
the instructions weren't rescheduled.  Although this gave a reproducible
order on my PowerPC machine, various buildbots of other architectures
produced codegen in a different order when cross-compiling to PowerPC.
Needless to say, the test is XFAILed until I can figure a way to test
this more safely.

All help appreciated!

Thanks,
Bill

N.B. I found the nondeterminacy at -O0 to be somewhat surprising...is
there some other option I need to avoid instruction reordering?

-- 
Bill Schmidt, Ph.D.
IBM Advance Toolchain for PowerLinux
IBM Linux Technology Center
wschmidt at us.ibm.com
wschmidt at linux.vnet.ibm.com

On 9/19/2012 12:06 PM, William J. Schmidt wrote:
> To test some recent changes, I need to verify that seven instructions
> are generated.  However, the order of those instructions doesn't matter
> (they are all independent loads from memory).  Is there a way to tell
> FileCheck to reset its scan position rather than assuming all CHECK:
> instructions must be in the given order?
See this thread:

Message-ID: <5049E695.2050508 at codeaurora.org>
Date: Fri, 07 Sep 2012 07:20:37 -0500
From: Matthew Curtis <mcurtis at codeaurora.org>
Subject: [LLVMdev] teaching FileCheck to handle variations in order


-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, 
hosted by The Linux Foundation

On 19 September 2012 13:06, William J. Schmidt
<wschmidt at linux.vnet.ibm.com> wrote:
> To test some recent changes, I need to verify that seven instructions
> are generated.  However, the order of those instructions doesn't matter
> (they are all independent loads from memory).  Is there a way to tell
> FileCheck to reset its scan position rather than assuming all CHECK:
> instructions must be in the given order?
>
> My initial version of the test was to use -O0, attempting to ensure that
> the instructions weren't rescheduled.  Although this gave a
reproducible
> order on my PowerPC machine, various buildbots of other architectures
> produced codegen in a different order when cross-compiling to PowerPC.
> Needless to say, the test is XFAILed until I can figure a way to test
> this more safely.
>
> All help appreciated!
The order should be deterministic, even if arbitrary. The most common
source of problem is some code iterating over a hash.
> Thanks,
> Bill
>
Cheers,
Rafael