[PATCH] Adds Cortex-A53 and Cortex-A57 subtargets.

Andrew Trick atrick at apple.com
Wed Mar 5 19:11:34 PST 2014


On Mar 5, 2014, at 3:22 PM, Dave Estes <cestes at codeaurora.org> wrote:

> On 03/05/2014 01:33 PM, Eric Christopher wrote:
>> On Tue, Mar 4, 2014 at 3:06 PM, Dave Estes <cestes at codeaurora.org> wrote:
>>> Hey Andy,
>>> 
>>> Chad committed the patch last night, and the new lit test in the patch broke
>>> one of the builders:
>>> http://lab.llvm.org:8013/builders/phase1%20-%20sanity/builds/17406
>>> 
>>> I've been trying to figure out why, but I'm at a loss because I can't
>>> reproduce the failure under Ubuntu or Darwin using make check-all or
>>> llvm-lit. I'm using a Release+Asserts build.
>>> 
>>> Initially I thought it might be a problem with the test case, but I find it
>>> less likely now. My test case is patterned off of another that you wrote for
>>> the ARM backend
>>> (llvm/test/CodeGen/ARM/misched-copy-arm.ll). It uses a lot of redirection to
>>> get the misched debug output piped to FileCheck. I've gone over that and ran
>>> it in a terminal on Ubuntu and Darwin, but no failures.
>>> 
>>> The builder output from FileCheck seems to indicate that llc might have
>>> crashed. If this is the case, then it's easy to conclude that this is the
>>> only test that failed because it's the only one that uses the Machine Model
>>> for Cortex-A53. I just need to figure out how to get it to fail. I've been
>>> digging through the artifacts tarball, but it only includes clang and I
>>> can't sort out the correct options to emulate an llc lit run.
>>> 
>>> Do you or anyone else have any ideas on how I might proceed?
>>> 
>> 
>> Weird, I'd almost try again to see if you can get it to fail. Another
>> thought would be to try building the tool up with asan and see if that
>> catches the problem (alternately run under valgrind etc).
>> 
>> -eric
>> 
> 
> Thanks, Eric. valgrind didn't report anything. I'm trying ASAN now. My Darwin setup is really lacking, so I'm working on Ubuntu. Not sure how valuable it'll be if the failure happened on Darwin.
> 
> Does anyone know if the builders at lab.llvm.org can test patches that haven't been submitted? A sort of try before you buy. :) I can try out a few things that might get me a better log, but I really don't want to submit a patch that's just going to need to be reverted again.

Hi Dave,

The error on Darwin is:

Assertion failed: (TM.getSubtarget<AArch64Subtarget>().isTargetELF() && "unknown subtarget type"), function createTLOF, file /s/sched/lib/Target/AArch64/AArch64ISelLowering.cpp, line 35.

Since AArch64 doesn’t appear to be supported on Darwin, you’ll have to add a triple, just like the other AArch64 test cases:
-mtriple=aarch64-none-linux-gnu

Thanks,
-Andy

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





More information about the llvm-commits mailing list