[llvm-commits] [llvm] r165126 - /llvm/trunk/lib/Transforms/Utils/BypassSlowDivision.cpp

[David Blaikie]

On Wed, Oct 3, 2012 at 10:58 AM, Nowicki, Tyler <tyler.nowicki at intel.com> wrote:
> Hi
>
> There are several tests for this optimization located in test\CodeGen\X86\atom-bypass-slow-division.ll. These test use LLC. When using LLC types are created from the global context. Bypass types are specified using the global context so this works fine. However, Clang does not use the global context. Prior to this patch, div/rem instruction types would not match those in the BypassTypeMap. I added tests to Clang to ensure this optimization is used by Clang.

Then this might necessitate an LLVM unit/API-level test if llc is insufficient.

> I am looking into Hal's question on why Clang does not use the global context.
>
> Thanks,
>
> Tyler
>
>> > Tests for this will be done as a separate patch to Clang.
>>
>> Ideally LLVM changes should have LLVM tests, otherwise we risk regressing
>> LLVM accidentally if we don't validate against Clang.
>>
>> Is there a reason this cannot be tested at the LLVM level?