[llvm-commits] [PATCH] RDRAND intrinsics

[Benjamin Kramer]

On 08.07.2012, at 13:56, Benjamin Kramer <benny.kra at gmail.com> wrote:

> 
> On 07.07.2012, at 11:12, Benjamin Kramer wrote:
> 
>> 
>> On 06.07.2012, at 23:30, Eli Friedman wrote:
>> 
>>> On Fri, Jul 6, 2012 at 2:03 PM, Benjamin Kramer <benny.kra at gmail.com> wrote:
>>>> 
>>>> On 06.07.2012, at 22:52, Eli Friedman wrote:
>>>> 
>>>>> On Fri, Jul 6, 2012 at 1:00 PM, Benjamin Kramer <benny.kra at gmail.com> wrote:
>>>>>> The attached patch adds support for intel's rdrand intrinsics to LLVM.
>>>>>> 
>>>>>> The tricky part is that the intrinsics have to return two values (the random value and the carry flag). This is solved with a pointer argument and a cmov to materialize CF, which matches GCC's output. This approach requires some custom lowering code, but I don't see a way to do it with tablegenerated patterns only.
>>>>> 
>>>>> It would be much more friendly to the optimizer to use multiple return
>>>>> values... but it might not be worth the hassle, considering that
>>>>> rdrand is relatively slow.
>>>> 
>>>> I tried this first as it also looks cleaner, but I couldn't find a way to smuggle it past legalization into the target specific lowering.
>>> 
>>> X86TargetLowering::ReplaceNodeResults?
>> 
>> Doesn't work because TLI.getOperationAction bails out early when it sees an extended type. The legalizer then tries to expand the node and aborts.
> 
> Apparently I made some mistake yesterday, it *does* work without applying any tricks as long as all return values are legal individually. I attached an updated patch which uses the following intrinsic signature, allowing the store to stay in IR.
> 
> declare { i16, i32 } @llvm.x86.rdrand.16()
> 
> Here i16 is the random value and i32 is the boolean indicating whether it is considered valid. I also attached the clang part of this patch and reverified that the code matches what ICC and GCC emit.
> 
> - Ben
> 
> <0001-Add-intrinsics-for-Ivy-Bridge-s-rdrand-instruction.patch><0001-Add-_rdrand-16-32-64-_step-intrinsics-to-immintrin.h.patch>

ping.