[llvm-dev] [cfe-dev] need help adding arch-specific builtin

[Mara Sophie Grosch via llvm-dev]

Changing to an i16 is not going to work, sadly. But I can think the
other things you wrote will help me going forward - thank you :)

Am Sun, Oct 24, 2021 at 09:45:02AM -0700 schrieb Craig Topper:
>The C to IR conversion happens in CGBuiltin.cpp where it calls
>Intrinsic::getIntrinsicForGCCBuiltin.
>
>The error your getting is coming from SelectionDAG. The error seems to
>indicate that i32 is not a legal type for the AVR and
>LegalizeIntegerTypes.cpp does know how to make it legal. It wants to split
>into 2 i16s but it doesn’t know how. Intrinsics with illegal types are
>usually handled by adding setOperationAction(ISD::INTRINSIC_W_CHAIN,
>MVT::i32, Custom) to the AVRTargetLowering constructor. And then adding a
>case for INTRINSIC_W_CHAIN to AVRTargetLowering::LowerOperation to detect
>the specific intrinsic and handle it.
>
>Alternatively you chang the intrinsic to use i16 instead of i32 if that
>meets your needs.
>
>On Sun, Oct 24, 2021 at 9:28 AM Mara Sophie Grosch <littlefox at lf-net.org>
>wrote:
>
>> that's good to know, thank you. Where is the code generation for those
>> intrinsics then?
>>
>> Am Sun, Oct 24, 2021 at 09:23:34AM -0700 schrieb Craig Topper:
>> >If you add GCCBuiltin to the intrinsic definition, clang will
>> automatically
>> >do the conversion from C to IR. EmitTargetArchBuiltin will only be called
>> >if no GCCBuiltin can be found.
>> >
>> >On Sun, Oct 24, 2021 at 6:48 AM Mara Sophie Grosch via cfe-dev <
>> >cfe-dev at lists.llvm.org> wrote:
>> >
>> >> Hi lists,
>> >>
>> >> I'm trying to add a new compiler builtin for AVR architecture,
>> >> `void __builtin_avr_delay_cycles(int32)`.
>> >>
>> >> I already made clang recognize this as a builtin with the correct type.
>> >> Now there are some assumptions on what I have to do:
>> >>
>> >> * I think have to map this to an arch-specific intrinsic in LLVM
>> >> * in clang/lib/CodeGen/CGBuiltin I'd have to add a function for AVR
>> >>    builtins and call it from EmitTargetArchBuiltinExpr
>> >> * in this new function, I would generate the actual instructions -
>> >>    building the "generate code from builtin call" part
>> >>
>> >> I added an intrinsic for this[1] and added the function in CGBuiltin and
>> >> call it for Arch == llvm::Triple::avr - but cannot get my new code to be
>> >> called. I even added an `assert(false)` in EmitTargetArchBuiltinExpr(),
>> >> which should cause the compiler to crash on that assert whenever it is
>> >> supposed to handle a builtin for AVR targets - but that assert never
>> >> fires.
>> >>
>> >> Instead, I get the error given below [2], which I cannot make a lot out
>> >> of it.
>> >>
>> >> With my not-yet-very-good knowledge how everything is combined together
>> >> in clang and llvm, I assume this might be because it tries to pass a i32
>> >> to the builtin (which is expected), but cannot do that because the arch
>> >> cannot pass i32s around in arguments? But that shouldn't be needed for a
>> >> builtin - that call to it shouldn't be in the resulting machine code
>> >> anyway.
>> >>
>> >> Any clues where to find more information on how to achieve what I'm
>> >> trying to do would be appreciated. I poked around at a lot of different
>> >> places in the code base by now, trying to find more info by myself, but
>> >> am stuck.
>> >>
>> >> The builtin in question is supposed to generate a loop taking the given
>> >> amount of clock cycles to complete - something commonly used on
>> >> microcontrollers to have timing. Frequency of the CPU is known at
>> >> compile time via preprocessor definition for that and the builtin
>> >> already exists on gcc, I just want to also have it in clang.
>> >>
>> >> Thanks for any help and comments
>> >> Mara
>> >> (I subscribed to cfe-dev in digest mode only - answering me directly too
>> >> would be appreciated)
>> >>
>> >>
>> >> [1] I added this into llvm/include/llvm/IR/IntrinsicsAVR.td, which I
>> >> include from llvm/include/llvm/IR/Intrinsics.td:
>> >>
>> >> let TargetPrefix = "avr" in {  // All intrinsics start with "llvm.avr."
>> >>    def int_avr_delay_cycles : GCCBuiltin<"__builtin_avr_delay_cycles">,
>> >>                Intrinsic<[], [llvm_i32_ty], [IntrInaccessibleMemOnly,
>> >> IntrCold, IntrNoDuplicate, IntrNoMerge, IntrHasSideEffects]>;
>> >> }
>> >>
>> >>
>> >> [2] Trying to compile code using this builtin results in this error:
>> >> ExpandIntegerOperand Op #2: t8: ch = llvm.avr.delay.cycles t5,
>> >> TargetConstant:i16<2445>, Constant:i32<8000000>
>> >> _______________________________________________
>> >> cfe-dev mailing list
>> >> cfe-dev at lists.llvm.org
>> >> https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>> >>
>> >--
>> >~Craig
>>
>-- 
>~Craig
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20211024/31259202/attachment.sig>