llvm dev - May 2020 - questions about legalize type for vector

Hi,


I am writing a new backend and I met some problem.
I want to support in IR: sdiv v32i16 %a, %b. But the target only support sdiv
v16i32 %a, %b natively.
Could I promote v32i16 to v32i32, split v32i32 to two v16i32 , truncate v16i32
to v16i16 and merge two v16i16 into one v32i16?
Is this way possible?
I saw the code fragment in TargetLowering.h:
=========================================================== 169  /// This enum
indicates whether a types are legal for a target, and if not,
 170  /// what action should be used to make them valid.
 171  enum LegalizeTypeAction : uint8_t {
 172    TypeLegal,           // The target natively supports this type.
 173    TypePromoteInteger,  // Replace this integer with a larger one.
 174    TypeExpandInteger,   // Split this integer into two of half the size.
 175    TypeSoftenFloat,     // Convert this float to a same size integer type.
 176    TypeExpandFloat,     // Split this float into two of half the size.
 177    TypeScalarizeVector, // Replace this one-element vector with its
element.
 178    TypeSplitVector,     // Split this vector into two of half the size.
 179    TypeWidenVector,     // This vector should be widened into a larger
vector.
 180    TypePromoteFloat,    // Replace this float with a larger one.
 181    TypeSoftPromoteHalf, // Soften half to i16 and use float to do
arithmetic.
 182  };
============================================================I don't quite
understand the action TypeWidenVector. What does it do?  Does this action do
like promote v32i16 to v32i32?
If this way is possible, where should I set the LegalizeTypeAction in the source
code?
Is it in the constructor of class XXXTargetLowering? And How could I set the
action to manipulate the process as I want?
Thanks!


Best regards,
Jerry

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200517/ec06448a/attachment.html>

Those are the actions for the type legalizer. They specify how to legalize
a type that has no supported operations and isn't assigned to any register
class by your target. The action will be applied to every operation on that
type.

I assume v32i16 is otherwise supported for your target and only division is
missing. If that's the case you'll need to define a Custom lowering
operation for your target. So you'll need to call
setOperationAction(ISD::SDIV, v32i16, Custom) from your TargetLowering
constructor.  Then write custom code in LowerOperation to custom handle
v32i16 ISD::SDIV. From there you can split the operands, SIGN_EXTEND each
to v16i32 do the SDIV, TRUNCATE the results back to v16i16, and use a
CONCAT_VECTORS to join them back to v32i16.

~Craig


On Sat, May 16, 2020 at 7:23 PM 林政宗 via llvm-dev <llvm-dev at
lists.llvm.org>
wrote:
> Hi,
>
> I am writing a new backend and I met some problem.
> I want to support in IR: sdiv v32i16 %a, %b. But the target only support
> sdiv v16i32 %a, %b natively.
> Could I promote v32i16 to v32i32, split v32i32 to two v16i32 , truncate
> v16i32 to v16i16 and merge two v16i16 into one v32i16?
> Is this way possible?
> I saw the code fragment in TargetLowering.h:
> ===========================================================>  169  ///
This enum indicates whether a types are legal for a target, and
> if not,
>  170  /// what action should be used to make them valid.
>  171  enum LegalizeTypeAction : uint8_t {
>  172    TypeLegal,           // The target natively supports this type.
>  173    TypePromoteInteger,  // Replace this integer with a larger one.
>  174    TypeExpandInteger,   // Split this integer into two of half the
> size.
>  175    TypeSoftenFloat,     // Convert this float to a same size integer
> type.
>  176    TypeExpandFloat,     // Split this float into two of half the size.
>  177    TypeScalarizeVector, // Replace this one-element vector with its
> element.
>  178    TypeSplitVector,     // Split this vector into two of half the
> size.
>  179    TypeWidenVector,     // This vector should be widened into a
> larger vector.
>  180    TypePromoteFloat,    // Replace this float with a larger one.
>  181    TypeSoftPromoteHalf, // Soften half to i16 and use float to do
> arithmetic.
>  182  };
> ============================================================> I
don't quite understand the action TypeWidenVector. What does it do?
> Does this action do like promote v32i16 to v32i32?
> If this way is possible, where should I set the LegalizeTypeAction in the
> source code?
> Is it in the constructor of class XXXTargetLowering? And How could I set
> the action to manipulate the process as I want?
> Thanks!
>
> Best regards,
> Jerry
>
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200516/e58e2f8a/attachment.html>