[HLSL] Add WaveActiveCountBits as Langugage builtin function for HLSL

[python3kgae]

Summary

One clang builtins are introduced
uint WaveActiveCountBits( bool bBit ) as Langugage builtin function for HLSL.

The detail for WaveActiveCountBits is at
https://github.com/microsoft/DirectXShaderCompiler/wiki/Wave-Intrinsics#uint-waveactivecountbits-bool-bbit-

This is only clang part change to make WaveActiveCountBits into AST.
llvm intrinsic for WaveActiveCountBits will be add in separate PR.

[Anastasia]

ok, so the reason you are adding this to clang is that it needs to be mapped into a target intrinsic?

[python3kgae]

In D126857#3554070, @Anastasia wrote:

ok, so the reason you are adding this to clang is that it needs to be mapped into a target intrinsic?

Yes. The plan is to create a target intrinsic for WaveActiveCountBits which map to this clang builtin directly.

[Anastasia]

clang/include/clang/Basic/Builtins.def

1697

FYI we most of time try to add a builtin name using a reserved identifier which is not part of the language standard (mostly prefixed by __). Then we add regular function that just calls the clang builtins. This way provides more flexibility to the implementers. However you might not need this... in which case using original name avoids an extra call.

clang/test/SemaHLSL/Wave.hlsl

8

if you are planning to add a proper CodeGen test later then here it might be sufficient to have a -verify test and check that the builtin is accepted with the right arguments... you might want to add a test checking that a diagnostic is provided when the arguments are wrong... although this is not strictly necessary as we already test clang builtins quite extensively.

[python3kgae]

Change to verify test.

clang/include/clang/Basic/Builtins.def

1697

Yes. For this one, it is without prefix to avoid extra call.
I'm following things like enqueue_kernel in opencl.
For other things support overload which has to make extra call, I'll add the prefix.

clang/test/SemaHLSL/Wave.hlsl

8

Changed to verify test.

[Anastasia]

clang/include/clang/Basic/Builtins.def

1697

Ok, although enqueue_kernel was implemented as a builtin because it has a weird variadic prototype that can't be implemented using normal features of C/C++ languages. Hence it is a builtin with custom SemaChecking.

clang/test/SemaHLSL/Wave.hlsl

8

ok, then also -ast-dump -o - can be removed from the Run line. Also you probably don't need -finclude-default-header here wither?

[python3kgae]

Simplify test.

[python3kgae]

clang/include/clang/Basic/Builtins.def

1697

I see.
Since HLSL also has things cannot implemented using HLSL itself, we cannot put all HLSL intrinsic in one place anyway.
So when it is possible to reduce an extra call, I just reduce it.

[Anastasia]

LGTM! Thanks

[beanz]

clang/include/clang/Basic/Builtins.def

1697

I don't think this was the right decision. It is trivial for the optimizer to remove a function that is more or less a tail-call of another function, and since HLSL has no call semantics anyways they will disappear.

It would be much better if this was a function that called the builtin, this could even be trivially implemented as a sub-header of hlsl.h as a wrapper function.

[beanz]

clang/include/clang/Basic/Builtins.def

1697

Let me expand a little here. I think we need to change this. We should name builtins as builtins (i.e. __builtin_hlsl_...), and we can implement a wrapper function in one of the headers from hlsl.h.

There are two reasons I think we should take this approach:

1. Matching existing builtin patterns makes it clear in the code
2. Clang builtins don't populate in the AST for AST queries, which impacts how the function for IDE tooling. Having the HLSL user-surfaced functions exist in the AST will allow for better IDE tooling for autocomplete.

If hlsl.h gets too slow to parse, there are other approaches we can take like using the external sema source, modules or PCH to speed it up later.

clang/include/clang/Basic/Builtins.h

enum LanguageID {
   MS_LANG = 0x10,            // builtin requires MS mode.
   OMP_LANG = 0x20,           // builtin requires OpenMP.
   CUDA_LANG = 0x40,          // builtin requires CUDA.
   COR_LANG = 0x80,           // builtin requires use of 'fcoroutine-ts' option.
   OCL_GAS = 0x100,           // builtin requires OpenCL generic address space.
   OCL_PIPE = 0x200,          // builtin requires OpenCL pipe.
   OCL_DSE = 0x400,           // builtin requires OpenCL device side enqueue.
   ALL_OCL_LANGUAGES = 0x800, // builtin for OCL languages.
+  HLSL_LANG = 0x1000,        // builtin requires HLSL.
   ALL_LANGUAGES = C_LANG | CXX_LANG | OBJC_LANG, // builtin for all languages.
   ALL_GNU_LANGUAGES = ALL_LANGUAGES | GNU_LANG,  // builtin requires GNU mode.
   ALL_MS_LANGUAGES = ALL_LANGUAGES | MS_LANG     // builtin requires MS mode.
 };
 
 namespace Builtin {
 enum ID {
   NotBuiltin  = 0,      // This is not a builtin function.

clang/include/clang/Basic/Builtins.def

 
 // Builtins for os_log/os_trace
 BUILTIN(__builtin_os_log_format_buffer_size, "zcC*.", "p:0:nut")
 BUILTIN(__builtin_os_log_format, "v*v*cC*.", "p:0:nt")
 
 // CUDA/HIP
 LANGBUILTIN(__builtin_get_device_side_mangled_name, "cC*.", "ncT", CUDA_LANG)
 
+// HLSL
+LANGBUILTIN(WaveActiveCountBits, "Uib", "nc", HLSL_LANG)
+
 // Builtins for XRay
 BUILTIN(__xray_customevent, "vcC*z", "")
 BUILTIN(__xray_typedevent, "vzcC*z", "")
 
 // Win64-compatible va_list functions
 BUILTIN(__builtin_ms_va_start, "vc*&.", "nt")
 BUILTIN(__builtin_ms_va_end, "vc*&", "n")
 BUILTIN(__builtin_ms_va_copy, "vc*&c*&", "n")
 
 // Arithmetic Fence: to prevent FP reordering and reassociation optimizations
 LANGBUILTIN(__arithmetic_fence, "v.", "t", ALL_LANGUAGES)
 
 #undef BUILTIN
 #undef LIBBUILTIN
 #undef LANGBUILTIN

clang/test/SemaHLSL/Wave.hlsl

+// RUN: %clang_cc1  -x hlsl -triple dxil--shadermodel6.7-library  %s  -verify
+
+// Make sure WaveActiveCountBits is accepted.
+
+// expected-no-diagnostics
+unsigned foo(bool b) {
+    return WaveActiveCountBits(b);
+}