[X86] Enable constexpr on POPCNT intrinsics (PR31446)

[RKSimon]

Summary

This is a RFC first step patch to enable constexpr support and testing to a large number of x86 intrinsics.

All I've done here is provide a DEFAULT_FN_ATTRS_CONSTEXPR variant to our existing DEFAULT_FN_ATTRS tag approach that adds constexpr on c++ builds. The clang cuda headers do something similar.

I've started with POPCNT mainly as its tiny and are wrappers to generic __builtin_* intrinsics which already act as constexpr.

[craig.topper]

clang/lib/Headers/popcntintrin.h

16

Do we need to check for C++11?

[RKSimon]

clang/lib/Headers/popcntintrin.h

16

I don't actually know - its trivial to add a __cplusplus >= 201103L check but I don't know what we are expected to support in compiler headers like these.

[arsenm]

clang/test/CodeGen/popcnt-builtins.cpp

3 ↗

(On Diff #286584)

Missing required register target?

[RKSimon]

add c++11 limiter

[craig.topper]

clang/test/CodeGen/popcnt-builtins.cpp

3 ↗

(On Diff #286584)

Do we need a registered target if we're stopping at IR generation?

[RKSimon]

clang/test/CodeGen/popcnt-builtins.cpp

3 ↗

(On Diff #286584)

sure, oddly we don't appear to do this for most of the x86 builtins codegen tests - I'll see if I can cleanup those soon

[RKSimon]

Require x86-registered-target

[RKSimon]

clang/test/CodeGen/popcnt-builtins.cpp

3 ↗

(On Diff #286584)

Not really - I'm not sure if its good practice or not, some targets seem to do this but x86 not so much.

[arsenm]

clang/test/CodeGen/popcnt-builtins.cpp

3 ↗

(On Diff #286584)

There's a vague desire to be able to omit the target intrinsics if the target isn't built, which would imply eventually making the clang parts conditionally compiled

[RKSimon]

clang/test/CodeGen/popcnt-builtins.cpp

3 ↗

(On Diff #286584)

I'm blanking on this (loooooong week) - but isn't there a way to tell the driver to treat this as a c program? Then we can either add run lines that test it as 'c' in the (renamed) cpp file or 'c++11' in the original c file - I don't mind which.

[erichkeane]

clang/test/CodeGen/popcnt-builtins.cpp

3 ↗

(On Diff #286584)

-x c does it.

[RKSimon]

Added -x c++ RUN lines to the existing popcnt-builtins.c file - thanks @erichkeane

[RKSimon]

I think that's everything - I'd prefer to handle the registered target fixes in separate commits.

[RKSimon]

Added release note entry - this will need to revised as we add more intrinsics, so is just an initial placeholder.

[craig.topper]

Should we test the intrinsics in a constexpr context? Something like https://godbolt.org/z/6o5zPn should work with the intrinsic instead of the builtin

[RKSimon]

Added constexpr tests - I went for the -ve array size approach as thats the approach in clang\test\Sema\constant-builtins-2.c etc.

[craig.topper]

LGTM

clang/docs/ReleaseNotes.rst

 
 - ...
 
 CUDA Support in Clang
 ---------------------
 
 - ...
 
+X86 Support in Clang
+--------------------
+
+- The x86 intrinsics ``_mm_popcnt_u32`` and ``_mm_popcnt_u64`` may now be used
+  within constexpr expressions.
+
 Internal API Changes
 --------------------
 
 These are major API changes that have happened since the 11.0.0 release of
 Clang. If upgrading an external codebase that uses Clang as a library,
 this section should help get you past the largest hurdles of upgrading.
 
 - ...

clang/lib/Headers/popcntintrin.h

 /*===---- popcntintrin.h - POPCNT intrinsics -------------------------------===
  *
  * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
  * See https://llvm.org/LICENSE.txt for license information.
  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
  *
  *===-----------------------------------------------------------------------===
  */
 
 #ifndef __POPCNTINTRIN_H
 #define __POPCNTINTRIN_H
 
 /* Define the default attributes for the functions in this file. */
 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("popcnt")))
 
+#if defined(__cplusplus) && (__cplusplus >= 201103L)
+#define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS constexpr
+#else
+#define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS
+#endif
+
 /// Counts the number of bits in the source operand having a value of 1.
 ///
 /// \headerfile <x86intrin.h>
 ///
 /// This intrinsic corresponds to the <c> POPCNT </c> instruction.
 ///
 /// \param __A
 ///    An unsigned 32-bit integer operand.
 /// \returns A 32-bit integer containing the number of bits with value 1 in the
 ///    source operand.
-static __inline__ int __DEFAULT_FN_ATTRS
+static __inline__ int __DEFAULT_FN_ATTRS_CONSTEXPR
 _mm_popcnt_u32(unsigned int __A)
 {
   return __builtin_popcount(__A);
 }
 
 #ifdef __x86_64__
 /// Counts the number of bits in the source operand having a value of 1.
 ///
 /// \headerfile <x86intrin.h>
 ///
 /// This intrinsic corresponds to the <c> POPCNT </c> instruction.
 ///
 /// \param __A
 ///    An unsigned 64-bit integer operand.
 /// \returns A 64-bit integer containing the number of bits with value 1 in the
 ///    source operand.
-static __inline__ long long __DEFAULT_FN_ATTRS
+static __inline__ long long __DEFAULT_FN_ATTRS_CONSTEXPR
 _mm_popcnt_u64(unsigned long long __A)
 {
   return __builtin_popcountll(__A);
 }
 #endif /* __x86_64__ */
 
 #undef __DEFAULT_FN_ATTRS
+#undef __DEFAULT_FN_ATTRS_CONSTEXPR
 
 #endif /* __POPCNTINTRIN_H */

clang/test/CodeGen/popcnt-builtins.c

-// RUN: %clang_cc1 -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +popcnt -emit-llvm -o - | FileCheck %s --check-prefixes=CHECK,CHECK-POPCNT
-// RUN: %clang_cc1 -ffreestanding %s -triple=x86_64-apple-darwin -emit-llvm -o - | FileCheck %s
-
+// RUN: %clang_cc1 -x c -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +popcnt -emit-llvm -o - | FileCheck %s --check-prefixes=CHECK,CHECK-POPCNT
+// RUN: %clang_cc1 -x c++ -std=c++11 -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +popcnt -emit-llvm -o - | FileCheck %s --check-prefixes=CHECK,CHECK-POPCNT
+// RUN: %clang_cc1 -x c -ffreestanding %s -triple=x86_64-apple-darwin -emit-llvm -o - | FileCheck %s
+// RUN: %clang_cc1 -x c++ -std=c++11 -ffreestanding %s -triple=x86_64-apple-darwin -emit-llvm -o - | FileCheck %s
 
 #include <x86intrin.h>
 
 #ifdef __POPCNT__
 int test_mm_popcnt_u32(unsigned int __X) {
   //CHECK-POPCNT: call i32 @llvm.ctpop.i32
   return _mm_popcnt_u32(__X);
 }
long long test_popcnt64(unsigned long long __X) {
   return _popcnt64(__X);
 }
 
 long long test__popcntq(unsigned long long __X) {
   //CHECK: call i64 @llvm.ctpop.i64
   return __popcntq(__X);
 }
 #endif
+
+// Test constexpr handling.
+#if defined(__cplusplus) && (__cplusplus >= 201103L)
+#if defined(__POPCNT__)
+char ctpop32_0[_mm_popcnt_u32(0x00000000) == 0 ? 1 : -1];
+char ctpop32_1[_mm_popcnt_u32(0x000000F0) == 4 ? 1 : -1];
+
+#ifdef __x86_64__
+char ctpop64_0[_mm_popcnt_u64(0x0000000000000000ULL) == 0 ? 1 : -1];
+char ctpop64_1[_mm_popcnt_u64(0xF000000000000001ULL) == 5 ? 1 : -1];
+#endif
+#endif
+#endif