[clang:frontend] Add `__builtin_elementwise_saturating_cast` intrinsic

[eisenwave]

llvm-project/libcxx/include/__numeric/saturation_arithmetic.h

Lines 105 to 117 in f6f1416

	template <__signed_or_unsigned_integer _Rp, __signed_or_unsigned_integer _Tp>
	_LIBCPP_HIDE_FROM_ABI constexpr _Rp __saturating_cast(_Tp __x) noexcept {
	// Saturation is impossible edge case when ((min _Rp) < (min _Tp) && (max _Rp) > (max _Tp)) and it is expected to be
	// optimized out by the compiler.
	// Handle overflow
	if (std::cmp_less(__x, std::numeric_limits<_Rp>::min()))
	return std::numeric_limits<_Rp>::min();
	if (std::cmp_greater(__x, std::numeric_limits<_Rp>::max()))
	return std::numeric_limits<_Rp>::max();
	// No overflow
	return static_cast<_Rp>(__x);
	}

Currently, we just have this software implementation, but future developments will necessitate an intrinsic:

• P2956R3: Allow std::simd overloads for saturating operations wants a SIMD version of the feature, and I'm not sure how easily the current implementation could be turned into vpmovsdw on x86.
• P4355R0: Floating-point std::saturating_cast (not yet mailed) extends std::saturating_cast with the ability to cast to float, which is exactly the existing behavior of llvm.fptoui.sat.*, except that it doesn't require the -fno-strict-float-cast-overflow flag to be specified. Irrespective of whether the proposal makes it in, it would be nice to have a way to emit fptoui.sat without needing to set a global flag.

[llvmorg-github-actions]

@llvm/issue-subscribers-clang-frontend

Author: Jan Schultke (eisenwave)

Details

https://github.com/llvm/llvm-project/blob/f6f14162f59d360509ad79ae96336223b724f44d/libcxx/include/__numeric/saturation_arithmetic.h#L105-L117

Currently, we just have this software implementation, but future developments will necessitate an intrinsic:

• P2956R3: Allow std::simd overloads for saturating operations wants a SIMD version of the feature, and I'm not sure how easily the current implementation could be turned into vpmovsdw on x86.
• P4355R0: Floating-point std::saturating_cast (not yet mailed) extends std::saturating_cast with the ability to cast to float, which is exactly the existing behavior of llvm.fptoui.sat.*, except that it doesn't require the -fno-strict-float-cast-overflow flag to be specified. Irrespective of whether the proposal makes it in, it would be nice to have a way to emit fptoui.sat without needing to set a global flag.