[clang] [Clang] [docs] [MSVC] Add sections on `__forceinline` and intrinsic behaviour differences between Clang and MSVC (PR #99426)
Max Winkler via cfe-commits
cfe-commits at lists.llvm.org
Wed Jul 17 21:02:05 PDT 2024
https://github.com/MaxEW707 created https://github.com/llvm/llvm-project/pull/99426
We have had quite a few issues created around how Clang treats intrinsics vs how MSVC treats intrinsics.
While I was writing this I also added some sections on behaviour changes that caught me while porting my MSVC codebase to clang-cl.
Hopefully we can point issues around intrinsics to this doc and hopefully it is useful to others who run into similar behaviour differences.
The behaviour differences highlighted here are differences, as far as I am aware, that we do not intend to change or fix for MSVC.
>From 37f6bb1a6bc061357f1971c59c101ef7e6b53141 Mon Sep 17 00:00:00 2001
From: MaxEW707 <max.enrico.winkler at gmail.com>
Date: Thu, 11 Jul 2024 23:07:35 -0700
Subject: [PATCH 1/3] Update MSVC compatibility docs
---
clang/docs/MSVCCompatibility.rst | 133 +++++++++++++++++++++++++++++++
1 file changed, 133 insertions(+)
diff --git a/clang/docs/MSVCCompatibility.rst b/clang/docs/MSVCCompatibility.rst
index b2486052abf9a..f56c6e0d13682 100644
--- a/clang/docs/MSVCCompatibility.rst
+++ b/clang/docs/MSVCCompatibility.rst
@@ -154,3 +154,136 @@ a hint suggesting how to fix the problem.
As of this writing, Clang is able to compile a simple ATL hello world
application. There are still issues parsing WRL headers for modern Windows 8
apps, but they should be addressed soon.
+
+__forceinline behavior
+======================
+
+``__forceinline`` behaves like ``[[clang::always_inline]]``.
+Inlining is always attempted regardless of optimization level.
+
+This differs from MSVC where ``__forceinline`` is only respected once inline expansion is enabled
+which allows any function marked implicitly or explicitly ``inline`` or ``__forceinline`` to be expanded.
+
+Therefore functions marked ``__forceinline`` will be expanded when the optimization level is ``/Od`` unlike
+MSVC where that is not the case.
+If this is an issue that cannot be easily worked around in your codebase by wrapping ``__forceinline`` behind
+a macro please file a bug report.
+
+SIMD and instruction set intrinsic behavior
+===========================================
+
+Clang follows the GCC model for intrinsics and not the MSVC model.
+There are currently no plans to support the MSVC model.
+
+MSVC intrinsics always emit the machine instruction the intrinsic models regardless of the compile time options specified.
+For example ``__popcnt`` always emits the x86 popcnt instruction even if the compiler does not have the option enabled to emit popcnt on its own volition.
+
+There are two common cases where code that compiles with MSVC will need reworking to build on clang.
+Assume the examples are only built with `-msse2` so we do not have the intrinsics at compile time.
+
+.. code-block:: c++
+
+ unsigned PopCnt(unsigned v) {
+ if (HavePopCnt)
+ return __popcnt(v);
+ else
+ return GenericPopCnt(v);
+ }
+
+.. code-block:: c++
+
+ __m128 dot4_sse3(__m128 v0, __m128 v1) {
+ __m128 r = _mm_mul_ps(v0, v1);
+ r = _mm_hadd_ps(r, r);
+ r = _mm_hadd_ps(r, r);
+ return r;
+ }
+
+Clang expects that either you have compile time support for the target features, `-msse3` and `-mpopcnt`, you mark the function with the expected target feature or use runtime detection with an indirect call.
+
+.. code-block:: c++
+
+ __attribute__((__target__("sse3"))) __m128 dot4_sse3(__m128 v0, __m128 v1) {
+ __m128 r = _mm_mul_ps(v0, v1);
+ r = _mm_hadd_ps(r, r);
+ r = _mm_hadd_ps(r, r);
+ return r;
+ }
+
+The SSE3 dot product can be easily fixed by either building the translation unit with SSE3 support or using `__target__` to compile that specific function with SSE3 support.
+
+.. code-block:: c++
+
+ unsigned PopCnt(unsigned v) {
+ if (HavePopCnt)
+ return __popcnt(v);
+ else
+ return GenericPopCnt(v);
+ }
+
+The above must be changed to work with clang. If we mark the function with `__target__("popcnt")` then the compiler is free to emit popcnt at will which we do not want. While this isn't a concern in our small example it is a concern in larger functions with surrounding code around the intrinsics. Similar goes for compiling the translation unit with `-mpopcnt`.
+We must split each branch into its own function that can be called indirectly instead of using the intrinsic directly.
+
+.. code-block:: c++
+
+ __attribute__((__target__("popcnt"))) unsigned hwPopCnt(unsigned v) { return __popcnt(v); }
+ unsigned (*PopCnt)(unsigned) = HavePopCnt ? hwPopCnt : GenericPopCnt;
+
+.. code-block:: c++
+
+ __attribute__((__target__("popcnt"))) unsigned hwPopCnt(unsigned v) { return __popcnt(v); }
+ unsigned PopCnt(unsigned v) {
+ if (HavePopCnt)
+ return hwPopCnt(v);
+ else
+ return GenericPopCnt(v);
+ }
+
+In the above example `hwPopCnt` will not be inlined into `PopCnt` since `PopCnt` doesn't have the popcnt target feature.
+With a larger function that does real work the function call overhead is negligible. However in our popcnt example there is the function call
+overhead. There is no analog for this specific MSVC behavior in clang.
+
+For clang we effectively have to create the dispatch function ourselves to each specfic implementation.
+
+SIMD vector types
+=================
+
+Clang's simd vector types are builtin types and not user defined types as in MSVC. This does have some observable behavior changes.
+We will look at the x86 `__m128` type for the examples below but the statements apply to all vector types imcluding ARM's `float32x4_t`.
+
+There are no members that can be accessed on the vector types. Vector types are not structs in clang.
+You cannot use ``__m128.m128_f32[0]`` to access the first element of the `__m128`.
+This also means struct initialization like ``__m128{ { 0.0f, 0.0f, 0.0f, 0.0f } }`` will not compile with clang.
+
+Since vector types are builtin types, clang implements operators on them natively.
+
+.. code-block:: c++
+
+ #ifdef _MSC_VER
+ __m128 operator+(__m128 a, __m128 b) { return _mm_add_ps(a, b); }
+ #endif
+
+The above code will fail to compile since overloaded 'operator+' must have at least one parameter of class or enumeration type.
+You will need to fix such code to have the check ``#if defined(_MSC_VER) && !defined(__clang__)``.
+
+Since `__m128` is not a class type in clang any overloads after a template definition will not be considered.
+
+.. code-block:: c++
+
+ template<class T>
+ void foo(T) {}
+
+ template<class T>
+ void bar(T t) {
+ foo(t);
+ }
+
+ void foo(__m128) {}
+
+ int main() {
+ bar(_mm_setzero_ps());
+ }
+
+With MSVC ``foo(__m128)`` will be selected but with clang ``foo<__m128>()`` will be selected since on clang `__m128` is a builtin type.
+
+In general the takeaway is `__m128` is a builtin type on clang while a class type on MSVC.
>From fb433060b0f2631d856f3567c8886dccbd389e5a Mon Sep 17 00:00:00 2001
From: MaxEW707 <max.enrico.winkler at gmail.com>
Date: Wed, 17 Jul 2024 20:38:16 -0700
Subject: [PATCH 2/3] Fix wording
---
clang/docs/MSVCCompatibility.rst | 9 +++------
1 file changed, 3 insertions(+), 6 deletions(-)
diff --git a/clang/docs/MSVCCompatibility.rst b/clang/docs/MSVCCompatibility.rst
index f56c6e0d13682..915e6ba602e1e 100644
--- a/clang/docs/MSVCCompatibility.rst
+++ b/clang/docs/MSVCCompatibility.rst
@@ -163,11 +163,8 @@ Inlining is always attempted regardless of optimization level.
This differs from MSVC where ``__forceinline`` is only respected once inline expansion is enabled
which allows any function marked implicitly or explicitly ``inline`` or ``__forceinline`` to be expanded.
-
Therefore functions marked ``__forceinline`` will be expanded when the optimization level is ``/Od`` unlike
-MSVC where that is not the case.
-If this is an issue that cannot be easily worked around in your codebase by wrapping ``__forceinline`` behind
-a macro please file a bug report.
+MSVC where ``__forceinline`` will not be expanded under ``/Od``.
SIMD and instruction set intrinsic behavior
===========================================
@@ -221,7 +218,7 @@ The SSE3 dot product can be easily fixed by either building the translation unit
return GenericPopCnt(v);
}
-The above must be changed to work with clang. If we mark the function with `__target__("popcnt")` then the compiler is free to emit popcnt at will which we do not want. While this isn't a concern in our small example it is a concern in larger functions with surrounding code around the intrinsics. Similar goes for compiling the translation unit with `-mpopcnt`.
+The above ``PopCnt`` example must be changed to work with clang. If we mark the function with `__target__("popcnt")` then the compiler is free to emit popcnt at will which we do not want. While this isn't a concern in our small example it is a concern in larger functions with surrounding code around the intrinsics. Similar goes for compiling the translation unit with `-mpopcnt`.
We must split each branch into its own function that can be called indirectly instead of using the intrinsic directly.
.. code-block:: c++
@@ -239,7 +236,7 @@ We must split each branch into its own function that can be called indirectly in
return GenericPopCnt(v);
}
-In the above example `hwPopCnt` will not be inlined into `PopCnt` since `PopCnt` doesn't have the popcnt target feature.
+In the above example ``hwPopCnt`` will not be inlined into ``PopCnt`` since ``PopCnt`` doesn't have the popcnt target feature.
With a larger function that does real work the function call overhead is negligible. However in our popcnt example there is the function call
overhead. There is no analog for this specific MSVC behavior in clang.
>From 85d68c5559e20f89a5a65fcd5e5819c1c3be8923 Mon Sep 17 00:00:00 2001
From: MaxEW707 <max.enrico.winkler at gmail.com>
Date: Wed, 17 Jul 2024 20:59:31 -0700
Subject: [PATCH 3/3] wording
---
clang/docs/MSVCCompatibility.rst | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/clang/docs/MSVCCompatibility.rst b/clang/docs/MSVCCompatibility.rst
index 915e6ba602e1e..cb5d89c6fd64f 100644
--- a/clang/docs/MSVCCompatibility.rst
+++ b/clang/docs/MSVCCompatibility.rst
@@ -218,7 +218,7 @@ The SSE3 dot product can be easily fixed by either building the translation unit
return GenericPopCnt(v);
}
-The above ``PopCnt`` example must be changed to work with clang. If we mark the function with `__target__("popcnt")` then the compiler is free to emit popcnt at will which we do not want. While this isn't a concern in our small example it is a concern in larger functions with surrounding code around the intrinsics. Similar goes for compiling the translation unit with `-mpopcnt`.
+The above ``PopCnt`` example must be changed to work with clang. If we mark the function with `__target__("popcnt")` then the compiler is free to emit popcnt at will which we do not want. While this isn't a concern in our small example it is a concern in larger functions with surrounding code around the intrinsics. Similar reasoning for compiling the translation unit with `-mpopcnt`.
We must split each branch into its own function that can be called indirectly instead of using the intrinsic directly.
.. code-block:: c++
More information about the cfe-commits
mailing list