cannot build with msvc x86 #38
Comments
I have tried it on windows 10 with vs2015. Everything is ok. Maybe you could try to add an undefine in xmake.lua as below: if is_plat("windows") then
add_cxflags("-Ox", "-fp:fast", "-EHsc")
add_files("**.cpp")
if is_arch("x64") then
add_files("co/context/context_x64.asm")
else
add_files("co/context/context_x86.asm")
add_undefines("_WIN64")
end
end |
|
As I can know, on 32 bit platform, #ifdef _WIN32
#pragma warning (disable:4800)
#include <intrin.h>
#ifndef _WIN64
#define WIN32_LEAN_AND_MEAN
#include <Windows.h>
#endifYou may simply add |
|
Hi, thank you for your feedback. This is strange, I tried your solution but still got the same error. Meanwhile, I check However, I check MSDN and it seems these
|
|
I have found it in LONGLONG
FORCEINLINE
_InterlockedIncrement64 (
_Inout_ _Interlocked_operand_ LONGLONG volatile *Addend
)
{
LONGLONG Old;
do {
Old = *Addend;
} while (InterlockedCompareExchange64(Addend,
Old + 1,
Old) != Old);
return Old + 1;
} |
|
Now I find it in |
|
Anyway, would it be possible to disable these |
|
Should also work on x86? May you have a test on it? |
|
Not work here 😂 Maybe we need more guys or more machines to test if this is a real issue. |
|
I got another computer and freshly installed visual studio 2019 and xmake. Still x86 cannot be built, neither with 2015 build toolset nor 2019 build toolset. I guess this is an actual problem. |
|
I have the same issue when compiling in visual studio 2019 for x86. #define __MACHINEARM_ARM64_X64 __MACHINEZ #define __MACHINEZ(X) /* NOTHING */ I guess izhengfan is right that x86 may not support it under certain versions, and they put /* NOTHING */ in their defination |
|
@izhengfan @dasuren |
|
@idealvin Need to add declaration through pragma intrinsic first. msvc will use it as a built-in function without links. __int64 _InterlockedExchangeAdd64(__int64 volatile* Destination, __int64 Value);
#pragma intrinsic(_InterlockedExchangeAdd64)You can see https://github.com/tboox/tbox/blob/master/src/tbox/platform/windows/atomic64.h |
|
@idealvin undef _MANAGED does not work. According to @waruqi 's comments, I added the lines below: diff --git a/base/win/atomic.h b/base/win/atomic.h
index d53e146..733832e 100644
--- a/base/win/atomic.h
+++ b/base/win/atomic.h
@@ -8,6 +8,20 @@
#ifndef _WIN64
#define WIN32_LEAN_AND_MEAN
#include <Windows.h>
+__int64 _InterlockedIncrement64(__int64 volatile* Dest);
+#pragma intrinsic(_InterlockedIncrement64)
+__int64 _InterlockedDecrement64(__int64 volatile* Dest);
+#pragma intrinsic(_InterlockedDecrement64)
+__int64 _InterlockedExchangeAdd64(__int64 volatile* Dest, __int64 value);
+#pragma intrinsic(_InterlockedExchangeAdd64)
+__int64 _InterlockedOr64(__int64 volatile* Dest, __int64 value);
+#pragma intrinsic(_InterlockedOr64)
+__int64 _InterlockedAnd64(__int64 volatile* Dest, __int64 value);
+#pragma intrinsic(_InterlockedAnd64)
+__int64 _InterlockedXor64(__int64 volatile* Dest, __int64 value);
+#pragma intrinsic(_InterlockedXor64)
+__int64 _InterlockedExchange64(__int64 volatile* Dest, __int64 value);
+#pragma intrinsic(_InterlockedExchange64)
#endifNow building of base, rpcgen and test pass, but unitest has link error: |
Try adding #ifdef __cplusplus
extern "C" {
#endif
__int64 _InterlockedExchangeAdd64(__int64 volatile* Destination, __int64 Value);
#pragma intrinsic(_InterlockedExchangeAdd64)
#ifdef __cplusplus
}
#endif |
|
Written like https://github.com/izhengfan/co/blob/dev-with-ci/base/win/atomic.h#L16, still not work. I made a tiny CI script, you can check the result at https://github.com/izhengfan/co/commit/4cf5ebe226c76b343752025212a2b4f0b17a63ca/checks?check_suite_id=375823384 |
|
Because only _InterlockedCompareExchange64 exists for x86/msvc, so we need wrap other atomic64 apis using _InterlockedCompareExchange64. |
|
Sorry I still do not understand how to do it, could you directly fix the code in this repo? |
|
Hi All, I moved all the 64 bits related functions to another block with #ifdef _WIN64. For 64 bit,it compiles fine and pass all the unit test.
which could be caused by the test cases like follows: where a _int64 parameter are passed to the 32bit application, where related code are skipped. Then I tried to replace _InterlockedExchange64 with some function which can be used under 32 bits and accept _int64 as parameter. However, it seems all the _int64 related functions are defined with _int64 parameter are end with **64. I remove all the 64 bit test case under x86 compiler, and now it pass all the test cased. So I have two questions: 1. do we need to use and test _int64 as input for x86 application? |
|
@dasuren I think this is a workaround, but not a good enough solution. For 32bit application, it is still desirable to have 64bit atomic operation. |
|
Need use _InterlockedCompareExchange64 to wrap other atomic64 apis for x86, because only _InterlockedCompareExchange64 exists for x86/msvc. for example: extern "C" {
__int64 _InterlockedCompareExchange64(__int64 volatile* Destination, __int64 Exchange, __int64 Comperand);
#pragma intrinsic(_InterlockedCompareExchange64)
}
static inline __int64 _InterlockedIncrement64 (__int64 volatile *Addend)
{
__int64 Old;
do {
Old = *Addend;
} while (_InterlockedCompareExchange64(Addend, Old + 1, Old) != Old);
return Old + 1;
} |
|
Thanks Waruqi for your suggestion and help. I did some research on "#pragma intrinsic" and FORCEINLINE. but still got confused by code where three parameter "declared" but only one is defined.And visual studio think it is legal....Guess I need to go back to the text books...wow, feels like I am using a fake C++ in the past several year.. :) |
|
Thanks to @waruqi 's instructions, I now successful fix this problem, like this: diff --git a/base/win/atomic.h b/base/win/atomic.h
index d53e146..a8e931f 100644
--- a/base/win/atomic.h
+++ b/base/win/atomic.h
@@ -8,6 +8,17 @@
#ifndef _WIN64
#define WIN32_LEAN_AND_MEAN
#include <Windows.h>
+extern "C" {
+__int64 _InterlockedCompareExchange64(__int64 volatile* Destination, __int64 Exchange, __int64 Comperand);
+#pragma intrinsic(_InterlockedCompareExchange64)
+}
+#define _InterlockedIncrement64 InterlockedIncrement64
+#define _InterlockedDecrement64 InterlockedDecrement64
+#define _InterlockedExchangeAdd64 InterlockedExchangeAdd64
+#define _InterlockedOr64 InterlockedOr64
+#define _InterlockedAnd64 InterlockedAnd64
+#define _InterlockedXor64 InterlockedXor64
+#define _InterlockedExchange64 InterlockedExchange64
#endif
namespace xx {Also, I notice that although x86 MSVC does not provide API like //@file winnt.h
LONGLONG
FORCEINLINE
_InlineInterlockedIncrement64 (
_Inout_ _Interlocked_operand_ LONGLONG volatile *Addend
)
{
LONGLONG Old;
do {
Old = *Addend;
} while (InterlockedCompareExchange64(Addend,
Old + 1,
Old) != Old);
return Old + 1;
}
#define InterlockedIncrement64 _InlineInterlockedIncrement64 |
|
@izhengfan //extern "C" {
//__int64 _InterlockedCompareExchange64(__int64 volatile* Destination, __int64 Exchange, __int64 Comperand);
//#pragma intrinsic(_InterlockedCompareExchange64)
//}Does it also work on your platform? |
|
@idealvin Yes, it still works. I guess win32 SDK has handled it well, so we don't need these lines. |
👌 Thanks for your work. |
Under
co/base/:The text was updated successfully, but these errors were encountered: