In order to get llvm to compile on Windows with MSVC 10, I have to disable
inline on utostr to avoid internal compiler failure, and disable _xgetbv call in
OsSupportsAVX because it's only defined in MSVC 2012 compiler. The first
patch (noinline) is known for many months now. I'm a casual llvm user, and
don't know the proper channels to go through to submit a patch. Could
someone tell me, where do I submit patches like this one below? Or maybe why I
shouldn't try to do it.
Sincerely,Sergiy
Index:
include/llvm/ADT/StringExtras.h===================================================================---
include/llvm/ADT/StringExtras.h (revision 179701)+++
include/llvm/ADT/StringExtras.h (working copy)@@ -84,7 +84,7 @@ return
std::string(BufPtr, Buffer+11); } -static inline std::string utostr(uint64_t X,
bool isNeg = false) {+static __declspec(noinline) inline std::string
utostr(uint64_t X, bool isNeg = false) { char Buffer[21]; char *BufPtr =
Buffer+21; Index:
lib/Support/Host.cpp===================================================================---
lib/Support/Host.cpp (revision 179701)+++ lib/Support/Host.cpp (working copy)@@
-119,7 +119,7 @@ // there is no easy way to conditionally compile based on the
assembler used. int rEAX, rEDX; __asm__ (".byte 0x0f, 0x01, 0xd0"
: "=a" (rEAX), "=d" (rEDX) : "c" (0));-#elif
defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 160040219+#elif
defined(_MSC_FULL_VER) && _MSC_FULL_VER > 160040219 unsigned long
long rEAX = _xgetbv(_XCR_XFEATURE_ENABLED_MASK); #else int rEAX = 0; //
Ensures we return false
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130417/6a8c8691/attachment.html>
On Wed, Apr 17, 2013 at 11:36 AM, Sergiy Migdalskiy <migdalskiy at hotmail.com> wrote:> In order to get llvm to compile on Windows with MSVC 10, I have to disable > inline on utostr to avoid internal compiler failure, and disable _xgetbv > call in OsSupportsAVX because it's only defined in MSVC 2012 compiler. The > first patch (noinline) is known for many months now. I'm a casual llvm user, > and don't know the proper channels to go through to submit a patch. Could > someone tell me, where do I submit patches like this one below?http://llvm.org/docs/DeveloperPolicy.html> Or maybe why > I shouldn't try to do it.We're generally happy to apply patches to ensure LLVM compiles on relevant compilers. For ICEs it might help to know what the precise problem is & to have a bug filed on the compiler.> > Sincerely, > Sergiy > > Index: include/llvm/ADT/StringExtras.h > ==================================================================> --- include/llvm/ADT/StringExtras.h (revision 179701) > +++ include/llvm/ADT/StringExtras.h (working copy) > @@ -84,7 +84,7 @@ > return std::string(BufPtr, Buffer+11); > } > > -static inline std::string utostr(uint64_t X, bool isNeg = false) { > +static __declspec(noinline) inline std::string utostr(uint64_t X, bool > isNeg = false) { > char Buffer[21]; > char *BufPtr = Buffer+21; > > Index: lib/Support/Host.cpp > ==================================================================> --- lib/Support/Host.cpp (revision 179701) > +++ lib/Support/Host.cpp (working copy) > @@ -119,7 +119,7 @@ > // there is no easy way to conditionally compile based on the assembler > used. > int rEAX, rEDX; > __asm__ (".byte 0x0f, 0x01, 0xd0" : "=a" (rEAX), "=d" (rEDX) : "c" (0)); > -#elif defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 160040219 > +#elif defined(_MSC_FULL_VER) && _MSC_FULL_VER > 160040219 > unsigned long long rEAX = _xgetbv(_XCR_XFEATURE_ENABLED_MASK); > #else > int rEAX = 0; // Ensures we return false > > > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev >
>From the DeveloperPolicy.html document I gathered I need to send a patch to this list (which I did); could you clarify if I misunderstood it?The second patch is really trivial and I think it's worth applying - it seems like a typo by someone who tested on MSVC 11 but not 10; or maybe my local instance of MSVC10 is somehow deficient, of course. The first patch is adding MSVC-specific __declspec(noinline) - that's the whole patch - so if someone experienced who has access to all the platforms LLVM compiles on could test it (and/or replace it with something cross-platform; or #ifdef it out everywhere non-MS), I think that would be worth applying to the trunk. I don't have those resources. The cause of ICE is pretty much right here - inlining utostr crashes MSVC cl.exe. I read about it a few months ago elsewhere and I think the bug was filed with MS (I'm not positive and I have no idea how/where to check it). I don't think I can find much more precise cause within reasonable time, ICE doesn't reproduce in simple cases and I'm not committed enough to do a lengthy binary search of the combination of factors that causes this particular ICE when I only have a lukewarm hope MS would ever give this bug a look. Either way, I thought I'd try to push this through. It's not high-value enough to try harder, but it'd be great if someone could help me submit it - it's a really simple couple of patches that make LLVM/clang compile on MSVC10 without further modifications. Regards,Sergiy> Date: Wed, 17 Apr 2013 21:56:43 -0700 > Subject: Re: [LLVMdev] Patch to compile LLVM with MSVC 2010 > From: dblaikie at gmail.com > To: migdalskiy at hotmail.com > CC: llvmdev at cs.uiuc.edu > > On Wed, Apr 17, 2013 at 11:36 AM, Sergiy Migdalskiy > <migdalskiy at hotmail.com> wrote: > > In order to get llvm to compile on Windows with MSVC 10, I have to disable > > inline on utostr to avoid internal compiler failure, and disable _xgetbv > > call in OsSupportsAVX because it's only defined in MSVC 2012 compiler. The > > first patch (noinline) is known for many months now. I'm a casual llvm user, > > and don't know the proper channels to go through to submit a patch. Could > > someone tell me, where do I submit patches like this one below? > > http://llvm.org/docs/DeveloperPolicy.html > > > Or maybe why > > I shouldn't try to do it. > > We're generally happy to apply patches to ensure LLVM compiles on > relevant compilers. For ICEs it might help to know what the precise > problem is & to have a bug filed on the compiler. > > > > > Sincerely, > > Sergiy > > > > Index: include/llvm/ADT/StringExtras.h > > ==================================================================> > --- include/llvm/ADT/StringExtras.h (revision 179701) > > +++ include/llvm/ADT/StringExtras.h (working copy) > > @@ -84,7 +84,7 @@ > > return std::string(BufPtr, Buffer+11); > > } > > > > -static inline std::string utostr(uint64_t X, bool isNeg = false) { > > +static __declspec(noinline) inline std::string utostr(uint64_t X, bool > > isNeg = false) { > > char Buffer[21]; > > char *BufPtr = Buffer+21; > > > > Index: lib/Support/Host.cpp > > ==================================================================> > --- lib/Support/Host.cpp (revision 179701) > > +++ lib/Support/Host.cpp (working copy) > > @@ -119,7 +119,7 @@ > > // there is no easy way to conditionally compile based on the assembler > > used. > > int rEAX, rEDX; > > __asm__ (".byte 0x0f, 0x01, 0xd0" : "=a" (rEAX), "=d" (rEDX) : "c" (0)); > > -#elif defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 160040219 > > +#elif defined(_MSC_FULL_VER) && _MSC_FULL_VER > 160040219 > > unsigned long long rEAX = _xgetbv(_XCR_XFEATURE_ENABLED_MASK); > > #else > > int rEAX = 0; // Ensures we return false > > > > > > _______________________________________________ > > LLVM Developers mailing list > > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130418/c72ffe8c/attachment.html>
On Wed, Apr 17, 2013 at 2:36 PM, Sergiy Migdalskiy <migdalskiy at hotmail.com> wrote:> In order to get llvm to compile on Windows with MSVC 10, I have to disable > inline on utostr to avoid internal compiler failure, and disable _xgetbv > call in OsSupportsAVX because it's only defined in MSVC 2012 compiler.The _xgetbv requires Visual Studio 2010 SP1 or greater, not VS 2012. http://software.intel.com/en-us/articles/introduction-to-intel-advanced-vector-extensions MSDN's documentation wasn't updated until 2012 though, so it's a bit confusing. That being said, it should still compile in a non-SP1 VS 2010, but I believe the original code does that because 160040219 is the identifier for 2010 SP1 ~Aaron
Apparently Analagous Threads
- [LLVMdev] Patch to compile LLVM with MSVC 2010
- [LLVMdev] Patch to compile LLVM with MSVC 2010
- [LLVMdev] Patch to compile LLVM with MSVC 2010
- [LLVMdev] [PATCH] Detect Haswell subarchitecture (i.e. using -march=native)
- [LLVMdev] [PATCH] Detect Haswell subarchitecture (i.e. using -march=native)