[X86] Fix the implementation of __readcr3/__writecr3 to work in 64-bit mode

[craig.topper]

Summary

We need to use a 64-bit type in 64-bit mode so a 64-bit register will get used in the generated assembly. I've also changed the constraints to just use "r" intead of "q". "q" forces to a only an a/b/c/d register in 32-bit mode, but I see no reason that would matter here. Still not sure why in 32-bit mode one uses long and the other uses int, but I'm just going to leave that alone.

Fixes Nico's note in PR19301 over 4 years ago.

[rnk]

What do you think of using __INTPTR_TYPE__ for this instead? If that doesn't work, we could define and undefine our own macro to use as a typedef. The resulting code should be clean enough that we can generalize it to __readcr0 etc.

It looks like long is the correct return type for 32-bit:
https://gcc.godbolt.org/z/6NzZWz
I'm not sure what the best device to figure out __writecr3 would be.

[craig.topper]

In D70101#1741452, @rnk wrote:

What do you think of using __INTPTR_TYPE__ for this instead? If that doesn't work, we could define and undefine our own macro to use as a typedef. The resulting code should be clean enough that we can generalize it to __readcr0 etc.

It looks like long is the correct return type for 32-bit:
https://gcc.godbolt.org/z/6NzZWz
I'm not sure what the best device to figure out __writecr3 would be.

I cobbled together some code from stackoverflow. I think the input type for __writecr3 really is int https://gcc.godbolt.org/z/dc8hQq

[craig.topper]

So in order to match the long and int behavior, I need two different macros to replace the argument? I think INTPTR_TYPE will return int on 32-bit targets so I can't use that for the long case. But maybe for the __int64?

[rnk]

In D70101#1741565, @craig.topper wrote:

So in order to match the long and int behavior, I need two different macros to replace the argument? I think INTPTR_TYPE will return int on 32-bit targets so I can't use that for the long case. But maybe for the __int64?

Right, we can use __INTPTR_TYPE__ for all __writecr* functions (they use int or long long), and our own macro (__LPTRINT_TYPE__?) for __readcr*. Make sure to undef it so it doesn't leak out of the header, similar to __DEFAULT_FN_ATTRS.

[craig.topper]

Add LPTRINT_TYPE and use that to qualify the readcr3. Use INTPTR_TYPE for writecr3

[rnk]

Lgtm

clang/lib/Headers/intrin.h

 void __outbyte(unsigned short, unsigned char);
 void __outbytestring(unsigned short, unsigned char *, unsigned long);
 void __outdword(unsigned short, unsigned long);
 void __outdwordstring(unsigned short, unsigned long *, unsigned long);
 void __outword(unsigned short, unsigned short);
 void __outwordstring(unsigned short, unsigned short *, unsigned long);
 unsigned long __readcr0(void);
 unsigned long __readcr2(void);
+#ifdef __x86_64__
+static __inline__
+unsigned __int64 __readcr3(void);
+#else
 static __inline__
 unsigned long __readcr3(void);
+#endif
 unsigned long __readcr4(void);
 unsigned long __readcr8(void);
 unsigned int __readdr(unsigned int);
 #ifdef __i386__
 static __inline__
 unsigned char __readfsbyte(unsigned long);
 static __inline__
 unsigned __int64 __readfsqword(unsigned long);

 void __svm_vmrun(size_t);
 void __svm_vmsave(size_t);
 void __ud2(void);
 unsigned __int64 __ull_rshift(unsigned __int64, int);
 void __vmx_off(void);
 void __vmx_vmptrst(unsigned __int64 *);
 void __wbinvd(void);
 void __writecr0(unsigned int);
+#ifdef __x86_64__
+static __inline__
+void __writecr3(unsigned __int64);
+#else
 static __inline__
 void __writecr3(unsigned int);
+#endif
 void __writecr4(unsigned int);
 void __writecr8(unsigned int);
 void __writedr(unsigned int, unsigned int);
 void __writefsbyte(unsigned long, unsigned char);
 void __writefsdword(unsigned long, unsigned long);
 void __writefsqword(unsigned long, unsigned __int64);
 void __writefsword(unsigned long, unsigned short);
 void __writemsr(unsigned long, unsigned __int64);
__readmsr(unsigned long __register) {
   // low-order 32 bits. If less than 64 bits are implemented in the MSR being
   // read, the values returned to EDX:EAX in unimplemented bit locations are
   // undefined.
   unsigned long __edx;
   unsigned long __eax;
   __asm__ ("rdmsr" : "=d"(__edx), "=a"(__eax) : "c"(__register));
   return (((unsigned __int64)__edx) << 32) | (unsigned __int64)__eax;
 }
+#endif
+
+#ifdef __x86_64__
+static __inline__ unsigned __int64 __DEFAULT_FN_ATTRS
+__readcr3(void) {
+  unsigned __int64 __cr3_val;
+  __asm__ __volatile__ ("mov %%cr3, %0" : "=r"(__cr3_val) : : "memory");
+  return __cr3_val;
+}
 
+static __inline__ void __DEFAULT_FN_ATTRS
+__writecr3(unsigned __int64 __cr3_val) {
+  __asm__ ("mov %0, %%cr3" : : "r"(__cr3_val) : "memory");
+}
+#else
 static __inline__ unsigned long __DEFAULT_FN_ATTRS
 __readcr3(void) {
   unsigned long __cr3_val;
-  __asm__ __volatile__ ("mov %%cr3, %0" : "=q"(__cr3_val) : : "memory");
+  __asm__ __volatile__ ("mov %%cr3, %0" : "=r"(__cr3_val) : : "memory");
   return __cr3_val;
 }
 
 static __inline__ void __DEFAULT_FN_ATTRS
 __writecr3(unsigned int __cr3_val) {
-  __asm__ ("mov %0, %%cr3" : : "q"(__cr3_val) : "memory");
+  __asm__ ("mov %0, %%cr3" : : "r"(__cr3_val) : "memory");
 }
 #endif
 
 #ifdef __cplusplus
 }
 #endif
 
 #undef __DEFAULT_FN_ATTRS
 
 #endif /* __INTRIN_H */
 #endif /* _MSC_VER */

clang/test/Headers/ms-intrin.cpp

#endif
   __cpuidex(info, 0, 0);
 #if defined(_M_X64) || defined(_M_IX86)
   _xgetbv(0);
 #endif
   __halt();
   __nop();
   __readmsr(0);
 
-  // FIXME: Call these in 64-bit too once the intrinsics have been fixed to
-  // work there, PR19301
-#ifndef _M_X64
   __readcr3();
   __writecr3(0);
-#endif
 
 #ifdef _M_ARM
   __dmb(_ARM_BARRIER_ISHST);
 #endif
 
 #ifdef _M_ARM64
   __dmb(_ARM64_BARRIER_SY);
 #endif
 }