Care to explain?
If it can be optimized the wrong way -- use a pragma or fix the compiler.
I hate these kinds of patches. I think Timo would agree.
Best regards,
Alex Ionescu
On Fri, Sep 30, 2011 at 5:03 PM,
<dgorbachev@svn.reactos.org> wrote:
Author: dgorbachev
Date: Fri Sep 30 21:03:29 2011
New Revision: 53907
URL: http://svn.reactos.org/svn/reactos?rev=53907&view=rev
Log:
[NTOSKRNL]
Use inline asm in KiIsNpxErrataPresent(). C code can be optimized in a wrong way.
Modified:
trunk/reactos/ntoskrnl/ke/i386/cpu.c
Modified: trunk/reactos/ntoskrnl/ke/i386/cpu.c
URL: http://svn.reactos.org/svn/reactos/trunk/reactos/ntoskrnl/ke/i386/cpu.c?rev=53907&r1=53906&r2=53907&view=diff
==============================================================================
--- trunk/reactos/ntoskrnl/ke/i386/cpu.c [iso-8859-1] (original)
+++ trunk/reactos/ntoskrnl/ke/i386/cpu.c [iso-8859-1] Fri Sep 30 21:03:29 2011
@@ -1254,9 +1254,9 @@
INIT_FUNCTION
KiIsNpxErrataPresent(VOID)
{
- BOOLEAN ErrataPresent;
+ static double Value1 = 4195835.0, Value2 = 3145727.0;
+ INT ErrataPresent;
ULONG Cr0;
- volatile double Value1, Value2;
/* Disable interrupts */
_disable();
@@ -1269,9 +1269,30 @@
Ke386FnInit();
/* Multiply the magic values and divide, we should get the result back */
- Value1 = 4195835.0;
- Value2 = 3145727.0;
- ErrataPresent = (Value1 * Value2 / 3145727.0) != 4195835.0;
+#ifdef __GNUC__
+ __asm__ __volatile__
+ (
+ "fldl %1\n\t"
+ "fdivl %2\n\t"
+ "fmull %2\n\t"
+ "fldl %1\n\t"
+ "fsubp\n\t"
+ "fistpl %0\n\t"
+ : "=m" (ErrataPresent)
+ : "m" (Value1),
+ "m" (Value2)
+ );
+#else
+ __asm
+ {
+ fld Value1
+ fdiv Value2
+ fmul Value2
+ fld Value1
+ fsubp
+ fistp ErrataPresent
+ };
+#endif
/* Restore CR0 */
__writecr0(Cr0);
@@ -1280,7 +1301,7 @@
_enable();
/* Return if there's an errata */
- return ErrataPresent;
+ return ErrataPresent != 0;
}
VOID