diff --git a/libcpuid/asm-bits.c b/libcpuid/asm-bits.c
index ca936f2..5e451ea 100644
--- a/libcpuid/asm-bits.c
+++ b/libcpuid/asm-bits.c
@@ -75,13 +75,13 @@ int cpuid_exists_by_eflags(void)
 #endif /* PLATFORM_X86 */
 }
 
+#ifdef INLINE_ASM_SUPPORTED
 /* 
  * with MSVC/AMD64, the exec_cpuid() and cpu_rdtsc() functions
  * are implemented in separate .asm files. Otherwise, use inline assembly
  */
 void exec_cpuid(uint32_t *regs)
 {
-#ifdef INLINE_ASM_SUPPORTED
 #ifdef COMPILER_GCC
 #	ifdef PLATFORM_X64
 	__asm __volatile(
@@ -166,8 +166,8 @@ void exec_cpuid(uint32_t *regs)
 #    error "Unsupported compiler"
 #  endif /* COMPILER_MICROSOFT */
 #endif
-#endif /* INLINE_ASSEMBLY_SUPPORTED */
 }
+#endif /* INLINE_ASSEMBLY_SUPPORTED */
 
 #ifdef INLINE_ASM_SUPPORTED
 void cpu_rdtsc(uint64_t* result)
diff --git a/libcpuid/masm-x64.asm b/libcpuid/masm-x64.asm
index 69bc933..51e9865 100644
--- a/libcpuid/masm-x64.asm
+++ b/libcpuid/masm-x64.asm
@@ -39,4 +39,321 @@ cpu_rdtsc Proc
 	ret
 cpu_rdtsc endp
 
+; procedure busy_sse_loop
+; Signature: void busy_sse_loop(int cycles)
+busy_sse_loop Proc
+	; save xmm6 & xmm7 into the shadow area, as Visual C++ 2008
+	; expects that we don't touch them:
+	movups	[rsp + 8],	xmm6
+	movups	[rsp + 24],	xmm7
+
+	xorps	xmm0,	xmm0
+	xorps	xmm1,	xmm1
+	xorps	xmm2,	xmm2
+	xorps	xmm3,	xmm3
+	xorps	xmm4,	xmm4
+	xorps	xmm5,	xmm5
+	xorps	xmm6,	xmm6
+	xorps	xmm7,	xmm7
+	; --
+	align 16
+bsLoop:
+	; 0:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 1:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 2:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 3:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 4:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 5:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 6:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 7:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 8:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 9:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 10:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 11:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 12:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 13:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 14:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 15:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 16:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 17:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 18:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 19:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 20:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 21:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 22:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 23:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 24:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 25:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 26:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 27:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 28:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 29:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 30:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; 31:
+	addps	xmm0,	xmm1
+	addps	xmm1,	xmm2
+	addps	xmm2,	xmm3
+	addps	xmm3,	xmm4
+	addps	xmm4,	xmm5
+	addps	xmm5,	xmm6
+	addps	xmm6,	xmm7
+	addps	xmm7,	xmm0
+	; ----------------------
+	dec		ecx
+	jnz		bsLoop
+
+	; restore xmm6 & xmm7:
+	movups	xmm6,	[rsp + 8]
+	movups	xmm7,	[rsp + 24]
+	ret
+busy_sse_loop endp
+
 END