Fix a floatround bug (#376)

* Fix floatround overflow

floatround would overflow for floats greater than 2^30 because internally
it would double the number (therefore anything greater than 2^30 results
in something greater than 2^31 which would cause overflow of course)

floatround behaviour is left exactly identical otherwise (although I find
it very weird and wrong to deliberately avoid banker's rounding, it would
be a bad idea to change this behaviour due to compatibility)

Remember to reassemble amxexecn and amxjitsn

* Update the compiled object files

amxmodx/amxexecn.asm

@@ -1492,22 +1492,20 @@ OP_FLOAT_ROUND:
 		;set the bits
 		or      ah,dl	;set bits 15,14 of FCW to rounding method
 		or      ah,3	;set precision to 64bit
-		mov     [ebp], eax
+
-		fldcw   [ebp]
 		;calculate
 		sub     esp,4
 		fld     dword [edi+ecx+4]
 		test    edx,edx
-		jz      .correct
+		jnz     .skip_correct
-		jmp     .skip_correct
+		;nearest mode
-	.correct:
+		;correct so as to AVOID bankers rounding
-		fadd    st0
+		or      ah, 4   ;set rounding mode to floor
 		fadd    dword [g_round_nearest]
-		fistp   dword [esp]
+
-		pop     eax
-		sar     eax,1
-		jmp     .done
 	.skip_correct:
+		mov     [ebp], eax
+		fldcw   [ebp]
 		frndint
 		fistp   dword [esp]
 		pop     eax

amxmodx/amxjitsn.asm

@@ -1979,22 +1979,20 @@ OP_FLOAT_ROUND:
 		;set the bits
 		or      ah,dl	;set bits 15,14 of FCW to rounding method
 		or      ah,3	;set precision to 64bit
-		mov     [ebp], eax
+
-		fldcw   [ebp]
 		;calculate
 		sub     esp,4
 		fld     dword [esi+4]
 		test    edx,edx
-		jz      .correct
+		jnz     .skip_correct
-		jmp     .skip_correct
+		;nearest mode
-	.correct:
+		;correct so as to AVOID bankers rounding
-		fadd    st0
+		or      ah, 4   ;set rounding mode to floor
 		fadd    dword [g_round_nearest]
-		fistp   dword [esp]
+
-		pop     eax
-		sar     eax,1
-		jmp     .done
 	.skip_correct:
+		mov     [ebp], eax
+		fldcw   [ebp]
 		frndint
 		fistp   dword [esp]
 		pop     eax