Assembly Language for Beginners

3.16. TOUPPER() FUNCTION

Thetmpvariable must be signed.

This makes two subtraction operations in case of a transformation plus one comparison.

In contrast the original algorithm uses two comparison operations plus one subtracting.

Optimizing GCC is even better, it gets rid of the jumps (which is good:2.10.1 on page 466) by using the

CMOVcc instruction:

Listing 3.73: Optimizing GCC 4.9 (x64)

1 toupper:
2 lea edx, [rdi-97] ; 0x61
3 lea eax, [rdi-32] ; 0x20
4 cmp dl, 25
5 cmova eax, edi
6 ret

At line 3 the code prepares the subtracted value in advance, as if the conversion will always happen.

Atline5thesubtractedvalueinEAXisreplacedbytheuntouchedinputvalueifaconversionisnotneeded.

And then this value (of course incorrect) is dropped.

Advance subtracting is a price the compiler pays for the absence of conditional jumps.

3.16.2 ARM

Optimizing Keil for ARM mode also generates only one comparison:

Listing 3.74: Optimizing Keil 6/2013 (ARM mode)

toupper PROC

SUB r1,r0,#0x61

CMP r1,#0x19

SUBLS r0,r0,#0x20

ANDLS r0,r0,#0xff

BX lr

ENDP

The SUBLS and ANDLS instructions are executed only if the value inR1is less than 0x19 (or equal). They

also do the actual conversion.

Optimizing Keil for Thumb mode generates only one comparison operation as well:

Listing 3.75: Optimizing Keil 6/2013 (Thumb mode)

toupper PROC

MOVS r1,r0

SUBS r1,r1,#0x61

CMP r1,#0x19

BHI |L0.14|

SUBS r0,r0,#0x20

LSLS r0,r0,#24

LSRS r0,r0,#24

|L0.14|

BX lr

ENDP

The last two LSLS and LSRS instructions work likeAND reg, 0xFF: they are equivalent to the C/C++-

expression(i<<24)>> 24.

Seems like that Keil for Thumb mode deduced that two 2-byte instructions are shorter than the code that

loads the 0xFF constant into a register plus an AND instruction.

GCC for ARM64

Listing 3.76: Non-optimizing GCC 4.9 (ARM64)

toupper:

sub sp, sp, #16