1.10. ACCESSING PASSED ARGUMENTS
The code saves its input arguments in the local stack, in case someone (or something) in this function
needs using theW0...W2registers. This prevents overwriting the original function arguments, which may
be needed again in the future.
This is calledRegister Save Area. ([Procedure Call Standard for the ARM 64-bit Architecture (AArch64),
(2013)]^85 ). The callee, however, is not obliged to save them. This is somewhat similar to “Shadow
Space”:1.10.2 on page 100.
WhydidtheoptimizingGCC4.9dropthisargumentsavingcode? Becauseitdidsomeadditionaloptimizing
workandconcludedthatthefunctionargumentswillnotbeneededinthefutureandalsothattheregisters
W0...W2will not be used.
We also see aMUL/ADDinstruction pair instead of single aMADD.
1.10.4 MIPS.
Listing 1.95: Optimizing GCC 4.4.5.text:00000000 f:
; $a0=a
; $a1=b
; $a2=c
.text:00000000 mult $a1, $a0
.text:00000004 mflo $v0
.text:00000008 jr $ra
.text:0000000C addu $v0, $a2, $v0 ; branch delay slot
; result is in $v0 upon return
.text:00000010 main:
.text:00000010
.text:00000010 var_10 = -0x10
.text:00000010 var_4 = -4
.text:00000010
.text:00000010 lui $gp, (gnu_local_gp >> 16)
.text:00000014 addiu $sp, -0x20
.text:00000018 la $gp, (gnu_local_gp & 0xFFFF)
.text:0000001C sw $ra, 0x20+var_4($sp)
.text:00000020 sw $gp, 0x20+var_10($sp)
; set c:
.text:00000024 li $a2, 3
; set a:
.text:00000028 li $a0, 1
.text:0000002C jal f
; set b:
.text:00000030 li $a1, 2 ; branch delay slot
; result in $v0 now
.text:00000034 lw $gp, 0x20+var_10($sp)
.text:00000038 lui $a0, ($LC0 >> 16)
.text:0000003C lw $t9, (printf & 0xFFFF)($gp)
.text:00000040 la $a0, ($LC0 & 0xFFFF)
.text:00000044 jalr $t9
; take result of f() function and pass it as a second argument to printf():
.text:00000048 move $a1, $v0 ; branch delay slot
.text:0000004C lw $ra, 0x20+var_4($sp)
.text:00000050 move $v0, $zero
.text:00000054 jr $ra
.text:00000058 addiu $sp, 0x20 ; branch delay slot
The first four function arguments are passed in four registers prefixed by A-.
TherearetwospecialregistersinMIPS:HIandLOwhicharefilledwiththe64-bitresultofthemultiplication
during the execution of theMULTinstruction.
These registers are accessible only by using theMFLOandMFHIinstructions.MFLOhere takes the low-part
of the multiplication result and stores it into $V0. So the high 32-bit part of the multiplication result is
dropped (the HI register content is not used). Indeed: we work with 32-bitintdata types here.
Finally,ADDU(“Add Unsigned”) adds the value of the third argument to the result.
(^85) Also available ashttp://go.yurichev.com/17287