ARM R4 Manuale Utente

ARM DDI 0363E

14-4

ID013010

ADD R3, R3, R1 LSL#6

;plus one because Register R1 is Early

The following sequence where R1 is a Late Reg takes two cycles:

LDR R1, [R2]

;Result latency two minus one cycles

STR R1, [R3]

;no penalty because R1 is a Late register

The following sequence where R1 is a Very Early Reg takes four cycles:

ADD R3, R1, R2

;Result latency one plus two cycles

LDR R4, [R3]

;plus two because register R3 is Very Early

Most instructions do not take more or fewer cycles to execute if they fail their condition codes. 
The exceptions to this are:
•

instructions that alter the PC, such as branches

•

integer divide instructions, which require only one execute cycle.

The result latency of most instructions that fail their condition codes is one. The exceptions to 
this are:
•

all load and store instructions, which have their result latency unaffected

•

integer divide instructions, which have a result latency of three.

Most instructions do not take more or fewer cycles to execute if they are flag-setting. The 
exceptions to this are certain multiply instructions.

Table 14-1 gives descriptions of cycle timing terms used in this chapter.

Memory Cycles

This is the number of cycles during which an instruction sends a memory access to the cache.

Cycles

This is the minimum number of cycles required to issue an instruction. Issue cycles that produce 
memory accesses to the cache are included, so Cycles is always greater than or equal to Memory 
Cycles.

Result Latency

This is the number of cycles before the result of this instruction is available to a Normal Reg of the 
following instruction. When the Result Latency of an instruction is greater than Cycles and the 
following instruction requires the result, the following instruction stalls for a number of cycles equal 
to Result Latency minus Cycles. 

 The Result Latency is counted from the first cycle of an instruction.

Normal Reg

The specified registers are required at the start of the Ex2 stage.

Late Reg

The specified registers are not required until the start of the Wr stage. Subtract one cycle from the 
Result Latency of the instruction producing this register.