Renesas SH7751 Group, Руководство пользователя

Страница: 277 / 1186

SH7751 Group, SH7751R Group Section 8 Pipelining
R01UH0457EJ0301 Rev. 3.01 Page 223 of 1128
Sep 24, 2013
The instruction execution sequence is expressed as a combination of the execution patterns shown
in figure 8.2. One instruction is separated from the next by the number of machine cycles for its
issue rate. Normally, execution, data access, and write-back stages cannot be overlapped onto the
same stages of another instruction; the only exception is when two instructions are executed in
parallel under parallel-executability conditions. Refer to (a) through (d) in figure 8.3 for some
simple examples.
Latency is the interval between issue and completion of an instruction, and is also the interval
between the execution of two instructions with an interdependent relationship. When there is
interdependency between two instructions fetched simultaneously, the latter of the two is stalled
for the following number of cycles:
•
(Latency) cycles when there is flow dependency (read-after-write)
•
(Latency - 1) or (latency - 2) cycles when there is output dependency (write-after-write)
⎯
Single/double-precision FDIV, FSQRT is the preceding instruction (latency – 1) cycles
⎯
The other FE group except above is the preceding instruction (latency – 2) cycles
•
5 or 2 cycles when there is anti-flow dependency (write-after-read), as in the following cases:
⎯
FTRV is the preceding instruction (5 cycles)
⎯
A double-precision FADD, FSUB, or FMUL is the preceding instruction (2 cycles)
In the case of flow dependency, latency may be exceptionally increased or decreased, depending
on the combination of sequential instructions (figure 8.3 (e)).
•
When a floating-point computation is followed by a floating-point register store, the latency of
the floating-point computation may be decreased by 1 cycle.
•
If there is a load of the shift amount immediately before an SHAD/SHLD instruction, the
latency of the load is increased by 1 cycle.
•
If an instruction with a latency of less than 2 cycles, including write-back to a floating-point
register, is followed by a double-precision floating-point instruction, FIPR, or FTRV, the
latency of the first instruction is increased to 2 cycles.
The number of cycles in a pipeline stall due to flow dependency will vary depending on the
combination of interdependent instructions or the fetch timing (see figure 8.3. (e)).
Output dependency occurs when the destination operands are the same in a preceding FE group
instruction and a following LS group instruction.
For the stall cycles of an instruction with output dependency, the longest latency to the last write-
back among all the destination operands must be applied instead of “latency” (see figure 8.3 (f)).
A stall due to output dependency with respect to FPSCR, which reflects the result of a floating-