On 11/07/2013 1:47 PM, Ed Jaffe wrote:
On 7/10/2013 11:54 AM, John Eells wrote:
And, finally, that the branch instruction itself does not stand
alone. One must load a register to use it as a base in order to
establish addressability, and load another to use it as a
displacement register, and Load instructions can cause real memory
(vs. cache) accesses. So, it seems to me there can be some
applicability to relative branch as well.
Branch prediction is a wonderful thing for optimizing loops and other
repetitive branch paths, but it makes it difficult to get meaningful
branch benchmarks. Relative branch, in addition to providing relief
from the oppressive 4K base register domain, should help a program
perform better because:
Great info. And it's easy to enable relative branching in your existing
code base simply by using the IEABRCX DEFINE macro.
1. Branch prediction occurs early in the pipeline when there is no
access to register contents. With relative branch there is no need for
the processor to "waste" time double-checking the branch target
address once the contents of the base register are actually known.
2. AGI (address generation interlock) can cause an instruction using a
base register to run *brutally* slowly (i.e., stop for a while) if it
executes shortly after the base register is loaded. This applies to
based branches as well as other types of instructions. Base registers
get loaded more often than people think e.g., LM when returning from a
subroutine. With relative branch there is ZERO possibility of AGI.
3. Freeing up code base registers for other uses allows program code
to be better interleaved by compilers, HLASM programmers that know
what they're doing, and OOO execution pipelines on the latest machines.
----------------------------------------------------------------------
For IBM-MAIN subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO IBM-MAIN
