[LLVMdev] llvm microblaze port - severe data hazards
System Error Message
gudiguda at gmail.com
Sat Aug 11 22:33:53 PDT 2012
Hi, i'm working on a microblaze port extremely similar to microblaze that only
has data hazards. The advice given to look at PPC and ARM is poor because
neither of them check for data dependencies and are entirely complicated and
different. The port has a 5 stage pipeline where the result of instruction can
only be used at stage 3 and 5 only whereas load and store instructions require 4
cycles. That means that if there was a code such as
code 1 code 2
lwi r3, r0, 4 addik r3, r4, 5
cmp r4, r3, 5 swi r3, r0, 12
swi r3, r0, 16
would cause cmp and swi r3,r0,12 to use the old value of r3.
It should be
lwi r3, r0, 4 addik r3, r4, 5
nop nop
nop swi r3, r0, 12
nop nop
cmp r4, r3, 5 swi r3, r0, 16
Due to the pipeline it is correct to have
lwi, r3, r0, 4
some instruction not writing to r3
some instruction not writing to r3
swi, r3, r0, 8 (will store the old r3 value)
cmp r4, r3, 5 (will use the new r3 value)
I've already tried adapting PPC hazard but i cant make heads or tails of it and
even make it work.
I'd also like to know if hazard recognisers are used when filling in delay
slots. The port fills in up to 3 delay slots the same way as it would fill 1
delay slot microblaze (moves instruction or nop if cant find any
) in order to prevent hazards.
I'd appreciate any help on this.
Thank you.
More information about the llvm-dev
mailing list