This additional cost is likely one of the predominant reasons that ILP has plateaued at 4 simultaneous instructions. To circumvent this issue, architects have tried to force the compiler to solve the dependency issue using VLIW (very long instruction word). To summarize VLIW, if a processor contains 5 independent execution units, the compiler will have 5 operations in the "very long instruction word" that the processor will map to the 5 execution units: https://en.wikipedia.org/wiki/Very_long_instruction_word. This way dependency checking is the responsibility of software and not hardware.
I am not sure if VLIW has helped significantly pushed the four simultaneous instruction threshold though. If somebody knows, please share.
kayvonf
Question: The key phrase on this slide is that a processor must execute instructions in a manner "appears" as if they were executed in program order. This is a key idea in this class.
What is program order?
And what does it mean for the results of a program's execution to appear as if instructions were executed in program order?
And finally... Why is the program order guarantee a useful one? (What if the results of execution were inconsistent with the results that would be obtained if the instructions were executed in program order?)
void
And what does it mean for the results of a program's execution to appear as if instructions were executed in program order?
A programmer might write something like the code below.
x = a + b
print(x)
y = c + d
print(y)
The programmer expects the computer to print the value of x before the value of y, otherwise it would be incorrect. However, there may be performance gains from reordering the instructions so that the two additions happen at the same time. As long as the two print statements happen in the right order and print the right values, then the programmer might never know that the instructions were reordered.
It is computationally expensive for the processor to determine dependencies between instructions. The following PPT (slides 9/10) provides an example of how the number of checks grows with the number of instructions that are simultaneously dispatched: http://www.cs.cmu.edu/afs/cs/academic/class/15740-f15/www/lectures/11-superscalar-pipelining.pdf
This additional cost is likely one of the predominant reasons that ILP has plateaued at 4 simultaneous instructions. To circumvent this issue, architects have tried to force the compiler to solve the dependency issue using VLIW (very long instruction word). To summarize VLIW, if a processor contains 5 independent execution units, the compiler will have 5 operations in the "very long instruction word" that the processor will map to the 5 execution units: https://en.wikipedia.org/wiki/Very_long_instruction_word. This way dependency checking is the responsibility of software and not hardware.
I am not sure if VLIW has helped significantly pushed the four simultaneous instruction threshold though. If somebody knows, please share.
Question: The key phrase on this slide is that a processor must execute instructions in a manner "appears" as if they were executed in program order. This is a key idea in this class.
What is program order?
And what does it mean for the results of a program's execution to appear as if instructions were executed in program order?
And finally... Why is the program order guarantee a useful one? (What if the results of execution were inconsistent with the results that would be obtained if the instructions were executed in program order?)
A programmer might write something like the code below.
The programmer expects the computer to print the value of
xbefore the value ofy, otherwise it would be incorrect. However, there may be performance gains from reordering the instructions so that the two additions happen at the same time. As long as the two print statements happen in the right order and print the right values, then the programmer might never know that the instructions were reordered.