We mentioned how it wasn't really possible for a compiler to maintain such a DAG for an entire C program. One language feature of C that would be extremely difficult to account for is computed pointers; for example, the fact that function A's 0xDEADBEEF must reference the same memory location as function B's 0xDEADBEEF, even if there is there is essentially no symbolic relationship between A and B. What are other large obstacles to automatic generation of such DAGs for full C programs?
This comment was marked helpful 0 times.
jedavis
I may be wrong here, but I'll give it a stab -
One problem with generating a DAG for a full C program might come up if the C program were a JIT compiler for another language. JITs (I think?) often exhibit the behavior of taking bytecode input, writing machine code out to a page in their address space, and then executing the machine code that they've just written. It seems like it would be impossible to generate a full DAG at compile-time for a program where part of the code depends on unknown user input and will not be generated until runtime.
This comment was marked helpful 0 times.
aakashr
This example is showing that the computing xx, yy and z*z are all independent instructions(ILP = 3), and can be computed in parallel using superscalar execution.
After this, you add any two of them, while the third one remains idle. There is no way to do more than this step at the same time, so the ILP = 1. And finally, we add the third one to the sum of the other two.
We mentioned how it wasn't really possible for a compiler to maintain such a DAG for an entire C program. One language feature of C that would be extremely difficult to account for is computed pointers; for example, the fact that function A's 0xDEADBEEF must reference the same memory location as function B's 0xDEADBEEF, even if there is there is essentially no symbolic relationship between A and B. What are other large obstacles to automatic generation of such DAGs for full C programs?
This comment was marked helpful 0 times.
I may be wrong here, but I'll give it a stab - One problem with generating a DAG for a full C program might come up if the C program were a JIT compiler for another language. JITs (I think?) often exhibit the behavior of taking bytecode input, writing machine code out to a page in their address space, and then executing the machine code that they've just written. It seems like it would be impossible to generate a full DAG at compile-time for a program where part of the code depends on unknown user input and will not be generated until runtime.
This comment was marked helpful 0 times.
This example is showing that the computing xx, yy and z*z are all independent instructions(ILP = 3), and can be computed in parallel using superscalar execution. After this, you add any two of them, while the third one remains idle. There is no way to do more than this step at the same time, so the ILP = 1. And finally, we add the third one to the sum of the other two.
This comment was marked helpful 0 times.