Basic ILP Techniques

• Dependences:
• Dependences are a property of programs , not of CPUs and pipelines.

 

• A dependence between two instructions will always exist unless the program is changed.

 

• The presence of the dependence indicates the potential for a hazard.

 

• However, the actual hazard and the length of any stall is a property of the pipeline.

 

• There are three types of dependencies:
• Data
• Name
• Control

Basic ILP Techniques

• Data dependences
• Instruction j is dependent on i if i produces a result used by j .

 

• Dependence is transitive: if j is dependent upon k and k is dependent upon i , then j is dependent on i , i.e.,

• Dependence chains can be arbitrarily long.
• When a compiler schedules instructions, it cannot move j before i if i depends upon j .

Basic ILP Techniques

• Data dependences
• Indicate the possibility of a hazard.
• Determine the order in which results must be generated.
• Place an upper bound on the amount of ILP available.

 

• Data dependences can be overcome in two ways:
• Maintaining the dependence but avoiding a hazard .
• Eliminating the dependence by transforming the code.

 

• Scheduling is the primary way to avoid hazards without altering dependences.

 

• We did this in the previous example with LD, ADDD and SD.
• We scheduled the code to avoid the hazard, but the dependence remained in the code.

Basic ILP Techniques

• Data dependences
• We also eliminated dependencies:

• Compiler removed this dependence by computing the intermediate values of R1 and folding the computation into an offset.

Basic ILP Techniques

• Data dependences
• Since eliminating data dependencies requires a fair amount of analysis, it is done by the compiler.

 

• Avoiding hazards through scheduling can be done in both hardware and software.

 

• One tricky data dependence that many people overlook is that of memory locations.

 

• Registers are easy to figure out, but memory dependences are much harder since they may not be known until runtime.

 

• For example, 100(R4) and 20(R6) may refer to the same memory location.

 

• We will look at hardware and software techniques that detect data dependencies that involve memory locations.

Basic ILP Techniques

• Name Dependences
• Occur when two instructions use the same register or memory location but there is NO flow of data between instructions that use that name.

 

• There are two types. Assume instruction i precedes instruction j :
• Antidependence (WAR)
• This occurs when two instructions use the same register in sequence, with the first reading it and the second writing it.

 

• Output dependence (WAW)
• This occurs when two instructions write to the same register.

Basic ILP Techniques

• Name Dependences
• These are NOT data dependences because no information is passed between the two instructions.

 

• The instructions could be executed out of order or in parallel if the CPU renamed the register or memory location involved.

 

• Register renaming can either be done by the compiler or the CPU .

 

• In our earlier unrolling example, the compiler renamed registers when it unrolled the loop.

Basic ILP Techniques

• Name Dependences
• Consider the unrolled loop with loop overhead removed but no renaming.

• Note that this forces almost a complete ordering of the code.
• Only LD following SD can be interchanged.

Basic ILP Techniques

• Control Dependences
• A control dependency determines the ordering of the instructions with respect to branch instructions.

 

• If an instruction depends on the outcome of an earlier branch then it is only executed on one of the two forks.

 

• For example:

 

• Obviously, we cannot:
• Move S1 before the if stmt or
• Move other instructions before the if stmt into the "then" block.

Basic ILP Techniques

• Control Dependences
• Consider the unrolled loop with the branches still in place (but the conditions complemented.)

 

• Previously, since we knew that the # of iterations was a multiple of 4, we changed the control dependencies and eliminated branches that were never taken.

Basic ILP Techniques

• Control Dependences
• Note that preserving control dependence is NOT a critical property.

 

• Program correctness is the critical property that needs to be preserved.

 

• We may be willing to violate control dependence if program correctness is preserved.

 

• The two properties critical to program correctness are:
• Preserving exception behavior
• Any changes in the ordering of instructions must NOT change how exceptions are raised.

 

• An instruction that should not have been executed, should not be allowed to cause an exception.

 

• Memory operations and floating point often cause problems like this.

Basic ILP Techniques

• Control Dependences
• The second property critical to program correctness is:
• Preserving data flow
• Branches make the flow of data between instructions dynamic :

• Preserving data dependence is not enough here (OR is dependent on both the ADD and SUB) !

 

• There may be other ways of preserving control dependence, i.e.,
• The CPU can cancel instructions that were wrongly executed.
• Or the compiler can cancel things out.

Basic ILP Techniques

• Control Dependences
• Sometimes violating control dependencies cannot affect exception behavior or data flow:

• We could move the SUB instruction before the BEQZ if we knew:
• The SUB instruction could not generate an exception.
• If R4 were not `live', i.e., used after the skipnext label.

 

• This type of scheduling is called speculation , i.e.,
• The compiler is betting that the branch will not be taken.

 

• More on this soon.

Basic ILP Techniques

• Control Dependences
• Control dependence is preserved by implementing control hazard detection .

 

• Control hazard detection causes control stalls .

 

• Control stalls can be avoided by:
• Scheduling instructions in delay slots.
• Loop unrolling.
• Conditional execution.
• Speculation (by both compiler and CPU).

 

• We will cover the latter two in the next two sections.