Improving performance of a processor having a defective cache

Abstract

In one embodiment, a method for improving performance of a processor having a defective cache includes accessing first object code and generating second object code from the first object code. The generation of the second object code takes into account one or more locations of one or more defects in a cache on a processor such that one or more instructions in the second object code are written only to nondefective locations in the cache.

Description

TECHNICAL FIELD OF THE INVENTION

This invention relates in general to memory systems and more particularly to improving performance of a processor having a defective cache.

BACKGROUND OF THE INVENTION

A processor (such as an embedded microprocessor) typically includes one or more caches that occupy a large percentage of the area of the processor. For example, approximately half of the total chip area of a StrongARM SA-110 processor is devoted to two 16 KB caches. Cache arrays are particularly susceptible to faults, and the tolerance of cache faults significantly affects the yield of processors that include caches.

Overview

Particular embodiments of the present invention may reduce or eliminate problems and disadvantages associated with processors having defective caches.

In one embodiment, a method for improving performance of a processor having a defective cache includes accessing first object code and generating second object code from the first object code. The generation of the second object code takes into account one or more locations of one or more defects in a cache on a processor such that one or more instructions in the second object code are written only to nondefective locations in the cache.

Particular embodiments of the present invention may provide one or more technical advantages. As an example and not by way of limitation, particular embodiments improve performance, e.g., speed up operation, of processors having defective caches. In particular embodiments, defect-aware code placement reduces performance degradation of a processor having a cache that is only partially good. Particular embodiments modify the placement of basic blocks, or functions, in an address space of a cache to reduce or even minimize the occurrence of cache misses in a defective cache. Particular embodiments may provide all, some, or none of these technical advantages. Particular embodiments may provide one or more other technical advantages, one or more of which may be readily apparent to those skilled in the art from the figures, descriptions, and claims herein.

BRIEF DESCRIPTION OF THE DRAWINGS

To provide a more complete understanding of the present invention and features and advantages thereof, reference is made to the following description, taken in conjunction with the accompanying drawings, in which:

FIG. 1 illustrates example code placement;

FIG. 2 illustrates an example flow of the technique;

FIG. 3 illustrates an example cache-locking function; and

FIG. 4 illustrates an example self-conflict miss.

DESCRIPTION OF EXAMPLE EMBODIMENTS

FIG. 1 illustrates example code placement. Consider a direct-mapped cache of size C (=2^m words) having a cache line size of L words, i.e., L consecutive words are fetched from the main memory on a cache read miss. In a direct-mapped cache, the cache line containing a word located at memory address M may be calculated by

$\lfloor \frac{M}{L}\rfloor \mathrm{mod}\frac{C}{L}.$ Therefore, memory locations M_i and M_j are mapped to the same cache line if:

$\begin{array}{cc}\left(\lfloor \frac{M_i}{L}\rfloor -\lfloor \frac{M_j}{L}\rfloor \right)\mathrm{mod}\frac{C}{L}=0& \left(1\right)\end{array}$ As an example and not by way of limitation, assume a direct-mapped cache having four cache lines that are each thirty-two bytes. Functions A, B, C and D may be put into the main memory as shown on the left side of FIG. 1. If the processor accesses functions A, B, and D in a loop, conflict misses occur because A and D are mapped to the same cache line. Swapping the locations of C and D as shown on the right side of FIG. 1 resolves the cache conflict. In particular embodiments, code placement modifies the placement of basic blocks or functions in an address space to reduce the total number of cache conflict misses.

FIG. 2 illustrates an example method for code placement. In particular embodiments, the method detects the locations of defects in a cache and then generates object code that keeps the number of cache misses below a predetermined number. Recompilation need occur only if the original object code fails to provide the performance required of a particular chip. If the original object code provides the required performance, the method uses the original object code. If the original object code does not provide the required performance, but object code generated to address a previous defect provides the required performance, the method uses the object code generated to address the previous defect. When the chip is turned on, the chip executes an initialization step during which, based on information collected during test, faulty cache-lines are marked using lock and valid bits. The chip then executes the compiled code. Particular embodiments generate object code while taking into account the locations of defective cache lines in a cache on a processor so that instructions in the object code are, where possible, written only to nondefective cache lines in the cache, which tends to reduce the cache misses caused by the defects. Particular embodiments also lock defective cache lines using existing flag bits in the cache to prevent the processor from accessing the defective cache lines when executing the object code.

FIG. 3 illustrates an example cache-locking function. Particular embodiments exploit an unused combination of existing flag bits, e.g., valid bit=0 and lock bit=1, to indicate a fault in a specific cache line. As an example and not by way of limitation, in a four-way set-associative cache including lock and valid bits, if the lock bit of Way1 in the fifth cache set is one and its valid bit is zero, as shown in FIG. 3, the corresponding cache line will not be used for replacement in case of a cache miss. If the valid bit of Way1 in the fifth cache set is zero, accessing the corresponding block will always cause a cache miss. Thus, in particular embodiments, the cache-locking function facilitates the correctness of the processor operation, even in presence of defects in tag or data memory. Some processors do not allow setting the lock bit for an invalid cache line. As a result, in some processors, valid bit=0 and lock bit=1 is an invalid combination. However, in particular embodiments, minor modification to the cache-control circuitry enables this combination of flag bits. For a set-associative cache, the associativity of the cache set that includes the faulty cache lines is reduced by one. In a direct-mapped cache, accesses to the cache set that includes the faulty cache line will cause a cache miss. Particular embodiments modify the placement of functions in the address space to compensate for the increased number of cache misses due to faulty cache lines. In particular embodiments, use of a cache-lock function obviates hardware overhead otherwise needed to improve performance of a processor having one or more defective caches.

Particular embodiments generate an instruction trace corresponding to an execution of the application program as follows:(o₀,o₁,o₃,o₅,o₆,o₇,o₃,o₅,o₇,o₃,o₅,o₆,o₇,o₃)  (2)Above, o_i represents the i^th instruction of the original object code. Assume a_i represents the address of o_i. Each instruction o_i may be mapped to a memory block having the address

$\lfloor \frac{a_i}{L}\rfloor ,$ where L is cache-line size in bytes. From the instruction trace, a trace of the memory-block addresses accessed, TMB=(b₁, b₁, b₂, b₃, . . . ), may be generated.

X(b_i) may be defined as follows:

$X\left(b_i\right)=\{\begin{array}{cc}1& \mathrm{if}\mathrm{there}\mathrm{are}\mathrm{at}\mathrm{least}{W}_s\mathrm{accesses}\\ & \mathrm{to}s=\left(b_i\mathrm{mod}S\right)\mathrm{between}\mathrm{an}\mathrm{access}\mathrm{to}\\ & b_i\mathrm{and}\mathrm{its}\mathrm{next}\mathrm{access}\\ 0& \mathrm{otherwise}.\end{array}$ W_s is the number of nonfaulty cache lines in the s^th cache set. The total number of misses can be calculated as follows:

$\begin{array}{cc}{M}_{\mathrm{total}}=\sum _{\forall b_i\in \mathrm{TMB}}X\left(b_i\right)& \left(3\right)\end{array}$ The above formula takes into account the effect of faulty cache lines. Particular embodiments generate gaps in the object code, which in turn increases memory requirements associated with the object code, but do not increase the size of the object code.

Particular embodiments generate a shorter trace that results in the same number of misses, but may be processed faster. FIG. 4 illustrates an example self-conflict miss. In FIG. 4, the function is smaller than the size of the cache. The left side of FIG. 4 shows a direct-mapped cache with four 32-byte lines. If a function having a size of 104 bytes is placed in the main memory, cache conflict misses occur in the first cache line because the first and last words of the function are mapped to the same cache line. If the size of the function is 100 bytes, no self-conflict miss occurs, regardless of where in the address space the function is placed. Therefore, the following is a condition of a function that never experiences self conflicts:Function Size≦L×(S−1)+I  (4)L, S, and I represent cache-line size in bytes, the number of cache sets, and instruction-word size in bytes, respectively. If a function satisfies the above condition, two instructions, o_i and o_j, of the function will never evict each other, regardless of where in the address space the function is placed. As a result, when a loop is executed, if all instructions in the loop belong to the function and the loop is executed n times, n−1 instances may be deleted from the trace.

The following notations may be used in the problem formulation in particular embodiments:

• • S: Number of cache sets.
  • L: Cache line size in bytes.
  • W_i: Number of nonfaulty cache ways in the i^th cache set.
  • T_c: Compacted trace of memory-block addresses accessed.
  • F: List of functions in the target program sorted in descending order of their execution counts.
  • M_total: Total number of cache misses.
  • M_bound: Upper bound of number of cache misses. The user may provide this parameter.In particular embodiments, the problem may be formally defined as follows: for a given S, L, a set of W_is, T_c, F, and M_bound, find an order of functions for which M_total is less than M_bound. Particular embodiments use the following algorithm:

Procedure Defect_Aware_Code_Placement
Input: S, L, a set of Wis, Tc, F, and Mbound
Output: order of functions in the optimized object code
Mmin = infinity;
repeat
for (t=0; t<|F|; t++) do
p = F[t];  BESTlocation = p;
for each p′ ∈ F and p′≠ p do
Insert function p in the place of p′;
Update Tc according to the locations of functions;
Calculate Mtotal using (3);
if (Mtotal ≦ Mmin) then
Mmin = Mtotal;
BESTlocation = p′;
end if
end for
Put function p in the place of BESTlocation
end for
until (Mmin < Mbound or Mmin stops decreasing)
Output order of functions
end Procedure

The above algorithm starts from an original object code and finds the optimal location of each function of the application program in the address space. This may be done by changing the order of placing functions in the address space and finding the best ordering. For each ordering, the algorithm updates the trace of memory block addresses executed (T_c) according to the locations of functions and calculates the total number of cache misses (M_total) using (3). The ordering yielding the minimum number of cache misses is selected. The algorithm continues as long as the number of cache misses reduces and is no less than M_bound. The computation time of the algorithm is quadratic in terms of the number of functions in the application program.

Particular embodiments have been used to describe the present invention. A person having skill in the art may comprehend one or more changes, substitutions, variations, alterations, or modifications to the particular embodiments used to describe the present invention that are within the scope of the appended claims. The present invention encompasses all such changes, substitutions, variations, alterations, and modifications.

Claims

1. A method for improving performance of a processor having a defective cache, the method comprising: accessing first object code of an application program; andgenerating second object code of the application program from the first object code, the generation of the second object code taking into account one or more locations of one or more defects in a cache on a processor such that one or more instructions in the second object code are written only to nondefective locations in the cache to keep a total number of cache misses below a predefined threshold.

2. The method of claim 1, further comprising: generating an original trace of addresses of memory locations accessed during execution of the first object code; andusing the original trace, calculating a number of cache misses.

3. The method of claim 2, further comprising generating a shorter trace of addresses of memory locations accessed during execution of the first object code, the shorter trace being shorter and processable faster than the original trace and operable to calculate the number of cache misses calculable by the original traces.

4. The method of claim 1, further comprising: loading the second object code into the cache; andlocking one or more of the defective locations such that accesses to the cache avoid one or more of the defective locations.

5. The method of claim 4, wherein locking one or more of the defective locations comprises locking one or more of the defective locations according to a locking function that uses a combination of otherwise unused flag bits in the cache.

6. The method of claim 5, wherein the combination of otherwise unused flag bits comprises a valid bit and a lock bit.

7. The method of claim 5, further comprising modifying circuitry associated with control of the cache to accommodate the locking function.

8. The method of claim 1, wherein the defective locations in the cache are cache lines.

9. The method of claim 1, wherein the cache is a set-associative cache.

10. The method of claim 1, wherein the cache is a direct-mapped cache.

11. A system for improving performance of a processor having a defective cache, the system comprising: means for accessing first object code of an application program; andmeans for generating second object code of the application program from the first object code, the generation of the second object code taking into account one or more locations of one or more defects in a cache on a processor such that one or more instructions in the second object code are written only to nondefective locations in the cache to keep a total number of cache misses below a predefined threshold.

12. A system for improving performance of a processor having a defective cache, the system comprising: means for detecting one or more locations of one or more defects in a cache on a fabricated processor;means for generating object code from an application program, generation of the object code taking into account one or more of the locations of one or more of the defects in the cache on the fabricated processor if previously generated object code from the application program provides less than a predetermined level of performance in the fabricated processor, generation of the object code comprising reusing the previously generated object code if the previously generated object code provides at least the predetermined level of performance in the fabricated processor,means for marking one or more of the locations of one or more of the defects in the cache on the fabricated processor; andmeans for executing the object code.