2.4.3 Instruction cache

Each CPU element contains a single L1 instruction cache that is on the code AHB interface of the Cortex®-M33 core.

The cache:

  • Reduces the code access fetches targeting the flash memory and therefore reduces activity on the flash and power.
  • Reduces the overall latency of code access. This permits increasing the latency of the instruction fetch from memory which might be necessary to deal with long timing paths.

The cache has the following features:

  • 2-way set associative.
  • 16-Byte cache lines.
  • Configurable size. The top-level parameters CPU0_ICACHESIZE and CPU1_ICACHESIZE set the size.
  • Configuration interface local to each processor.
  • Supports uncached bypass operation.

For guidance on cache size configuration, see the Arm® CoreLink™ SSE-200 Subsystem for Embedded Analysis Report.

Note:

  • The instruction code fetches in the region from 0x0000_0000 to 0x1FFF_FFFF are cached. All accesses to other memory regions are not cached.
  • Each cache includes a configuration interface that is only accessible to the processor that is connected to it and it resides at address 0x5001_0000. This configuration interface is not accessible to other masters in the system.

The cache supports Secure and Non-secure segregation of contents by retaining the security attributes of the cached contents. If the SAU or the IDAU configurations change, the cache might contain contents that do not match the new security settings. To maintain security, you must therefore disable and invalidate the cache before modifying the SAU or IDAU.

Because this is an instruction cache, only read accesses are subject to caching. All write accesses bypass the cache. If the INVMAT option of the cache is enabled, when a cacheable write access occurs to a line that exists in the cache, that cache line is invalidated. When INVMAT is not used, any (cacheable) code memory change must be re-aligned with cache-disable & invalidation.

Note:

  • If a cache line exists in the cache that matches a non-cacheable write access, the invalidation does not occur. Therefore, when changing the attribute of an address region from cacheable to non-cacheable that the instruction cache accesses, you must invalidate the cache.
  • Write access from the Debugger could invalidate the related cacheline, even when it is a non-cacheable attribute. Such access must go through the relevant (INVMAT enabled) active cache.

The following table lists the instruction cache configuration parameters.

Table 2-3 Instruction cache configurations

Parameter Processor 0 configuration Processor 1 configuration Processor 0 default value Processor 1 default value Description
CSIZE CPU0_ICACHESIZE CPU1_ICACHESIZE 2KB 2KB

Define the cache size. Supported cache sizes:

  • 9 = 512B.
  • 10 = 1KB.
  • 11 = 2KB.
  • 12 = 4KB.
  • 13 = 8KB.
  • 14 = 16KB.
  • Other = Reserved.
DMA CPU0_ICACHEDMA CPU1_ICACHEDMA 0 0 Defines the existence of micro DMA capability and line locking capability.

Supported value : 0.

INVMAT CPU0_ICACHEINVMAT CPU1_ICACHEINVMAT 0 0 Invalidate on Write Match. When set to 1, any cacheable writes to a line that also exists in the instruction cache results in the cache line being invalidated.
COFFSET 0 0 0 0 Cacheable Region Offset Upper bits to compare against. This parameter is fixed.
COFFSIZE 3 3 3 3 Cacheable Region Size. Value defines the number of upper address bits to compare. Therefore, the addressable space is 2(32-COFFSIZE).
STATS CPU0_ICACHESTATS CPU1_ICACHESTATS 1 1 1 = include statistics functionality.
XOM CPU0_XOM CPU1_XOM 0 0

Enable Execute Only Memory support.

When set to 1, the HRUSER[0] signal on the AHB5 Master Expansion Code Interface is used to indicate if current data being returned is Execute Only. If the data type access arriving at the instruction cache targets a XOM location, the instruction cache masks the data.

When set to 0, this HRUSER[0] is ignored by the associated instruction cache.

REDUCE_READS ICACHERRDS 1 1 Reduce instruction cache Tag Reads. When set to 1, it masks off an access to the TAG RAM if this set has been previously accessed and the RAM data is valid and does not need re-accessing.

This feature can be used for power saving. If used, RAM can hold last read information while the power state does not change.

Non-ConfidentialPDF file icon PDF version101104_0200_00_en
Copyright © 2016–2018 Arm Limited or its affiliates. All rights reserved.