GitLab

This patch moves the assembly exclusive lock library code
`spinlock.S` into architecture specific folder `aarch64`.
A stub file which includes the file from new location is
retained at the original location for compatibility. The BL
makefiles are also modified to include the file from the new
location.

Change-Id: Ide0b601b79c439e390c3a017d93220a66be73543

This patch fixes the translation table library for wraparound cases. These
cases are not expected to occur on AArch64 platforms because only the
48 bits of the 64 bit address space are used. But it is a possibility for
AArch32 platforms.

Change-Id: Ie7735f7ba2977019381e1c124800381471381499

A PSCI CPU_SUSPEND request to place a CPU in retention states at power levels
higher than the CPU power level is subject to the same state coordination as a
power down state. A CPU could implement multiple retention states at a
particular power level. When exiting WFI, the non-CPU power levels may be in a
different retention state to what was initially requested, therefore each CPU
should refresh its view of the states of all power levels.

Previously, a CPU re-used the state of the power levels when it entered the
retention state. This patch fixes this issue by ensuring that a CPU upon exit
from retention reads the state of each power level afresh.

Change-Id: I93b5f5065c63400c6fd2598dbaafac385748f989

This patch adds a runtime check that psci_find_target_suspend_lvl()
returns a valid value back to psci_cpu_suspend() and psci_get_stat().
If it is invalid, BL31 will now panic.

Note that on the PSCI CPU suspend path there is already a debug
assertion checking the validity of the target composite power state,
which effectively also checks the validity of the target suspend level.
Therefore, the error condition would already be caught in debug builds,
but in a release build this assertion would be compiled out.

On the PSCI stat path, there is currently no debug assertion checking
the validity of the power state before using it as an index into
the power domain state array.

Although BL31 platforms ports are responsible for validating the
power state parameter, the security impact (i.e. an out-of-bounds
array access) of a potential platform port bug in this code would
be quite high, given that this parameter comes from an untrusted
source. The cost of checking this in runtime generic code is low.

Change-Id: Icea85b8020e39928ac03ec0cd49805b5857b3906

This patch introduces the PSCI Library interface. The major changes
introduced are as follows:

* Earlier BL31 was responsible for Architectural initialization during cold
boot via bl31_arch_setup() whereas PSCI was responsible for the same during
warm boot. This functionality is now consolidated by the PSCI library
and it does Architectural initialization via psci_arch_setup() during both
cold and warm boots.

* Earlier the warm boot entry point was always `psci_entrypoint()`. This was
not flexible enough as a library interface. Now PSCI expects the runtime
firmware to provide the entry point via `psci_setup()`. A new function
`bl31_warm_entrypoint` is introduced in BL31 and the previous
`psci_entrypoint()` is deprecated.

* The `smc_helpers.h` is reorganized to separate the SMC Calling Convention
defines from the Trusted Firmware SMC helpers. The former is now in a new
header file `smcc.h` and the SMC helpers are moved to Architecture specific
header.

* The CPU context is used by PSCI for context initialization and
restoration after power down (PSCI Context). It is also used by BL31 for SMC
handling and context management during Normal-Secure world switch (SMC
Context). The `psci_smc_handler()` interface is redefined to not use SMC
helper macros thus enabling to decouple the PSCI context from EL3 runtime
firmware SMC context. This enables PSCI to be integrated with other runtime
firmware using a different SMC context.

NOTE: With this patch the architectural setup done in `bl31_arch_setup()`
is done as part of `psci_setup()` and hence `bl31_platform_setup()` will be
invoked prior to architectural setup. It is highly unlikely that the platform
setup will depend on architectural setup and cause any failure. Please be
be aware of this change in sequence.

Change-Id: I7f497a08d33be234bbb822c28146250cb20dab73

This patch moves the PSCI services and BL31 frameworks like context
management and per-cpu data into new library components `PSCI` and
`el3_runtime` respectively. This enables PSCI to be built independently from
BL31. A new `psci_lib.mk` makefile is introduced which adds the relevant
PSCI library sources and gets included by `bl31.mk`. Other changes which
are done as part of this patch are:

* The runtime services framework is now moved to the `common/` folder to
  enable reuse.
* The `asm_macros.S` and `assert_macros.S` helpers are moved to architecture
  specific folder.
* The `plat_psci_common.c` is moved from the `plat/common/aarch64/` folder
  to `plat/common` folder. The original file location now has a stub which
  just includes the file from new location to maintain platform compatibility.

Most of the changes wouldn't affect platform builds as they just involve
changes to the generic bl1.mk and bl31.mk makefiles.

NOTE: THE `plat_psci_common.c` FILE HAS MOVED LOCATION AND THE STUB FILE AT
THE ORIGINAL LOCATION IS NOW DEPRECATED. PLATFORMS SHOULD MODIFY THEIR
MAKEFILES TO INCLUDE THE FILE FROM THE NEW LOCATION.

Change-Id: I6bd87d5b59424995c6a65ef8076d4fda91ad5e86

This patch reworks type usage in generic code, drivers and ARM platform files
to make it more portable. The major changes done with respect to
type usage are as listed below:

* Use uintptr_t for storing address instead of uint64_t or unsigned long.
* Review usage of unsigned long as it can no longer be assumed to be 64 bit.
* Use u_register_t for register values whose width varies depending on
  whether AArch64 or AArch32.
* Use generic C types where-ever possible.

In addition to the above changes, this patch also modifies format specifiers
in print invocations so that they are AArch64/AArch32 agnostic. Only files
related to upcoming feature development have been reworked.

Change-Id: I9f8c78347c5a52ba7027ff389791f1dad63ee5f8

This patch adds cpumerrsr_el1 and l2merrsr_el1 to the register dump on
error for applicable CPUs.

These registers hold the ECC errors on L1 and L2 caches.

This patch updates the A53, A57, A72, A73 (l2merrsr_el1 only) CPU libraries.
Signed-off-by: Naga Sureshkumar Relli <nagasure@xilinx.com>

This patch introduces a new header file: include/lib/utils.h.
Its purpose is to provide generic macros and helper functions that
are independent of any BL image, architecture, platform and even
not specific to Trusted Firmware.

For now, it contains only 2 macros: ARRAY_SIZE() and
IS_POWER_OF_TWO(). These were previously defined in bl_common.h and
xlat_tables.c respectively.

bl_common.h includes utils.h to retain compatibility for platforms
that relied on bl_common.h for the ARRAY_SIZE() macro. Upstream
platform ports that use this macro have been updated to include
utils.h.

Change-Id: I960450f54134f25d1710bfbdc4184f12c049a9a9

This patch introduces the MT_EXECUTE/MT_EXECUTE_NEVER memory mapping
attributes in the translation table library to specify the
access permissions for instruction execution of a memory region.
These new attributes should be used only for normal, read-only
memory regions. For other types of memory, the translation table
library still enforces the following rules, regardless of the
MT_EXECUTE/MT_EXECUTE_NEVER attribute:

 - Device memory is always marked as execute-never.
 - Read-write normal memory is always marked as execute-never.

Change-Id: I8bd27800a8c1d8ac1559910caf4a4840cf25b8b0

This patch clarifies the mmap_desc() function by adding some comments
and reorganising its code. No functional change has been introduced.

Change-Id: I873493be17b4e60a89c1dc087dd908b425065401

This patch adds Performance Measurement Framework(PMF) in the
ARM Trusted Firmware. PMF is implemented as a library and the
SMC interface is provided through ARM SiP service.

The PMF provides capturing, storing, dumping and retrieving the
time-stamps, by enabling the development of services by different
providers, that can be easily integrated into ARM Trusted Firmware.
The PMF capture and retrieval APIs can also do appropriate cache
maintenance operations to the timestamp memory when the caller
indicates so.

`pmf_main.c` consists of core functions that implement service
registration, initialization, storing, dumping and retrieving
the time-stamp.
`pmf_smc.c` consists SMC handling for registered PMF services.
`pmf.h` consists of the macros that can be used by the PMF service
providers to register service and declare time-stamp functions.
`pmf_helpers.h` consists of internal macros that are used by `pmf.h`

By default this feature is disabled in the ARM trusted firmware.
To enable it set the boolean flag `ENABLE_PMF` to 1.

NOTE: The caller is responsible for specifying the appropriate cache
maintenance flags and for acquiring/releasing appropriate locks
before/after capturing/retrieving the time-stamps.

Change-Id: Ib45219ac07c2a81b9726ef6bd9c190cc55e81854

* Move libfdt API headers to include/lib/libfdt
* Add libfdt.mk helper makefile
* Remove unused libfdt files
* Minor changes to fdt.h and libfdt.h to make them C99 compliant
Co-Authored-By: Jens Wiklander <jens.wiklander@linaro.org>

Change-Id: I425842c2b111dcd5fb6908cc698064de4f77220e

Imports libfdt code from https://git.kernel.org/cgit/utils/dtc/dtc.git
tag "v1.4.1" commit 302fca9f4c283e1994cf0a5a9ce1cf43ca15e6d2.

Change-Id: Ia0d966058beee55a9047e80d8a05bbe4f71d8446

* Move stdlib header files from include/stdlib to include/lib/stdlib for
  consistency with other library headers.
* Fix checkpatch paths to continue excluding stdlib files.
* Create stdlib.mk to define the stdlib source files and include directories.
* Include stdlib.mk from the top level Makefile.
* Update stdlib header path in the fip_create Makefile.
* Update porting-guide.md with the new paths.

Change-Id: Ia92c2dc572e9efb54a783e306b5ceb2ce24d27fa

This patch adds ARM Cortex-A73 MPCore Processor support
in the CPU specific operations framework. It also includes
this support for the Base FVP port.

Change-Id: I0e26b594f2ec1d28eb815db9810c682e3885716d

This patch fixes the computation of the bitmask used to isolate
the level 1 field of a virtual address. The whole computation needs
to work on 64-bit values to produce the correct bitmask value.
XLAT_TABLE_ENTRIES_MASK being a C constant, it is a 32-bit value
so it needs to be extended to a 64-bit value before it takes part
in any other computation.

This patch fixes this bug by casting XLAT_TABLE_ENTRIES_MASK as
an unsigned long long.

Note that this bug doesn't manifest itself in practice because
address spaces larger than 39 bits are not yet supported in the
Trusted Firmware.

Change-Id: I955fd263ecb691ca94b29b9c9f576008ce1d87ee

Change-Id: I86ac81ffd7cd094ce68c4cceb01c16563671a063

Change-Id: Icaacd19c4cef9c10d02adcc2f84a4d7c97d4bcfa

Change-Id: Ia2ce8aa752efb090cfc734c1895c8f2539e82439

Change-Id: I632a8c5bb517ff89c69268e865be33101059be7d

Change-Id: I45641551474f4c58c638aff8c42c0ab9a8ec78b4

The CPU errata build flags don't enable errata, they enable errata
workarounds.

Change-Id: Ica65689d1205fc54eee9081a73442144b973400f

The only case in which regions can now overlap is if they are
identity mapped or they have the same virtual to physical address
offset (identity mapping is just a particular case of the latter).
They must overlap completely (i.e. one of them must be completely
inside the other one) and not cover the same area.

This allow future enhancements to the xlat_tables library without
having to support unnecessarily complex edge cases.

Outer regions are now sorted by mmap_add_region() before inner
regions with the same base virtual address for consistency: all
regions contained inside another one must be placed after the outer
one in the list.

If an inner region has the same attributes as the outer ones it will
be merged when creating the tables with init_xlation_table(). This
cannot be done as regions are added because there may be cases where
adding a region makes previously mergeable regions no longer
mergeable.

If the attributes of an inner region are different than the outer
region, new pages will be generated regardless of how "restrictive"
they are. For example, RO memory is more restrictive than RW. The
old implementation would give priority to RO if there is an overlap,
the new one doesn't.

NOTE: THIS IS THEORETICALLY A COMPATABILITY BREAK FOR PLATFORMS THAT
USE THE XLAT_TABLES LIBRARY IN AN UNEXPECTED WAY. PLEASE RAISE A
TF-ISSUE IF YOUR PLATFORM IS AFFECTED.

Change-Id: I75fba5cf6db627c2ead70da3feb3cc648c4fe2af

The AArch32 long descriptor format and the AArch64 descriptor format
correspond to each other which allows possible sharing of xlat_tables
library code between AArch64 and AArch32. This patch refactors the
xlat_tables library code to seperate the common functionality from
architecture specific code. Prior to this patch, all of the xlat_tables
library code were in `lib/aarch64/xlat_tables.c` file. The refactored code
is now in `lib/xlat_tables/` directory. The AArch64 specific programming
for xlat_tables is in `lib/xlat_tables/aarch64/xlat_tables.c` and the rest
of the code common to AArch64 and AArch32 is in
`lib/xlat_tables/xlat_tables_common.c`. Also the data types used in
xlat_tables library APIs are reworked to make it compatible between AArch64
and AArch32.

The `lib/aarch64/xlat_tables.c` file now includes the new xlat_tables
library files to retain compatibility for existing platform ports.
The macros related to xlat_tables library are also moved from
`include/lib/aarch64/arch.h` to the header `include/lib/xlat_tables.h`.

NOTE: THE `lib/aarch64/xlat_tables.c` FILE IS DEPRECATED AND PLATFORM PORTS
ARE EXPECTED TO INCLUDE THE NEW XLAT_TABLES LIBRARY FILES IN THEIR MAKEFILES.

Change-Id: I3d17217d24aaf3a05a4685d642a31d4d56255a0f

lib/aarch64/xlat_helpers.c defines helper functions to build
translation descriptors, but no common code or upstream platform
port uses them. As the rest of the xlat_tables code evolves, there
may be conflicts with these helpers, therefore this code should be
removed.

Change-Id: I9f5be99720f929264818af33db8dada785368711

The assembler helper function `print_revision_warning` is used when a
CPU specific operation is enabled in the debug build (e.g. an errata
workaround) but doesn't apply to the executing CPU's revision/part number.
However, in some cases the system integrator may want a single binary to
support multiple platforms with different IP versions, only some of which
contain a specific erratum.  In this case, the warning can be emitted very
frequently when CPUs are being powered on/off.

This patch modifies this warning print behaviour so that it is emitted only
when LOG_LEVEL >= LOG_LEVEL_VERBOSE. The `debug.h` header file now contains
guard macros so that it can be included in assembly code.

Change-Id: Ic6e7a07f128dcdb8498a5bfdae920a8feeea1345

The current translation table code maps in a series of regions, zeroing
the unmapped table entries before and in between the mapped regions. It
doesn't, however, zero the unmapped entries after the last mapped
region, leaving those entries at whatever value that memory has
initially.

This is bad because those values can look like valid translation table
entries, pointing to valid physical addresses. The CPU is allowed to do
speculative reads from any such addresses. If the addresses point to
device memory, the results can be unpredictable.

This patch zeroes the translation table entries following the last
mapped region, ensuring all table entries are either valid or zero
(invalid).

In addition, it limits the value of ADDR_SPACE_SIZE to those allowed by
the architecture and supported by the current code (see D4.2.5 in the
Architecture Reference Manual). This simplifies this patch a lot and
ensures existing code doesn't do unexpected things.

Change-Id: Ic28b6c3f89d73ef58fa80319a9466bb2c7131c21

At the moment, the memory translation library allows to create memory
mappings of 2 types:

 - Device nGnRE memory (named MT_DEVICE in the library);

 - Normal, Inner Write-back non-transient, Outer Write-back
   non-transient memory (named MT_MEMORY in the library).

As a consequence, the library code treats the memory type field as a
boolean: everything that is not device memory is normal memory and
vice-versa.

In reality, the ARMv8 architecture allows up to 8 types of memory to
be used at a single time for a given exception level. This patch
reworks the memory attributes such that the memory type is now defined
as an integer ranging from 0 to 7 instead of a boolean. This makes it
possible to extend the list of memory types supported by the memory
translation library.

The priority system dictating memory attributes for overlapping
memory regions has been extended to cope with these changes but the
algorithm at its core has been preserved. When a memory region is
re-mapped with different memory attributes, the memory translation
library examines the former attributes and updates them only if
the new attributes create a more restrictive mapping. This behaviour
is unchanged, only the manipulation of the value has been modified
to cope with the new format.

This patch also introduces a new type of memory mapping in the memory
translation library: MT_NON_CACHEABLE, meaning Normal, Inner
Non-cacheable, Outer Non-cacheable memory. This can be useful to map
a non-cacheable memory region, such as a DMA buffer for example.

The rules around the Execute-Never (XN) bit in a translation table
for an MT_NON_CACHEABLE memory mapping have been aligned on the rules
used for MT_MEMORY mappings:
 - If the memory is read-only then it is also executable (XN = 0);
 - If the memory is read-write then it is not executable (XN = 1).

The shareability field for MT_NON_CACHEABLE mappings is always set as
'Outer-Shareable'. Note that this is not strictly needed since
shareability is only relevant if the memory is a Normal Cacheable
memory type, but this is to align with the existing device memory
mappings setup. All Device and Normal Non-cacheable memory regions
are always treated as Outer Shareable, regardless of the translation
table shareability attributes.

This patch also removes the 'ATTR_SO' and 'ATTR_SO_INDEX' #defines.
They were introduced to map memory as Device nGnRnE (formerly called
"Strongly-Ordered" memory in the ARMv7 architecture) but were not
used anywhere in the code base. Removing them avoids any confusion
about the memory types supported by the library.

Upstream platforms do not currently use the MT_NON_CACHEABLE memory
type.

NOTE: THIS CHANGE IS SOURCE COMPATIBLE BUT PLATFORMS THAT RELY ON THE
BINARY VALUES OF `mmap_attr_t` or the `attr` argument of
`mmap_add_region()` MAY BE BROKEN.

Change-Id: I717d6ed79b4c845a04e34132432f98b93d661d79

All C files of stdlib were included into std.c, which was the file
that the Makefile actually compiled. This is a poor way of compiling
all the files and, while it may work fine most times, it's
discouraged.

In this particular case, each C file included its own headers, which
were later included into std.c. For example, this caused problems
because a duplicated typedef of u_short in both subr_prf.c and
types.h. While that may require an issue on its own, this kind of
problems are avoided if all C files are as independent as possible.

Change-Id: I9a7833fd2933003f19a5d7db921ed8542ea2d04a

In the Cortex-A35/A53/A57 CPUs library code, some of the CPU specific
reset operations are skipped if they have already been applied in a
previous invocation of the reset handler. This precaution is not
required, as all these operations can be reapplied safely.

This patch removes the unneeded test-before-set instructions in
the reset handler for these CPUs.

Change-Id: Ib175952c814dc51f1b5125f76ed6c06a22b95167

The LDNP/STNP instructions as implemented on Cortex-A53 and
Cortex-A57 do not behave in a way most programmers expect, and will
most probably result in a significant speed degradation to any code
that employs them. The ARMv8-A architecture (see Document ARM DDI
0487A.h, section D3.4.3) allows cores to ignore the non-temporal hint
and treat LDNP/STNP as LDP/STP instead.

This patch introduces 2 new build flags:
A53_DISABLE_NON_TEMPORAL_HINT and A57_DISABLE_NON_TEMPORAL_HINT
to enforce this behaviour on Cortex-A53 and Cortex-A57. They are
enabled by default.

The string printed in debug builds when a specific CPU errata
workaround is compiled in but skipped at runtime has been
generalised, so that it can be reused for the non-temporal hint use
case as well.

Change-Id: I3e354f4797fd5d3959872a678e160322b13867a1

The debug prints used to debug translation table setup in xlat_tables.c
used the `printf()` standard library function instead of the stack
optimized `tf_printf()` API. DEBUG_XLAT_TABLE option was used to enable
debug logs within xlat_tables.c and it configured a much larger stack
size for the platform in case it was enabled. This patch modifies these
debug prints within xlat_tables.c to use tf_printf() and modifies the format
specifiers to be compatible with tf_printf(). The debug prints are now enabled
if the VERBOSE prints are enabled in Trusted Firmware via LOG_LEVEL build
option.

The much larger stack size definition when DEBUG_XLAT_TABLE is defined
is no longer required and the platform ports are modified to remove this
stack size definition.

Change-Id: I2f7d77ea12a04b827fa15e2adc3125b1175e4c23

Migrate all direct usage of __attribute__ to usage of their
corresponding macros from cdefs.h.
e.g.:
 - __attribute__((unused)) -> __unused
Signed-off-by: Soren Brinkmann <soren.brinkmann@xilinx.com>

This patch adds support for ARM Cortex-A35 processor in the CPU
specific framework, as described in the Cortex-A35 TRM (r0p0).

Change-Id: Ief930a0bdf6cd82f6cb1c3b106f591a71c883464

On the ARMv8 architecture, cache maintenance operations by set/way on the last
level of integrated cache do not affect the system cache. This means that such a
flush or clean operation could result in the data being pushed out to the system
cache rather than main memory. Another CPU could access this data before it
enables its data cache or MMU. Such accesses could be serviced from the main
memory instead of the system cache. If the data in the sysem cache has not yet
been flushed or evicted to main memory then there could be a loss of
coherency. The only mechanism to guarantee that the main memory will be updated
is to use cache maintenance operations to the PoC by MVA(See section D3.4.11
(System level caches) of ARMv8-A Reference Manual (Issue A.g/ARM DDI0487A.G).

This patch removes the reliance of Trusted Firmware on the flush by set/way
operation to ensure visibility of data in the main memory. Cache maintenance
operations by MVA are now used instead. The following are the broad category of
changes:

1. The RW areas of BL2/BL31/BL32 are invalidated by MVA before the C runtime is
   initialised. This ensures that any stale cache lines at any level of cache
   are removed.

2. Updates to global data in runtime firmware (BL31) by the primary CPU are made
   visible to secondary CPUs using a cache clean operation by MVA.

3. Cache maintenance by set/way operations are only used prior to power down.

NOTE: NON-UPSTREAM TRUSTED FIRMWARE CODE SHOULD MAKE EQUIVALENT CHANGES IN
ORDER TO FUNCTION CORRECTLY ON PLATFORMS WITH SUPPORT FOR SYSTEM CACHES.

Fixes ARM-software/tf-issues#205

Change-Id: I64f1b398de0432813a0e0881d70f8337681f6e9a

This patch unifies the bakery lock api's across coherent and normal
memory implementation of locks by using same data type `bakery_lock_t`
and similar arguments to functions.

A separate section `bakery_lock` has been created and used to allocate
memory for bakery locks using `DEFINE_BAKERY_LOCK`. When locks are
allocated in normal memory, each lock for a core has to spread
across multiple cache lines. By using the total size allocated in a
separate cache line for a single core at compile time, the memory for
other core locks is allocated at link time by multiplying the single
core locks size with (PLATFORM_CORE_COUNT - 1). The normal memory lock
algorithm now uses lock address instead of the `id` in the per_cpu_data.
For locks allocated in coherent memory, it moves locks from
tzfw_coherent_memory to bakery_lock section.

The bakery locks are allocated as part of bss or in coherent memory
depending on usage of coherent memory. Both these regions are
initialised to zero as part of run_time_init before locks are used.
Hence, bakery_lock_init() is made an empty function as the lock memory
is already initialised to zero.

The above design lead to the removal of psci bakery locks from
non_cpu_power_pd_node to psci_locks.

NOTE: THE BAKERY LOCK API WHEN USE_COHERENT_MEM IS NOT SET HAS CHANGED.
THIS IS A BREAKING CHANGE FOR ALL PLATFORM PORTS THAT ALLOCATE BAKERY
LOCKS IN NORMAL MEMORY.

Change-Id: Ic3751c0066b8032dcbf9d88f1d4dc73d15f61d8b

This patch migrates the rest of Trusted Firmware excluding Secure Payload and
the dispatchers to the new platform and context management API. The per-cpu
data framework APIs which took MPIDRs as their arguments are deleted and only
the ones which take core index as parameter are retained.

Change-Id: I839d05ad995df34d2163a1cfed6baa768a5a595d

- Apply a53 errata #826319 to revision <= r0p2
- Apply a53 errata #836870 to revision <= r0p3
- Update docs/cpu-specific-build-macros.md for newly added errata build flags

Change-Id: I44918e36b47dca1fa29695b68700ff9bf888865e
Signed-off-by: Jimmy Huang <jimmy.huang@mediatek.com>

Denver is NVIDIA's own custom-designed, 64-bit, dual-core CPU which is
fully ARMv8 architecture compatible.  Each of the two Denver cores
implements a 7-way superscalar microarchitecture (up to 7 concurrent
micro-ops can be executed per clock), and includes a 128KB 4-way L1
instruction cache, a 64KB 4-way L1 data cache, and a 2MB 16-way L2
cache, which services both cores.

Denver implements an innovative process called Dynamic Code Optimization,
which optimizes frequently used software routines at runtime into dense,
highly tuned microcode-equivalent routines. These are stored in a
dedicated, 128MB main-memory-based optimization cache. After being read
into the instruction cache, the optimized micro-ops are executed,
re-fetched and executed from the instruction cache as long as needed and
capacity allows.

Effectively, this reduces the need to re-optimize the software routines.
Instead of using hardware to extract the instruction-level parallelism
(ILP) inherent in the code, Denver extracts the ILP once via software
techniques, and then executes those routines repeatedly, thus amortizing
the cost of ILP extraction over the many execution instances.

Denver also features new low latency power-state transitions, in addition
to extensive power-gating and dynamic voltage and clock scaling based on
workloads.
Signed-off-by: Varun Wadekar <vwadekar@nvidia.com>