This patch adds am654_sdhci_set_control_reg to am654_sdhci.
This is required to fix UHS_MODE_SELECT for TI K3 boards.
If any of HIGH_SPEED_ENA, V1P8_SIGNAL_ENA, UHS_MODE_SELECT
are set, then data will be launched on the pos-edge of the
clock.
Since K3 SoCs did not meet timing requirements for High Speed
SDR mode at rising clock edge, none of these three should be
set, therefore limit UHS_MODE_SELECT to only be set for modes
above MMC_HS_52.
This fixes MMC write issue on am64x evm at mode High Speed
SDR.
Signed-off-by: Judith Mendez <jm@ti.com>
Reviewed-by: Bryan Brattlof <bb@ti.com>
Jerome Forissier <jerome.forissier@linaro.org> says:
This series introduces threads and uses them to improve the performance
of the USB bus scanning code and to implement background jobs in the
shell via two new commands: 'spawn' and 'wait'.
The threading framework is called 'uthread' and is inspired from the
barebox threads [2]. setjmp() and longjmp() are used to save and
restore contexts, as well as a non-standard extension called initjmp().
This new function is added in several patches, one for each
architecture that supports HAVE_SETJMP. A new symbol is defined:
HAVE_INITJMP. Two tests, one for initjmp() and one for the uthread
scheduling, are added to the lib suite.
After introducing threads and making schedule() and udelay() a thread
re-scheduling point, the USB stack initialization is modified to benefit
from concurrency when UTHREAD is enabled, where uthreads are used in
usb_init() to initialize and scan multiple busses at the same time.
The code was tested on arm64 and arm QEMU with 4 simulated XHCI buses
and some devices. On this platform the USB scan takes 2.2 s instead of
5.6 s. Tested on i.MX93 EVK with two USB hubs, one ethernet adapter and
one webcam on each, "usb start" takes 2.4 s instead of 4.6 s.
Finally, the spawn and wait commands are introduced, allowing the use of
threads from the shell. Tested on the i.MX93 EVK with a spinning HDD
connected to USB1 and the network connected to ENET1. The USB plus DHCP
init sequence "spawn usb start; spawn dhcp; wait" takes 4.5 seconds
instead of 8 seconds for "usb start; dhcp".
[1] https://patchwork.ozlabs.org/project/uboot/list/?series=446674
[2] https://github.com/barebox/barebox/blob/master/common/bthread.c
Link: https://lore.kernel.org/r/20250418141114.2056981-1-jerome.forissier@linaro.org
Make the schedule() call from the CYCLIC framework a uthread scheduling
point too. This makes sense since schedule() is called from a lot of
places where uthread_schedule() needs to be called.
Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Reviewed-by: Stefan Roese <sr@denx.de>
Add struct uthread_mutex and uthread_mutex_lock(),
uthread_mutex_trylock(), uthread_mutex_unlock() to protect shared data
structures from concurrent modifications.
Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Add a new internal API called uthread (Kconfig symbol: UTHREAD) which
provides cooperative multi-tasking. The goal is to be able to improve
the performance of some parts of U-Boot by overlapping lengthy
operations, and also implement background jobs in the U-Boot shell.
Each uthread has its own stack allocated on the heap. The default stack
size is defined by the UTHREAD_STACK_SIZE symbol and is used when
uthread_create() receives zero for the stack_sz argument.
The implementation is based on context-switching via initjmp()/setjmp()/
longjmp() and is inspired from barebox threads [1]. A notion of thread
group helps with dependencies, such as when a thread needs to block
until a number of other threads have returned.
The name "uthread" comes from "user-space threads" because the
scheduling happens with no help from a higher privileged mode, contrary
to more complex models where kernel threads are defined. But the 'u'
may as well stand for 'U-Boot' since the bootloader may actually be
running at any privilege level and the notion of user vs. kernel may
not make much sense in this context.
[1] https://github.com/barebox/barebox/blob/master/common/bthread.c
Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Add the HAVE_INIJMP symbol to be set by architectures that support
initjmp(), a non-standard extension to setjmp()/longjmp() allowing to
initialize a jump buffer with a function pointer and a stack pointer.
This will be useful to later introduce threads. With this new function
it becomes possible to longjmp() to a particular function pointer
(rather than to a point previously reached during program execution as
is the case with setjmp()), and with a custom stack. Both things are
needed to spin off a new thread. Then the usual setjmp()/longjmp() pair
is enough to save and restore a context, i.e., switch thread.
Add the initjmp() prototype to <include/setjmp.h> since it is common to
all architectures.
Add an entry to the API documentation: doc/api/setjmp.rst.
Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
- Introducing back send_init_stream for omap_hsmmc
to perform the 74 clocks cycle sequence
- Move scmi regulator subnode hack to scmi_regulator
- Typo fix
The ROC-RK3576-PC is a SBC made by Firefly, designed around the RK3576
SoC. This adds the needed board infrastructure and config for it.
Signed-off-by: Heiko Stuebner <heiko@sntech.de>
Reviewed-by: Jonas Karlman <jonas@kwiboo.se>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
The Rockchip RK3576 is a ARM-based SoC with quad-core Cortex-A72
and quad-core Cortex-A53 including 6TOPS NPU, Mali-G52 MC3, HDMI Out,
DP, eDP, MIPI DSI, MIPI CSI2, LPDDR4/4X/5, eMMC5.1, SD3.0/MMC4.5, UFS,
USB OTG 3.0, Type-C, USB 2.0, PCIe 2.1, SATA 3, Ethernet, SDIO3.0, I2C,
UART, SPI, GPIO and PWM.
Add arch core support for it.
Signed-off-by: Xuhui Lin <xuhui.lin@rock-chips.com>
[adapted for mainline u-boot]
Signed-off-by: Heiko Stuebner <heiko@sntech.de>
Reviewed-by: Jonas Karlman <jonas@kwiboo.se>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
Rockchip RK3528 is a ARM-based SoC with quad-core Cortex-A53.
Add initial arch support for the RK3528 SoC.
Signed-off-by: Jonas Karlman <jonas@kwiboo.se>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
Ensure rk3588_common.h can be used by boards directly by defining a
blank ROCKCHIP_DEVICE_SETTINGS unless it already is defined.
Add a default SYS_CONFIG_NAME to include rk3588_common.h unless a board
target overrides it in its board Kconfig.
Signed-off-by: Jonas Karlman <jonas@kwiboo.se>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
Ensure rk3568_common.h can be used by boards directly by defining a
blank ROCKCHIP_DEVICE_SETTINGS unless it already is defined.
Add a default SYS_CONFIG_NAME to include rk3568_common.h unless a board
target overrides it in its board Kconfig.
Signed-off-by: Jonas Karlman <jonas@kwiboo.se>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
Ensure ROCKCHIP_DEVICE_SETTINGS is defined before including
rk3588_common.h in board include/configs files.
Signed-off-by: Jonas Karlman <jonas@kwiboo.se>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
Ensure ROCKCHIP_DEVICE_SETTINGS is defined before including
rk3568_common.h in board include/configs files.
Signed-off-by: Jonas Karlman <jonas@kwiboo.se>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
The partitions env variable is using an outdated partition layout that
is typically expected to be used with older vendor miniloader blobs.
Rockchip devices will run fine using any partition layout if the first
16 MiB of MMC storage is ignored/skipped.
Remove the partitions env variable to stop encourage users a continued
use of this outdated partition layout on RK3588 devices.
Signed-off-by: Jonas Karlman <jonas@kwiboo.se>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
The partitions env variable is using an outdated partition layout that
is typically expected to be used with older vendor miniloader blobs.
Rockchip devices will run fine using any partition layout if the first
16 MiB of MMC storage is ignored/skipped.
Remove the partitions env variable to stop encourage users a continued
use of this outdated partition layout on RK356x devices.
Signed-off-by: Jonas Karlman <jonas@kwiboo.se>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
In order to be able to use Device Tree Overlays, the fdtoverlay_addr_r
needs to be specified.
Follow what's been done for other Rockchip SoCs and leave 1MiB for the
base DTB before the address for the overlay.
Signed-off-by: Quentin Schulz <quentin.schulz@cherry.de>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
Make the do_ping() function in cmd/net.c a global one by getting rid of
the static qualifier, and move the prototype declaration from net-lwip.h
to net-common.h. This makes the function available to other parts of
U-Boot when CONFIG_NET=y, as was already the case when
CONFIG_NET_LWIP=y.
This is a peparation step to make the sandbox tests use a common API
between NET and NET_LWIP.
Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reviewed-by: Simon Glass <sjg@chromium.org>
Implement net_start_again() when NET_LWIP=y in a very similar way to
NET. This will be used in a future commit to determine if a failed
ping needs to be tried again on a different interface.
Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reviewed-by: Simon Glass <sjg@chromium.org>
The things that are done prior to executing a network command with
NET_LWIP are not consistent with what is done with NET. It impacts the
selection of the current device, and more precisely if the active device
is invalid NET would return an error while NET_LWIP would try to pick a
new device. This incorrect behavior was detected thanks to the eth_rotate
sandbox test (dm_test_eth_rotate()).
Fix it by re-using a sequence similar to what NET has in net_loop().
This piece of code is inserted in a function called net_lwip_eth_start()
renamed from net_lwip_eth_set_current().
Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reviewed-by: Simon Glass <sjg@chromium.org>
In IPv6, the broadcast MAC address is not used. Instead, it should use
the multicast address (see RFC RFC2464).
Add IPV6_ALL_NODE_ETH_ADDR macro for clarity.
Signed-off-by: Sean Edmond <seanedmond@microsoft.com>
Reviewed-by: Jerome Forissier <jerome.forissier@linaro.org>
f_acm calls usb_ep_disable(f_acm->ep_notify) unconditionally in
acm_start_ctrl(), even if the USB endpoint was never enabled before. This
causes crashes for some UDC drivers (e.g. ci_udc), because they dereference
data structures that are assigned only after having called usb_ep_enable().
The f_acm driver in U-Boot is similar to the Linux driver, where this issue
does not occur because usb_ep_disable() and usb_ep_enable() internally
track the enabled state. In Linux this change was made in commit
b0bac2581c19 ("usb: gadget: introduce 'enabled' flag in struct usb_ep") by
Robert Baldyga.
Fix the crashes for f_acm by making the same change in U-Boot. This makes
the API less bug-prone and avoids introducing crashes when adapting new
gadget drivers from Linux.
Signed-off-by: Stephan Gerhold <stephan.gerhold@linaro.org>
Reviewed-by: Mattijs Korpershoek <mkorpershoek@kernel.org>
Link: https://lore.kernel.org/r/20250407-acm-fixes-v1-3-e3dcb592d6d6@linaro.org
Signed-off-by: Mattijs Korpershoek <mkorpershoek@kernel.org>
Add a new callback function *send_init_stream* which start a sequence of
at least 74 clock cycles.
The mmc core uses *mmc_send_init_stream* in order to invoke the callback
function. This will be used during power cycle where the specification
requires such a sequence after power up.
Signed-off-by: Jean-Jacques Hiblot <jjhiblot@ti.com>
Reviewed-by: Peng Fan <peng.fan@nxp.com>
Signed-off-by: Peng Fan <peng.fan@nxp.com>
Jiaxun Yang <jiaxun.yang@flygoat.com> says:
This is a huge series which promoted MIPS/Boston target into a
usable state, with fixes to drivers and general framework issues
I found in this process.
I also converted the target to OF_UPSTREAM.
This target is covered by QEMU, to test on QEMU:
```
make boston64r6el_defconfig
make
qemu-system-mips64el -M boston -cpu I6500 -bios ./u-boot.bin -nographic
```
Link: https://lore.kernel.org/r/20240517-boston-v3-0-1ea7d23f4a1d@flygoat.com
Ensure that we are using correct physical/virtual address for
DMA buffer write and hardware register settings.
The convention is: in ahci_ioports all pointers are virtual,
that will be converted to physical address when writing to
hardware registers or into sg/cmd_tbl.
Also fixed 64bit physical address support for dwc_ahsata, ensure
higher bits are written into registers/sg properly.
Use memalign for allocating aligned buffer in dwc_ahsata so we
don't have to do our own alignment in driver.
Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
This pull request contains updates for the SoCFPGA platform, targeting
the 2025.07 release cycle. Highlights include enhancements to Agilex5
support, improvements in DDR error handling, and bridge reset handling
for SoC64 devices.
Key updates:
Agilex5 platform enhancements:
* New MMU region mappings and memory layout updates using
LMB_ARCH_MEM_MAP.
* Fixes for bloblist configuration, kernel FIT image generation, and
VAB flow enablement.
* GPIO pin control added for SDIO selection.
* Marvell PHY driver enabled in defconfig.
Agilex5 / SoC64 DDR subsystem:
* Added ECC debug improvements for IOSSM.
* Introduced LPDDR inline ECC support.
* Resolved size calculation overflow in memory driver.
SoC64 improvements:
* Enhanced mailbox communication with the SDM to reflect various
boot stage transitions.
* Implemented F2S bridge reset support and updated related reset
manager registers.
* Expanded SoC64 CPU info reporting.
General maintenance:
* Additional peripherals released from reset for Arria10.
* Cleanup of legacy or incorrect Kconfig implications.
This patch set has been tested on Agilex 5 devkit.
Passing all pipeline tests at:
https://source.denx.de/u-boot/custodians/u-boot-socfpga/-/pipelines/25867
Add reset manager registers in preparation for F2S bridge reset
support as well as the mask support to enable/disable the bridges.
Mask value:
BIT0: soc2fpga
BIT1: lwhps2fpga
BIT2: fpga2soc
These bridges are available only in Stratix10:
BIT3: f2sdram0
BIT4: f2sdram1
BIT5: f2sdram2
Signed-off-by: Alif Zakuan Yuslaimi <alif.zakuan.yuslaimi@altera.com>
Reviewed-by: Tien Fong Chee <tien.fong.chee@altera.com>
Implement exfat_fs_rename() to rename or move files. This is used
by the 'mv' generic FS interface command. The rename implementation
for other filesystems was added recently and was not part of exfat
porting layer due to merge issue, which made 'mv' command crash,
fix this by adding the missing implementation.
Fixes: b86a651b646c ("fs: exfat: Add U-Boot porting layer")
Signed-off-by: Marek Vasut <marex@denx.de>
A common device attached to SPI are SPI NAND and some device might
require to have info on the attached NAND to know the flash page size
and spare size.
To support this, introduce setup_for_spinand() that pass the attached
spinand info from manufacturer.
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
[trini: Switch to forward declaration of struct spinand_info]
Signed-off-by: Tom Rini <trini@konsulko.com>
Drop unneeded spi.h header include from spinand.h, nothing included by
spi.h is actually used in this header and .c should correctly included
spi.h if actually needed.
Replace spi.h with linux/bitops.h as this is what is actually required
for spinand.h
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
Port Linux kernel regmap_set/clear_bits shorthands to set and clear bits
in a regmap. These are handy if only specific bits needs to be applied
or cleared and makes it easier to port regmap based driver from kernel
upstream.
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
This commit adds support for the brcp1, brsmarc2, brcp150 and brcp170
boards. This boards are based on the Xilinx Zynq SoC.
Signed-off-by: Bernhard Messerklinger <bernhard.messerklinger@br-automation.com>
Acked-by: Michal Simek <michal.simek@amd.com>
Link: https://lore.kernel.org/r/20250404072819.69642-5-bernhard.messerklinger@br-automation.com
Signed-off-by: Michal Simek <michal.simek@amd.com>
Add support for loading the secure & non-secure pdi images and PL
bitstream on the Versal Gen2 platform. The FPGA driver is enabled
to load the bitstream in PDI format on the AMD Versal Gen2 device.
PDI is the new programmable device image format for Versal Gen2,
and the bitstream for the Versal Gen2 platform is generated exclusively
in this format.
With the enhanced SMC format in TF-A ensuring transparent payload
forwarding for Versal Gen2, the u-boot driver must now handle the
word swapping of PDI address that was previously done in TF-A for
this API. The source code for the Versal2 loadpdi command and the
CONFIG_CMD_VERSAL2 configuration has been removed. It now utilizes
the fpga load <dev> <address> <length> command to load secure &
non-secure pdi images.
Signed-off-by: Prasad Kummari <prasad.kummari@amd.com>
Link: https://lore.kernel.org/r/20250327105200.1262615-3-prasad.kummari@amd.com
Signed-off-by: Michal Simek <michal.simek@amd.com>
Add support for loading the secure & non-secure pdi images and
PL bitstream on the Versal NET platform. The FPGA driver is enabled
to load the bitstream in PDI format on the AMD Versal NET device.
PDI is the new programmable device image format for Versal NET,
and the bitstream for the Versal NET platform is generated exclusively
in this format.
The source code for the versalnet loadpdi command and the
CONFIG_CMD_VERSAL_NET configuration has been removed. It now utilizes
the fpga load <dev> <address> <length> command to load secure &
non-secure pdi images.
Signed-off-by: Prasad Kummari <prasad.kummari@amd.com>
Link: https://lore.kernel.org/r/20250327105200.1262615-2-prasad.kummari@amd.com
Signed-off-by: Michal Simek <michal.simek@amd.com>
Update the firmware driver UFS APIs zynqmp_pm_ufs_* to directly
read/write to the pmc_iou_slcr and efuse_cache registers. Replace
these raw reads/writes with the xilinx_pm_request() API with the
correct arguments once the PM related changes are done.
Signed-off-by: Venkatesh Yadav Abbarapu <venkatesh.abbarapu@amd.com>
Signed-off-by: Michal Simek <michal.simek@amd.com>
Link: https://lore.kernel.org/r/ee2d1ad2e07e96f1948ab6ffe8f3c50a3b8f9be9.1742462001.git.michal.simek@amd.com
Adjust configuration and devicetree so the topic-miami board actually
boots.
Replace the custom scripting and just use distro_boot. Override the
standard zynq routines.
The board attempts to boot from SD card first, and falls back to booting
UBIFS from the QSPI NOR flash.
Signed-off-by: Mike Looijmans <mike.looijmans@topic.nl>
Link: https://lore.kernel.org/r/20250312153741.24007-2-mike.looijmans@topic.nl
Signed-off-by: Michal Simek <michal.simek@amd.com>
Allow config headers that include zynq-common.h to provide their own
(distro) boot strategies. This is implemented by skipping the section
when BOOT_ENV has already been defined.
Signed-off-by: Mike Looijmans <mike.looijmans@topic.nl>
Link: https://lore.kernel.org/r/20250312153741.24007-1-mike.looijmans@topic.nl
Signed-off-by: Michal Simek <michal.simek@amd.com>
Added extended support for retrieving the PMC muti boot mode
register via the firmware interface, which is preferred when
U-Boot runs in EL2 and cannot directly access PMC registers
via raw reads. Ideally, all secure registers should be accessed
via xilinx_pm_request(). Introduced the secure
zynqmp_pm_get_pmc_multi_boot_reg() call, which uses
xilinx_pm_request() to read the PMC multi boot mode register.
BootROM increments the MultiBoot register (PMC_MULTI_BOOT) read
address offset by 32 KB and retries. For SD and eMMC boot modes,
it can search up to 8191 FAT files for the identification string.
A 13-bit mask (0x1FFF) is applied to PMC_MULTI_BOOT_MASK to obtain
the correct values in BootROM.
Signed-off-by: Prasad Kummari <prasad.kummari@amd.com>
Link: https://lore.kernel.org/r/20250305134845.3182193-1-prasad.kummari@amd.com
Signed-off-by: Michal Simek <michal.simek@amd.com>
Added extended support for retrieving the boot mode register
via the firmware interface, which is preferred when U-Boot
runs in EL2 and cannot directly access CRP registers via raw
reads. Ideally, all secure registers should be accessed via
xilinx_pm_request(). Introduced the secure zynqmp_pm_get_bootmode_reg()
call, which uses xilinx_pm_request() to read the boot mode register.
When CONFIG_ZYNQMP_FIRMWARE is enabled, the secure
zynqmp_pm_get_bootmode_reg() call is used; otherwise,
direct raw reads are performed in the case of mini U-Boot.
Signed-off-by: Prasad Kummari <prasad.kummari@amd.com>
Link: https://lore.kernel.org/r/20250219115301.3661036-1-prasad.kummari@amd.com
Signed-off-by: Michal Simek <michal.simek@amd.com>
On the AM3359 ICE we have two modes of operation CPSW mode or PRU-ICSS
mode.
For PRU-ICSS mode, connect Pin2 and Pin3 of J18 and J19 and for CPSW mode,
connect Pin1 and Pin2 of J18 and J19.
This patch adds support for PRUSS mode boot strapping from uboot.
Co-developed-by: Basharath Hussain Khaja <basharath@couthit.com>
Signed-off-by: Basharath Hussain Khaja <basharath@couthit.com>
Signed-off-by: Parvathi Pudi <parvathi@couthit.com>
When FIT image with multiple dtbs are involved for R5 boot process,
R5 SPL starts off with the first instance of dtb to probe the
eeprom, then once we have identified the type of board, invocation
of setup_multi_dtb_fit will replace the gd->fdt_blob with the proper
board dtb match. However, when we do this, two things happen:
a) Prior to the invocation of setup_multi_dtb_fit, as part of the eeprom
discovery process, i2c controller device is already probed and marked
as exclusive with the match of the very first tisci match (from the
original boot dtb). This list is stored in the info->dev_list of the
first probe.
b) When the second dtb is loaded, tisci is probed again (since this is a
new node) and the new info->dev_list is empty.
At this stage, the exclusive devices such as i2c instances used to
probe the board information is left in the old info->dev_list that is
no longer used actively by the system using the replaced dtb.
As a result of this, the cleanup we intend to do with
ti_sci_cmd_release_exclusive_devices is no longer complete and
leaves the instances such as i2c for eeprom marked used as we scan just
the new info->dev_list.
This creates a problem when Device Manager(DM) firmware starts up later
on in the boot process and identifies that this instance of i2c is
already marked active, so it assumes this can no longer be controlled
by software and is marked internally as reserved and HLOS can no
longer control these instances. This defeated the purpose of
ti_sci_cmd_release_exclusive_devices.
NOTE: This scheme works just fine if the FIT has just a single dtb as
the info->dev_list is upto date.
To fix this, let us make ti_sci_cmd_release_exclusive_devices scan the
all registrations of tisci instances and cleanup all exclusive devices
that have ever been registered.
As part of this, change the prototype of release_exclusive_devices to
drop the handle since that has no further meaning now.
Though this issue was identified on AM64-sk, this can be present in
other builds which use multi-fit-dtb for R5 SPL startup.
Fixes: 9566b777ae0a ("firmware: ti_sci: Add a command for releasing all exclusive devices")
Signed-off-by: Nishanth Menon <nm@ti.com>
Reviewed-by: Dhruva Gole <d-gole@ti.com>
Reviewed-by: Neha Malcom Francis <n-francis@ti.com>
Jerome Forissier <jerome.forissier@linaro.org> says:
This series replaces the dynamic initcalls (with function pointers) with
static calls, and gets rid of initcall_run_list(), init_sequence_f,
init_sequence_f_r and init_sequence_r. This makes the code simpler and the
binary slighlty smaller: -2281 bytes/-0.21 % with LTO enabled and -510
bytes/-0.05 % with LTO disabled (xilinx_zynqmp_kria_defconfig).
Execution time doesn't seem to change noticeably. There is no impact on
the SPL.
The inline assembly fixes, although they look unrelated, are triggered
on some platforms with LTO enabled. For example: kirkwood_defconfig.
CI: https://source.denx.de/u-boot/custodians/u-boot-net/-/pipelines/25514
Link: https://lore.kernel.org/r/20250404135038.2134570-1-jerome.forissier@linaro.org
