[PATCH] LD Fix for STMicroelectronics hardware erratum STM32L4XX 629360.

[Christophe MONAT <christophe dot monat at st dot com>]

Hi list,

This is a proposal for a ld patch that implements a work-around to a 
STMicroelectronics hardware erratum.

* Description of the problem:
The STM32L4x6xx family of microcontroller suffers from a memory
controller ('FMC') limitation, described in the publicly available
document:
<www.st.com/st-web-ui/static/active/en/resource/technical/document/errata_sheet/DM00111498.pdf>
See : "Read burst access of 9 words or more is not supported by
FMC" for the full the description.
In a few words, the multiple loads that are issued from the code
(ldm, vldm) through the FMC may receive corrupt data when more
than 9 words are involved. Access should be limited to chunks of 8
words.

* Impact of the problem:
The STM32L476 product is available to anybody worldwide either
directly through ST or through the distributors.  As there is no
plan for a revision of that silicon and it will live for at least
10 years, in the end tens of millions of parts or even more are
potentially affected.

* Possible solutions to the problem:
The compilation tools can be changed to limit their code
generation to appropriate multiple load sizes, but this solution
does not account for existing libraries and third-party software
libraries that cannot be recompiled.

An alternative solution is to create a specific treatment at
the linker level to identify and divert the faulty instruction
and divert it to a correct sequence. Thus it can deal with any
pre-existing .o or .a contribution that may expose the issue
and produce a fixed, final executable.

* Purpose of the patch:
The patch is disabled by default and has no observable
impact. When it is enabled, it checks if the expected architecture
is used and warns otherwise. It scans the code sections for faulty
ldm and vldm instructions and when one is detected it creates a
veneer that contains a fixed sequence and a branch back to the
following instruction, and replaces the original instruction to a
branch to that veneer.

The veneer contents are emitted based on the original
instruction details, depending on the instruction family (9
cases for ldm, 3 for vldm) and various other details (PC being
or not being part of the set of restored registers, write-back
of the base register, etc..)

* Implementation:
The implementation is strongly inspired by the so-called VFPV11
work-around, and works in a similar way with regard to the veneer
management. The only specificities are the specific conditions
detection, and the emitted code supported by quite a few
instruction encoding helpers.

A deviation with regard to VFPV11 is that the veneers are not
created in a dedicated section (as it was for .vfp11_veneer),
but in a .text.stm32l4xx_veneer that is naturally treated by
the linker scripts.

* Testing:
- Unit tests have been written to cover the veneer emission
   code for all ldm and vldm sub-cases, the IT block logic
   detection, the detection of specific IT block issues, the
   detection of branch overflows when creating veneers; and are
   part of the proposed change.
- No regression on ld test suite under arm-none-eabi and
   arm-linux-gnueabi configuration
- Our standard QA has been run with a 4_9-2015q2-20150609
   Linaro embedded compiler patched with this linker changes,
   in two variants: default linker settings and fix enabled,
   forcing the linker warnings to be turned into hard
   errors. Our QA contains various commercial tests suites,
   industrial benchmarks and proprietary ST applications. We
   have not observed behavioral changes with and without the
   fixes, nor detected the occurrence of unhandled sequences
   (IT blocks limitations), nor observed branch limitations.
- Hardware testing has been done on as STM32L476 "Discovery
   kit".

* Impact:
- The code size increase has been measured to 1.5% geometric
   mean increase on our whole set of QA applications, with a
   peak at 3.2%.

* Limitations:
- The veneer sizes are dimensioned to the maximum sequence
   that might be generated from a given instruction class (4
   words for ldm, 6 for vldm), but the actual sequence may be
   smaller.
- Faulty ldm or vldm that are part of IT blocks are
   transformed if and only if they appear as the last
   instruction of the IT block, where the creation of a jump is
   the only legitimate place, otherwise the linker emits a
   warning and does nothing.
- In ldmdb cases when the pc is involved in the restored
   register list, the memory access order is reversed compared
   to what it was originally.

Changelogs:

bfd/ChangeLog:

2015-09-17  Laurent Alfonsi <laurent.alfonsi@st.com>
            Christophe Monat <christophe.monat@st.com>

     * bfd-in2.h: Regenerate.
     * bfd-in.h (bfd_arm_stm32l4xx_fix): New enum. Specify how
     STM32L4XX instruction scanning should be done.
     (bfd_elf32_arm_set_stm32l4xx_fix)
     (bfd_elf32_arm_stm32l4xx_erratum_scan)
     (bfd_elf32_arm_stm32l4xx_fix_veneer_locations): Add prototypes.
     (bfd_elf32_arm_set_target_relocs): Add stm32l4xx fix type argument
     to prototype.
     * elf32-arm.c (STM32L4XX_ERRATUM_VENEER_SECTION_NAME)
     (STM32L4XX_ERRATUM_VENEER_ENTRY_NAME): Define macros.
     (elf32_stm32l4xx_erratum_type): New enum.
     (elf32_stm32l4xx_erratum_list): New struct. List of veneers or
     jumps to veneers.
     (_arm_elf_section_data): Add stm32l4xx_erratumcount,
     stm32l4xx_erratumlist.
     (elf32_arm_link_hash_table): Add stm32l4xx_erratum_glue_size,
     stm32l4xx_fix and num_stm32l4xx_fixes fields.
     (ctz): New function.
     (popcount): New function.
     (elf32_arm_link_hash_table_create): Initialize stm32l4xx_fix.
     (put_thumb2_insn): New function.
     (STM32L4XX_ERRATUM_LDM_VENEER_SIZE): Define. Size of a veneer for
     LDM instructions.
     (STM32L4XX_ERRATUM_VLDM_VENEER_SIZE): Define. Size of a veneer for
     VLDM instructions.
     (bfd_elf32_arm_allocate_interworking_sections): Initialise erratum
     glue section.
     (record_stm32l4xx_erratum_veneer) : New function. Create a single
     veneer, and its associated symbols.
     (bfd_elf32_arm_add_glue_sections_to_bfd): Add STM32L4XX erratum glue.
     (bfd_elf32_arm_set_stm32l4xx_fix): New function. Set the type of
     erratum workaround required.
     (bfd_elf32_arm_stm32l4xx_fix_veneer_locations): New function. Find
     out where veneers and branches to veneers have been placed in
     virtual memory after layout.
     (is_thumb2_ldmia): New function.
     (is_thumb2_ldmdb): Likewise.
     (is_thumb2_vldm ): Likewise.
     (stm32l4xx_need_create_replacing_stub): New function. Decide if a
     veneer must be emitted.
     (bfd_elf32_arm_stm32l4xx_erratum_scan): Scan the sections of an
     input BFD for potential erratum-triggering insns. Record results.
     (bfd_elf32_arm_set_target_relocs): Set stm32l4xx_fix field in
     global hash table.
     (elf32_arm_size_dynamic_sections): Collect glue information.
     (create_instruction_branch_absolute): New function.
     (create_instruction_ldmia): Likewise.
     (create_instruction_ldmdb): Likewise.
     (create_instruction_mov): Likewise.
     (create_instruction_sub): Likewise.
     (create_instruction_vldmia): Likewise.
     (create_instruction_vldmdb): Likewise.
     (create_instruction_udf_w): Likewise.
     (create_instruction_udf): Likewise.
     (push_thumb2_insn32): Likewise.
     (push_thumb2_insn16): Likewise.
     (stm32l4xx_fill_stub_udf): Likewise.
     (stm32l4xx_create_replacing_stub_ldmia): New function. Expands the
     replacing stub for ldmia instructions.
     (stm32l4xx_create_replacing_stub_ldmdb): Likewise for ldmdb.
     (stm32l4xx_create_replacing_stub_vldm): Likewise for vldm.
     (stm32l4xx_create_replacing_stub): New function. Dispatches the
     stub emission to the appropriate functions.
     (elf32_arm_write_section): Output veneers, and branches to veneers.

ld/ChangeLog:

2015-09-17  Laurent Alfonsi <laurent.alfonsi@st.com>
            Christophe Monat <christophe.monat@st.com>

     * ld.texinfo: Description of the STM32L4xx erratum workaround.
     * emultempl/armelf.em (stm32l4xx_fix): New.
     (arm_elf_before_allocation): Choose the type of fix, scan for
     erratum.
     (gld${EMULATION_NAME}_finish): Fix veneer locations.
     (arm_elf_create_output_section_statements): Propagate
     stm32l4xx_fix value.
     (PARSE_AND_LIST_PROLOGUE): Define OPTION_STM32L4XX_FIX.
     (PARSE_AND_LIST_LONGOPTS): Add entry for handling
     --fix-stm32l4xx-629360.
     (PARSE_AND_LIST_OPTION): Add entry for helping on
     --fix-stm32l4xx-629360.
     (PARSE_AND_LIST_ARGS_CASES): Treat OPTION_STM32L4XX_FIX.

ld/testsuite/ChangeLog:

2015-09-17  Laurent Alfonsi <laurent.alfonsi@st.com>
            Christophe Monat <christophe.monat@st.com>

     * ld-arm/arm-elf.exp (armelftests_common): Add STM32L4XX
       tests.
     * ld-arm/stm32l4xx-cannot-fix-far-ldm.d: New.
     * ld-arm/stm32l4xx-cannot-fix-far-ldm.s: Likewise.
     * ld-arm/stm32l4xx-cannot-fix-it-block.d: Likewise.
     * ld-arm/stm32l4xx-cannot-fix-it-block.s: Likewise.
     * ld-arm/stm32l4xx-fix-all.d: Likewise.
     * ld-arm/stm32l4xx-fix-all.s: Likewise.
     * ld-arm/stm32l4xx-fix-it-block.d: Likewise.
     * ld-arm/stm32l4xx-fix-it-block.s: Likewise.
     * ld-arm/stm32l4xx-fix-ldm.d: Likewise.
     * ld-arm/stm32l4xx-fix-ldm.s: Likewise.
     * ld-arm/stm32l4xx-fix-vldm.d: Likewise.
     * ld-arm/stm32l4xx-fix-vldm.s: Likewise.

Patch attached, waiting for your review.
--C

Attachment: 0001-Fix-for-STMicroelectronics-hardware-erratum-STM32L4X.patch
Description: Text document