Fix Thumb2 Curve25519 asm to do full reduce.
Change ARM32 to simpler carry/overflow processing.
Minor optimizations - use ubfx, no need to move register into temporary, cache value instead of loading again later.
Reduce the register push and pops in Thumb2 generated code.
Fix Thumb2 to have values less than 64 in decimal.
Refactors the output format of generated assembly files across all
platforms
(x86_64, ARM AArch64, ARMv8-32, Thumb2, PowerPC) for consistency and
correctness.
Changes
Data constant consolidation
- Pack multiple values per directive line (e.g., 4× .long or 8× .short
per
line) instead of one value per line, reducing file sizes significantly
- Normalize hex literal formatting: 64-bit values use full 8-byte
zero-padded
form (e.g., 0x0000000003ffffff instead of 0x3ffffff)
x86_64 assembly
- Use decimal immediate values for shift counts (e.g., $1 instead of
$0x01)
- .asm (MASM): use hex notation consistently for data constants;
update ALIGN
values to match data width (e.g., ALIGN 32 for 256-bit aligned data)
ARM .S files
- Move .type directive before .section for data objects (correct
ordering per
ELF convention)
ARM/Thumb2 inline C (_c.c) files
- Replace asm( with __asm__( for register variable constraints (better
portability)
- Add XALIGNED(8) attribute to constant lookup tables used in inline
asm
- Remove redundant #include <stdint.h> and
<wolfssl/wolfcrypt/libwolfssl_sources.h> headers
Files affected: 71 assembly and companion C files across
wolfcrypt/src/,
wolfcrypt/src/port/arm/, covering AES, ChaCha, Poly1305,
SHA-256/512/3,
Curve25519, ML-KEM, ML-DSA, and SP math routines.
ARM32/Thumb2: Generated code now omits the frame pointer attribute on
each function. Remove global use in configure.ac.
RISC-V 64: Omit the frame pointer on the one function that uses the
register 's0'.
Unroll loops and use larger types.
Allow benchmark to run each kyber parameter separately.
Allow benchmark to have -ml-dsa specified which runs all parameters.
Fix thumb2 ASM C code to not have duplicate includes and ifdef checks.
Fix thumb2 ASM C code to include error-crypt.h to ensure no empty
translation unit.
Check for WOLFSSL_SHA3 before including Thumb2 SHA-3 assembly code.
When functions are inlined, the labels need to be unique.
Putting '%=' on the end of the label ensures that the compilers appends
a unique number to the end.
Make a separate AES for IAR that has AES_encrypt_block and
AES_decrypt_block inlined. Default code is relying on compiler to use
specific registers and not modify others.
Improve performance of small SP ASM code for RSA.
IAR doesn't parse register variable declarations with specified
registers. IAR doesn't even honor the register keyword.
Can use small negative but IAR doesn't like it.
Specify the positive value instead.
Add a small code size version of mont_reduce_full using umlal and umaal.
Make 'asm' usage in variables use keyword '__asm__'.
Explicitly don't inline some functions when compiling with IAR.
Thumb2/ARM32: use umaal instruction to speed up mulitplication, squaring
and reduction. umaal not always available so use umlal instead.
Implementations for architectures less than 4 still kept - no umull.
Cleanup point arithmetic to not take shortcuts.
Make masking for constant time faster.
Add alternate asm for different compilers.
Get ARMv4 and ARMv6 compiling.
Remove whitespace at end of lines.
Add support for compiling ASM for Thumb2
Add Curve25519 ASM for Thumb2
Limit assembly code compiled when Ed25519 not required.
Rework all assembly implementations to replace ge_*() functions instead
of having fe_ge_*() versions that take many parameters.
Get ARM32 inline asm working.
Sean Parkinson
JacobBarthelmeh
Daniel Pouzzner
David Garske