As an optimization to improve performance, the assembler attempts to align branch targets so they do not cross instruction fetch boundaries. (Xtensa processors can be configured with either 32-bit or 64-bit instruction fetch widths.) An instruction immediately following a call is treated as a branch target in this context, because it will be the target of a return from the call. This alignment has the potential to reduce branch penalties at some expense in code size. This optimization is enabled by default. You can disable it with the ‘--no-target-align’ command-line option (see Command Line Options).
The target alignment optimization is done without adding instructions that could increase the execution time of the program. If there are density instructions in the code preceding a target, the assembler can change the target alignment by widening some of those instructions to the equivalent 24-bit instructions. Extra bytes of padding can be inserted immediately following unconditional jump and return instructions. This approach is usually successful in aligning many, but not all, branch targets.
LOOP family of instructions must be aligned such that the
first instruction in the loop body does not cross an instruction fetch
boundary (e.g., with a 32-bit fetch width, a
must be on either a 1 or 2 mod 4 byte boundary). The assembler knows
about this restriction and inserts the minimal number of 2 or 3 byte
no-op instructions to satisfy it. When no-op instructions are added,
any label immediately preceding the original loop will be moved in order
to refer to the loop instruction, not the newly generated no-op
instruction. To preserve binary compatibility across processors with
different fetch widths, the assembler conservatively assumes a 32-bit
fetch width when aligning
LOOP instructions (except if the first
instruction in the loop is a 64-bit instruction).
Previous versions of the assembler automatically aligned
instructions to 4-byte boundaries, but that alignment is now the