[Thanatos9404]

• Mark this if you consider it ready to merge
• I've added tests (optional)
• I wrote some lines in the book (optional)

Description

Fixes #8389. mips_i encoded a branch target as target / 4 - 1 (and 0 for targets up to 4), which is the offset only for a branch at address 0. mips_assemble already gets pc but never used it for branches. So:

• beq t0, zero, 0x100 at 0x100 jumped forward instead of to itself;
• branches back to an earlier label, like the loop in the issue, could not be assembled.

The offset is now (target - (pc + 4)) / 4, counted from the instruction after the branch and computed in 32 bits. A misaligned or out-of-range target now fails with the usual "Cannot assemble" error instead of being rounded.

While labels are sized, a label defined further down is still unknown, so encode reports the 4 bytes of every MIPS instruction in that pass instead of failing; otherwise every later label would be placed 4 bytes short.

The issue's loop now assembles to ffff0810feff0814, which disassembles to the same two branches back to 0.

A numeric target is now an address, like the ones pd prints and GNU as takes. Before, at any address other than 0 it acted as a distance from the branch: b 0x8 at 0x100 went to 0x108.

Testing

• test/db/asm/mips_32 (new) has assemble lines at different addresses for beq, bne, b and bal. Four of them fail on master; the one for a branch at address 0 passes on master too.
• test/db/tools/rasm2 covers the issue's backward loop, a forward label, and two targets that do not fit.
• All MIPS asm, anal and esil tests pass, as do db/tools/rasm2 and db/cmd/archs. sys/lint.py --all is clean.