TL;DR cset code generated by assembler uses different condition codes than ARM manual.
I do not understand how the conditions are encoded for the cset instruction.
I have troubles understanding why the assembler generates different binary code than what I would expect after looking at the ARM Architecture Reference Manual - ARMv8.
The GNU assembler on my pi3 generates following code for different cset checks (I confirmed that the code works as expected):
9a 9f 17 e0 cset x0, eq // eq = none
9a 9f 07 e0 cset x0, ne // ne = any
9a 9f b7 e0 cset x0, ge // ge = tcont
9a 9f a7 e0 cset x0, lt // lt = tstop
9a 9f d7 e0 cset x0, gt
9a 9f c7 e0 cset x0, le
The interesting part here are bit 12-15 (the 5th nibble from the left) it encodes the condition (cond), as shown in the manual:
Using the cset documentation and the assembler output one can extract following table:
+------+----------+
| cond | mnemonic |
|------+----------|
| 0001 | EQ |
| 0000 | NE |
| 1011 | GE |
| 1010 | LT |
| 1101 | GT |
| 1100 | LE |
+------+----------+
The confusion now starts when one looks at the condition codes in the ARM manual:
They are the opposite!
+-----------------+----------+
| ASSEMBLER | MANUAL |
+------+----------+----------+
| cond | mnemonic | mnemonic |
|------+----------+----------|
| 0001 | EQ | NE |
| 0000 | NE | EQ |
| 1011 | GE | LT |
| 1010 | LT | GE |
| 1101 | GT | LE |
| 1100 | LE | GT |
+------+----------+----------+
Could someone please point out to me the 'obvious' error I am making?