Possible bug in pico_float in current Pico-SDK

[XFer012]

Just a head up: a user has apparently found a bug in current Pico-SDK (see last comment) which affects at least sin() and cos() in a nontrivial way:

raspberrypi/pico-sdk#2406

Maybe it's for real, maybe just user error, but...

[earlephilhower]

It looks like we're not using those files with the develop branch here. I've just run on a Pico 2

void setup() {
}
void loop() {
float angle = 1.56525815f;
float sin_angle = sin(angle);
float cos_angle = cos(angle);
Serial.printf("%f %f\n", sin_angle, cos_angle);
delay(1000);
}

And get the expected +1, 0 output:

0.999985 0.005538

[XFer012]

Excellent!
By the way: does the core automatically takes advantage of Pico2 hardware FP?

[earlephilhower]

Yes, single precision FP uses the HW:

float x = 1.0;
void setup() {
  x *= 2.1f;
}

Builds to:

(gdb) disassemble  setup
Dump of assembler code for function setup():
   0x10003274 <+0>:	ldr	r3, [pc, #16]	@ (0x10003288 <setup()+20>)
   0x10003276 <+2>:	vldr	s14, [pc, #20]	@ 0x1000328c <setup()+24>
   0x1000327a <+6>:	vldr	s15, [r3]
   0x1000327e <+10>:	vmul.f32	s15, s15, s14
   0x10003282 <+14>:	vstr	s15, [r3]
   0x10003286 <+18>:	bx	lr
   0x10003288 <+20>:	asrs	r0, r0, #5
   0x1000328a <+22>:	movs	r0, #0
   0x1000328c <+24>:	str	r6, [r4, #100]	@ 0x64
   0x1000328e <+26>:	ands	r6, r0
End of assembler dump.

[earlephilhower]

Looks like the bug may actually exist in this core, but it's not something that always fails. It seems to depend on the flash symbol layout so isn't guaranteed reproducible with a simple test. So the above test was fine, but it may be possible with certain other compiles that it will fail.

raspberrypi/pico-sdk#2429

That said, when they merge to develop there we can easily pull it in. Although I'd much rather have them do a real 2.1.2 release and avoid the non-tagged-version...