timers: Increase test variation, increase EFM32 flexibility

[chrysn]

Contribution description

Timers appear to be written with a strong focus on providing for XTimer; requirements have changed with the advent of ZTimer. In particular, EFM32 timers encoded a few of these choices rather hardly.

This PR contains small changes for making them test- and usable in a more versatile manner:

• Timer test: Run with all the typical frequencies. It's OK if some can not be configured on the timer; this is reflected both in the output text and the success assessment.
• EFM32: Expose initialization's faillibility: An assert is turned into a negative return. (The same commit also returns an error on an invalid configuration, but that is later fully handled anyway; expecting things to work without a prescaler was an interesting debug journey).
• EFM32: Allow timers without prescale timers. They have a relatively narrow band of frequencies they can use (their clock / 2**{0..=10}), but that may be precisely what you need.

Testing procedure

• Run the periph_timer test, preferably with an LETIMER board like sltb001a.
• Before, only one frequency was tested, now all are. Without only the first commit, rather than merely showing an error on exotic frequencies, the system would run into an assert (but didn't because we only tested the preferred frequency anyway).

To test the last commit, we currently have no boards that are configured with single timers, so apply this and run on stk3700:

diff --git a/boards/stk3700/include/periph_conf.h b/boards/stk3700/include/periph_conf.h
index c66959c2dd..68e44469ec 100644
--- a/boards/stk3700/include/periph_conf.h
+++ b/boards/stk3700/include/periph_conf.h
@@ -222,12 +222,21 @@ static const timer_conf_t timer_config[] = {
         },
         .irq = LETIMER0_IRQn,
         .channel_numof = 2
-    }
+    },
+    {
+        .timer = {
+            .dev = TIMER2,
+            .cmu = cmuClock_TIMER2
+        },
+        .irq = TIMER2_IRQn,
+        .channel_numof = 3
+    },
 };
 
 #define TIMER_NUMOF         ARRAY_SIZE(timer_config)
 #define TIMER_0_ISR         isr_timer1
 #define TIMER_1_ISR         isr_letimer0
+#define TIMER_2_ISR         isr_timer2
 /** @} */
 
 /**
diff --git a/tests/periph_timer/main.c b/tests/periph_timer/main.c
index 7a8b1af7c6..1019bf6f7f 100644
--- a/tests/periph_timer/main.c
+++ b/tests/periph_timer/main.c
@@ -125,7 +125,7 @@ int main(void)
         res += test_timer(i, TIMER_SPEED);
     }
 
-    unsigned speeds[] = {32768, 250000, 500000, 1000000};
+    unsigned speeds[] = {32768, 250000, 500000, 1000000, 3000000};
     for (unsigned s = 0; s < ARRAY_SIZE(speeds); ++s) {
         unsigned speed = speeds[s];
         if (speed == TIMER_SPEED) {

and then run the periph_timers test. It will formally fail (because for automatable testing we'd need to give a per-timer known-good frequency, but there are no such boards right now so it's pointless anyway), but you can observe that timer 2 fails the 32kiHz test and the others, but then passes the 3MHz test.

Issues/PRs references

(I previously had a lot of potential follow-ups in here that are now all in this PR).

[maribu]

Maybe I can interest you in #16349 - but I think currently it is waiting for me to fix some issues.

[chrysn]

It's a good addition (esp. for testing), but I think the pressing need for that is more at build time. If it becomes unblocked, I'll be happy to adjust the new tests to that instead.

[fjmolinas]

I ran the test on a set of BOARD, results are here, I did not look into the details of any test, but in master only openmote-b and nucleo-l552ze-q did not pass, while on this PR also: arduino-uno, samr21-xpro, atmega256rfr2-xpro, nucleo-l4r5z, z1, i-nucleo-lrwan1 and arduino-mega2560

[chrysn]

The build failures were due to some unsigned where I should have said uint32_t; fixuped. Oddly, some of the failing ones are not AVR (and thus 16bit), investigating. (And sorry for the force push: I forgot to push earlier and thought all was in sync).

[chrysn]

I think I found what samr21-xpro tripped over (another assert similar to the removed EFM32 one); could you trigger that test run again?

[chrysn]

(Belaying tests until ... WAAH can we please have a version of C in which the number hierarchy and literals is not messed up?)

[kfessel]

#17654 might be of interest adding a second periph-timer (LP/LE however one wants to call it) to the ztimer initialization

[chrysn]

The only remaining test failure that does not appear to be a fluke is run_test/tests/periph_timer/esp32-wroom-32:gnu, where the timer0 initialization fails -- odd, because I didn't touch the ESP32 code.

And the number of other tests failing "just so" is worrying:

compile/examples/hello-world/esp8266-esp-12x:gnu
run_test/tests/periph_timer/esp32-wroom-32:gnu
run_test/tests/thread_float/samr21-xpro:gnu
run_test/tests/sys_sched_round_robin/esp32-wroom-32:gnu
compile/tests/ztimer_overhead/native:gnu