initVariant() function not being linked in 1.6.6

[ffissore]

From @kmclifton on November 23, 2015 3:5

We have some custom init code that needs to run to initialise our hardware (not Arduino). In version 1.6.5 of the IDE we simply included a file called initvariant.cpp (defining void initVariant() ) alongside pins_arduino.h under variants/standard in our package directory. See contents of initvariant.cpp below. With the upgrade to 1.6.6 that seems to have broken. The init code no longer runs at all. It isn't linked in.
Possibly related to this issue: aee32ff

#include <Wire.h>
#include <Arduino.h>
#include <pins_arduino.h>

void initVariant() 
{ 
    Wire.begin();
    pinMode(EXPANDER_RSTN, OUTPUT);
    digitalWrite(EXPANDER_RSTN, LOW);
    delay(10);
    digitalWrite(EXPANDER_RSTN, HIGH);

    delay(10);
}

Copied from original issue: arduino/Arduino#4205

[ffissore]

@kmclifton When I compile it I get

/tmp/arduino_a5e387bb8d8dac73c8f4a56a09b8171a/sketch_nov23a.ino: In function 'void initVariant()':
sketch_nov23a:8: error: 'EXPANDER_RSTN' was not declared in this scope
     pinMode(EXPANDER_RSTN, OUTPUT);
             ^

Is there some missing piece? Core, libs...?

[kmclifton]

Apologies, @ffissore. Yes, there is a missing piece. We have the following definition in our pins_arduino.h file (located in variants/standard in the installed package directory).

static const uint8_t EXPANDER_RSTN = 6;

[kmclifton]

@ffissore Some more clues after experimenting a bit:

One difference between 1.6.5 and 1.6.6 relates to the command that generates the elf file.

The comparison:

1.6.5
C:\Program Files (x86)\Arduino\hardware\tools\avr/bin/avr-gcc -w -Os -Wl,--gc-sections -mmcu=atmega328p -o C:\Users\AppData\Local\Temp\build7760252863079702224.tmp/sketch_nov24a.cpp.elf C:\Users\AppData\Local\Temp\build7760252863079702224.tmp\sketch_nov24a.cpp.o C:\Users\AppData\Local\Temp\build7760252863079702224.tmp\initvariant.cpp.o C:\Users\AppData\Local\Temp\build7760252863079702224.tmp/core.a -LC:\Users\AppData\Local\Temp\build7760252863079702224.tmp -lm

1.6.6
"C:\Program Files (x86)\Arduino\hardware\tools\avr/bin/avr-gcc" -w -Os -Wl,--gc-sections -mmcu=atmega328p -o "C:\Users\AppData\Local\Temp\build11ee33954a5ebb625d1463b873c77548.tmp/sketch_nov24a.ino.elf" "C:\Users\AppData\Local\Temp\build11ee33954a5ebb625d1463b873c77548.tmp\sketch\sketch_nov24a.ino.cpp.o" "C:\Users\AppData\Local\Temp\build11ee33954a5ebb625d1463b873c77548.tmp/core\core.a" "-LC:\Users\AppData\Local\Temp\build11ee33954a5ebb625d1463b873c77548.tmp" -lm

Note how in 1.6.5, initvariant.cpp.o is included, but it's not there in 1.6.6.

Using 1.6.6 - if I add initvariant.cpp explicitly to the sketch (Sketch | Add File), then when I compile, it's listed in the command that generates the elf file, and we're back to working again.

[ffissore]

Please provide the actual sketch, core and libs and error message you're getting: I can't reproduce your issue.
What I've attempted to do is:

• added initvariant.cpp to leonardo "variants" folder. code is https://gist.github.com/ffissore/0c3f04c54fc0620c6980
• added static const uint8_t EXPANDER_RSTN = 6; to pins_arduino.h of leonardo
• compiled an empty sketch with a call to initVariant(); in setup

Everything compiles as expected. Maybe a bug we've already solved in hourly builds? http://www.arduino.cc/en/Main/Software#hourly

[kmclifton]

Package at link below, including core etc. Library referenced in sketch also at link.
https://drive.google.com/folderview?id=0B1zBLRlluQd3eGFLZTNxSVdVdGs&usp=sharing
There is no error. The sketch compiles and uploads, but our initVariant() function doesn't get linked in. A snippet of the .lst file below shows that the initVariant() function that does get linked in is the default empty function:

 56a:        61 e0               ldi        r22, 0x01        ; 1
 56c:        8a e0               ldi        r24, 0x0A        ; 10
 56e:        0c 94 32 03         jmp        0x664        ; 0x664 <digitalWrite>

00000572 <initVariant>:
 572:        08 95               ret

00000574 <main>:
 574:        0e 94 f7 01         call        0x3ee        ; 0x3ee <init>
 578:        0e 94 b9 02         call        0x572        ; 0x572 <initVariant>
 57c:        0e 94 88 00         call        0x110        ; 0x110 <setup>

Sketch:

#include <dP_LedButton.h>

dP_LedButton myLedButton(2);

void setup() {
  myLedButton.begin();
}

void loop() {
  // TEST LED 1 (D1) and SWITCH 1 (SW1)
  if(myLedButton.readSwitch(1))
  {
    myLedButton.setLed(1, dP_LedButton::GREEN);
    delay(1000);    // 1 second delay
    myLedButton.setLed(1, dP_LedButton::RED);
    delay(1000);
    myLedButton.setLed(1, dP_LedButton::YELLOW);
    delay(1000);
    myLedButton.setLed(1, dP_LedButton::OFF);
    delay(1000);  
  }


}

[feilipu]

See this similar issue, and the resolution.

[matthijskooijman]

So the forum post suggests that calling a weak function from within the same compilation unit (cpp file) as where the empty weak stub is defined, causes problems. Presumably the compiler then inlines the function, or otherwise directly references it, without involving the linker, so the call will not use the other, strong, definition of the function. To fix this this, moving the weak empty initVariant into another cpp file (the forum post suggests hooks.cpp) would help.

I'm not 100% sure this analysis is correct, especially since the disassembler dump shows that no inlining takes place. I wonder if somehow the linking order, and/or .a archive linking affects which initVariant function is selected as well.

[feilipu]

To be honest, I'm not sure of the technical reason why this works.
I don't have a deep enough understanding of the vagaries of the gcc compiler and linker.

What I can say is that I spent hours on this issue, including moving the linker order and putting the external library before the ArduinoCore library, with no outcome. The .mpp file just showed the empty initVariant() function, not inlined.

Simply moving the attribute((weak)) function definitions to another file solved the problem.
And to be honest, putting all the empty hooks into the hooks.c file is a nicer solution anyway.

Still, it would be nice to know why this solution works.

[ffissore]

Since the problem, atm, seems to lie in Arduino core files, I'm assigning this to @cmaglie /cc @facchinm . If you guys come up with a fix at arduino-builder level, like reordering linker params, pls reassign this to me

[ffissore]

Turned out the problem was in a wrong porting from java to go. Variant files were archived in core.a while they MUST be passed to the linker one by one

[ffissore]

A fix is available with latest IDE hourly build http://www.arduino.cc/en/Main/Software#hourly

[kmclifton]

Thanks, @ffissore. That seems to have fixed things. I assume that this will then make it into 1.6.7?

[ffissore]

Yes