linking succeeds with arm-none-eabi-g++ but not arm-none-eabi-gcc

Question

I am using the Launchpad Arm compiler tools. Specifically,

arm-none-eabi-g++ and arm-none-eabi-gcc from:

(GNU Tools for ARM Embedded Processors) 5.2.1 20151202 (release) [ARM/embedded-5-branch revision 231848]

I have a simple program targeted at an STM32F103 processor that has no purpose except to prove that I can write the hardware and call a function from the math library. This is all it is:

#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include "stm32f10x.h"

void hardwareTest(void){
   // Turn on the clock for PortB
   RCC->APB2ENR = RCC_APB2ENR_IOPBEN; // Turn on IO Port B
   // Put PB0 into push pull 50 MHz mode
   GPIOB->CRL = 0x03;
   // Turn PB0 on
   GPIOB->ODR = 1;
}

volatile int x; // force call to sqrt() later

int main(void) {
   x = sqrt(100.0f);
   x = sqrt(x);
   hardwareTest();
   return (x);
}

When I tried to build this, I got a linker error telling me that there is an undefined reference to sqrt. The build was done with arm-none-eabi-gcc. By chance I discovered that, if the build is done with arm-none-eabi-g++, using the same command line arguments, the linking is performed successfully.

I wrote a Makefile to demonstrate the difference:

PROJECT = minimal
SOURCES = src/startup_stm32f10x_hd.s \
          src/system_stm32f10x.c \
          src/main.c
OUTPUT = ./out
print-%:
    @echo '$*=$($*)'

TOOLCHAIN = arm-none-eabi-

CXX = $(TOOLCHAIN)g++
CC = $(TOOLCHAIN)gcc
AR = $(TOOLCHAIN)ar
AS = $(TOOLCHAIN)gcc -c -x assembler-with-cpp
LD =  $(TOOLCHAIN)gcc
OBJCOPY = $(TOOLCHAIN)objcopy
OBJDUMP = $(TOOLCHAIN)objdump
SIZE = $(TOOLCHAIN)size
RM = rm -f

CFLAGS  = -O
CFLAGS += -nostartfiles

CXXFLAGS  = -O
CXXFLAGS += -nostartfiles

ARCH = -mcpu=cortex-m3 -mthumb -DSTM32F10X_HD 
LDFLAGS = 

all: clean $(PROJECT).elf $(PROJECT).gcc $(PROJECT).bin

$(PROJECT).bin: $(PROJECT).elf 
    @echo ' ======== '
    @echo ' Generating binaries'
    $(OBJCOPY) -O binary $(OUTPUT)/$< $(OUTPUT)/$(PROJECT).bin
    $(OBJCOPY) -O ihex   $(OUTPUT)/$< $(OUTPUT)/$(PROJECT).hex
    @echo ' ======== '

$(PROJECT).elf: $(SOURCES)
    @echo ' ======== '
    @echo ' Successful build uses g++'
    @echo ' CXXFLAGS = $(CXXFLAGS)'
    @echo ' LDFLAGS = $(LDFLAGS)'
    @echo ' ARCH = $(ARCH)'
    $(CXX) -o $(OUTPUT)/$@ $(ARCH)  $(CXXFLAGS) $(LDFLAGS) -Wl,-Tld_script/stm32.ld,-lm  $^
    @echo ' ======== '

$(PROJECT).gcc: $(SOURCES)
    @echo ' ======== '
    @echo ' Broken build uses gcc'
    @echo ' CFLAGS = $(CFLAGS)'
    @echo ' LDFLAGS = $(LDFLAGS)'
    @echo ' ARCH = $(ARCH)'
    $(CC) -o $(OUTPUT)/$@ $(ARCH)   $(CFLAGS) $(LDFLAGS) -Wl,-Tld_script/stm32.ld,-lm  $^
    @echo ' ======== '

$(PROJECT).gxx: $(SOURCES)
    @echo ' ======== '
    @echo ' build with g++'
    $(CXX) -o $(OUTPUT)/$@ $(ARCH)  $(CXXFLAGS) $(LDFLAGS) -Wl,-Tld_script/stm32.ld  $^
    @echo ' ======== '

# Program the binary to the board using the builtin serial bootloader
program:
    stm32loader.py -p /dev/ttyUSB0 -ewv $(OUTPUT)/$(PROJECT).bin

# Remove the temporary files
clean:
    @echo ' '
    @echo ' Cleaning up: '
    $(RM) $(OUTPUT)/* *.o *.elf *.bin *.hex *.gcc *.gxx *.g++
    @echo ' ======== '

It give the following results:

Cleaning up: 
rm -f ./out/* *.o *.elf *.bin *.hex *.gcc *.gxx *.g++
======== 
======== 
 Successful build uses g++
 CXXFLAGS = -O -nostartfiles
 LDFLAGS = 
 ARCH = -mcpu=cortex-m3 -mthumb -DSTM32F10X_HD 
arm-none-eabi-g++ -o ./out/minimal.elf -mcpu=cortex-m3 -mthumb -DSTM32F10X_HD   -O -nostartfiles  -Wl,-Tld_script/stm32.ld,-lm  src/startup_stm32f10x_hd.s src/system_stm32f10x.c src/main.c
 ======== 
 ======== 
 Broken build uses gcc
 CFLAGS = -O -nostartfiles
 LDFLAGS = 
 ARCH = -mcpu=cortex-m3 -mthumb -DSTM32F10X_HD 
arm-none-eabi-gcc -o ./out/minimal.gcc -mcpu=cortex-m3 -mthumb -DSTM32F10X_HD    -O -nostartfiles  -Wl,-Tld_script/stm32.ld,-lm  src/startup_stm32f10x_hd.s src/system_stm32f10x.c src/main.c
/var/folders/t4/dv7b46055cjgknp4nndn1_zr0000gn/T//ccbl4swG.o: In function `main':
main.c:(.text+0x28): undefined reference to `sqrt'
collect2: error: ld returned 1 exit status
make: *** [minimal.gcc] Error 1
 ======== 
 Generating binaries
arm-none-eabi-objcopy -O binary ./out/minimal.elf ./out/minimal.bin
arm-none-eabi-objcopy -O ihex   ./out/minimal.elf ./out/minimal.hex
make: Target `all' not remade because of errors.

So can anyone tell me why the two compilers behave differently? What simple thing have I overlooked? How should I ensure proper linking with libm and others if I want to use arm-none-eabi-gcc?

I have looked at Freddie Chopin's makefiles but they are too complicated for me to unravel.

C and C++ are different languages. Would you expect a Java compiler to behave identical? Identical syntax&grammar does not imply identical semantics. — too honest for this site
Where's the problem specifying -lm as linker flags for gcc? The behaviour is as expected, afaics. — Ctx

rici rici · Accepted Answer · 2016-01-25T13:48:36

C++ requires that the math functions be part of the basic runtime while C allows them to be in a library. The GCC implementation achieved this by automatically linking libm in a C++ build.

There are many other differences in the link phase; C++ linking will consistently fail if a C linker is used.

For a C link, use the C linker and specify -lm to make libm available.

linking succeeds with arm-none-eabi-g++ but not arm-none-eabi-gcc

1 Answers