[llvm-dev] Should CPPFLAGS and CFLAGS still be used when compiling .ll to .o?

[Serge Guelton via llvm-dev]

On Thu, Jan 24, 2019 at 12:23:01PM -0600, Peng Yu via llvm-dev wrote:
> I am trying to use GNU make generating .o from .ll (which is converted from .c)
> 
> GNU use the following variable and rule to generate .o from .c. I'd
> like to follow its convention.
> 
> OUTPUT_OPTION = -o $@
> COMPILE.c = $(CC) $(CFLAGS) $(CPPFLAGS) $(TARGET_ARCH) -c
> 
> %.o: %.c
> #  commands to execute (built-in):
>   $(COMPILE.c) $(OUTPUT_OPTION) $<
> 
> My understanding is that since $(CFLAGS) $(CPPFLAGS) $(TARGET_ARCH)
> already has been used for .c -> .ll conversion, .ll ->.o conversions
> doesn't need $(CFLAGS) $(CPPFLAGS) $(TARGET_ARCH) any more. Is it so?

That's my feeling too. Moreover,

- as .ll files already embed machine-specific information, it could lead to invalid code
- if the .ll file was generated with -O0, it will carry optnone attribute and won't be sensible to extra optimisations anyway

> So the following rule should be sufficient?
> 
> %.o: %.ll
> #  commands to execute (built-in):
>   $(CC) $(OUTPUT_OPTION) $<

Beware that .ll files may be generated from something else than C code, I'd be wary of $(CC) and maybe introduce $(LLC) and set $(LLCFLAGS) to -filetype=obj ?
That would also avoid stupid situation where a user set CC=gcc

hope it helps,

serge