build : enable more non-default compiler warnings

[cebtenzzre]

I compiled llama.cpp with clang's -Weverything option, and found some warning flags that I think we should use.

The new warnings are -Wmissing-noreturn -Wextra-semi for g++ and clang++, -Wunreachable-code-break -Wunreachable-code-return for clang and clang++, and -Wrange-loop-bind-reference for clang++.

I made a new GGML_UNREACHABLE() macro to cover some of the cases where it should be explicit that code is unreachable. It's basically assert(false), but with the added benefit that it will compile to __builtin_unreachable() when assertions are disabled, both to enable compiler optimizations and to make sure no -Wreturn-type warnings appear because of the lack of a return statement at the end of a function.

Other included changes:

• Add missing 'static' specifiers
• Use -Werror=implicit-function-declaration for C
• Don't pass -Wdouble-promotion to old clang
• Build q8dot and benchmark-matmult in the Makefile by default

[Green-Sky]

did you check which version of the compilers support the flags?

[cebtenzzre]

did you check which version of the compilers support the flags?

It looks like -Wextra-semi was only added in GCC 8.1. master compiles fine with GCC 5. So I'll have to add some version checks to the build scripts.

[cebtenzzre]

I added some compiler version checks. I've tested with several compilers, including LLVM clang 3.7.0, Apple clang 8.0.0, and gcc 5.5.0.

[ggerganov]

Should merge after resolving conflicts

[ggerganov]

With CUDA builds, I think the -Wpedantic now spams the build with the following warning for every line:

...

In file included from tmpxft_001370c3_00000000-7_ggml-cuda.cudafe1.stub.c:1:
/home/ubuntu/ggerganov/llama.cpp/ggml-cuda.cu:1295:3: warning: style of line directive is a GCC extension
 1295 | 
      |   ^
In file included from tmpxft_001370c3_00000000-7_ggml-cuda.cudafe1.stub.c:1:
/tmp/tmpxft_001370c3_00000000-7_ggml-cuda.cudafe1.stub.c:5:3: warning: style of line directive is a GCC extension
    5 | #if !defined(__CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__)
      |   ^~~~
In file included from tmpxft_001370c3_00000000-7_ggml-cuda.cudafe1.stub.c:1:
/home/ubuntu/ggerganov/llama.cpp/ggml-cuda.cu:1297:3: warning: style of line directive is a GCC extension
 1297 | 
      |   ^   
In file included from tmpxft_001370c3_00000000-7_ggml-cuda.cudafe1.stub.c:1:
/home/ubuntu/ggerganov/llama.cpp/ggml-cuda.cu:1297:3: warning: style of line directive is a GCC extension
 1297 | 
      |   ^   
In file included from tmpxft_001370c3_00000000-7_ggml-cuda.cudafe1.stub.c:1:
/home/ubuntu/ggerganov/llama.cpp/ggml-cuda.cu:1411:3: warning: style of line directive is a GCC extension
 1411 | 
      |   ^
In file included from tmpxft_001370c3_00000000-7_ggml-cuda.cudafe1.stub.c:1:
/tmp/tmpxft_001370c3_00000000-7_ggml-cuda.cudafe1.stub.c:5:3: warning: style of line directive is a GCC extension
    5 | #if !defined(__CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__)
      |   ^~~~
In file included from tmpxft_001370c3_00000000-7_ggml-cuda.cudafe1.stub.c:1:
/home/ubuntu/ggerganov/llama.cpp/ggml-cuda.cu:1413:3: warning: style of line directive is a GCC extension
 1413 | 

...

[slaren]

I also see that when building with cmake with CUDA enabled. The Makefile build disables -Wpendatic with nvcc to avoid this issue:

llama.cpp/Makefile

Line 484 in 40e07a6

NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"

[cebtenzzre]

I could try and change the structure of the flags to match the Makefile better. I wasn't really thinking about CUDA when I was porting my changes to CMake.