master does not compile from cc()

[jangorecki]

We should add a test to CI that code compiles from cc()

cc()

/home/jan/git/data.table/src 
Running command:
MAKEFLAGS='-j CC=gcc CFLAGS=-fopenmp\ -std=c11\ -O3\ -pipe\ -Wall\ -pedantic\ -Wstrict-prototypes\ -isystem\ /usr/share/R/include\ \ -fno-common' R CMD SHLIB -o data_table.so *.c
using C compiler: ‘gcc (GCC) 15.1.1 20250521 (Red Hat 15.1.1-2)’
gcc -I"/usr/include/R" -DNDEBUG   -I/usr/local/include    -fpic  -fopenmp -std=c11 -O3 -pipe -Wall -pedantic -Wstrict-prototypes -isystem /usr/share/R/include  -fno-common -c fread.c -o fread.o
In file included from /usr/include/R/R.h:70,
                 from freadR.h:6,
                 from fread.h:12,
                 from fread.c:1:
/usr/include/R/R_ext/Boolean.h:65:16: warning: ISO C does not support specifying ‘enum’ underlying types before C23 [-Wpedantic]
   65 |   typedef enum :int { FALSE = 0, TRUE } Rboolean;  // so NOT NA
      |                ^
fread.c: In function ‘copyFile’:
fread.c:120:37: error: implicit declaration of function ‘typeof’ [-Wimplicit-function-declaration]
  120 | #define OFFSET_POINTER(x, offset) ((typeof(x))((char*)x + (offset)))
      |                                     ^~~~~~
fread.c:454:17: note: in expansion of macro ‘OFFSET_POINTER’
  454 |   eof = (char *)OFFSET_POINTER(mmp_copy, fileSize);
      |                 ^~~~~~~~~~~~~~
fread.c:120:37: error: called object is not a function or function pointer
  120 | #define OFFSET_POINTER(x, offset) ((typeof(x))((char*)x + (offset)))
      |                                    ~^~~~~~~~~~
fread.c:454:17: note: in expansion of macro ‘OFFSET_POINTER’
  454 |   eof = (char *)OFFSET_POINTER(mmp_copy, fileSize);
      |                 ^~~~~~~~~~~~~~
fread.c: In function ‘freadMain’:
fread.c:120:37: error: called object is not a function or function pointer
  120 | #define OFFSET_POINTER(x, offset) ((typeof(x))((char*)x + (offset)))
      |                                    ~^~~~~~~~~~
fread.c:2411:31: note: in expansion of macro ‘OFFSET_POINTER’
 2411 |             fctx.targets[8] = OFFSET_POINTER(ctx.buff8, myNrow * rowSize8);
      |                               ^~~~~~~~~~~~~~
fread.c:120:37: error: called object is not a function or function pointer
  120 | #define OFFSET_POINTER(x, offset) ((typeof(x))((char*)x + (offset)))
      |                                    ~^~~~~~~~~~
fread.c:2412:31: note: in expansion of macro ‘OFFSET_POINTER’
 2412 |             fctx.targets[4] = OFFSET_POINTER(ctx.buff4, myNrow * rowSize4);
      |                               ^~~~~~~~~~~~~~
fread.c:120:37: error: called object is not a function or function pointer
  120 | #define OFFSET_POINTER(x, offset) ((typeof(x))((char*)x + (offset)))
      |                                    ~^~~~~~~~~~
fread.c:2413:31: note: in expansion of macro ‘OFFSET_POINTER’
 2413 |             fctx.targets[1] = OFFSET_POINTER(ctx.buff1, myNrow * rowSize1);
      |                               ^~~~~~~~~~~~~~
make: *** [/usr/lib64/R/etc/Makeconf:202: fread.o] Error 1
NULL

R version 4.5.0 (2025-04-11)
Platform: x86_64-redhat-linux-gnu
Running under: Fedora Linux 42 (Workstation Edition)

Matrix products: default
BLAS/LAPACK: FlexiBLAS OPENBLAS-OPENMP;  LAPACK version 3.12.0

locale:
 [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
 [3] LC_TIME=en_US.UTF-8        LC_COLLATE=en_US.UTF-8    
 [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
 [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
 [9] LC_ADDRESS=C               LC_TELEPHONE=C            
[11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

loaded via a namespace (and not attached):
[1] compiler_4.5.0 tools_4.5.0   

[aitap]

Sigh. Would you like a GitHub Action, or a GitLab CI check?

As for the problem itself, typeof is a widely supported GCC extension that made its way into C23. When compiling in strict standard compliance mode below C23, the extension keyword is disabled, but is instead available as __typeof__.

Potential solutions:

1. Note that OFFSET_POINTER is only ever used on void * pointers (because pointer arithmetic on void * is indeed forbidden by the standard). Drop the cast altogether and let the char * be implicitly converted back into void *, which shouldn't be raising any warnings:

--- a/src/fread.c
+++ b/src/fread.c
@@ -117,7 +117,7 @@ static void Field(FieldParseContext *ctx);
 #define IGNORE_BUMP(x) abs(x)
 #define TOGGLE_BUMP(x) (-x)

-#define OFFSET_POINTER(x, offset) ((typeof(x))((char*)x + (offset)))
+#define OFFSET_PVOID(x, offset) ((char*)(x) + (offset))

 //=================================================================================================
 //
@@ -451,7 +451,7 @@ double copyFile(size_t fileSize)  // only called in very very rare cases
     return -1.0; // # nocov
   memcpy(mmp_copy, mmp, fileSize);
   sof = mmp_copy;
-  eof = (char *)OFFSET_POINTER(mmp_copy, fileSize);
+  eof = sof + fileSize;
   return wallclock()-tt;
 }

@@ -2408,9 +2408,9 @@ int freadMain(freadMainArgs _args) {
               break;
             }
             // shift current buffer positions, since `myBuffX`s were probably moved by realloc
-            fctx.targets[8] = OFFSET_POINTER(ctx.buff8, myNrow * rowSize8);
-            fctx.targets[4] = OFFSET_POINTER(ctx.buff4, myNrow * rowSize4);
-            fctx.targets[1] = OFFSET_POINTER(ctx.buff1, myNrow * rowSize1);
+            fctx.targets[8] = OFFSET_PVOID(ctx.buff8, myNrow * rowSize8);
+            fctx.targets[4] = OFFSET_PVOID(ctx.buff4, myNrow * rowSize4);
+            fctx.targets[1] = OFFSET_PVOID(ctx.buff1, myNrow * rowSize1);
           }
           tLineStart = tch;  // for error message
           const char *fieldStart = tch;

2. Use __typeof__ when compiling for C < 23. GCC documentation says that "other C compilers won’t accept these alternative keywords", but other compilers consider it so important that they support it anyway, even MSVC. All of Rtools34, clang in Debian oldoldstable, and the latest Intel C compiler understand __typeof__.

--- a/src/fread.c
+++ b/src/fread.c
@@ -117,6 +117,9 @@ static void Field(FieldParseContext *ctx);
 #define IGNORE_BUMP(x) abs(x)
 #define TOGGLE_BUMP(x) (-x)

+#if !defined(__STDC_VERSION__) || __STDC_VERSION__ < 202311L
+#  define typeof(x) __typeof__(x) // nonstandard but widely supported extension
+#endif
 #define OFFSET_POINTER(x, offset) ((typeof(x))((char*)x + (offset)))

 //=================================================================================================

3. There's probably no point in trying to construct anything with _Generic, which is from C11 anyway.
4. There may be a way to use char * instead of void * and avoid OFFSET_POINTER altogether, but that diff would be much larger and take more time to produce and verify.

[jangorecki]

Maybe we can just add some missing flag inside cc() so that it will magically compile it same way as R CMD INSTALL does?

[aitap]

The -std=c11 is added here:

data.table/.dev/cc.R

Lines 94 to 96 in 5bbc4d5

	cmd = sprintf(R"(MAKEFLAGS='-j CC=%s PKG_CFLAGS=-f% CFLAGS=-std=c11\ -O0\ -ggdb\ %s\ -pedantic' R CMD SHLIB -d -o data_table.so *.c)", CC, OMP, W32)
	} else {
	cmd = sprintf(R"(MAKEFLAGS='-j CC=%s CFLAGS=-f%s\ -std=c11\ -O3\ -pipe\ -Wall\ -pedantic\ -Wstrict-prototypes\ -isystem\ /usr/share/R/include\ %s\ -fno-common' R CMD SHLIB -o data_table.so *.c)", CC, OMP, W32)

The !debug branch has a workaround for R's own headers causing -Wstrict-prototypes warnings, but maybe it's not needed any more? Can anyone reproduce them by running PKG_CFLAGS=-Wstrict-prototypes R CMD check on a modern version of R? Then cc()'s own flags can be added to PKG_CFLAGS as usual, without spending extra effort to put R includes into -isystem or to find out R's own flags using R CMD config CFLAGS.

[jangorecki]

I never use debug. Cannot we simply remove -std=c11? Or change it to c23?

[aitap]

Yes, it should be fine to remove -std=c11. It may have been required in the past to add -std=c99 on a compiler that used an older version of the standard by default, but it shouldn't be required now. Still, the code is broken for --std=... below C23. I wonder how prevalent this configuration is.

[aitap]

See also: #6471. Do we want do support the more strict C standard, or rely on GNU extensions? Compilers on popular platforms all support typeof by default, including the latest Intel compiler. Supporting the more strict C standard means applying one of the fixes above in order to avoid typeof() and avoiding typeof() in the future until data.table depends on an R version that requires C23. Past decisions, such as 98aa9d1, favoured avoiding language extensions.