improve OpenMP detection on macOS (#6034)

* improve OpenMP detection on macOS

* detect LLVM OpenMP builds

* remove some comments

* define _OPENMP when using LLVM libomp

* fix nesting for 'echo no'

* let openmp warning stand out a bit more

* Always redirect to config.log, not /dev/null

* NEWS

* Restore deletion of all test-omp* artefacts

* try and reduce diff by restoring comment

* reduce diff: delete in same order

* annotate 'fi' for readability

---------

Co-authored-by: Michael Chirico <michaelchirico4@gmail.com>

NEWS.md

@@ -54,6 +54,8 @@
 
 7. Updated a test relying on `>` working for comparing language objects to a string, which will be deprecated by R, [#5977](https://github.com/Rdatatable/data.table/issues/5977); no user-facing effect. Thanks to R-core for continuously improving the language.
 
+8. OpenMP detection when building from source on Mac is improved, [#4348](https://github.com/Rdatatable/data.table/issues/4348). Thanks @jameshester and @kevinushey for the request and @kevinushey for the PR, @jameslamb for the advice and @s-u of R-core for ensuring CRAN machines are configured to support the uxpected setup.
+
 # data.table [v1.15.0](https://github.com/Rdatatable/data.table/milestone/29)  (30 Jan 2024)
 
 ## BREAKING CHANGE

configure

@@ -21,7 +21,7 @@ esac
 
 msg=0
 NOZLIB=1 # if pkg-config is not available then zlib will be disabled for higher chance of compilation success
-pkg-config --version >/dev/null 2>&1
+pkg-config --version >> config.log 2>&1
 if [ $? -ne 0 ]; then
   echo "*** pkg-config is not installed."
   msg=1
@@ -33,9 +33,9 @@ else
   else
     NOZLIB=0
     lib=`pkg-config --libs zlib`
-    expr -- "$lib" : ".*-lz$" >/dev/null  # -- for FreeBSD, #4652
+    expr -- "$lib" : ".*-lz$" >> config.log  # -- for FreeBSD, #4652
     if [ $? -ne 0 ]; then
-      expr -- "$lib" : ".*-lz " >/dev/null
+      expr -- "$lib" : ".*-lz " >> config.log
       # would use \b in one expr but MacOS does not support \b
       if [ $? -ne 0 ]; then
         echo "*** pkg-config is installed and 'pkg-config --exists zlib' succeeds but"
@@ -81,36 +81,93 @@ int main() {
 }
 EOF
 
-# First, try R CMD SHLIB to see if R can already compile
-# things using OpenMP without any extra help from data.table
-"${R_HOME}/bin/R" CMD SHLIB test-omp.c >/dev/null 2>&1 || R_NO_OPENMP=1
+detect_openmp () {
 
-if [ "$R_NO_OPENMP" = "1" ]; then
-  # Compilation failed -- try forcing -fopenmp instead.
-  R_NO_OPENMP=0
-  "${CC}" "${CFLAGS}" -fopenmp test-omp.c || R_NO_OPENMP=1
-  # TODO: and then nothing seems to be done with this outcome
-else
-  echo "R CMD SHLIB supports OpenMP without any extra hint"
-fi
+  if [ "$(uname)" = "Linux" ]; then
+
+    printf "%s" "* checking if R installation supports OpenMP without any extra hints... "
+    if "${R_HOME}/bin/R" CMD SHLIB test-omp.c >> config.log 2>&1; then
+      echo "yes"
+      export R_OPENMP_ENABLED=1
+      return
+    else
+      echo "no"
+    fi
+
+
+    printf "%s" "* checking if R installation supports openmp with \"-fopenmp\" flag... "
+    if ${CC} ${CFLAGS} -fopenmp test-omp.c >> config.log 2>&1; then
+      echo "yes"
+      export PKG_CFLAGS="${PKG_CFLAGS} -fopenmp"
+      export R_OPENMP_ENABLED=1
+      return
+    else
+      echo "no"
+    fi
+  fi # uname=Linux
+
+  if [ "$(uname)" = "Darwin" ]; then
+
+    # https://mac.r-project.org/openmp
+    printf "%s" "* checking if R installation supports OpenMP with \"-Xclang -fopenmp\" ... "
+    if CPPFLAGS="${CPPFLAGS} -Xclang -fopenmp" LDFLAGS="${LDFLAGS} -lomp" "${R_HOME}/bin/R" CMD SHLIB test-omp.c >> config.log 2>&1; then
+      echo "yes"
+      export PKG_CFLAGS="${PKG_CFLAGS} -Xclang -fopenmp"
+      export PKG_LIBS="${PKG_LIBS} -lomp"
+      export R_OPENMP_ENABLED=1
+      return
+    else
+      echo "no"
+    fi
 
+    if [ "$(uname -m)" = "arm64" ]; then
+      HOMEBREW_PREFIX=/opt/homebrew
+    else
+      HOMEBREW_PREFIX=/usr/local
+    fi
+
+    if [ -e "${HOMEBREW_PREFIX}/opt/libomp" ]; then
+      printf "%s" "* checking if libomp installation at ${HOMEBREW_PREFIX}/opt/libomp can be used... "
+      LIBOMP_INCLUDE="-I${HOMEBREW_PREFIX}/opt/libomp/include -Xclang -fopenmp"
+      LIBOMP_LINK="-L${HOMEBREW_PREFIX}/opt/libomp/lib -lomp"
+      if ${CC} ${CFLAGS} ${LIBOMP_INCLUDE} ${LIBOMP_LINK} test-omp.c >> config.log 2>&1; then
+        echo "yes"
+        export PKG_CFLAGS="${PKG_CFLAGS} ${LIBOMP_INCLUDE}"
+        export PKG_LIBS="${PKG_LIBS} ${LIBOMP_LINK}"
+        export R_OPENMP_ENABLED=1
+        return
+      else
+        echo "no"
+      fi
+    fi
+
+  fi # uname=Darwin
+
+  # No support for OpenMP available
+  export R_OPENMP_ENABLED=0
+}
+
+detect_openmp
 # Clean up.
 rm -f test-omp.* a.out
 
-# Write to Makevars
-if [ "$R_NO_OPENMP" = "1" ]; then
+if [ "${R_OPENMP_ENABLED}" = "0" ]; then
+  echo "***"
   echo "*** OpenMP not supported! data.table uses OpenMP to automatically"
   echo "***   parallelize operations like sorting, grouping, file reading, etc."
   echo "*** For details on how to install the necessary toolchains on your OS see:"
   echo "***   https://github.com/Rdatatable/data.table/wiki/Installation"
   echo "*** Continuing installation without OpenMP support..."
+  echo "***"
   sed -e "s|@openmp_cflags@||" src/Makevars.in > src/Makevars
 else
   sed -e "s|@openmp_cflags@|\$(SHLIB_OPENMP_CFLAGS)|" src/Makevars.in > src/Makevars
 fi
+
 # retain user supplied PKG_ env variables, #4664. See comments in Makevars.in too.
 sed -e "s|@PKG_CFLAGS@|$PKG_CFLAGS|" src/Makevars > src/Makevars.tmp && mv src/Makevars.tmp src/Makevars
 sed -e "s|@PKG_LIBS@|$PKG_LIBS|" src/Makevars > src/Makevars.tmp && mv src/Makevars.tmp src/Makevars
+
 # optional dependency on zlib
 if [ "$NOZLIB" = "1" ]; then
   echo "*** Compilation without compression support in fwrite"

src/init.c

@@ -321,15 +321,18 @@ int GetVerbose(void) {
 
 // # nocov start
 SEXP hasOpenMP(void) {
-  // Just for use by onAttach (hence nocov) to avoid an RPRINTF from C level which isn't suppressable by CRAN
-  // There is now a 'grep' in CRAN_Release.cmd to detect any use of RPRINTF in init.c, which is
-  // why RPRINTF is capitalized in this comment to avoid that grep.
-  // .Platform or .Machine in R itself does not contain whether OpenMP is available because compiler and flags are per-package.
-  #ifdef _OPENMP
+
+#if defined(_OPENMP)
+  // gcc build of libomp
   return ScalarInteger(_OPENMP); // return the version; e.g. 201511 (i.e. 4.5)
-  #else
-  return ScalarInteger(0);       // 0 rather than NA so that if() can be used on the result
-  #endif
+#elif defined(KMP_VERSION_BUILD)
+  // LLVM builds of libomp
+  return ScalarInteger(KMP_VERSION_BUILD);
+#else
+  // no OpenMP support detected
+  return ScalarInteger(0);
+#endif
+
 }
 // # nocov end
 

src/myomp.h

@@ -1,3 +1,11 @@
+
+// Compatibility define for LLVM builds of libomp.
+#ifdef KMP_VERSION_BUILD
+# ifndef _OPENMP
+#  define _OPENMP KMP_VERSION_BUILD
+# endif
+#endif
+
 #ifdef _OPENMP
   #include <omp.h>
   #if _OPENMP >= 201511