Support optimization for the AMD aocc compiler

Make sure we use the optimized maths library and interprocedural optimization. Both are needed to get optimization that works as well as other clang based-compilers.

Note in this MR we enable interprocedural optimization by default, unless the --enable-debug or --disable-optimization flags are used.

configure.ac

@@ -97,16 +97,33 @@ fi
 AM_CONDITIONAL([HAVECSDS],[test $with_csds = "yes"])
 
 
+# Use best known optimization for the current architecture. Actual optimization
+# happens later so we can avoid any issues it introduces with the compiler,
+# but we need to know if that will happen now.
+AC_ARG_ENABLE([optimization],
+   [AS_HELP_STRING([--enable-optimization],
+     [Enable compile time optimization flags for host @<:@yes/no@:>@]
+   )],
+   [enable_opt="$enableval"],
+   [enable_opt="yes"]
+)
 
-# Interprocedural optimization support. Needs special handling for linking and
-# archiving as well as compilation with Intels, needs to be done before
-# libtool is configured (to use correct LD).
+# Interprocedural optimization support. Can need special handling for linking
+# and archiving as well as compilation. Needs to be done before libtool is
+# configured so we use the correct LD. It can give good improvements for
+# clang based compilers, so the default is enabled, but we make that
+# disabled when not optimizing or debugging support is enabled.
+enable_ipo_default="yes";
+if test "x$ax_enable_debug" = "xyes" -o "x$enable_opt" != "xyes"; then
+   enable_ipo_default="no";
+   AC_MSG_WARN([Interprocedural optimization support default is changed to false])
+fi
 AC_ARG_ENABLE([ipo],
    [AS_HELP_STRING([--enable-ipo],
-     [Enable interprocedural optimization @<:@no/yes@:>@]
+     [Enable interprocedural optimization [default=yes unless debugging]]
    )],
    [enable_ipo="$enableval"],
-   [enable_ipo="no"]
+   [enable_ipo="$enable_ipo_default"]
 )
 
 if test "$enable_ipo" = "yes"; then
@@ -132,8 +149,8 @@ if test "$enable_ipo" = "yes"; then
                           ], [:] )
       AC_MSG_RESULT([added GCC interprocedural optimization support])
    elif test "$ax_cv_c_compiler_vendor" = "clang"; then
-      CFLAGS="$CFLAGS -flto=thin"
-      LDFLAGS="$LDFLAGS -flto=thin"
+      CFLAGS="$CFLAGS -flto"
+      LDFLAGS="$LDFLAGS -flto"
       : ${RANLIB="llvm-ranlib"}
       AC_MSG_RESULT([added LLVM interprocedural optimization support])
    else
@@ -231,7 +248,7 @@ AC_C_INLINE
 
 
 # If debugging try to show inlined functions.
-if test "x$enable_debug" = "xyes"; then
+if test "x$ax_enable_debug" = "xyes"; then
    #  Show inlined functions.
    if test "$ax_cv_c_compiler_vendor" = "gnu"; then
       # Would like to use -gdwarf and let the compiler pick a good version
@@ -548,16 +565,6 @@ fi
 # Define HAVE_POSIX_MEMALIGN if it works.
 AX_FUNC_POSIX_MEMALIGN
 
-# Only optimize if allowed, otherwise assume user will set CFLAGS as
-# appropriate.
-AC_ARG_ENABLE([optimization],
-   [AS_HELP_STRING([--enable-optimization],
-     [Enable compile time optimization flags for host @<:@yes/no@:>@]
-   )],
-   [enable_opt="$enableval"],
-   [enable_opt="yes"]
-)
-
 #  Disable vectorisation for known compilers. This switches off optimizations
 #  that could be enabled above, so in general should be appended. Slightly odd
 #  implementation as want to describe as --disable-vec, but macro is enable
@@ -582,6 +589,9 @@ AC_ARG_ENABLE([hand-vec],
 
 HAVEVECTORIZATION=0
 
+# Only optimize if allowed, otherwise assume user will set CFLAGS as
+# appropriate. Note argument check is done earlier so we can configure
+# other options related to optimization.
 if test "$enable_opt" = "yes" ; then
 
    # Choose the best flags for this compiler and architecture
@@ -603,7 +613,7 @@ if test "$enable_opt" = "yes" ; then
       esac
    elif test "$ax_cv_c_compiler_vendor" = "gnu"; then
       if test "$gcc_handles_avx512" = "yes"; then
-         case "$ax_gcc_arch" in
+         case "$ax_cv_gcc_archflag" in
 	    *skylake-avx512*)
                GRAVITY_CFLAGS="$GRAVITY_CFLAGS -mprefer-vector-width=512"
 	       ;;
@@ -614,6 +624,21 @@ if test "$enable_opt" = "yes" ; then
       else
          AC_MSG_NOTICE([No additional flags needed for gravity on this platform])
       fi
+   elif test "$ax_cv_c_compiler_vendor" = "clang"; then
+      #  Could be a number of compilers. Check for aocc specific flags we want
+      #  to use.
+      AX_CHECK_COMPILE_FLAG("-zopt", [GRAVITY_CFLAGS="$GRAVITY_CFLAGS -fvectorize -zopt"])
+      case "$ax_cv_gcc_archflag" in
+	    *skylake-avx512*)
+               GRAVITY_CFLAGS="$GRAVITY_CFLAGS -mprefer-vector-width=512"
+	       ;;
+	    *znver[[4-9]])
+               GRAVITY_CFLAGS="$GRAVITY_CFLAGS -mprefer-vector-width=512"
+               ;;
+            *)
+               :
+            ;;
+       esac
    else
       AC_MSG_WARN([Do not know what best gravity vectorization flags to choose for this compiler])
    fi
@@ -650,7 +675,6 @@ if test "$enable_opt" = "yes" ; then
 fi
 AM_CONDITIONAL([HAVEVECTORIZATION],[test -n "$HAVEVECTORIZATION"])
 
-
 # Add address sanitizer options to flags, if requested. Only useful for GCC
 # version 4.8 and later and clang.
 AC_ARG_ENABLE([sanitizer],
@@ -1727,12 +1751,12 @@ AC_MSG_RESULT($rtc_ok)
 # Special timers for the ARM v7 platforms (taken from FFTW-3 to match their cycle.h)
 AC_ARG_ENABLE(armv7a-cntvct, [AS_HELP_STRING([--enable-armv7a-cntvct],[enable the cycle counter on Armv7a via the CNTVCT register])], have_armv7acntvct=$enableval)
 if test "$have_armv7acntvct"x = "yes"x; then
-	AC_DEFINE(HAVE_ARMV7A_CNTVCT,1,[Define if you have enabled the CNTVCT cycle counter on ARMv7a])
+      AC_DEFINE(HAVE_ARMV7A_CNTVCT,1,[Define if you have enabled the CNTVCT cycle counter on ARMv7a])
 fi
 
 AC_ARG_ENABLE(armv7a-pmccntr, [AS_HELP_STRING([--enable-armv7a-pmccntr],[enable the cycle counter on Armv7a via the PMCCNTR register])], have_armv7apmccntr=$enableval)
 if test "$have_armv7apmccntr"x = "yes"x; then
-	AC_DEFINE(HAVE_ARMV7A_PMCCNTR,1,[Define if you have enabled the PMCCNTR cycle counter on ARMv7a])
+      AC_DEFINE(HAVE_ARMV7A_PMCCNTR,1,[Define if you have enabled the PMCCNTR cycle counter on ARMv7a])
 fi
 
 # Check if we have native exp10 and exp10f functions. If not failback to our
@@ -1754,6 +1778,19 @@ if test "$ax_cv_c_compiler_vendor" = "clang"; then
       AC_CHECK_LIB([m],[__sincos], [AC_DEFINE([HAVE___SINCOS],1,[The __sincos function is present.])])
       AC_CHECK_LIB([m],[__sincosf], [AC_DEFINE([HAVE___SINCOSF],1,[The __sincosf function is present.])])
 fi
+ 
+# The aocc compiler has optimized maths libraries that we should use. Check
+# any clang for this support. Note do this after the basic check for maths
+# as we need to make sure -lm follows. Also note needs -Ofast or -ffast-math
+# so only when optimizing.
+if test "$enable_opt" = "yes" -a "$ax_cv_c_compiler_vendor" = "clang"; then
+   have_almfast="yes"
+   AC_CHECK_LIB([almfast],[amd_fastexp],[LIBS="-fveclib=AMDLIBM -fsclrlib=AMDLIBM -lalmfast -lamdlibm $LIBS"],[have_almfast="no"],[-lamdlibm -lm])
+   if test "$have_almfast" = "no"; then
+      # Less optimized version.
+      AC_CHECK_LIB([amdlibm],[sqrt],,,[-lm])
+   fi
+fi
 
 # Check for glibc extension backtrace().
 AC_CHECK_FUNCS([backtrace backtrace_symbols])