Merge pull request #8722 from hjelmn/exterminate_all_32_bit_atomic_platform_support

Exterminate all 32 bit platform support from opal atomic (IA32, ARM32, PowerPC32)

Author: hjelmn

README.md

@@ -201,9 +201,13 @@ base as of this writing (April 2020):
 
 * Open MPI requires a C99-capable compiler to build.
 
-* On platforms other than x86-64, ARM, and PPC, Open MPI requires a
-  compiler that either supports C11 atomics or the GCC `__atomic`
-  atomics (e.g., GCC >= v4.7.2).
+* On platforms other than x86-64, AArch64 (64-bit ARM), and PPC64,
+  Open MPI requires a compiler that either supports C11 atomics or
+  the GCC `__atomic` atomics (e.g., GCC >= v4.7.2).
+
+* 32-bit platforms are only supported with a recent compiler that
+  supports C11 atomics. This includes gcc 4.9.x+ (gcc 6.x or newer
+  recommened), icc 16, clang 3.1+, etc.
 
 * Mixing compilers from different vendors when building Open MPI
   (e.g., using the C/C++ compiler from one vendor and the Fortran

config/opal_config_asm.m4

@@ -880,53 +880,6 @@ AC_DEFUN([OPAL_CHECK_POWERPC_REG],[
                        [Whether r notation is used for ppc registers])
 ])dnl
 
-
-dnl #################################################################
-dnl
-dnl OPAL_CHECK_POWERPC_64BIT
-dnl
-dnl On some powerpc chips (the PPC970 or G5), the OS usually runs in
-dnl 32 bit mode, even though the hardware can do 64bit things.  If
-dnl the compiler will let us, emit code for 64bit test and set type
-dnl operations (on a long long).
-dnl
-dnl #################################################################
-AC_DEFUN([OPAL_CHECK_POWERPC_64BIT],[
-    if test "$ac_cv_sizeof_long" != "4" ; then
-        # this function should only be called in the 32 bit case
-        AC_MSG_ERROR([CHECK_POWERPC_64BIT called on 64 bit platform.  Internal error.])
-    fi
-    AC_MSG_CHECKING([for 64-bit PowerPC assembly support])
-        case $host in
-            *-darwin*)
-                ppc64_result=0
-                if test "$opal_cv_asm_powerpc_r_reg" = "1" ; then
-                   ldarx_asm="        ldarx r1,r1,r1";
-                else
-                   ldarx_asm="        ldarx 1,1,1";
-                fi
-                OPAL_TRY_ASSEMBLE([$opal_cv_asm_text
-        $ldarx_asm],
-                    [ppc64_result=1],
-                    [ppc64_result=0])
-            ;;
-            *)
-                ppc64_result=0
-            ;;
-        esac
-
-    if test "$ppc64_result" = "1" ; then
-        AC_MSG_RESULT([yes])
-        ifelse([$1],,:,[$1])
-    else
-        AC_MSG_RESULT([no])
-        ifelse([$2],,:,[$2])
-    fi
-
-    unset ppc64_result ldarx_asm
-])dnl
-
-
 dnl #################################################################
 dnl
 dnl OPAL_CHECK_CMPXCHG16B
@@ -1108,7 +1061,7 @@ AC_DEFUN([OPAL_CONFIG_ASM],[
         AC_MSG_ERROR([Cannot continue])
     elif test "$enable_builtin_atomics" = "yes" ; then
 	if test $opal_cv_have___atomic = "yes" ; then
-	   opal_cv_asm_builtin="BUILTIN_GCC"
+	    opal_cv_asm_builtin="BUILTIN_GCC"
 	else
 	    AC_MSG_WARN([GCC built-in atomics requested but not found.])
 	    AC_MSG_ERROR([Cannot continue])
@@ -1129,63 +1082,43 @@ AC_DEFUN([OPAL_CONFIG_ASM],[
         # find our architecture for purposes of assembly stuff
         opal_cv_asm_arch="UNSUPPORTED"
         OPAL_GCC_INLINE_ASSIGN=""
-	if test "$opal_cv_have___atomic_64" ; then
-            OPAL_ASM_SUPPORT_64BIT=1
-	else
-	    OPAL_ASM_SUPPORT_64BIT=0
-	fi
 
         case "${host}" in
-        x86_64-*x32)
-            opal_cv_asm_arch="X86_64"
-            OPAL_ASM_SUPPORT_64BIT=1
-            OPAL_GCC_INLINE_ASSIGN='"xaddl %1,%0" : "=m"(ret), "+r"(negone) : "m"(ret)'
-            ;;
-        i?86-*|x86_64*|amd64*)
+        x86_64-*x32|i?86-*|x86_64*|amd64*)
             if test "$ac_cv_sizeof_long" = "4" ; then
-                opal_cv_asm_arch="IA32"
-            else
-                opal_cv_asm_arch="X86_64"
-            fi
-            OPAL_ASM_SUPPORT_64BIT=1
+		if test $opal_cv_asm_builtin = BUILTIN_NO ; then
+		    AC_MSG_ERROR([IA32 atomics are no longer supported. Use a C11 compiler])
+		fi
+		opal_cv_asm_arch="IA32"
+	    else
+		opal_cv_asm_arch="X86_64"
+		OPAL_CHECK_CMPXCHG16B
+	    fi
             OPAL_GCC_INLINE_ASSIGN='"xaddl %1,%0" : "=m"(ret), "+r"(negone) : "m"(ret)'
-            OPAL_CHECK_CMPXCHG16B
             ;;
 
         aarch64*)
             opal_cv_asm_arch="ARM64"
-            OPAL_ASM_SUPPORT_64BIT=1
-            OPAL_ASM_ARM_VERSION=8
             OPAL_GCC_INLINE_ASSIGN='"mov %0, #0" : "=&r"(ret)'
             ;;
 
-        armv7*|arm-*-linux-gnueabihf)
-            opal_cv_asm_arch="ARM"
-            OPAL_ASM_SUPPORT_64BIT=1
-            OPAL_ASM_ARM_VERSION=7
-            OPAL_GCC_INLINE_ASSIGN='"mov %0, #0" : "=&r"(ret)'
-            ;;
+        armv7*|arm-*-linux-gnueabihf|armv6*)
+            if test $opal_cv_asm_builtin = BUILTIN_NO ; then
+		AC_MSG_ERROR([32-bit ARM atomics are no longer supported. Use a C11 compiler])
+	    fi
 
-        armv6*)
             opal_cv_asm_arch="ARM"
-            OPAL_ASM_SUPPORT_64BIT=0
-            OPAL_ASM_ARM_VERSION=6
-            CCASFLAGS="$CCASFLAGS -march=armv7-a"
             OPAL_GCC_INLINE_ASSIGN='"mov %0, #0" : "=&r"(ret)'
             ;;
 
         powerpc-*|powerpc64-*|powerpcle-*|powerpc64le-*|rs6000-*|ppc-*)
             OPAL_CHECK_POWERPC_REG
             if test "$ac_cv_sizeof_long" = "4" ; then
-                opal_cv_asm_arch="POWERPC32"
-
-                # Note that on some platforms (Apple G5), even if we are
-                # compiling in 32 bit mode (and therefore should assume
-                # sizeof(long) == 4), we can use the 64 bit test and set
-                # operations.
-                OPAL_CHECK_POWERPC_64BIT(OPAL_ASM_SUPPORT_64BIT=1)
-            elif test "$ac_cv_sizeof_long" = "8" ; then
-                OPAL_ASM_SUPPORT_64BIT=1
+		if test $opal_cv_asm_builtin = BUILTIN_NO ; then
+		    AC_MSG_ERROR([PowerPC 32-bit atomics are no longer supported. Use a C11 compiler])
+		fi
+		opal_cv_asm_arch="POWERPC32"
+	    elif test "$ac_cv_sizeof_long" = "8" ; then
                 opal_cv_asm_arch="POWERPC64"
             else
                 AC_MSG_ERROR([Could not determine PowerPC word size: $ac_cv_sizeof_long])
@@ -1213,20 +1146,10 @@ AC_DEFUN([OPAL_CONFIG_ASM],[
 	    ;;
         esac
 
-	if ! test -z "$OPAL_ASM_ARM_VERSION" ; then
-	    AC_DEFINE_UNQUOTED([OPAL_ASM_ARM_VERSION], [$OPAL_ASM_ARM_VERSION],
-                               [What ARM assembly version to use])
-	fi
-
 	if test "$opal_cv_asm_builtin" = "BUILTIN_GCC" ; then
          AC_DEFINE([OPAL_C_GCC_INLINE_ASSEMBLY], [1],
 		   [Whether C compiler supports GCC style inline assembly])
 	else
-         AC_DEFINE_UNQUOTED([OPAL_ASM_SUPPORT_64BIT],
-			    [$OPAL_ASM_SUPPORT_64BIT],
-			    [Whether we can do 64bit assembly operations or not.  Should not be used outside of the assembly header files])
-         AC_SUBST([OPAL_ASM_SUPPORT_64BIT])
-
          opal_cv_asm_inline_supported="no"
          # now that we know our architecture, try to inline assemble
          OPAL_CHECK_INLINE_C_GCC([$OPAL_GCC_INLINE_ASSIGN])
@@ -1239,12 +1162,12 @@ AC_DEFUN([OPAL_CONFIG_ASM],[
          asm_format="${asm_format}-${opal_cv_asm_lsym}"
          asm_format="${asm_format}-${opal_cv_asm_type}-${opal_asm_size}"
          asm_format="${asm_format}-${opal_asm_align_log_result}"
-         if test "$opal_cv_asm_arch" = "POWERPC32" || test "$opal_cv_asm_arch" = "POWERPC64" ; then
+         if test "$opal_cv_asm_arch" = "POWERPC64" ; then
              asm_format="${asm_format}-${opal_cv_asm_powerpc_r_reg}"
          else
              asm_format="${asm_format}-1"
          fi
-         asm_format="${asm_format}-${OPAL_ASM_SUPPORT_64BIT}"
+         asm_format="${asm_format}-1"
          opal_cv_asm_format="${asm_format}-${opal_cv_asm_gnu_stack}"
          # For the Makefile, need to escape the $ as $$.  Don't display
          # this version, but make sure the Makefile gives the right thing
@@ -1256,7 +1179,7 @@ AC_DEFUN([OPAL_CONFIG_ASM],[
         AC_DEFINE_UNQUOTED([OPAL_ASSEMBLY_FORMAT], ["$OPAL_ASSEMBLY_FORMAT"],
                            [Format of assembly file])
         AC_SUBST([OPAL_ASSEMBLY_FORMAT])
-      fi # if opal_cv_asm_builtin = BUILTIN_GCC
+    fi # if opal_cv_asm_builtin = BUILTIN_GCC
 
     result="OPAL_$opal_cv_asm_arch"
     OPAL_ASSEMBLY_ARCH="$opal_cv_asm_arch"

ompi/mca/osc/rdma/osc_rdma_frag.h

@@ -25,11 +25,7 @@ static inline void ompi_osc_rdma_frag_complete (ompi_osc_rdma_frag_t *frag)
         opal_atomic_rmb ();
 
         (void) opal_atomic_swap_32 (&frag->pending, 1);
-#if OPAL_HAVE_ATOMIC_MATH_64
         (void) opal_atomic_swap_64 (&frag->curr_index, 0);
-#else
-        (void) opal_atomic_swap_32 (&frag->curr_index, 0);
-#endif
     }
 }
 
@@ -88,11 +84,7 @@ static inline int ompi_osc_rdma_frag_alloc (ompi_osc_rdma_module_t *module, size
     OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_INFO, "allocating frag. pending = %d", curr->pending);
     OPAL_THREAD_ADD_FETCH32(&curr->pending, 1);
 
-#if OPAL_HAVE_ATOMIC_MATH_64
     my_index = opal_atomic_fetch_add_64 (&curr->curr_index, request_len);
-#else
-    my_index = opal_atomic_fetch_add_32 (&curr->curr_index, request_len);
-#endif
     if (my_index + request_len > mca_osc_rdma_component.buffer_size) {
         if (my_index <= mca_osc_rdma_component.buffer_size) {
             /* this thread caused the buffer to spill over */

ompi/mca/osc/rdma/osc_rdma_types.h

@@ -21,28 +21,13 @@ struct ompi_osc_rdma_frag_t;
 struct ompi_osc_rdma_sync_t;
 struct ompi_osc_rdma_peer_t;
 
-#if OPAL_HAVE_ATOMIC_MATH_64
-
 typedef int64_t osc_rdma_base_t;
 typedef uint64_t osc_rdma_size_t;
 typedef int64_t osc_rdma_counter_t;
 typedef opal_atomic_int64_t osc_rdma_atomic_counter_t;
 
 #define ompi_osc_rdma_counter_add opal_atomic_add_fetch_64
 
-#else
-
-typedef int32_t osc_rdma_base_t;
-typedef uint32_t osc_rdma_size_t;
-typedef int32_t osc_rdma_counter_t;
-typedef opal_atomic_int32_t osc_rdma_atomic_counter_t;
-
-#define ompi_osc_rdma_counter_add opal_atomic_add_fetch_32
-
-#endif
-
-#if OPAL_HAVE_ATOMIC_MATH_64
-
 #define OMPI_OSC_RDMA_LOCK_EXCLUSIVE   0x8000000000000000l
 
 typedef int64_t  ompi_osc_rdma_lock_t;
@@ -70,38 +55,6 @@ static inline int ompi_osc_rdma_lock_compare_exchange (opal_atomic_int64_t *p, i
     return ret;
 }
 
-#else
-
-#define OMPI_OSC_RDMA_LOCK_EXCLUSIVE 0x80000000l
-
-typedef int32_t  ompi_osc_rdma_lock_t;
-typedef opal_atomic_int32_t  ompi_osc_rdma_atomic_lock_t;
-
-static inline int32_t ompi_osc_rdma_lock_add (opal_atomic_int32_t *p, int32_t value)
-{
-    int32_t new;
-
-    opal_atomic_mb ();
-    /* opal_atomic_add_fetch_32 differs from normal atomics in that is returns the new value */
-    new = opal_atomic_add_fetch_32 (p, value) - value;
-    opal_atomic_mb ();
-
-    return new;
-}
-
-static inline int ompi_osc_rdma_lock_compare_exchange (opal_atomic_int32_t *p, int32_t *comp, int32_t value)
-{
-    int ret;
-
-    opal_atomic_mb ();
-    ret = opal_atomic_compare_exchange_strong_32 (p, comp, value);
-    opal_atomic_mb ();
-
-    return ret;
-}
-
-#endif /* OPAL_HAVE_ATOMIC_MATH_64 */
-
 /**
  * @brief structure describing a window memory region
  */
@@ -219,11 +172,7 @@ struct ompi_osc_rdma_frag_t {
 
     /* Number of operations which have started writing into the frag, but not yet completed doing so */
     opal_atomic_int32_t pending;
-#if OPAL_HAVE_ATOMIC_MATH_64
     opal_atomic_int64_t curr_index;
-#else
-    opal_atomic_int32_t curr_index;
-#endif
 
     struct ompi_osc_rdma_module_t *module;
     mca_btl_base_registration_handle_t *handle;

ompi/mca/osc/sm/osc_sm.h

@@ -18,22 +18,11 @@
 
 #include "opal/mca/shmem/base/base.h"
 
-#if OPAL_HAVE_ATOMIC_MATH_64
-
 typedef uint64_t osc_sm_post_type_t;
 typedef opal_atomic_uint64_t osc_sm_post_atomic_type_t;
 #define OSC_SM_POST_BITS 6
 #define OSC_SM_POST_MASK 0x3f
 
-#else
-
-typedef uint32_t osc_sm_post_type_t;
-typedef opal_atomic_uint32_t osc_sm_post_atomic_type_t;
-#define OSC_SM_POST_BITS 5
-#define OSC_SM_POST_MASK 0x1f
-
-#endif
-
 /* data shared across all peers */
 struct ompi_osc_sm_global_state_t {
     int use_barrier_for_fence;

ompi/mca/osc/sm/osc_sm_active_target.c

@@ -161,11 +161,7 @@ ompi_osc_sm_start(struct ompi_group_t *group,
 
             opal_atomic_rmb ();
 
-#if OPAL_HAVE_ATOMIC_MATH_64
             (void) opal_atomic_fetch_xor_64 ((opal_atomic_int64_t *) module->posts[my_rank] + rank_byte, rank_bit);
-#else
-            (void) opal_atomic_fetch_xor_32 ((opal_atomic_int32_t *) module->posts[my_rank] + rank_byte, rank_bit);
-#endif
        }
 
         free (ranks);
@@ -247,11 +243,7 @@ ompi_osc_sm_post(struct ompi_group_t *group,
 
         gsize = ompi_group_size(module->post_group);
         for (int i = 0 ; i < gsize ; ++i) {
-#if OPAL_HAVE_ATOMIC_MATH_64
             (void) opal_atomic_fetch_add_64 ((opal_atomic_int64_t *) module->posts[ranks[i]] + my_byte, my_bit);
-#else
-            (void) opal_atomic_fetch_add_32 ((opal_atomic_int32_t *) module->posts[ranks[i]] + my_byte, my_bit);
-#endif
         }
 
         opal_atomic_wmb ();

opal/include/opal/sys/Makefile.am

@@ -15,7 +15,7 @@
 #                         reserved.
 # Copyright (c) 2017      Research Organization for Information Science
 #                         and Technology (RIST). All rights reserved.
-# Copyright (c) 2020      Google, LLC. All rights reserved.
+# Copyright (c) 2020-2021 Google, LLC. All rights reserved.
 # $COPYRIGHT$
 #
 # Additional copyrights may follow
@@ -34,8 +34,6 @@ headers += \
         opal/sys/cma.h
 
 include opal/sys/x86_64/Makefile.am
-include opal/sys/arm/Makefile.am
 include opal/sys/arm64/Makefile.am
-include opal/sys/ia32/Makefile.am
 include opal/sys/powerpc/Makefile.am
 include opal/sys/gcc_builtin/Makefile.am