Change system() calls in llvm_codegen() to vfork() + execv().

Add a new subroutine pocl_run_command in pocl-util.c,
and make both pocl-hsa and pocl-llvm-codegen use it.

CMakeLists.txt

@@ -685,6 +685,10 @@ set(EXTRA_HOST_LLC_FLAGS "" CACHE STRING "Extra parameters to llc for code gener
 
 set(HOST_AS_FLAGS "${DEFAULT_HOST_AS_FLAGS} ${EXTRA_HOST_AS_FLAGS}")
 set(HOST_LD_FLAGS "${DEFAULT_HOST_LD_FLAGS} ${EXTRA_HOST_LD_FLAGS}" )
+string(STRIP "${HOST_LD_FLAGS}" HOST_LD_FLAGS_STRIPPED)
+string(REGEX REPLACE "[\r\n\t ]+" "\", \"" HOST_LD_FLAGS_ARRAY "${HOST_LD_FLAGS_STRIPPED}")
+# string(REPLACE "###, ###" " oo \", \" oo " HOST_LD_FLAGS_ARRAY "${HOST_LD_FLAGS_ARRAY_1}")
+
 set(HOST_CLANG_FLAGS "${DEFAULT_HOST_CLANG_FLAGS} ${EXTRA_HOST_CLANG_FLAGS}")
 set(HOST_LLC_FLAGS "${DEFAULT_HOST_LLC_FLAGS} ${EXTRA_HOST_LLC_FLAGS}")
 
@@ -974,6 +978,10 @@ set(PACKAGE_VERSION "${POCL_VERSION}")
 
 configure_file("config.h.in.cmake" "config.h.new" ESCAPE_QUOTES)
 rename_if_different("${CMAKE_BINARY_DIR}/config.h.new" "${CMAKE_BINARY_DIR}/config.h")
+
+configure_file("config2.h.in.cmake" "config2.h.new")
+rename_if_different("${CMAKE_BINARY_DIR}/config2.h.new" "${CMAKE_BINARY_DIR}/config2.h")
+
 include_directories("${CMAKE_BINARY_DIR}")
 
 # autotools compat

config2.h.in.cmake

@@ -0,0 +1,5 @@
+/* this config file is for values NOT escaped for C/C++
+ * required e.g. for values with doublequotes, like C string arrays */
+
+#define HOST_LD_FLAGS_ARRAY "@HOST_LD_FLAGS_ARRAY@"
+

lib/CL/devices/common.c

@@ -40,21 +40,23 @@
 #endif
 
 #include "config.h"
-#include "pocl_image_util.h"
-#include "pocl_file_util.h"
-#include "pocl_util.h"
-#include "pocl_cache.h"
+#include "config2.h"
 #include "devices.h"
+#include "pocl_cache.h"
+#include "pocl_debug.h"
+#include "pocl_file_util.h"
+#include "pocl_image_util.h"
 #include "pocl_mem_management.h"
 #include "pocl_runtime_config.h"
-#include "pocl_debug.h"
+#include "pocl_util.h"
 
 #ifdef OCS_AVAILABLE
 #include "pocl_llvm.h"
 #endif
 
 #include "_kernel_constants.h"
 
+
 #define COMMAND_LENGTH 2048
 
 /**
@@ -73,7 +75,6 @@ llvm_codegen (const char* tmpdir, cl_kernel kernel, cl_device_id device,
               size_t local_x, size_t local_y, size_t local_z)
 {
 
-  char command[COMMAND_LENGTH];
   char bytecode[POCL_FILENAME_LENGTH];
   char objfile[POCL_FILENAME_LENGTH];
   /* strlen of / .so 4+1 */
@@ -120,22 +121,28 @@ llvm_codegen (const char* tmpdir, cl_kernel kernel, cl_device_id device,
   assert (error == 0);
 
   /* clang is used as the linker driver in LINK_CMD */
-  error = snprintf (command, COMMAND_LENGTH,
+
+  POCL_MSG_PRINT_INFO ("Linking final module\n");
+  char *const args1[]
 #ifndef POCL_ANDROID
-                    LINK_COMMAND " " HOST_LD_FLAGS " -o %s %s",
+      = { LINK_COMMAND,
+          HOST_LD_FLAGS_ARRAY,
+          "-o",
+          tmp_module,
+          objfile,
+          NULL };
 #else
-                    POCL_ANDROID_PREFIX"/bin/ld " HOST_LD_FLAGS " -o %s %s ",
+      = { POCL_ANDROID_PREFIX "/bin/ld",
+          HOST_LD_FLAGS_ARRAY,
+          "-o",
+          tmp_module,
+          objfile,
+          NULL };
 #endif
-                    tmp_module, objfile);
-  assert (error >= 0);
-
-  POCL_MSG_PRINT_INFO ("executing [%s]\n", command);
-  error = system (command);
+  error = pocl_run_command (args1);
   assert (error == 0);
 
-  error = snprintf (command, COMMAND_LENGTH, "mv %s %s", tmp_module, module);
-  assert (error >= 0);
-  error = system (command);
+  error = pocl_rename (tmp_module, module);
   assert (error == 0);
 
   /* Save space in kernel cache */

lib/CL/devices/hsa/pocl-hsa.c

@@ -941,41 +941,6 @@ setup_kernel_args (pocl_hsa_device_data_t *d,
 #endif
 }
 
-/*
- * This replaces a simple system(), because system() was causing issues
- * (gpu lockups) when compiling code (via compile_parallel_bc_to_brig)
- * with OpenCL 2.0 atomics (like CalcPie from AMD SDK).
- * The reason of lockups is unknown (yet).
- */
-static int
-run_command(char* args[])
-{
-  POCL_MSG_PRINT_INFO("Launching: %s\n", args[0]);
-#ifdef HAVE_VFORK
-  pid_t p = vfork();
-#elif defined(HAVE_FORK)
-  pid_t p = fork();
-#else
-#error Must have fork() or vfork() system calls for HSA
-#endif
-  if (p == 0)
-    {
-      return execv(args[0], args);
-    }
-  else
-    {
-      if (p < 0)
-        return -1;
-      int status;
-      if (waitpid(p, &status, 0) < 0)
-        POCL_ABORT("pocl-hsa: waitpid() itself failed.\n");
-      if (WIFEXITED(status))
-        return WEXITSTATUS(status);
-      else
-        return -2;
-    }
-}
-
 static int
 compile_parallel_bc_to_brig(char* brigfile, cl_kernel kernel,
                             cl_device_id device) {
@@ -1005,14 +970,14 @@ compile_parallel_bc_to_brig(char* brigfile, cl_kernel kernel,
 
       char* args1[] = { LLVM_LLC, "-O2", "-march=hsail64", "-filetype=asm",
                         "-o", hsailfile, parallel_bc_path, NULL };
-      if ((error = run_command(args1)))
+      if ((error = pocl_run_command (args1)))
         {
           POCL_MSG_PRINT_INFO("pocl-hsa: llc exit status %i\n", error);
           return error;
         }
 
       char* args2[] = { HSAIL_ASM, "-o", brigfile, hsailfile, NULL };
-      if ((error = run_command(args2)))
+      if ((error = pocl_run_command (args2)))
         {
           POCL_MSG_PRINT_INFO("pocl-hsa: HSAILasm exit status %i\n", error);
           return error;

lib/CL/pocl_util.c

@@ -26,14 +26,18 @@
 #include <stdlib.h>
 #include <string.h>
 
-#include <sys/types.h>
-#include <sys/stat.h>
 #include <time.h>
 
 #ifndef _MSC_VER
-#  include <dirent.h>
-#  include <unistd.h>
-#  include <utime.h>
+#include <dirent.h>
+#include <sys/resource.h>
+#include <sys/stat.h>
+#include <sys/time.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+#include <unistd.h>
+#include <utime.h>
 #else
 #  include "vccompat.hpp"
 #endif
@@ -855,3 +859,48 @@ pocl_command_to_str (cl_command_type cmd)
 
   return "unknown";
 }
+
+/*
+ * This replaces a simple system(), because:
+ *
+ * 1) system() was causing issues (gpu lockups) with HSA when
+ * compiling code (via compile_parallel_bc_to_brig)
+ * with OpenCL 2.0 atomics (like CalcPie from AMD SDK).
+ * The reason of lockups is unknown (yet).
+ *
+ * 2) system() uses fork() which copies page table maps, and runs
+ * out of AS when pocl has already allocated huge buffers in memory.
+ * this happened in llvm_codegen()
+ *
+ * vfork() does not copy pagetables.
+ */
+int
+pocl_run_command (char *const *args)
+{
+  POCL_MSG_PRINT_INFO ("Launching: %s\n", args[0]);
+#ifdef HAVE_VFORK
+  pid_t p = vfork ();
+#elif defined(HAVE_FORK)
+  pid_t p = fork ();
+#else
+#error Must have fork() or vfork() system calls for HSA
+#endif
+  if (p == 0)
+    {
+      return execv (args[0], args);
+    }
+  else
+    {
+      if (p < 0)
+        return EXIT_FAILURE;
+      int status;
+      if (waitpid (p, &status, 0) < 0)
+        POCL_ABORT ("pocl: waitpid() failed.\n");
+      if (WIFEXITED (status))
+        return WEXITSTATUS (status);
+      else if (WIFSIGNALED (status))
+        return WTERMSIG (status);
+      else
+        return EXIT_FAILURE;
+    }
+}

lib/CL/pocl_util.h

@@ -151,6 +151,9 @@ pocl_status_to_str (int status);
 const char *
 pocl_command_to_str (cl_command_type cmd);
 
+int
+pocl_run_command(char * const *args);
+
 #ifdef __cplusplus
 }
 #endif