[SYCL][libclc] Remangler tool for consistent mangling

buildbot/configure.py

@@ -27,6 +27,7 @@ def do_configure(args):
     llvm_targets_to_build = 'X86'
     llvm_enable_projects = 'clang;' + llvm_external_projects
     libclc_targets_to_build = ''
+    libclc_gen_remangled_variants = 'OFF'
     sycl_build_pi_cuda = 'OFF'
     sycl_build_pi_esimd_cpu = 'ON'
     sycl_build_pi_rocm = 'OFF'
@@ -53,6 +54,7 @@ def do_configure(args):
     if args.cuda:
         llvm_targets_to_build += ';NVPTX'
         libclc_targets_to_build = 'nvptx64--;nvptx64--nvidiacl'
+        libclc_gen_remangled_variants = 'ON'
         sycl_build_pi_cuda = 'ON'
 
     if args.rocm:
@@ -62,6 +64,7 @@ def do_configure(args):
         elif args.rocm_platform == 'NVIDIA' and not args.cuda:
             llvm_targets_to_build += ';NVPTX'
             libclc_targets_to_build += ';nvptx64--;nvptx64--nvidiacl'
+        libclc_gen_remangled_variants = 'ON'
 
         sycl_build_pi_rocm_platform = args.rocm_platform
         sycl_build_pi_rocm = 'ON'
@@ -99,6 +102,7 @@ def do_configure(args):
         "-DLLVM_EXTERNAL_LIBDEVICE_SOURCE_DIR={}".format(libdevice_dir),
         "-DLLVM_ENABLE_PROJECTS={}".format(llvm_enable_projects),
         "-DLIBCLC_TARGETS_TO_BUILD={}".format(libclc_targets_to_build),
+        "-DLIBCLC_GENERATE_REMANGLED_VARIANTS={}".format(libclc_gen_remangled_variants),
         "-DSYCL_BUILD_PI_CUDA={}".format(sycl_build_pi_cuda),
         "-DSYCL_BUILD_PI_ROCM={}".format(sycl_build_pi_rocm),
         "-DSYCL_BUILD_PI_ROCM_PLATFORM={}".format(sycl_build_pi_rocm_platform),

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -722,7 +722,8 @@ void CudaToolChain::addClangTargetOptions(
       llvm::sys::path::append(WithInstallPath, Twine("../../../share/clc"));
       LibraryPaths.emplace_back(WithInstallPath.c_str());
 
-      std::string LibSpirvTargetName = "libspirv-nvptx64--nvidiacl.bc";
+      std::string LibSpirvTargetName =
+          "remangled-l64-signed_char.libspirv-nvptx64--nvidiacl.bc";
       for (StringRef LibraryPath : LibraryPaths) {
         SmallString<128> LibSpirvTargetFile(LibraryPath);
         llvm::sys::path::append(LibSpirvTargetFile, LibSpirvTargetName);

clang/lib/Driver/ToolChains/HIP.cpp

@@ -309,7 +309,8 @@ void HIPToolChain::addClangTargetOptions(
       llvm::sys::path::append(WithInstallPath, Twine("../../../share/clc"));
       LibraryPaths.emplace_back(WithInstallPath.c_str());
 
-      std::string LibSpirvTargetName = "libspirv-amdgcn--amdhsa.bc";
+      std::string LibSpirvTargetName =
+          "remangled-l64-signed_char.libspirv-amdgcn--amdhsa.bc";
       for (StringRef LibraryPath : LibraryPaths) {
         SmallString<128> LibSpirvTargetFile(LibraryPath);
         llvm::sys::path::append(LibSpirvTargetFile, LibSpirvTargetName);

clang/lib/Sema/SPIRVBuiltins.td

@@ -321,9 +321,8 @@ class ConstOCLSPVBuiltin<string _Name, list<Type> _Signature> :
 
 // OpenCL v1.0/1.2/2.0 s6.1.1: Built-in Scalar Data Types.
 def Bool      : IntType<"bool",   QualType<"Context.BoolTy">, 1>;
-def TrueChar  : IntType<"_char",      QualType<"Context.CharTy", 0, 1>, 8>;
-def Char      : IntType<"char",      QualType<"Context.SignedCharTy", 0, 1>, 8>;
-def SChar     : IntType<"schar",     QualType<"Context.SignedCharTy", 0, 1>, 8>;
+def Char      : IntType<"char",      QualType<"Context.CharTy", 0, 1>, 8>;
+def SChar     : IntType<"schar",      QualType<"Context.SignedCharTy", 0, 1>, 8>;
 def UChar     : UIntType<"uchar",     QualType<"Context.UnsignedCharTy">, 8>;
 def Short     : IntType<"short",     QualType<"Context.ShortTy", 0, 1>, 16>;
 def UShort    : UIntType<"ushort",    QualType<"Context.UnsignedShortTy">, 16>;
@@ -356,15 +355,13 @@ def Vec16       : IntList<"Vec16", [16]>;
 def Vec1234     : IntList<"Vec1234", [1, 2, 3, 4]>;
 
 // Type lists.
-def TLAll           : TypeList<[Char,  UChar, Short,  UShort, Int,  UInt, Long,  ULong, Float, Double, Half]>;
+def TLAll           : TypeList<[Char, UChar, Short,  UShort, Int,  UInt, Long,  ULong, Float, Double, Half]>;
 def TLAllUnsigned   : TypeList<[UChar, UChar, UShort, UShort, UInt, UInt, ULong, ULong, UInt,  ULong,  UShort]>;
 def TLFloat         : TypeList<[Float, Double, Half]>;
-// FIXME: handle properly char (signed or unsigned depending on host)
 def TLSignedInts    : TypeList<[Char, Short, Int, Long]>;
 def TLUnsignedInts  : TypeList<[UChar, UShort, UInt, ULong]>;
 
 // Signed to Unsigned conversion
-// FIXME: handle properly char (signed or unsigned depending on host)
 def TLSToUSignedInts    : TypeList<[Char, Short, Int, Long]>;
 def TLSToUUnsignedInts  : TypeList<[UChar, UShort, UInt, ULong]>;
 
@@ -375,7 +372,7 @@ def TLIntLongFloats : TypeList<[Int, UInt, Long, ULong, Float, Double, Half]>;
 // uchar abs(uchar).
 def TLAllUIntsTwice : TypeList<[UChar, UChar, UChar, UShort, UShort, UInt, UInt, ULong, ULong]>;
 
-def TLAllInts       : TypeList<[Char, UChar, Short, UShort, Int, UInt, Long, ULong]>;
+def TLAllInts       : TypeList<[Char, SChar, UChar, Short, UShort, Int, UInt, Long, ULong]>;
 
 // GenType definitions for multiple base types (e.g. all floating point types,
 // or all integer types).
@@ -402,7 +399,7 @@ def IntLongFloatGenType1   : GenericType<"IntLongFloatGenType1", TLIntLongFloats
 
 // GenType definitions for every single base type (e.g. fp32 only).
 // Names are like: GenTypeFloatVecAndScalar.
-foreach Type = [Char, UChar, SChar, Short, UShort,
+foreach Type = [Char, SChar, UChar, Short, UShort,
                 Int, UInt, Long, ULong,
                 Float, Double, Half] in {
   foreach VecSizes = [VecAndScalar, VecNoScalar] in {
@@ -557,6 +554,7 @@ foreach name = ["s_clamp", "s_mad_hi", "s_mad_sat"] in {
 
 foreach name = ["s_upsample"] in {
   def : ConstOCLSPVBuiltin<name, [GenTypeShortVecAndScalar, GenTypeCharVecAndScalar, GenTypeUCharVecAndScalar]>;
+  def : ConstOCLSPVBuiltin<name, [GenTypeShortVecAndScalar, GenTypeSCharVecAndScalar, GenTypeUCharVecAndScalar]>;
   def : ConstOCLSPVBuiltin<name, [GenTypeIntVecAndScalar, GenTypeShortVecAndScalar, GenTypeUShortVecAndScalar]>;
   def : ConstOCLSPVBuiltin<name, [GenTypeLongVecAndScalar, GenTypeIntVecAndScalar, GenTypeUIntVecAndScalar]>;
 }
@@ -724,7 +722,7 @@ foreach VSize1 = [Vec2, Vec4, Vec8, Vec16] in {
 
 let IsVariadic = 1 in {
   foreach name = ["printf"] in {
-    def : OCLSPVBuiltin<name, [Int, PointerType<ConstType<TrueChar>, ConstantAS>]>;
+    def : OCLSPVBuiltin<name, [Int, PointerType<ConstType<Char>, ConstantAS>]>;
   }
 }
 

libclc/CMakeLists.txt

@@ -65,6 +65,10 @@ set( LIBCLC_TARGETS_TO_BUILD "all"
 option( ENABLE_RUNTIME_SUBNORMAL "Enable runtime linking of subnormal support."
 OFF )
 
+option( LIBCLC_GENERATE_REMANGLED_VARIANTS
+				"Generate remangled variants of enabled libclc targets."
+				OFF )
+
 # mesa3d environment is only available since LLVM 4.0
 if( ${LLVM_VERSION} VERSION_GREATER "3.9.0" )
 	set( LIBCLC_TARGETS_ALL ${LIBCLC_TARGETS_ALL} amdgcn-mesa-mesa3d )
@@ -119,15 +123,19 @@ find_program( LLVM_AS llvm-as PATHS ${LLVM_BINDIR} NO_DEFAULT_PATH )
 find_program( LLVM_LINK llvm-link PATHS ${LLVM_BINDIR} NO_DEFAULT_PATH )
 find_program( LLVM_OPT opt PATHS ${LLVM_BINDIR} NO_DEFAULT_PATH )
 find_program( LLVM_SPIRV llvm-spirv PATHS ${LLVM_BINDIR} NO_DEFAULT_PATH )
+find_program( LIBCLC_REMANGLER libclc-remangler PATHS ${LLVM_BINDIR}
+							NO_DEFAULT_PATH )
 
 # Print toolchain
 message( "clang: ${LLVM_CLANG}" )
 message( "llvm-as: ${LLVM_AS}" )
 message( "llvm-link: ${LLVM_LINK}" )
 message( "opt: ${LLVM_OPT}" )
 message( "llvm-spirv: ${LLVM_SPIRV}" )
+message( "libclc-remangler: ${LIBCLC_REMANGLER}" )
 message( "" )
-if( NOT LLVM_CLANG OR NOT LLVM_OPT OR NOT LLVM_AS OR NOT LLVM_LINK )
+if( NOT LLVM_CLANG OR NOT LLVM_OPT OR NOT LLVM_AS OR NOT LLVM_LINK
+		OR NOT LIBCLC_REMANGLER )
 	message( FATAL_ERROR "toolchain incomplete!" )
 endif()
 

libclc/cmake/modules/AddLibclc.cmake

@@ -105,6 +105,40 @@ macro(add_libclc_builtin_set arch_suffix)
   install(
     FILES ${LIBCLC_LIBRARY_OUTPUT_INTDIR}/${obj_suffix}
     DESTINATION ${CMAKE_INSTALL_DATADIR}/clc )
+
+  # Generate remangled variants if requested
+  if( LIBCLC_GENERATE_REMANGLED_VARIANTS )
+    set(long_widths l32 l64)
+    set(char_signedness signed unsigned)
+    # All permutations of [l32, l64] and [signed, unsigned]
+    foreach(long_width ${long_widths})
+      foreach(signedness ${char_signedness})
+        # Remangle
+        set( builtins_remangle_path
+            "${LIBCLC_LIBRARY_OUTPUT_INTDIR}/remangled-${long_width}-${signedness}_char.${obj_suffix}" )
+        add_custom_command( OUTPUT "${builtins_remangle_path}"
+          COMMAND libclc-remangler
+          -o "${builtins_remangle_path}"
+          --long-width=${long_width}
+          --char-signedness=${signedness}
+          "$<TARGET_PROPERTY:prepare-${obj_suffix},TARGET_FILE>"
+          DEPENDS "prepare-${obj_suffix}" libclc-remangler )
+        add_custom_target( "remangled-${long_width}-${signedness}_char.${obj_suffix}" ALL
+          DEPENDS "${builtins_remangle_path}" )
+        set_target_properties("remangled-${long_width}-${signedness}_char.${obj_suffix}"
+          PROPERTIES TARGET_FILE "${builtins_remangle_path}")
+
+        # Add dependency to top-level pseudo target to ease making other
+        # targets dependent on libclc.
+        add_dependencies(${ARG_PARENT_TARGET} "remangled-${long_width}-${signedness}_char.${obj_suffix}")
+
+        # Keep remangled variants
+        install(
+          FILES ${builtins_remangle_path}
+          DESTINATION ${CMAKE_INSTALL_DATADIR}/clc )
+      endforeach()
+    endforeach()
+  endif()
 
   # nvptx-- targets don't include workitem builtins
   if( NOT ${t} MATCHES ".*ptx.*--$" )

libclc/cmake/modules/HandleInLLVMTree.cmake

@@ -3,6 +3,7 @@ macro(configure_in_llvm_tree)
   set(LLVM_AS ${LLVM_RUNTIME_OUTPUT_INTDIR}/llvm-as)
   set(LLVM_LINK ${LLVM_RUNTIME_OUTPUT_INTDIR}/llvm-link)
   set(LLVM_OPT ${LLVM_RUNTIME_OUTPUT_INTDIR}/opt)
+  set(LIBCLC_REMANGLER ${LLVM_RUNTIME_OUTPUT_INTDIR}/libclc-remangler)
 
   if (NOT EXISTS ${LLVM_RUNTIME_OUTPUT_INTDIR}/clang)
     file(WRITE ${LLVM_RUNTIME_OUTPUT_INTDIR}/clang "" )
@@ -16,6 +17,9 @@ macro(configure_in_llvm_tree)
   if (NOT EXISTS ${LLVM_RUNTIME_OUTPUT_INTDIR}/opt)
     file(WRITE ${LLVM_RUNTIME_OUTPUT_INTDIR}/opt "" )
   endif (NOT EXISTS ${LLVM_RUNTIME_OUTPUT_INTDIR}/opt)
+  if (NOT EXISTS ${LLVM_RUNTIME_OUTPUT_INTDIR}/libclc-remangler)
+    file(WRITE ${LLVM_RUNTIME_OUTPUT_INTDIR}/libclc-remangler "" )
+  endif (NOT EXISTS ${LLVM_RUNTIME_OUTPUT_INTDIR}/libclc-remangler)
 
   # Assume all works well
   # We can't test the compilers as they haven't been built yet

libclc/cmake/modules/HandleOutOfTreeLLVM.cmake

@@ -46,6 +46,7 @@ macro(configure_out_of_tree_llvm)
   get_property(LLVM_AS TARGET llvm-as PROPERTY LOCATION)
   get_property(LLVM_LINK TARGET llvm-link PROPERTY LOCATION)
   get_property(LLVM_OPT TARGET opt PROPERTY LOCATION)
+  get_property(LIBCLC_REMANGLER TARGET libclc-remangler PROPERTY LOCATION)
 
   set(LLVM_ENABLE_PIC OFF)
 

libclc/generic/gen_convert_common.py

@@ -1,10 +1,10 @@
 # This file contains common variables and helper functions used by the
 # `gen_convert.py` in both the libclc and libspirv libraries.
 
-types = ['char', 'schar', 'uchar', 'short', 'ushort', 'int', 'uint', 'long', 'ulong', 'half', 'float', 'double']
-int_types = ['char', 'schar', 'uchar', 'short', 'ushort', 'int', 'uint', 'long', 'ulong']
+types = ['char', 'uchar', 'short', 'ushort', 'int', 'uint', 'long', 'ulong', 'half', 'float', 'double']
+int_types = ['char', 'uchar', 'short', 'ushort', 'int', 'uint', 'long', 'ulong']
 unsigned_types = ['uchar', 'ushort', 'uint', 'ulong']
-signed_types = ['char', 'schar', 'short', 'int', 'long']
+signed_types = ['char', 'short', 'int', 'long']
 float_types = ['half', 'float', 'double']
 int64_types = ['long', 'ulong']
 float64_types = ['double']
@@ -18,8 +18,7 @@
 float_suffix = {'float':'f', 'double':''}
 
 bool_type = {'char'   : 'char',
-             'schar'  : 'schar',
-             'uchar'  : 'schar',
+             'uchar'  : 'char',
              'short'  : 'short',
              'ushort' : 'short',
              'int'    : 'int',
@@ -31,7 +30,6 @@
              'double' : 'long'}
 
 unsigned_type = {'char'  : 'uchar',
-                 'schar' : 'uchar',
                  'uchar' : 'uchar',
                  'short' : 'ushort',
                  'ushort': 'ushort',
@@ -40,15 +38,14 @@
                  'long'  : 'ulong',
                  'ulong' : 'ulong'}
 
-sizeof_type = {'char'  : 1, 'schar'  : 1, 'uchar'  : 1,
+sizeof_type = {'char'  : 1, 'uchar'  : 1,
                'short' : 2, 'ushort' : 2,
                'int'   : 4, 'uint'   : 4,
                'long'  : 8, 'ulong'  : 8,
                'half'  : 2, 'float'  : 4,
                'double': 8}
 
 limit_max = {'char'  : 'CHAR_MAX',
-             'schar' : 'CHAR_MAX',
              'uchar' : 'UCHAR_MAX',
              'short' : 'SHRT_MAX',
              'ushort': 'USHRT_MAX',
@@ -58,7 +55,6 @@
              'ulong' : 'ULONG_MAX'}
 
 limit_min = {'char'  : 'CHAR_MIN',
-             'schar' : 'CHAR_MIN',
              'uchar' : '0',
              'short' : 'SHRT_MIN',
              'ushort': '0',

libclc/generic/include/as_type.h

@@ -2,7 +2,6 @@
 #define CLC_AS_TYPE
 
 #define as_char(x) __builtin_astype(x, char)
-#define as_schar(x) __builtin_astype(x, schar)
 #define as_uchar(x) __builtin_astype(x, uchar)
 #define as_short(x) __builtin_astype(x, short)
 #define as_ushort(x) __builtin_astype(x, ushort)
@@ -13,7 +12,6 @@
 #define as_float(x) __builtin_astype(x, float)
 
 #define as_char2(x) __builtin_astype(x, char2)
-#define as_schar2(x) __builtin_astype(x, schar2)
 #define as_uchar2(x) __builtin_astype(x, uchar2)
 #define as_short2(x) __builtin_astype(x, short2)
 #define as_ushort2(x) __builtin_astype(x, ushort2)
@@ -24,7 +22,6 @@
 #define as_float2(x) __builtin_astype(x, float2)
 
 #define as_char3(x) __builtin_astype(x, char3)
-#define as_schar3(x) __builtin_astype(x, schar3)
 #define as_uchar3(x) __builtin_astype(x, uchar3)
 #define as_short3(x) __builtin_astype(x, short3)
 #define as_ushort3(x) __builtin_astype(x, ushort3)
@@ -35,7 +32,6 @@
 #define as_float3(x) __builtin_astype(x, float3)
 
 #define as_char4(x) __builtin_astype(x, char4)
-#define as_schar4(x) __builtin_astype(x, schar4)
 #define as_uchar4(x) __builtin_astype(x, uchar4)
 #define as_short4(x) __builtin_astype(x, short4)
 #define as_ushort4(x) __builtin_astype(x, ushort4)
@@ -46,7 +42,6 @@
 #define as_float4(x) __builtin_astype(x, float4)
 
 #define as_char8(x) __builtin_astype(x, char8)
-#define as_schar8(x) __builtin_astype(x, schar8)
 #define as_uchar8(x) __builtin_astype(x, uchar8)
 #define as_short8(x) __builtin_astype(x, short8)
 #define as_ushort8(x) __builtin_astype(x, ushort8)
@@ -57,7 +52,6 @@
 #define as_float8(x) __builtin_astype(x, float8)
 
 #define as_char16(x) __builtin_astype(x, char16)
-#define as_schar16(x) __builtin_astype(x, schar16)
 #define as_uchar16(x) __builtin_astype(x, uchar16)
 #define as_short16(x) __builtin_astype(x, short16)
 #define as_ushort16(x) __builtin_astype(x, ushort16)

libclc/generic/include/clc/async/gentype.inc

@@ -39,46 +39,6 @@
 #undef __CLC_GENTYPE_MANGLED
 #undef __CLC_GENTYPE
 
-#ifndef __CLC_NO_SCHAR
-#define __CLC_GENTYPE schar
-#define __CLC_GENTYPE_MANGLED a
-#include __CLC_BODY
-#undef __CLC_GENTYPE_MANGLED
-#undef __CLC_GENTYPE
-
-#define __CLC_GENTYPE schar2
-#define __CLC_GENTYPE_MANGLED Dv2_a
-#include __CLC_BODY
-#undef __CLC_GENTYPE_MANGLED
-#undef __CLC_GENTYPE
-
-#ifdef __CLC_GEN_VEC3
-#define __CLC_GENTYPE schar3
-#define __CLC_GENTYPE_MANGLED Dv3_a
-#include __CLC_BODY
-#undef __CLC_GENTYPE_MANGLED
-#undef __CLC_GENTYPE
-#endif
-
-#define __CLC_GENTYPE schar4
-#define __CLC_GENTYPE_MANGLED Dv4_a
-#include __CLC_BODY
-#undef __CLC_GENTYPE_MANGLED
-#undef __CLC_GENTYPE
-
-#define __CLC_GENTYPE schar8
-#define __CLC_GENTYPE_MANGLED Dv8_a
-#include __CLC_BODY
-#undef __CLC_GENTYPE_MANGLED
-#undef __CLC_GENTYPE
-
-#define __CLC_GENTYPE schar16
-#define __CLC_GENTYPE_MANGLED Dv16_a
-#include __CLC_BODY
-#undef __CLC_GENTYPE_MANGLED
-#undef __CLC_GENTYPE
-#endif
-
 #define __CLC_GENTYPE uchar
 #define __CLC_GENTYPE_MANGLED h
 #include __CLC_BODY

libclc/generic/include/clc/clc.h

@@ -12,8 +12,6 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 #endif
 
-#define __CLC_NO_SCHAR
-
 /* Function Attributes */
 #include <func.h>