Merge pull request #215 from andrjohns/tbb-winarm64

Support building TBB on Windows arm64

Author: kevinushey

.Rbuildignore

@@ -17,3 +17,4 @@
 ^tests/testthat/pkg/RcppParallelTest/src/.*\.s?o$
 ^tools/tbb$
 ^\.github$
+^patches

R/tbb.R

@@ -49,8 +49,14 @@ tbbCxxFlags <- function() {
    flags <- character()
 
    # opt-in to TBB on Windows
-   if (is_windows())
+   if (is_windows()) {
       flags <- c(flags, "-DRCPP_PARALLEL_USE_TBB=1")
+      if (R.version$arch == "aarch64") {
+         # TBB does not have assembly code for Windows ARM64
+         # so we need to use compiler builtins
+         flags <- c(flags, "-DTBB_USE_GCC_BUILTINS")
+      }
+   }
 
    # if TBB_INC is set, apply those library paths
    tbbInc <- Sys.getenv("TBB_INC", unset = TBB_INC)
@@ -81,10 +87,6 @@ tbbLdFlags <- function() {
       return(sprintf(fmt, asBuildPath(tbbLib)))
    }
 
-   # on Aarch64 builds, use the version of TBB provided by Rtools
-   if (is_windows() && R.version$arch == "aarch64")
-      return("-ltbb12 -ltbbmalloc")
-   
    # on Mac, Windows and Solaris, we need to explicitly link (#206)
    needsExplicitFlags <- is_mac() || is_windows() || (is_solaris() && !is_sparc())
    if (needsExplicitFlags) {

R/zzz.R

@@ -27,12 +27,8 @@ loadTbbLibrary <- function(name) {
 
 .onLoad <- function(libname, pkgname) {
 
-   tbbLibraryName <- "tbb"
-   if (is_windows() && R.version$arch == "aarch64")
-      tbbLibraryName <- "tbb12"
-   
    # load tbb, tbbmalloc
-   .tbbDllInfo       <<- loadTbbLibrary(tbbLibraryName)
+   .tbbDllInfo       <<- loadTbbLibrary("tbb")
    .tbbMallocDllInfo <<- loadTbbLibrary("tbbmalloc")
 
    # load tbbmalloc_proxy, but only if requested

inst/include/RcppParallel.h

@@ -18,6 +18,9 @@
 #endif
 
 #if RCPP_PARALLEL_USE_TBB
+#if defined(WINNT) && defined(__aarch64__) && !defined(TBB_USE_GCC_BUILTINS)
+#define TBB_USE_GCC_BUILTINS 1
+#endif
 # include "RcppParallel/TBB.h"
 #endif
 

patches/windows_arm64.diff

@@ -0,0 +1,47 @@
+diff --git a/src/tbb/build/Makefile.tbb b/src/tbb/build/Makefile.tbb
+index 8d155f80..c58f4fb1 100644
+--- a/src/tbb/build/Makefile.tbb
++++ b/src/tbb/build/Makefile.tbb
+@@ -91,7 +91,11 @@ ifneq (,$(TBB.DEF))
+ tbb.def: $(TBB.DEF) $(TBB.LST)
+ 	$(CPLUS) $(PREPROC_ONLY) $< $(CPLUS_FLAGS) $(INCLUDES) > $@
+ 
+-LIB_LINK_FLAGS += $(EXPORT_KEY)tbb.def
++# LLVM on Windows doesn't need --version-script export
++# https://reviews.llvm.org/D63743
++ifeq (, $(WINARM64_CLANG))
++  LIB_LINK_FLAGS += $(EXPORT_KEY)tbb.def
++endif
+ $(TBB.DLL): tbb.def
+ endif
+ 
+diff --git a/src/tbb/build/Makefile.tbbmalloc b/src/tbb/build/Makefile.tbbmalloc
+index 421e95c5..e7c38fa4 100644
+--- a/src/tbb/build/Makefile.tbbmalloc
++++ b/src/tbb/build/Makefile.tbbmalloc
+@@ -74,7 +74,11 @@ ifneq (,$(MALLOC.DEF))
+ tbbmalloc.def: $(MALLOC.DEF)
+ 	$(CPLUS) $(PREPROC_ONLY) $< $(M_CPLUS_FLAGS) $(WARNING_SUPPRESS) $(INCLUDES) > $@
+ 
+-MALLOC_LINK_FLAGS += $(EXPORT_KEY)tbbmalloc.def
++# LLVM on Windows doesn't need --version-script export
++# https://reviews.llvm.org/D63743
++ifeq (, $(WINARM64_CLANG))
++	MALLOC_LINK_FLAGS += $(EXPORT_KEY)tbbmalloc.def
++endif
+ $(MALLOC.DLL): tbbmalloc.def
+ endif
+ 
+diff --git a/src/tbb/src/tbbmalloc/TypeDefinitions.h b/src/tbb/src/tbbmalloc/TypeDefinitions.h
+index 3178442e..fd4b7956 100644
+--- a/src/tbb/src/tbbmalloc/TypeDefinitions.h
++++ b/src/tbb/src/tbbmalloc/TypeDefinitions.h
+@@ -25,7 +25,7 @@
+ #       define __ARCH_ipf 1
+ #   elif defined(_M_IX86)||defined(__i386__) // the latter for MinGW support
+ #       define __ARCH_x86_32 1
+-#   elif defined(_M_ARM)
++#   elif defined(_M_ARM) || defined(__aarch64__)
+ #       define __ARCH_other 1
+ #   else
+ #       error Unknown processor architecture for Windows

src/Makevars.in

@@ -21,16 +21,28 @@ ifeq ($(OS), Windows_NT)
 	USE_TBB=Windows
 	TBB_COPY_PATTERN=tbb*.dll
 
+	ARCH=$(shell "${R_HOME}/bin/R" --vanilla -s -e 'cat(R.version$$arch)')
+	TBB_CXXFLAGS = @CXX11FLAGS@ -DTBB_NO_LEGACY=1
+	ifeq "$(ARCH)" "aarch64"
+		PKG_CPPFLAGS += -DTBB_USE_GCC_BUILTINS
+		TBB_CXXFLAGS += -DTBB_USE_GCC_BUILTINS
+		CLANG_CHECK := $(shell echo | $(CC) -E -dM -  | findstr __clang__)
+		ifneq ($(CLANG_CHECK), )
+			WINARM64_CLANG=true
+		endif
+	endif
+
 	MAKE = make
 	MAKEFLAGS = -e -j1
 	MAKE_CMD =                                    \
 		MSYS2_ARG_CONV_EXCL="*"                   \
 		CYGWIN=nodosfilewarning                   \
 		CONLY="@WINDOWS_CC@"                      \
 		CPLUS="@WINDOWS_CXX11@"                   \
-		CXXFLAGS="@CXX11FLAGS@ -DTBB_NO_LEGACY=1" \
+		CXXFLAGS="$(TBB_CXXFLAGS)" 				  \
 		PIC_KEY="@CXX11PICFLAGS@"                 \
 		WARNING_SUPPRESS=""                       \
+		WINARM64_CLANG="$(WINARM64_CLANG)"        \
 		$(MAKE)
 
 else
@@ -77,26 +89,21 @@ ifeq ($(USE_TBB), Windows)
 	# rtools: turn on hacks to compensate for make and shell differences rtools<=>MinGW
 	# compiler: overwrite default (which is cl = MS compiler)
 	MAKE_ARGS += rtools=true compiler=gcc
-	ifeq ($(WIN), 64)
-		MAKE_ARGS += arch=intel64 runtime=mingw
-		ARCH_DIR=x64/
-	else
-		MAKE_ARGS += arch=ia32 runtime=mingw
-		ARCH_DIR=i386/
+	# TBB configure will detect mingw runtime with unknown arch on WINARM64_CLANG but not an 
+	#   issue as we are using compiler built-ins instead of arch-specific code
+	ifneq ($(WINARM64_CLANG), true)
+		ifeq ($(WIN), 64)
+			MAKE_ARGS += arch=intel64 runtime=mingw
+			ARCH_DIR=x64/
+		else
+			MAKE_ARGS += arch=ia32 runtime=mingw
+			ARCH_DIR=i386/
+		endif
 	endif
 
 	# Linker needs access to the tbb dll; otherwise you get errors such as:
 	# "undefined reference to `tbb::task_scheduler_init::terminate()'"
 	PKG_LIBS += -Ltbb/build/lib_release -ltbb -ltbbmalloc
-	
-	# override for aarch64 (experimental) to use Rtools TBB
-	ARCH=$(shell "${R_HOME}/bin/R" --vanilla -s -e 'cat(R.version$$arch)')
-	ifeq "$(ARCH)" "aarch64"
-		TBB_LIB = ${R_TOOLS_SOFT}
-		TBB_INC = ${R_TOOLS_SOFT}
-		PKG_LIBS = -ltbb12 -ltbbmalloc
-	endif
-
 endif
 
 # write compiler if set

src/tbb/build/Makefile.tbb

@@ -91,7 +91,11 @@ ifneq (,$(TBB.DEF))
 tbb.def: $(TBB.DEF) $(TBB.LST)
 	$(CPLUS) $(PREPROC_ONLY) $< $(CPLUS_FLAGS) $(INCLUDES) > $@
 
-LIB_LINK_FLAGS += $(EXPORT_KEY)tbb.def
+# LLVM on Windows doesn't need --version-script export
+# https://reviews.llvm.org/D63743
+ifeq (, $(WINARM64_CLANG))
+  LIB_LINK_FLAGS += $(EXPORT_KEY)tbb.def
+endif
 $(TBB.DLL): tbb.def
 endif
 

src/tbb/build/Makefile.tbbmalloc

@@ -74,7 +74,11 @@ ifneq (,$(MALLOC.DEF))
 tbbmalloc.def: $(MALLOC.DEF)
 	$(CPLUS) $(PREPROC_ONLY) $< $(M_CPLUS_FLAGS) $(WARNING_SUPPRESS) $(INCLUDES) > $@
 
-MALLOC_LINK_FLAGS += $(EXPORT_KEY)tbbmalloc.def
+# LLVM on Windows doesn't need --version-script export
+# https://reviews.llvm.org/D63743
+ifeq (, $(WINARM64_CLANG))
+	MALLOC_LINK_FLAGS += $(EXPORT_KEY)tbbmalloc.def
+endif
 $(MALLOC.DLL): tbbmalloc.def
 endif
 

src/tbb/src/tbbmalloc/TypeDefinitions.h

@@ -25,7 +25,7 @@
 #       define __ARCH_ipf 1
 #   elif defined(_M_IX86)||defined(__i386__) // the latter for MinGW support
 #       define __ARCH_x86_32 1
-#   elif defined(_M_ARM)
+#   elif defined(_M_ARM) || defined(__aarch64__)
 #       define __ARCH_other 1
 #   else
 #       error Unknown processor architecture for Windows