[Driver][SYCL] Add support for large device code via a linker script (#6584)

Add -fsycl-huge-device-code as a driver option. When enabled, the driver
generates a linker script which directs the linker to place device code
later in the binary, which makes it less likely to create a distance
larger than that which a PC32 relocation can span.

For example, if a 3GB __clang_offload_bundle__ section is placed betwen
.text and .rodata, a PC32 relocation to .rodata in the .text is not
possible. With this option enabled the device code will be placed after
both rather than between them.

Author: tcreech-intel

clang/include/clang/Driver/Options.td

@@ -2867,6 +2867,12 @@ defm sycl_instrument_device_code
           BothFlags<[CC1Option, CoreOption], " Instrumentation and Tracing "
                     "Technology (ITT) instrumentation intrinsics calls "
                     "(experimental)">>;
+def fsycl_link_huge_device_code : Flag<["-"], "fsycl-link-huge-device-code">,
+  Group<sycl_Group>, HelpText<"Generate and use a custom linker script for huge"
+  " device code sections">;
+def fno_sycl_link_huge_device_code : Flag<["-"], "fno-sycl-link-huge-device-code">,
+  Group<sycl_Group>, HelpText<"Do not generate or use a custom linker script"
+  " for huge device code sections (default)">;
 defm sycl_id_queries_fit_in_int: BoolFOption<"sycl-id-queries-fit-in-int",
   LangOpts<"SYCLValueFitInMaxInt">, DefaultTrue,
   PosFlag<SetTrue, [], "Assume">, NegFlag<SetFalse, [], "Do not assume">,

clang/lib/Driver/ToolChains/Gnu.cpp

@@ -660,6 +660,56 @@ void tools::gnutools::Linker::ConstructJob(Compilation &C, const JobAction &JA,
     CmdArgs.push_back("-lm");
   }
 
+  // If requested, use a custom linker script to handle very large device code
+  // sections.
+  if (Args.hasArg(options::OPT_fsycl) &&
+      Args.hasFlag(options::OPT_fsycl_link_huge_device_code,
+                   options::OPT_fno_sycl_link_huge_device_code, false)) {
+    // Create temporary linker script. Keep it if save-temps is enabled.
+    const char *LKS;
+    SmallString<256> Name = llvm::sys::path::filename(Output.getFilename());
+    if (C.getDriver().isSaveTempsEnabled()) {
+      llvm::sys::path::replace_extension(Name, "ld");
+      LKS = C.getArgs().MakeArgString(Name.c_str());
+    } else {
+      llvm::sys::path::replace_extension(Name, "");
+      Name = C.getDriver().GetTemporaryPath(Name, "ld");
+      LKS = C.addTempFile(C.getArgs().MakeArgString(Name.c_str()));
+    }
+
+    // Add linker script option to the command.
+    CmdArgs.push_back("-T");
+    CmdArgs.push_back(LKS);
+
+    // If this is not a dry run, create the linker script file.
+    if (!C.getArgs().hasArg(options::OPT__HASH_HASH_HASH)) {
+      std::error_code EC;
+      llvm::raw_fd_ostream ScriptOS(LKS, EC, llvm::sys::fs::OF_None);
+
+      if (EC) {
+        C.getDriver().Diag(clang::diag::err_unable_to_make_temp)
+            << EC.message();
+      } else {
+        ScriptOS
+            << "/*\n"
+               " * This linker script allows huge (>3GB) device code\n"
+               " * sections. It has been auto-generated by the SYCL driver.\n"
+               " */\n"
+               "SECTIONS\n"
+               "{\n"
+               "  . = SEGMENT_START(\"sycl-device-code\", .);\n"
+               "  SYCL_DEVICE_CODE ALIGN(CONSTANT (MAXPAGESIZE)) + (. & "
+               "(CONSTANT (MAXPAGESIZE) - 1)) :\n"
+               "  {\n"
+               "    *(__CLANG_OFFLOAD_BUNDLE__*)\n"
+               "  }\n"
+               "}\n"
+               "INSERT AFTER .bss\n";
+        ScriptOS.close();
+      }
+    }
+  }
+
   if (!Args.hasArg(options::OPT_nostdlib, options::OPT_r)) {
     if (!Args.hasArg(options::OPT_nodefaultlibs)) {
       if (IsStatic || IsStaticPIE)

clang/test/Driver/sycl-huge-device-code-linker-script.cpp

@@ -0,0 +1,12 @@
+// UNSUPPORTED: system-windows
+// Test the addition of a custom linker script for huge device code.
+
+// RUN: %clangxx -### -fsycl -fsycl-link-huge-device-code %s 2>&1 | \
+// RUN:       FileCheck --check-prefix=CHECK-LINKER-SCRIPT %s
+// CHECK-LINKER-SCRIPT: "-T" "{{.*}}.ld"
+
+// Also check that a user-provided linker script may be used:
+// RUN: %clangxx -### -fsycl -fsycl-link-huge-device-code %s \
+// RUN:       -T custom-user-script.ld 2>&1 | \
+// RUN:       FileCheck --check-prefixes=CHECK-USER-SCRIPT %s
+// CHECK-USER-SCRIPT: "-T" "custom-user-script.ld"

sycl/doc/UsersManual.md

@@ -217,6 +217,14 @@ and not recommended to use in production environment.
     various events inside JIT generated kernels. These device libraries are
     linked in by default.
 
+**`-f[no-]sycl-link-huge-device-code`**
+
+    Place device code later in the linked binary in order to avoid precluding
+    32-bit PC relative relocations between surrounding ELF sections when device
+    code is larger than 2GiB. This is disabled by default.
+
+    NOTE: This option is currently only supported on Linux.
+
 ## Intel FPGA specific options
 
 **`-fintelfpga`**