Support the LLVM backend in HiPE

Add flags to enable and use the LLVM backend:
  * to_llvm: use the LLVM pipeline for compilation (default optimization level
  is O3),
  * llvm_save_temps: save the intermediate files in current directory in order
  to be able to debug or optimize the LLVM assembly,
  * {to_llvm, O}: set the optimization level of LLVM opt and llc tools.

Add some debug support to the loader; no semantic change intented.

4 files changed, 87 insertions, 25 deletions

diff --git a/lib/hipe/main/hipe.app.src b/lib/hipe/main/hipe.app.src
index bcdfcb0e03..a63c261719 100644
--- a/lib/hipe/main/hipe.app.src
+++ b/lib/hipe/main/hipe.app.src
@@ -30,6 +30,7 @@
 	     cerl_prettypr,
 	     cerl_to_icode,
 	     cerl_typean,
+	     elf_format,
 	     erl_bif_types,
 	     erl_types,
 	     hipe,
@@ -108,6 +109,10 @@
 	     hipe_ig,
 	     hipe_ig_moves,
 	     hipe_jit,
+	     hipe_llvm,
+	     hipe_llvm_liveness,
+	     hipe_llvm_main,
+	     hipe_llvm_merge,
 	     hipe_ls_regalloc,
 	     hipe_main,
 	     hipe_moves,
@@ -159,6 +164,7 @@
 	     hipe_rtl_symbolic,
 	     hipe_rtl_to_amd64,
 	     hipe_rtl_to_arm,
+	     hipe_rtl_to_llvm,
 	     hipe_rtl_to_ppc,
 	     hipe_rtl_to_sparc,
 	     hipe_rtl_to_x86,
diff --git a/lib/hipe/main/hipe.erl b/lib/hipe/main/hipe.erl
index 434d5c3061..d47eced6d8 100644
--- a/lib/hipe/main/hipe.erl
+++ b/lib/hipe/main/hipe.erl
@@ -821,7 +821,9 @@ finalize_fun_sequential({MFA, Icode}, Opts, Servers) ->
 		   ?debug_msg("Compiled ~w in ~.2f s\n", [MFA,(T2-T1)/1000])),
       {MFA, Code};
     {rtl, LinearRtl} ->
-      {MFA, LinearRtl}
+      {MFA, LinearRtl};
+    {llvm_binary, Binary} ->
+      {MFA, Binary}
   catch
     error:Error ->
       ?when_option(verbose, Opts, ?debug_untagged_msg("\n", [])),
@@ -890,21 +892,27 @@ do_load(Mod, Bin, BeamBinOrPath) when is_binary(BeamBinOrPath);
   end.
 
 assemble(CompiledCode, Closures, Exports, Options) ->
-  case get(hipe_target_arch) of
-    ultrasparc ->
-      hipe_sparc_assemble:assemble(CompiledCode, Closures, Exports, Options);
-    powerpc ->
-      hipe_ppc_assemble:assemble(CompiledCode, Closures, Exports, Options);
-    ppc64 ->
-      hipe_ppc_assemble:assemble(CompiledCode, Closures, Exports, Options);
-    arm ->
-      hipe_arm_assemble:assemble(CompiledCode, Closures, Exports, Options);
-    x86 ->
-      hipe_x86_assemble:assemble(CompiledCode, Closures, Exports, Options);
-    amd64 ->
-      hipe_amd64_assemble:assemble(CompiledCode, Closures, Exports, Options);
-    Arch ->
-      ?EXIT({executing_on_an_unsupported_architecture, Arch})
+  case proplists:get_bool(to_llvm, Options) of
+    false ->
+      case get(hipe_target_arch) of
+        ultrasparc ->
+          hipe_sparc_assemble:assemble(CompiledCode, Closures, Exports, Options);
+        powerpc ->
+          hipe_ppc_assemble:assemble(CompiledCode, Closures, Exports, Options);
+        ppc64 ->
+          hipe_ppc_assemble:assemble(CompiledCode, Closures, Exports, Options);
+        arm ->
+          hipe_arm_assemble:assemble(CompiledCode, Closures, Exports, Options);
+        x86 ->
+          hipe_x86_assemble:assemble(CompiledCode, Closures, Exports, Options);
+        amd64 ->
+          hipe_amd64_assemble:assemble(CompiledCode, Closures, Exports, Options);
+        Arch ->
+          ?EXIT({executing_on_an_unsupported_architecture, Arch})
+      end;
+    true ->
+      %% Merge already compiled code (per MFA) to a single binary.
+      hipe_llvm_merge:finalize(CompiledCode, Closures, Exports)
   end.
 
 %% --------------------------------------------------------------------
@@ -1330,6 +1338,11 @@ opt_keys() ->
      timeregalloc,
      timers,
      to_rtl,
+     to_llvm, % Use the LLVM backend for compilation.
+     llvm_save_temps, % Save the LLVM intermediate files in the current
+                      % directory; useful for debugging.
+     llvm_llc, % Specify llc optimization-level: o1, o2, o3, undefined.
+     llvm_opt, % Specify opt optimization-level: o1, o2, o3, undefined.
      use_indexing,
      use_inline_atom_search,
      use_callgraph,
@@ -1468,11 +1481,19 @@ opt_expansions() ->
   [{o1, o1_opts()},
    {o2, o2_opts()},
    {o3, o3_opts()},
+   {to_llvm, llvm_opts(o3)},
+   {{to_llvm, o0}, llvm_opts(o0)},
+   {{to_llvm, o1}, llvm_opts(o1)},
+   {{to_llvm, o2}, llvm_opts(o2)},
+   {{to_llvm, o3}, llvm_opts(o3)},
    {x87, [x87, inline_fp]},
    {inline_fp, case get(hipe_target_arch) of %% XXX: Temporary until x86
 		 x86 -> [x87, inline_fp];    %%       has sse2
 		 _ -> [inline_fp] end}].
 
+llvm_opts(O) ->
+  [to_llvm, {llvm_opt, O}, {llvm_llc, O}].
+
 %% This expands "basic" options, which may be tested early and cannot be
 %% in conflict with options found in the source code.
 
diff --git a/lib/hipe/main/hipe_main.erl b/lib/hipe/main/hipe_main.erl
index 99028cc3c1..89b79998be 100644
--- a/lib/hipe/main/hipe_main.erl
+++ b/lib/hipe/main/hipe_main.erl
@@ -49,7 +49,7 @@
 %%=====================================================================
 
 -type comp_icode_ret() :: {'native',hipe_architecture(),{'unprofiled',_}}
-			| {'rtl',tuple()}.
+			| {'rtl',tuple()} | {'llvm_binary',term()}.
 
 %%=====================================================================
 
@@ -115,11 +115,18 @@ compile_icode(MFA, LinearIcode0, Options, Servers, DebugState) ->
   pp(IcodeCfg7, MFA, icode_liveness, pp_icode_liveness, Options, Servers),
   FinalIcode = hipe_icode_cfg:cfg_to_linear(IcodeCfg7),
   ?opt_stop_timer("Icode"),
-  LinearRTL = ?option_time(icode_to_rtl(MFA,FinalIcode,Options, Servers), 
-			   "RTL", Options),
+  {LinearRTL, Roots} = ?option_time(icode_to_rtl(MFA, FinalIcode, Options, Servers),
+          "RTL", Options),
   case proplists:get_bool(to_rtl, Options) of
     false ->
-      rtl_to_native(MFA, LinearRTL, Options, DebugState);
+      case proplists:get_bool(to_llvm, Options) of
+        false ->
+          rtl_to_native(MFA, LinearRTL, Options, DebugState);
+        true ->
+          %% The LLVM backend returns binary code, unlike the rest of the HiPE
+          %% backends which return native assembly.
+          rtl_to_llvm_to_binary(MFA, LinearRTL, Roots, Options, DebugState)
+      end;
     true ->
       put(hipe_debug, DebugState),
       {rtl, LinearRTL}
@@ -385,11 +392,21 @@ icode_to_rtl(MFA, Icode, Options, Servers) ->
   %% hipe_rtl_cfg:pp(RtlCfg3),
   pp(RtlCfg3, MFA, rtl_liveness, pp_rtl_liveness, Options, Servers),
   RtlCfg4 = rtl_lcm(RtlCfg3, Options),
-  pp(RtlCfg4, MFA, rtl, pp_rtl, Options, Servers),
-  LinearRTL1 = hipe_rtl_cfg:linearize(RtlCfg4),
+  %% LLVM: A liveness analysis on RTL must be performed in order to find the GC
+  %% roots and explicitly mark them (in RTL) when they go out of scope (only
+  %% when the LLVM backend is used).
+  {RtlCfg5, Roots} =
+    case proplists:get_bool(to_llvm, Options) of
+      false ->
+        {RtlCfg4, []};
+      true ->
+        hipe_llvm_liveness:analyze(RtlCfg4)
+    end,
+  pp(RtlCfg5, MFA, rtl, pp_rtl, Options, Servers),
+  LinearRTL1 = hipe_rtl_cfg:linearize(RtlCfg5),
   LinearRTL2 = hipe_rtl_cleanup_const:cleanup(LinearRTL1),
   %% hipe_rtl:pp(standard_io, LinearRTL2),
-  LinearRTL2.
+  {LinearRTL2, Roots}.
 
 translate_to_rtl(Icode, Options) ->
   %% GC tests should have been added in the conversion to Icode.
@@ -540,6 +557,17 @@ rtl_to_native(MFA, LinearRTL, Options, DebugState) ->
   put(hipe_debug, DebugState),
   LinearNativeCode.
 
+%% Translate Linear RTL to binary code using LLVM.
+rtl_to_llvm_to_binary(MFA, LinearRTL, Roots, Options, DebugState) ->
+  ?opt_start_timer("LLVM native code"),
+  %% BinaryCode is a tuple, as defined in llvm/hipe_llvm_main module, which
+  %% contains the binary code together with info needed by the loader, e.g.
+  %% ConstTab, Refs, LabelMap, etc.
+  BinaryCode = hipe_llvm_main:rtl_to_native(MFA, LinearRTL, Roots, Options),
+  ?opt_stop_timer("LLVM native code"),
+  put(hipe_debug, DebugState),
+  {llvm_binary, BinaryCode}.
+
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
 %% Debugging stuff ...
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
diff --git a/lib/kernel/src/hipe_unified_loader.erl b/lib/kernel/src/hipe_unified_loader.erl
index 976d5e35cb..e5928c7b63 100644
--- a/lib/kernel/src/hipe_unified_loader.erl
+++ b/lib/kernel/src/hipe_unified_loader.erl
@@ -194,6 +194,13 @@ load_common(Mod, Bin, Beam, OldReferencesToPatch) ->
    CodeSize,  CodeBinary,  Refs,
    0,[] % ColdSize, CRrefs
   ] = binary_to_term(Bin),
+  ?debug_msg("***** ErLLVM *****~nVersion: ~s~nCheckSum: ~w~nConstAlign: ~w~n" ++
+    "ConstSize: ~w~nConstMap: ~w~nLabelMap: ~w~nExportMap ~w~nRefs ~w~n",
+    [Version, CheckSum, ConstAlign, ConstSize, ConstMap, LabelMap, ExportMap,
+      Refs]),
+  %% Write HiPE binary code to a file in the current directory in order to
+  %% debug by disassembling.
+  %% file:write_file("erl.o", CodeBinary, [binary]),
   %% Check that we are loading up-to-date code.
   version_check(Version, Mod),
   case hipe_bifs:check_crc(CheckSum) of
@@ -221,6 +228,7 @@ load_common(Mod, Bin, Beam, OldReferencesToPatch) ->
       {MFAs,Addresses} = exports(ExportMap, CodeAddress),
       %% Remove references to old versions of the module.
       ReferencesToPatch = get_refs_from(MFAs, []),
+      %% io:format("References to patch: ~w~n", [ReferencesToPatch]),
       ok = remove_refs_from(MFAs),
       %% Patch all dynamic references in the code.
       %%  Function calls, Atoms, Constants, System calls
@@ -246,8 +254,7 @@ load_common(Mod, Bin, Beam, OldReferencesToPatch) ->
 	  AddressesOfClosuresToPatch =
 	    calculate_addresses(ClosurePatches, CodeAddress, Addresses),
 	  export_funs(Addresses),
-	  export_funs(Mod, BeamBinary, Addresses, AddressesOfClosuresToPatch),
-	  ok
+	  export_funs(Mod, BeamBinary, Addresses, AddressesOfClosuresToPatch)
       end,
       %% Redirect references to the old module to the new module's BEAM stub.
       patch_to_emu_step2(OldReferencesToPatch),