[LTO] Add an option to enable fast-isel

[qcolombet]

Hi Davide,

Shouldn’t fast isel be an attribute inherited from each individual compiler invocation to produce the bitcode files?
I.e., shouldn’t it be a function attribute?

Note: I am not very familiar of how LTO is supposed to work so you can just ignore my remarks :).

Cheers,
-Quentin

[• rafael]

Is this something you expect to be generally useful or something that is
only needed to run some experiments.

Is there an equivalent option in the gold plugin or in lib lto?

[davide]

In D19845#420177, @rafael wrote:

Is this something you expect to be generally useful or something that is
only needed to run some experiments.

Is there an equivalent option in the gold plugin or in lib lto?

There's no equivalent. Yes, I'm using it to run experiments in pair with paralellCodeGen.
I can keep this locally if you like.

[mehdi_amini]

Isn't there a cl::opt that forces to use fast-isel?

ELF/Config.h

struct Configuration {
   bool DiscardLocals;
   bool DiscardNone;
   bool EhFrameHdr;
   bool EnableNewDtags;
   bool ExportDynamic;
   bool GcSections;
   bool GnuHash = false;
   bool ICF;
+  bool LtoFastIsel;
   bool Mips64EL = false;
   bool NoGnuUnique;
   bool NoUndefined;
   bool NoinhibitExec;
   bool Pic;
   bool Pie;
   bool PrintGcSections;
   bool Rela;

ELF/Driver.cpp

void LinkerDriver::readConfigs(opt::InputArgList &Args) {
   Config->DiscardAll = Args.hasArg(OPT_discard_all);
   Config->DiscardLocals = Args.hasArg(OPT_discard_locals);
   Config->DiscardNone = Args.hasArg(OPT_discard_none);
   Config->EhFrameHdr = Args.hasArg(OPT_eh_frame_hdr);
   Config->EnableNewDtags = !Args.hasArg(OPT_disable_new_dtags);
   Config->ExportDynamic = Args.hasArg(OPT_export_dynamic);
   Config->GcSections = Args.hasArg(OPT_gc_sections);
   Config->ICF = Args.hasArg(OPT_icf);
+  Config->LtoFastIsel = Args.hasArg(OPT_lto_fast_isel);
   Config->NoGnuUnique = Args.hasArg(OPT_no_gnu_unique);
   Config->NoUndefined = Args.hasArg(OPT_no_undefined);
   Config->NoinhibitExec = Args.hasArg(OPT_noinhibit_exec);
   Config->Pie = Args.hasArg(OPT_pie);
   Config->PrintGcSections = Args.hasArg(OPT_print_gc_sections);
   Config->Relocatable = Args.hasArg(OPT_relocatable);
   Config->SaveTemps = Args.hasArg(OPT_save_temps);
   Config->Shared = Args.hasArg(OPT_shared);

ELF/LTO.cpp

std::vector<std::unique_ptr<InputFile>> BitcodeCompiler::compile() {
   std::string Msg;
   const Target *T = TargetRegistry::lookupTarget(TheTriple, Msg);
   if (!T)
     fatal("target not found: " + Msg);
   TargetOptions Options = InitTargetOptionsFromCodeGenFlags();
   Reloc::Model R = Config->Pic ? Reloc::PIC_ : Reloc::Static;
 
   auto CreateTargetMachine = [&]() {
-    return std::unique_ptr<TargetMachine>(
+    auto TM = std::unique_ptr<TargetMachine>(
         T->createTargetMachine(TheTriple, "", "", Options, R));
+    TM->setFastISel(Config->LtoFastIsel);
+    return TM;
   };
 
   std::unique_ptr<TargetMachine> TM = CreateTargetMachine();
   runLTOPasses(*Combined, *TM);
 
   return runSplitCodegen(CreateTargetMachine);
 }

ELF/Options.td

 
 // Aliases for ignored options
 def alias_define_common_d : Flag<["-"], "d">, Alias<define_common>;
 def alias_define_common_dc : Flag<["-"], "dc">, Alias<define_common>;
 def alias_define_common_dp : Flag<["-"], "dp">, Alias<define_common>;
 def alias_version_script_version_script : Joined<["--"], "version-script=">, Alias<version_script>;
 
 // LTO-related options.
+def lto_fast_isel : Flag<["--"], "lto-fast-isel">;
 def lto_jobs : Joined<["--"], "lto-jobs=">,
   HelpText<"Number of threads to run codegen">;
 def disable_verify : Flag<["-"], "disable-verify">;
 def mllvm : Separate<["-"], "mllvm">;
 def save_temps : Flag<["-"], "save-temps">;

test/ELF/lto/fast-isel.ll

+; Test the -lto-fast-isel option.
+
+; REQUIRES: x86
+; RUN: rm -f %t.so.lto.bc %t.so.lto.opt.bc %t.so.lto.o
+; RUN: llvm-as %s -o %t.o
+; RUN: ld.lld -m elf_x86_64 %t.o -o %t.so -save-temps --lto-fast-isel -shared
+; RUN: llvm-objdump -d %t.so | FileCheck %s
+; RUN: ld.lld -m elf_x86_64 %t.o -o %t2.so -save-temps -shared
+; RUN: llvm-objdump -d %t2.so | FileCheck %s --check-prefix=NOISEL
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-none-linux"
+
+define i32* @foo(i32* %p, i32* %q, i32** %z) {
+entry:
+  %r = load i32, i32* %p
+  %s = load i32, i32* %q
+  %y = load i32*, i32** %z
+
+  %t0 = add i32 %r, %s
+  %t1 = add i32 %t0, 1
+  %t2 = getelementptr i32, i32* %y, i32 1
+  %t3 = getelementptr i32, i32* %t2, i32 %t1
+
+  ret i32* %t3
+}
+
+; CHECK:      addl    $1, %eax
+; CHECK-NEXT: addq    $4, %rcx
+; CHECK-NEXT: cltq
+; CHECK-NEXT: shlq    $2, %rax
+; CHECK-NEXT: addq    %rcx, %rax
+; CHECK-NEXT: retq
+
+; NOISEL:      leal    1(%rax,%rcx), %eax
+; NOISEL-NEXT: cltq
+; NOISEL-NEXT: leaq    4(%rdx,%rax,4), %rax
+; NOISEL-NEXT: retq