[CodeGen][NewPM] Plug greedy RA in codegen pipeline

Author: optimisan

llvm/include/llvm/Passes/CodeGenPassBuilder.h

@@ -1062,7 +1062,9 @@ void CodeGenPassBuilder<Derived, TargetMachineT>::addMachineSSAOptimization(
 ///
 /// A target that uses the standard regalloc pass order for fast or optimized
 /// allocation may still override this for per-target regalloc
-/// selection. But -regalloc=... always takes precedence.
+/// selection. But -regalloc-npm=... always takes precedence.
+/// If a target does not want to allow users to set -regalloc-npm=... at all,
+/// check if Opt.RegAlloc == RegAllocType::Unset.
 template <typename Derived, typename TargetMachineT>
 void CodeGenPassBuilder<Derived, TargetMachineT>::addTargetRegisterAllocator(
     AddMachinePass &addPass, bool Optimized) const {
@@ -1075,10 +1077,29 @@ void CodeGenPassBuilder<Derived, TargetMachineT>::addTargetRegisterAllocator(
 /// Find and instantiate the register allocation pass requested by this target
 /// at the current optimization level.  Different register allocators are
 /// defined as separate passes because they may require different analysis.
+///
+/// This helper ensures that the -regalloc-npm= option is always available,
+/// even for targets that override the default allocator.
 template <typename Derived, typename TargetMachineT>
 void CodeGenPassBuilder<Derived, TargetMachineT>::addRegAllocPass(
     AddMachinePass &addPass, bool Optimized) const {
-  // TODO: Parse Opt.RegAlloc to add register allocator.
+  // Use the specified -regalloc-npm={basic|greedy|fast|pbqp}
+  if (Opt.RegAlloc > RegAllocType::Default) {
+    switch (Opt.RegAlloc) {
+    case RegAllocType::Fast:
+      addPass(RegAllocFastPass());
+      break;
+    case RegAllocType::Greedy:
+      addPass(RAGreedyPass());
+      break;
+    default:
+      report_fatal_error("register allocator not supported yet.", false);
+    }
+    return;
+  }
+  // -regalloc=default or unspecified, so pick based on the optimization level
+  // or ask the target for the regalloc pass.
+  derived().addTargetRegisterAllocator(addPass, Optimized);
 }
 
 template <typename Derived, typename TargetMachineT>
@@ -1149,20 +1170,22 @@ void CodeGenPassBuilder<Derived, TargetMachineT>::addOptimizedRegAlloc(
   // PreRA instruction scheduling.
   addPass(MachineSchedulerPass(&TM));
 
-  if (derived().addRegAssignmentOptimized(addPass)) {
-    // Allow targets to expand pseudo instructions depending on the choice of
-    // registers before MachineCopyPropagation.
-    derived().addPostRewrite(addPass);
+  if (auto E = derived().addRegAssignmentOptimized(addPass)) {
+    // addRegAssignmentOptimized did not add a reg alloc pass, so do nothing.
+    return;
+  }
+  // Allow targets to expand pseudo instructions depending on the choice of
+  // registers before MachineCopyPropagation.
+  derived().addPostRewrite(addPass);
 
-    // Copy propagate to forward register uses and try to eliminate COPYs that
-    // were not coalesced.
-    addPass(MachineCopyPropagationPass());
+  // Copy propagate to forward register uses and try to eliminate COPYs that
+  // were not coalesced.
+  addPass(MachineCopyPropagationPass());
 
-    // Run post-ra machine LICM to hoist reloads / remats.
-    //
-    // FIXME: can this move into MachineLateOptimization?
-    addPass(MachineLICMPass());
-  }
+  // Run post-ra machine LICM to hoist reloads / remats.
+  //
+  // FIXME: can this move into MachineLateOptimization?
+  addPass(MachineLICMPass());
 }
 
 //===---------------------------------------------------------------------===//

llvm/include/llvm/Passes/MachinePassRegistry.def

@@ -195,12 +195,12 @@ MACHINE_FUNCTION_PASS_WITH_PARAMS(
     },
     "filter=reg-filter;no-clear-vregs")
 
+// 'all' is the default filter
 MACHINE_FUNCTION_PASS_WITH_PARAMS(
     "greedy", "RAGreedyPass",
     [](RAGreedyPass::Options Opts) { return RAGreedyPass(Opts); },
     [PB = this](StringRef Params) {
-      // TODO: parseRegAllocGreedyFilterFunc(*PB, Params);
-      return Expected<RAGreedyPass::Options>(RAGreedyPass::Options{});
+      return parseRegAllocGreedyFilterFunc(*PB, Params);
     }, "reg-filter"
 )
 #undef MACHINE_FUNCTION_PASS_WITH_PARAMS

llvm/include/llvm/Target/CGPassBuilderOption.h

@@ -20,7 +20,7 @@
 namespace llvm {
 
 enum class RunOutliner { TargetDefault, AlwaysOutline, NeverOutline };
-enum class RegAllocType { Default, Basic, Fast, Greedy, PBQP };
+enum class RegAllocType { Unset, Default, Basic, Fast, Greedy, PBQP };
 
 // Not one-on-one but mostly corresponding to commandline options in
 // TargetPassConfig.cpp.
@@ -52,7 +52,7 @@ struct CGPassBuilderOption {
   bool RequiresCodeGenSCCOrder = false;
 
   RunOutliner EnableMachineOutliner = RunOutliner::TargetDefault;
-  StringRef RegAlloc = "default";
+  RegAllocType RegAlloc = RegAllocType::Unset;
   std::optional<GlobalISelAbortMode> EnableGlobalISelAbort;
   std::string FSProfileFile;
   std::string FSRemappingFile;

llvm/lib/Passes/PassBuilder.cpp

@@ -1414,6 +1414,20 @@ parseBoundsCheckingOptions(StringRef Params) {
   return Options;
 }
 
+Expected<RAGreedyPass::Options>
+parseRegAllocGreedyFilterFunc(PassBuilder &PB, StringRef Params) {
+  if (Params.empty() || Params == "all") {
+    return RAGreedyPass::Options();
+  }
+  std::optional<RegAllocFilterFunc> Filter = PB.parseRegAllocFilter(Params);
+  if (!Filter) {
+    return make_error<StringError>(
+        formatv("invalid regallocgreedy register filter '{0}' ", Params).str(),
+        inconvertibleErrorCode());
+  }
+  return RAGreedyPass::Options{*Filter, Params};
+}
+
 } // namespace
 
 /// Tests whether a pass name starts with a valid prefix for a default pipeline

llvm/test/CodeGen/AArch64/implicit-def-remat-requires-impdef-check.mir

@@ -1,5 +1,6 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 4
 # RUN: llc -mtriple=arm64-apple-macosx -mcpu=apple-m1 -stress-regalloc=4 -verify-regalloc -run-pass=greedy -o - %s | FileCheck %s
+# RUN: llc -mtriple=arm64-apple-macosx -mcpu=apple-m1 -stress-regalloc=4 -verify-regalloc -passes=greedy -o - %s | FileCheck %s
 
 --- |
   define void @inst_stores_to_dead_spill_implicit_def_impdef() {

llvm/test/CodeGen/AArch64/implicit-def-with-impdef-greedy-assert.mir

@@ -1,5 +1,6 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 3
 # RUN: llc -mtriple=arm64-apple-ios -run-pass=greedy -o - %s | FileCheck %s
+# RUN: llc -mtriple=arm64-apple-ios -passes=greedy -o - %s | FileCheck %s
 
 ---
 name:            widget

llvm/test/CodeGen/AArch64/pr51516.mir

@@ -1,4 +1,5 @@
 # RUN: llc -mtriple=aarch64-unknown-fuchsia -run-pass=greedy -verify-machineinstrs -o - %s | FileCheck %s
+# RUN: llc -mtriple=aarch64-unknown-fuchsia -passes=greedy -verify-machineinstrs -o - %s | FileCheck %s
 
 # Check that we spill %31 and do not rematerialize it since the use operand
 # of ADDXri is killed by the STRXui in this block.

llvm/test/CodeGen/AArch64/spill-fold.mir

@@ -1,5 +1,7 @@
 # RUN: llc -mtriple=aarch64-none-linux-gnu -run-pass greedy -verify-machineinstrs  -o - %s | FileCheck %s
 # RUN: llc -mtriple=aarch64_be-none-linux-gnu -run-pass greedy -verify-machineinstrs  -o - %s | FileCheck %s
+# RUN: llc -mtriple=aarch64-none-linux-gnu -passes=greedy -o - %s | FileCheck %s
+# RUN: llc -mtriple=aarch64_be-none-linux-gnu -passes=greedy -o - %s | FileCheck %s
 --- |
   define i64 @test_subreg_spill_fold() { ret i64 0 }
   define i64 @test_subreg_spill_fold2() { ret i64 0 }

llvm/test/CodeGen/MIR/Generic/runPass.mir

@@ -2,6 +2,7 @@
 # RUN: llc -run-pass=regallocbasic -debug-pass=Arguments -o - %s | FileCheck %s
 # RUN: llc -run-pass=regallocfast -debug-pass=Arguments -o - %s | FileCheck %s
 # RUN: llc -passes=regallocfast -o - %s | FileCheck %s
+# RUN: llc -passes=greedy -o - %s | FileCheck %s
 
 # Check that passes are initialized correctly, so that it's possible to
 # use -run-pass.

llvm/test/CodeGen/SystemZ/clear-liverange-spillreg.mir

@@ -1,4 +1,5 @@
 #RUN: llc -o - %s -mtriple=s390x-ibm-linux -run-pass=greedy
+#RUN: llc -o - %s -mtriple=s390x-ibm-linux -passes=greedy
 #PR34502. Check HoistSpill works properly after the live range of spilled
 #virtual register is cleared.
 --- |

llvm/test/CodeGen/Thumb/high-reg-clobber.mir

@@ -3,6 +3,7 @@
 # RUN: llc -mtriple thumbv6m-arm-none-eabi -run-pass greedy %s -o - | FileCheck %s
 # RUN: llc -mtriple thumbv6m-arm-none-eabi -run-pass regallocfast %s -o - | FileCheck %s --check-prefix=FAST
 # RUN: llc -mtriple thumbv6m-arm-none-eabi -passes=regallocfast %s -o - | FileCheck %s --check-prefix=FAST
+# RUN: llc -mtriple thumbv6m-arm-none-eabi -passes=greedy %s -o - | FileCheck %s
 
 ...
 ---

llvm/test/CodeGen/X86/limit-split-cost.mir

@@ -1,5 +1,6 @@
 # REQUIRES: asserts
 # RUN: llc -mtriple=x86_64-- -run-pass=greedy %s -debug-only=regalloc -huge-size-for-split=0 -o /dev/null 2>&1 | FileCheck %s
+# RUN: llc -mtriple=x86_64-- -passes=greedy %s -debug-only=regalloc -huge-size-for-split=0 -o /dev/null 2>&1 | FileCheck %s
 # Check no global region split is needed because the live range to split is trivially rematerializable.
 # CHECK-NOT: Compact region bundles
 --- |

llvm/tools/llc/NewPMDriver.cpp

@@ -48,10 +48,17 @@
 
 using namespace llvm;
 
-static cl::opt<std::string>
-    RegAlloc("regalloc-npm",
-             cl::desc("Register allocator to use for new pass manager"),
-             cl::Hidden, cl::init("default"));
+static cl::opt<RegAllocType> RegAlloc(
+    "regalloc-npm", cl::desc("Register allocator to use for new pass manager"),
+    cl::Hidden, cl::init(RegAllocType::Unset),
+    cl::values(
+        clEnumValN(RegAllocType::Default, "default",
+                   "Default register allocator"),
+        clEnumValN(RegAllocType::PBQP, "pbqp", "PBQP register allocator"),
+        clEnumValN(RegAllocType::Fast, "fast", "Fast register allocator"),
+        clEnumValN(RegAllocType::Basic, "basic", "Basic register allocator"),
+        clEnumValN(RegAllocType::Greedy, "greedy",
+                   "Greedy register allocator")));
 
 static cl::opt<bool>
     DebugPM("debug-pass-manager", cl::Hidden,