[RISCV] Implement Statepoint and Patchpoint lowering to call instructions (#77337)

This patch adds stackmap support for RISC-V with call targets.

Based on patch from https://reviews.llvm.org/D129848.

Author: Zeavee

llvm/lib/Target/RISCV/RISCVAsmPrinter.cpp

@@ -14,6 +14,7 @@
 #include "MCTargetDesc/RISCVBaseInfo.h"
 #include "MCTargetDesc/RISCVInstPrinter.h"
 #include "MCTargetDesc/RISCVMCExpr.h"
+#include "MCTargetDesc/RISCVMatInt.h"
 #include "MCTargetDesc/RISCVTargetStreamer.h"
 #include "RISCV.h"
 #include "RISCVMachineFunctionInfo.h"
@@ -153,8 +154,35 @@ void RISCVAsmPrinter::LowerPATCHPOINT(MCStreamer &OutStreamer, StackMaps &SM,
 
   PatchPointOpers Opers(&MI);
 
+  const MachineOperand &CalleeMO = Opers.getCallTarget();
   unsigned EncodedBytes = 0;
 
+  if (CalleeMO.isImm()) {
+    uint64_t CallTarget = CalleeMO.getImm();
+    if (CallTarget) {
+      assert((CallTarget & 0xFFFF'FFFF'FFFF) == CallTarget &&
+             "High 16 bits of call target should be zero.");
+      // Materialize the jump address:
+      SmallVector<MCInst, 8> Seq;
+      RISCVMatInt::generateMCInstSeq(CallTarget, *STI, RISCV::X1, Seq);
+      for (MCInst &Inst : Seq) {
+        bool Compressed = EmitToStreamer(OutStreamer, Inst);
+        EncodedBytes += Compressed ? 2 : 4;
+      }
+      bool Compressed = EmitToStreamer(OutStreamer, MCInstBuilder(RISCV::JALR)
+                                                        .addReg(RISCV::X1)
+                                                        .addReg(RISCV::X1)
+                                                        .addImm(0));
+      EncodedBytes += Compressed ? 2 : 4;
+    }
+  } else if (CalleeMO.isGlobal()) {
+    MCOperand CallTargetMCOp;
+    lowerOperand(CalleeMO, CallTargetMCOp);
+    EmitToStreamer(OutStreamer,
+                   MCInstBuilder(RISCV::PseudoCALL).addOperand(CallTargetMCOp));
+    EncodedBytes += 8;
+  }
+
   // Emit padding.
   unsigned NumBytes = Opers.getNumPatchBytes();
   assert(NumBytes >= EncodedBytes &&
@@ -173,6 +201,35 @@ void RISCVAsmPrinter::LowerSTATEPOINT(MCStreamer &OutStreamer, StackMaps &SM,
     assert(PatchBytes % NOPBytes == 0 &&
            "Invalid number of NOP bytes requested!");
     emitNops(PatchBytes / NOPBytes);
+  } else {
+    // Lower call target and choose correct opcode
+    const MachineOperand &CallTarget = SOpers.getCallTarget();
+    MCOperand CallTargetMCOp;
+    switch (CallTarget.getType()) {
+    case MachineOperand::MO_GlobalAddress:
+    case MachineOperand::MO_ExternalSymbol:
+      lowerOperand(CallTarget, CallTargetMCOp);
+      EmitToStreamer(
+          OutStreamer,
+          MCInstBuilder(RISCV::PseudoCALL).addOperand(CallTargetMCOp));
+      break;
+    case MachineOperand::MO_Immediate:
+      CallTargetMCOp = MCOperand::createImm(CallTarget.getImm());
+      EmitToStreamer(OutStreamer, MCInstBuilder(RISCV::JAL)
+                                      .addReg(RISCV::X1)
+                                      .addOperand(CallTargetMCOp));
+      break;
+    case MachineOperand::MO_Register:
+      CallTargetMCOp = MCOperand::createReg(CallTarget.getReg());
+      EmitToStreamer(OutStreamer, MCInstBuilder(RISCV::JALR)
+                                      .addReg(RISCV::X1)
+                                      .addOperand(CallTargetMCOp)
+                                      .addImm(0));
+      break;
+    default:
+      llvm_unreachable("Unsupported operand type in statepoint call target");
+      break;
+    }
   }
 
   auto &Ctx = OutStreamer.getContext();

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -17910,6 +17910,17 @@ RISCVTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
   case RISCV::PseudoFROUND_D_IN32X:
     return emitFROUND(MI, BB, Subtarget);
   case TargetOpcode::STATEPOINT:
+    // STATEPOINT is a pseudo instruction which has no implicit defs/uses
+    // while jal call instruction (where statepoint will be lowered at the end)
+    // has implicit def. This def is early-clobber as it will be set at
+    // the moment of the call and earlier than any use is read.
+    // Add this implicit dead def here as a workaround.
+    MI.addOperand(*MI.getMF(),
+                  MachineOperand::CreateReg(
+                      RISCV::X1, /*isDef*/ true,
+                      /*isImp*/ true, /*isKill*/ false, /*isDead*/ true,
+                      /*isUndef*/ false, /*isEarlyClobber*/ true));
+    [[fallthrough]];
   case TargetOpcode::STACKMAP:
   case TargetOpcode::PATCHPOINT:
     if (!Subtarget.is64Bit())

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

@@ -1469,9 +1469,12 @@ unsigned RISCVInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
   case TargetOpcode::PATCHPOINT:
     // The size of the patchpoint intrinsic is the number of bytes requested
     return PatchPointOpers(&MI).getNumPatchBytes();
-  case TargetOpcode::STATEPOINT:
+  case TargetOpcode::STATEPOINT: {
     // The size of the statepoint intrinsic is the number of bytes requested
-    return StatepointOpers(&MI).getNumPatchBytes();
+    unsigned NumBytes = StatepointOpers(&MI).getNumPatchBytes();
+    // No patch bytes means at most a PseudoCall is emitted
+    return std::max(NumBytes, 8U);
+  }
   default:
     return get(Opcode).getSize();
   }

llvm/test/CodeGen/RISCV/rv64-patchpoint.ll

@@ -1,12 +1,56 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=riscv64 -debug-entry-values -enable-misched=0 < %s | FileCheck %s
 
+; Trivial patchpoint codegen
+;
+define i64 @trivial_patchpoint_codegen(i64 %p1, i64 %p2, i64 %p3, i64 %p4) {
+; CHECK-LABEL: trivial_patchpoint_codegen:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    addi sp, sp, -16
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    sd s0, 8(sp) # 8-byte Folded Spill
+; CHECK-NEXT:    sd s1, 0(sp) # 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_offset s0, -8
+; CHECK-NEXT:    .cfi_offset s1, -16
+; CHECK-NEXT:    mv s0, a0
+; CHECK-NEXT:  .Ltmp0:
+; CHECK-NEXT:    lui ra, 3563
+; CHECK-NEXT:    addiw ra, ra, -577
+; CHECK-NEXT:    slli ra, ra, 12
+; CHECK-NEXT:    addi ra, ra, -259
+; CHECK-NEXT:    slli ra, ra, 12
+; CHECK-NEXT:    addi ra, ra, -1282
+; CHECK-NEXT:    jalr ra
+; CHECK-NEXT:    mv s1, a0
+; CHECK-NEXT:    mv a0, s0
+; CHECK-NEXT:    mv a1, s1
+; CHECK-NEXT:  .Ltmp1:
+; CHECK-NEXT:    lui ra, 3563
+; CHECK-NEXT:    addiw ra, ra, -577
+; CHECK-NEXT:    slli ra, ra, 12
+; CHECK-NEXT:    addi ra, ra, -259
+; CHECK-NEXT:    slli ra, ra, 12
+; CHECK-NEXT:    addi ra, ra, -1281
+; CHECK-NEXT:    jalr ra
+; CHECK-NEXT:    mv a0, s1
+; CHECK-NEXT:    ld s0, 8(sp) # 8-byte Folded Reload
+; CHECK-NEXT:    ld s1, 0(sp) # 8-byte Folded Reload
+; CHECK-NEXT:    addi sp, sp, 16
+; CHECK-NEXT:    ret
+entry:
+  %resolveCall2 = inttoptr i64 244837814094590 to i8*
+  %result = tail call i64 (i64, i32, i8*, i32, ...) @llvm.experimental.patchpoint.i64(i64 2, i32 28, i8* %resolveCall2, i32 4, i64 %p1, i64 %p2, i64 %p3, i64 %p4)
+  %resolveCall3 = inttoptr i64 244837814094591 to i8*
+  tail call void (i64, i32, i8*, i32, ...) @llvm.experimental.patchpoint.void(i64 3, i32 28, i8* %resolveCall3, i32 2, i64 %p1, i64 %result)
+  ret i64 %result
+}
+
 ; Test small patchpoints that don't emit calls.
 define void @small_patchpoint_codegen(i64 %p1, i64 %p2, i64 %p3, i64 %p4) {
 ; CHECK-LABEL: small_patchpoint_codegen:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    .cfi_def_cfa_offset 0
-; CHECK-NEXT:  .Ltmp0:
+; CHECK-NEXT:  .Ltmp2:
 ; CHECK-NEXT:    nop
 ; CHECK-NEXT:    nop
 ; CHECK-NEXT:    nop

llvm/test/CodeGen/RISCV/rv64-statepoint-call-lowering-x1.ll

@@ -0,0 +1,16 @@
+; RUN: llc -mtriple riscv64 -verify-machineinstrs -stop-after=prologepilog < %s | FileCheck %s
+
+; Check that STATEPOINT instruction has an early clobber implicit def for LR.
+target datalayout = "e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "riscv64"
+
+define void @test() "frame-pointer"="all" gc "statepoint-example" {
+entry:
+  %safepoint_token = tail call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) @return_i1, i32 0, i32 0, i32 0, i32 0) ["gc-live" ()]
+; CHECK: STATEPOINT 0, 0, 0, target-flags(riscv-call) @return_i1, 2, 0, 2, 0, 2, 0, 2, 0, 2, 0, 2, 0, csr_ilp32_lp64, implicit-def $x2, implicit-def dead early-clobber $x1
+  ret void
+}
+
+
+declare void @return_i1()
+declare token @llvm.experimental.gc.statepoint.p0(i64, i32, ptr, i32, i32, ...)

llvm/test/CodeGen/RISCV/rv64-statepoint-call-lowering-x2.ll

@@ -0,0 +1,23 @@
+; RUN: llc -mtriple riscv64 -verify-machineinstrs -stop-after=prologepilog < %s | FileCheck %s
+
+; Check that STATEPOINT instruction prefer to use x2 in presense of x8.
+target datalayout = "e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "riscv64"
+
+declare void @consume(ptr addrspace(1) %obj)
+
+define i1 @test(ptr addrspace(1) %a) "frame-pointer"="all" gc "statepoint-example" {
+entry:
+  %safepoint_token = tail call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(i1 ()) @return_i1, i32 0, i32 0, i32 0, i32 0) ["gc-live" (ptr addrspace(1) %a)]
+; CHECK: STATEPOINT 0, 0, 0, target-flags(riscv-call) @return_i1, 2, 0, 2, 0, 2, 0, 2, 1, 1, 8, $x8, -32, 2, 0, 2, 1, 0, 0
+  %call1 = call ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token,  i32 0, i32 0)
+  %call2 = call zeroext i1 @llvm.experimental.gc.result.i1(token %safepoint_token)
+  call void @consume(ptr addrspace(1) %call1)
+  ret i1 %call2
+}
+
+
+declare i1 @return_i1()
+declare token @llvm.experimental.gc.statepoint.p0(i64, i32, ptr, i32, i32, ...)
+declare ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token, i32, i32)
+declare i1 @llvm.experimental.gc.result.i1(token)