[GlobalISel] convergence control tokens and intrinsics (#67006)

[GlobalISel] Implement convergence control tokens and intrinsics in GMIR

In the IR translator, convert the LLVM token type to LLT::token(), which is an
alias for the s0 type. These show up as implicit uses on convergent operations.

Differential Revision: https://reviews.llvm.org/D158147

Author: ssahasra

llvm/include/llvm/CodeGen/GlobalISel/CallLowering.h

@@ -117,6 +117,9 @@ class CallLowering {
     /// vreg that the swifterror should be copied into after the call.
     Register SwiftErrorVReg;
 
+    /// Valid if the call is a controlled convergent operation.
+    Register ConvergenceCtrlToken;
+
     /// Original IR callsite corresponding to this call, if available.
     const CallBase *CB = nullptr;
 
@@ -584,6 +587,7 @@ class CallLowering {
   bool lowerCall(MachineIRBuilder &MIRBuilder, const CallBase &Call,
                  ArrayRef<Register> ResRegs,
                  ArrayRef<ArrayRef<Register>> ArgRegs, Register SwiftErrorVReg,
+                 Register ConvergenceCtrlToken,
                  std::function<unsigned()> GetCalleeReg) const;
 
   /// For targets which want to use big-endian can enable it with

llvm/include/llvm/CodeGen/GlobalISel/IRTranslator.h

@@ -587,6 +587,10 @@ class IRTranslator : public MachineFunctionPass {
     return false;
   }
 
+  bool translateConvergenceControlIntrinsic(const CallInst &CI,
+                                            Intrinsic::ID ID,
+                                            MachineIRBuilder &MIRBuilder);
+
   /// @}
 
   // Builder for machine instruction a la IRBuilder.
@@ -705,6 +709,23 @@ class IRTranslator : public MachineFunctionPass {
     return Regs[0];
   }
 
+  Register getOrCreateConvergenceTokenVReg(const Value &Token) {
+    assert(Token.getType()->isTokenTy());
+    auto &Regs = *VMap.getVRegs(Token);
+    if (!Regs.empty()) {
+      assert(Regs.size() == 1 &&
+             "Expected a single register for convergence tokens.");
+      return Regs[0];
+    }
+
+    auto Reg = MRI->createGenericVirtualRegister(LLT::token());
+    Regs.push_back(Reg);
+    auto &Offsets = *VMap.getOffsets(Token);
+    if (Offsets.empty())
+      Offsets.push_back(0);
+    return Reg;
+  }
+
   /// Allocate some vregs and offsets in the VMap. Then populate just the
   /// offsets while leaving the vregs empty.
   ValueToVRegInfo::VRegListT &allocateVRegs(const Value &Val);

llvm/lib/CodeGen/GlobalISel/CallLowering.cpp

@@ -21,6 +21,7 @@
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/TargetLowering.h"
 #include "llvm/IR/DataLayout.h"
+#include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/IR/Module.h"
 #include "llvm/Target/TargetMachine.h"
@@ -91,6 +92,7 @@ bool CallLowering::lowerCall(MachineIRBuilder &MIRBuilder, const CallBase &CB,
                              ArrayRef<Register> ResRegs,
                              ArrayRef<ArrayRef<Register>> ArgRegs,
                              Register SwiftErrorVReg,
+                             Register ConvergenceCtrlToken,
                              std::function<unsigned()> GetCalleeReg) const {
   CallLoweringInfo Info;
   const DataLayout &DL = MIRBuilder.getDataLayout();
@@ -121,7 +123,6 @@ bool CallLowering::lowerCall(MachineIRBuilder &MIRBuilder, const CallBase &CB,
     CanBeTailCalled = false;
   }
 
-
   // First step is to marshall all the function's parameters into the correct
   // physregs and memory locations. Gather the sequence of argument types that
   // we'll pass to the assigner function.
@@ -187,6 +188,7 @@ bool CallLowering::lowerCall(MachineIRBuilder &MIRBuilder, const CallBase &CB,
   Info.KnownCallees = CB.getMetadata(LLVMContext::MD_callees);
   Info.CallConv = CallConv;
   Info.SwiftErrorVReg = SwiftErrorVReg;
+  Info.ConvergenceCtrlToken = ConvergenceCtrlToken;
   Info.IsMustTailCall = CB.isMustTailCall();
   Info.IsTailCall = CanBeTailCalled;
   Info.IsVarArg = IsVarArg;

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

@@ -214,8 +214,9 @@ ArrayRef<Register> IRTranslator::getOrCreateVRegs(const Value &Val) {
   auto *VRegs = VMap.getVRegs(Val);
   auto *Offsets = VMap.getOffsets(Val);
 
-  assert(Val.getType()->isSized() &&
-         "Don't know how to create an empty vreg");
+  if (!Val.getType()->isTokenTy())
+    assert(Val.getType()->isSized() &&
+           "Don't know how to create an empty vreg");
 
   SmallVector<LLT, 4> SplitTys;
   computeValueLLTs(*DL, *Val.getType(), SplitTys,
@@ -2074,6 +2075,36 @@ bool IRTranslator::translateIfEntryValueArgument(bool isDeclare, Value *Val,
   return true;
 }
 
+static unsigned getConvOpcode(Intrinsic::ID ID) {
+  switch (ID) {
+  default:
+    llvm_unreachable("Unexpected intrinsic");
+  case Intrinsic::experimental_convergence_anchor:
+    return TargetOpcode::CONVERGENCECTRL_ANCHOR;
+  case Intrinsic::experimental_convergence_entry:
+    return TargetOpcode::CONVERGENCECTRL_ENTRY;
+  case Intrinsic::experimental_convergence_loop:
+    return TargetOpcode::CONVERGENCECTRL_LOOP;
+  }
+}
+
+bool IRTranslator::translateConvergenceControlIntrinsic(
+    const CallInst &CI, Intrinsic::ID ID, MachineIRBuilder &MIRBuilder) {
+  MachineInstrBuilder MIB = MIRBuilder.buildInstr(getConvOpcode(ID));
+  Register OutputReg = getOrCreateConvergenceTokenVReg(CI);
+  MIB.addDef(OutputReg);
+
+  if (ID == Intrinsic::experimental_convergence_loop) {
+    auto Bundle = CI.getOperandBundle(LLVMContext::OB_convergencectrl);
+    assert(Bundle && "Expected a convergence control token.");
+    Register InputReg =
+        getOrCreateConvergenceTokenVReg(*Bundle->Inputs[0].get());
+    MIB.addUse(InputReg);
+  }
+
+  return true;
+}
+
 bool IRTranslator::translateKnownIntrinsic(const CallInst &CI, Intrinsic::ID ID,
                                            MachineIRBuilder &MIRBuilder) {
   if (auto *MI = dyn_cast<AnyMemIntrinsic>(&CI)) {
@@ -2530,7 +2561,10 @@ bool IRTranslator::translateKnownIntrinsic(const CallInst &CI, Intrinsic::ID ID,
 #include "llvm/IR/ConstrainedOps.def"
     return translateConstrainedFPIntrinsic(cast<ConstrainedFPIntrinsic>(CI),
                                            MIRBuilder);
-
+  case Intrinsic::experimental_convergence_anchor:
+  case Intrinsic::experimental_convergence_entry:
+  case Intrinsic::experimental_convergence_loop:
+    return translateConvergenceControlIntrinsic(CI, ID, MIRBuilder);
   }
   return false;
 }
@@ -2581,12 +2615,18 @@ bool IRTranslator::translateCallBase(const CallBase &CB,
     }
   }
 
+  Register ConvergenceCtrlToken = 0;
+  if (auto Bundle = CB.getOperandBundle(LLVMContext::OB_convergencectrl)) {
+    const auto &Token = *Bundle->Inputs[0].get();
+    ConvergenceCtrlToken = getOrCreateConvergenceTokenVReg(Token);
+  }
+
   // We don't set HasCalls on MFI here yet because call lowering may decide to
   // optimize into tail calls. Instead, we defer that to selection where a final
   // scan is done to check if any instructions are calls.
-  bool Success =
-      CLI->lowerCall(MIRBuilder, CB, Res, Args, SwiftErrorVReg,
-                     [&]() { return getOrCreateVReg(*CB.getCalledOperand()); });
+  bool Success = CLI->lowerCall(
+      MIRBuilder, CB, Res, Args, SwiftErrorVReg, ConvergenceCtrlToken,
+      [&]() { return getOrCreateVReg(*CB.getCalledOperand()); });
 
   // Check if we just inserted a tail call.
   if (Success) {
@@ -2700,6 +2740,14 @@ bool IRTranslator::translateCall(const User &U, MachineIRBuilder &MIRBuilder) {
         MF->getMachineMemOperand(MPI, Info.flags, MemTy, Alignment, CI.getAAMetadata()));
   }
 
+  if (CI.isConvergent()) {
+    if (auto Bundle = CI.getOperandBundle(LLVMContext::OB_convergencectrl)) {
+      auto *Token = Bundle->Inputs[0].get();
+      Register TokenReg = getOrCreateVReg(*Token);
+      MIB.addUse(TokenReg, RegState::Implicit);
+    }
+  }
+
   return true;
 }
 

llvm/lib/CodeGen/GlobalISel/InlineAsmLowering.cpp

@@ -538,6 +538,14 @@ bool InlineAsmLowering::lowerInlineAsm(
     }
   }
 
+  if (auto Bundle = Call.getOperandBundle(LLVMContext::OB_convergencectrl)) {
+    auto *Token = Bundle->Inputs[0].get();
+    ArrayRef<Register> SourceRegs = GetOrCreateVRegs(*Token);
+    assert(SourceRegs.size() == 1 &&
+           "Expected the control token to fit into a single virtual register");
+    Inst.addUse(SourceRegs[0], RegState::Implicit);
+  }
+
   if (const MDNode *SrcLoc = Call.getMetadata("srcloc"))
     Inst.addMetadata(SrcLoc);
 

llvm/lib/CodeGen/MIRParser/MIParser.cpp

@@ -1919,10 +1919,13 @@ bool MIParser::parseLowLevelType(StringRef::iterator Loc, LLT &Ty) {
 
   if (Token.range().front() == 's') {
     auto ScalarSize = APSInt(Token.range().drop_front()).getZExtValue();
-    if (!verifyScalarSize(ScalarSize))
-      return error("invalid size for scalar type");
-
-    Ty = LLT::scalar(ScalarSize);
+    if (ScalarSize) {
+      if (!verifyScalarSize(ScalarSize))
+        return error("invalid size for scalar type");
+      Ty = LLT::scalar(ScalarSize);
+    } else {
+      Ty = LLT::token();
+    }
     lex();
     return false;
   } else if (Token.range().front() == 'p') {
@@ -1980,7 +1983,7 @@ bool MIParser::parseLowLevelType(StringRef::iterator Loc, LLT &Ty) {
   if (Token.range().front() == 's') {
     auto ScalarSize = APSInt(Token.range().drop_front()).getZExtValue();
     if (!verifyScalarSize(ScalarSize))
-      return error("invalid size for scalar type");
+      return error("invalid size for scalar element in vector");
     Ty = LLT::scalar(ScalarSize);
   } else if (Token.range().front() == 'p') {
     const DataLayout &DL = MF.getDataLayout();

llvm/lib/IR/ConvergenceVerifier.cpp

@@ -75,14 +75,14 @@ GenericConvergenceVerifier<SSAContext>::findAndCheckConvergenceTokenUsed(
 
 template <>
 bool GenericConvergenceVerifier<SSAContext>::isInsideConvergentFunction(
-    const InstructionT &I) {
+    const Instruction &I) {
   auto *F = I.getFunction();
   return F->isConvergent();
 }
 
 template <>
 bool GenericConvergenceVerifier<SSAContext>::isConvergent(
-    const InstructionT &I) {
+    const Instruction &I) {
   if (auto *CB = dyn_cast<CallBase>(&I)) {
     return CB->isConvergent();
   }

llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp

@@ -1301,6 +1301,9 @@ bool AMDGPUCallLowering::lowerTailCall(
   if (!handleAssignments(Handler, OutArgs, CCInfo, ArgLocs, MIRBuilder))
     return false;
 
+  if (Info.ConvergenceCtrlToken) {
+    MIB.addUse(Info.ConvergenceCtrlToken, RegState::Implicit);
+  }
   handleImplicitCallArguments(MIRBuilder, MIB, ST, *FuncInfo, CalleeCC,
                               ImplicitArgRegs);
 
@@ -1483,6 +1486,9 @@ bool AMDGPUCallLowering::lowerCall(MachineIRBuilder &MIRBuilder,
 
   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
 
+  if (Info.ConvergenceCtrlToken) {
+    MIB.addUse(Info.ConvergenceCtrlToken, RegState::Implicit);
+  }
   handleImplicitCallArguments(MIRBuilder, MIB, ST, *MFI, Info.CallConv,
                               ImplicitArgRegs);
 

llvm/test/CodeGen/AMDGPU/convergence-tokens.ll

@@ -1,21 +1,24 @@
 ; RUN: llc --amdgpu-disable-structurizer -stop-after=amdgpu-isel -mtriple=amdgcn-- -mcpu=gfx900 -verify-machineinstrs -o - %s | FileCheck --check-prefixes=CHECK,ISEL %s
 ; RUN: llc --amdgpu-disable-structurizer -stop-after=dead-mi-elimination -mtriple=amdgcn-- -mcpu=gfx900 -verify-machineinstrs -o - %s | FileCheck --check-prefixes=CHECK,DEADMI %s
+; RUN: llc --amdgpu-disable-structurizer -global-isel -stop-after=irtranslator -mtriple=amdgcn-- -mcpu=gfx900 -verify-machineinstrs -o - %s | FileCheck %s --check-prefixes=CHECK,GISEL
 
 ; CHECK-LABEL: name:            basic_call
-;       CHECK:    [[TOKEN:%[0-9]+]]:sreg_64 = CONVERGENCECTRL_ENTRY
+;       CHECK:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ENTRY
 ;        ISEL:    {{.*}} SI_CALL_ISEL {{.*}}, @foo, [[TOKEN]], csr_amdgpu, {{.*}}
 ;      DEADMI:    {{.*}} SI_CALL {{.*}}, @foo, csr_amdgpu, {{.*}}, implicit [[TOKEN]]
+;       GISEL:    {{.*}} G_SI_CALL {{.*}}, @foo, csr_amdgpu, {{.*}}, implicit [[TOKEN]]
 define i32 @basic_call(i32 %src) #0 {
   %t = call token @llvm.experimental.convergence.entry()
   %r = call i32 @foo(i32 %src) [ "convergencectrl"(token %t) ]
   ret i32 %r
 }
 
 ; CHECK-LABEL: name:            basic_intrinsic
-;       CHECK:    [[TOKEN:%[0-9]+]]:sreg_64 = CONVERGENCECTRL_ANCHOR
+;       CHECK:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ANCHOR
 ;        ISEL:    CONVERGENCECTRL_GLUE [[TOKEN]]
 ;  DEADMI-NOT:    CONVERGENCECTRL_GLUE
-;       CHECK:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[TOKEN]]
+;        ISEL:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[TOKEN]]
+;       GISEL:    {{.*}} = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane){{.*}}, implicit [[TOKEN]]
 define i32 @basic_intrinsic(i32 %src) #0 {
   %t = call token @llvm.experimental.convergence.anchor()
   %r = call i32 @llvm.amdgcn.readfirstlane(i32 %src) [ "convergencectrl"(token %t) ]
@@ -30,12 +33,13 @@ define i32 @uncontrolled_call(i32 %src) #0 {
 }
 
 ; CHECK-LABEL: name:            basic_branch
-;       CHECK:  bb.0.entry:
-;       CHECK:    [[TOKEN:%[0-9]+]]:sreg_64 = CONVERGENCECTRL_ANCHOR
-;       CHECK:  bb.1.then:
+;       CHECK:  bb.[[#]].entry:
+;       CHECK:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ANCHOR
+;       CHECK:  bb.[[#]].then:
 ;        ISEL:    CONVERGENCECTRL_GLUE [[TOKEN]]
 ;  DEADMI-NOT:    CONVERGENCECTRL_GLUE
-;       CHECK:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[TOKEN]]
+;        ISEL:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[TOKEN]]
+;       GISEL:    {{.*}} = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane){{.*}}, implicit [[TOKEN]]
 define i32 @basic_branch(i32 %src, i1 %cond) #0 {
 entry:
   %t = call token @llvm.experimental.convergence.anchor()
@@ -52,12 +56,13 @@ else:
 }
 
 ; CHECK-LABEL: name:            basic_loop
-;       CHECK:    [[TOKEN:%[0-9]+]]:sreg_64 = CONVERGENCECTRL_ANCHOR
-;       CHECK:  bb.1.loop:
-;       CHECK:    [[LOOP:%[0-9]+]]:sreg_64 = CONVERGENCECTRL_LOOP [[TOKEN]]
+;       CHECK:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ANCHOR
+;       CHECK:  bb.[[#]].loop:
+;       CHECK:    [[LOOP:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_LOOP [[TOKEN]]
 ;        ISEL:    CONVERGENCECTRL_GLUE [[LOOP]]
 ;  DEADMI-NOT:    CONVERGENCECTRL_GLUE
-;       CHECK:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[LOOP]]
+;        ISEL:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[LOOP]]
+;       GISEL:    {{.*}} = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane){{.*}}, implicit [[LOOP]]
 define i32 @basic_loop(i32 %src, i1 %cond) #0 {
   %t1 = call token @llvm.experimental.convergence.anchor()
   br label %loop
@@ -71,6 +76,38 @@ end:
   ret i32 %r
 }
 
+; CHECK-LABEL: name: nested
+;       CHECK:    [[ENTRY:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ENTRY
+;       CHECK:    [[ANCHOR:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ANCHOR
+;        ISEL:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[ANCHOR]]
+;       GISEL:    {{.*}} = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane){{.*}}, implicit [[ANCHOR]]
+;        ISEL:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[ENTRY]]
+;       GISEL:    {{.*}} = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane){{.*}}, implicit [[ENTRY]]
+define i32 @nested(i32 %src) #0 {
+  %t1 = call token @llvm.experimental.convergence.entry()
+  %t2 = call token @llvm.experimental.convergence.anchor()
+  %r2 = call i32 @llvm.amdgcn.readfirstlane(i32 %src) [ "convergencectrl"(token %t2) ]
+  %r1 = call i32 @llvm.amdgcn.readfirstlane(i32 %src) [ "convergencectrl"(token %t1) ]
+  %sum = add i32 %r1, %r2
+  ret i32 %sum
+}
+
+; COM: FIXME: Tokens on tail-call have not been implemented for SelectionDAG
+; COM:        yet; the corresponding checks have been commented out.
+;
+; CHECK-LABEL: name:            tail_call_void_func_void
+;       GISEL:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ENTRY
+; COM:  CHECK:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ENTRY
+; COM:   ISEL:    {{.*}} SI_CALL_ISEL {{.*}}, @external_void_func_void, [[TOKEN]], csr_amdgpu, {{.*}}
+; COM: DEADMI:    {{.*}} SI_CALL {{.*}}, @external_void_func_void, csr_amdgpu, {{.*}}, implicit [[TOKEN]]
+;       GISEL:    {{.*}} SI_TCRETURN {{.*}}, @external_void_func_void, 0, csr_amdgpu, implicit [[TOKEN]]
+define void @tail_call_void_func_void() #0 {
+  %t1 = call token @llvm.experimental.convergence.entry()
+  tail call void @external_void_func_void() [ "convergencectrl"(token %t1) ]
+  ret void
+}
+
+declare hidden void @external_void_func_void() #0
 declare i32 @foo(i32 %x) #0
 
 declare i32 @llvm.amdgcn.readfirstlane(i32) #0