[PoC][DO_NOT_MERGE] Add llc option to switch on and off hasMultipleConditionRegisters feature from a target.

[PaoloS]

Summary

This patch is currently for display only.
The aim of the patch is to make it easier to test the effectiveness of the feature hasMultipleConditionRegisters in a target as such feature is set statically in the target lowering info and influences optimization in several places.
This patch adds the llc option -has-multiple-condition-registers=0/1.
It doesn't change the codegen behaviour.

[PaoloS]

Fixed the test to reflect the effects of the hasMultipleConditionRegisters on the lowering of select and the sinking of icmp.

llvm/include/llvm/CodeGen/TargetLowering.h

protected:
   void setStackPointerRegisterToSaveRestore(Register R) {
     StackPointerRegisterToSaveRestore = R;
   }
 
   /// Tells the code generator that the target has multiple (allocatable)
   /// condition registers that can be used to store the results of comparisons
   /// for use by selects and conditional branches. With multiple condition
   /// registers, the code generator will not aggressively sink comparisons into
   /// the blocks of their users.
-  void setHasMultipleConditionRegisters(bool hasManyRegs = true) {
-    HasMultipleConditionRegisters = hasManyRegs;
-  }
+  void setHasMultipleConditionRegisters(bool hasManyRegs = true);
 
   /// Tells the code generator that the target has BitExtract instructions.
   /// The code generator will aggressively sink "shift"s into the blocks of
   /// their users if the users will generate "and" instructions which can be
   /// combined with "shift" to BitExtract instructions.
   void setHasExtractBitsInsn(bool hasExtractInsn = true) {
     HasExtractBitsInsn = hasExtractInsn;
   }

llvm/lib/CodeGen/TargetLoweringBase.cpp

 // FIXME: This option is only to test if the strict fp operation processed
 // correctly by preventing mutating strict fp operation to normal fp operation
 // during development. When the backend supports strict float operation, this
 // option will be meaningless.
 static cl::opt<bool> DisableStrictNodeMutation("disable-strictnode-mutation",
        cl::desc("Don't mutate strict-float node to a legalize node"),
        cl::init(false), cl::Hidden);
 
+static cl::opt<bool> HasMultipleConditionRegistersOverride(
+    "has-multiple-condition-registers", cl::init(false),
+    cl::desc("The target has multiple condition registers and doesn't need to "
+       "sink compare instructions to save virtual registers."),
+    cl::Hidden);
+
 static bool darwinHasSinCos(const Triple &TT) {
   assert(TT.isOSDarwin() && "should be called with darwin triple");
   // Don't bother with 32 bit x86.
   if (TT.getArch() == Triple::x86)
     return false;
   // Macos < 10.9 has no sincos_stret.
   if (TT.isMacOSX())
     return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit();
TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) {
   initActions();
 
   // Perform these initializations only once.
   MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove =
       MaxLoadsPerMemcmp = 8;
   MaxGluedStoresPerMemcpy = 0;
   MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize =
       MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4;
-  HasMultipleConditionRegisters = false;
+  HasMultipleConditionRegisters = HasMultipleConditionRegistersOverride;
   HasExtractBitsInsn = false;
   JumpIsExpensive = JumpIsExpensiveOverride;
   PredictableSelectIsExpensive = false;
   EnableExtLdPromotion = false;
   StackPointerRegisterToSaveRestore = 0;
   BooleanContents = UndefinedBooleanContent;
   BooleanFloatContents = UndefinedBooleanContent;
   BooleanVectorContents = UndefinedBooleanContent;

 }
 
 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
   // If the command-line option was specified, ignore this request.
   if (!JumpIsExpensiveOverride.getNumOccurrences())
     JumpIsExpensive = isExpensive;
 }
 
+void TargetLoweringBase::setHasMultipleConditionRegisters(bool hasManyRegs) {
+  // If the command-line option was specified, ignore this request.
+  if (!HasMultipleConditionRegistersOverride.getNumOccurrences())
+    HasMultipleConditionRegisters = hasManyRegs;
+}
+
 TargetLoweringBase::LegalizeKind
 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
   // If this is a simple type, use the ComputeRegisterProp mechanism.
   if (VT.isSimple()) {
     MVT SVT = VT.getSimpleVT();
     assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
     MVT NVT = TransformToType[SVT.SimpleTy];
     LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);

llvm/test/CodeGen/RISCV/multiple-cond-regs.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=riscv32 -has-multiple-condition-registers=0 | FileCheck %s --check-prefix=RV32-NO-MCR
+; RUN: llc < %s -mtriple=riscv32 -has-multiple-condition-registers=1 | FileCheck %s --check-prefix=RV32-MCR
+
+; Setting -has-multiple-condition-registers=1/0 switches on and off the hasMultipleConditionRegisters in
+; target lowering. Such feature controls how select is lowered and whether compare instructions are sinked.
+; The following tests show how enabling the feature achieves better sinking of compares but generates more
+; branches when lowering select.
+
+define void @foo_select(i32 addrspace(1)* %out, i32 %a, i32 %b, i32 %c, i32 %x, i32 %y) nounwind {
+; RV32-NO-MCR-LABEL: foo_select:
+; RV32-NO-MCR:       # %bb.0:
+; RV32-NO-MCR-NEXT:    beq a1, a3, .LBB0_3
+; RV32-NO-MCR-NEXT:  # %bb.1:
+; RV32-NO-MCR-NEXT:    beq a1, a2, .LBB0_4
+; RV32-NO-MCR-NEXT:  .LBB0_2:
+; RV32-NO-MCR-NEXT:    sw a4, 0(a0)
+; RV32-NO-MCR-NEXT:    ret
+; RV32-NO-MCR-NEXT:  .LBB0_3:
+; RV32-NO-MCR-NEXT:    mv a4, a5
+; RV32-NO-MCR-NEXT:    bne a1, a2, .LBB0_2
+; RV32-NO-MCR-NEXT:  .LBB0_4:
+; RV32-NO-MCR-NEXT:    mv a4, a5
+; RV32-NO-MCR-NEXT:    sw a4, 0(a0)
+; RV32-NO-MCR-NEXT:    ret
+;
+; RV32-MCR-LABEL: foo_select:
+; RV32-MCR:       # %bb.0:
+; RV32-MCR-NEXT:    xor a2, a1, a2
+; RV32-MCR-NEXT:    snez a2, a2
+; RV32-MCR-NEXT:    xor a1, a1, a3
+; RV32-MCR-NEXT:    snez a1, a1
+; RV32-MCR-NEXT:    and a1, a2, a1
+; RV32-MCR-NEXT:    bnez a1, .LBB0_2
+; RV32-MCR-NEXT:  # %bb.1:
+; RV32-MCR-NEXT:    mv a4, a5
+; RV32-MCR-NEXT:  .LBB0_2:
+; RV32-MCR-NEXT:    sw a4, 0(a0)
+; RV32-MCR-NEXT:    ret
+  %icmp0 = icmp ne i32 %a, %b
+  %icmp1 = icmp ne i32 %a, %c
+  %and = and i1 %icmp0, %icmp1
+  %select = select i1 %and, i32 %x, i32 %y
+  store i32 %select, i32 addrspace(1)* %out
+  ret void
+}
+
+define dso_local signext i16 @foo_cmp(i16* %a, i16* %b) nounwind {
+; RV32-NO-MCR-LABEL: foo_cmp:
+; RV32-NO-MCR:       # %bb.0: # %entry
+; RV32-NO-MCR-NEXT:    lh a0, 0(a0)
+; RV32-NO-MCR-NEXT:    bltz a0, .LBB1_3
+; RV32-NO-MCR-NEXT:  # %bb.1: # %.LBB0_1
+; RV32-NO-MCR-NEXT:    beqz a1, .LBB1_3
+; RV32-NO-MCR-NEXT:  # %bb.2: # %.LBB0_2
+; RV32-NO-MCR-NEXT:    ret
+; RV32-NO-MCR-NEXT:  .LBB1_3: # %return
+; RV32-NO-MCR-NEXT:    mv a0, zero
+; RV32-NO-MCR-NEXT:    ret
+;
+; RV32-MCR-LABEL: foo_cmp:
+; RV32-MCR:       # %bb.0: # %entry
+; RV32-MCR-NEXT:    lh a0, 0(a0)
+; RV32-MCR-NEXT:    bltz a0, .LBB1_3
+; RV32-MCR-NEXT:  # %bb.1: # %.LBB0_1
+; RV32-MCR-NEXT:    seqz a1, a1
+; RV32-MCR-NEXT:    bnez a1, .LBB1_3
+; RV32-MCR-NEXT:  # %bb.2: # %.LBB0_2
+; RV32-MCR-NEXT:    ret
+; RV32-MCR-NEXT:  .LBB1_3: # %return
+; RV32-MCR-NEXT:    mv a0, zero
+; RV32-MCR-NEXT:    ret
+entry:
+  %0 = load i16, i16* %a
+  %cmp = icmp sgt i16 %0, -1
+  %tobool.not = icmp eq i16* %b, null
+  br i1 %cmp, label %.LBB0_1, label %return
+
+.LBB0_1:
+  br i1 %tobool.not, label %return, label %.LBB0_2
+
+.LBB0_2:
+  ret i16 %0
+
+return:
+  ret i16 0
+}
+