[PowerPC] Add the support for __builtin_setrnd()

[ZhangKang]

Summary

PowerPC64/PowerPC64le supports the builtin function __builtin_setrnd to set the floating point rounding mode. This function will use the least significant two bits of integer argument to set the floating point rounding mode.

double __builtin_setrnd(int mode);

The effective values for mode are:

0 - round to nearest
1 - round to zero
2 - round to +infinity
3 - round to -infinity

Note that the mode argument will modulo 4, so if the int argument is greater than 3, it will only use the least significant two bits of the mode. Namely, builtin_setrnd(102)) is equal to builtin_setrnd(2).

[ZhangKang]

Add the test file.

[jsji]

LGTM, with some comment change please. Thanks.

llvm/include/llvm/IR/IntrinsicsPowerPC.td

1170 ↗

(On Diff #190779)

this is set only? remove get in comment.

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

10996 ↗

(On Diff #190779)

Typo? two G8RCRegClass here!

11022 ↗

(On Diff #190779)

Load from the stack which SrcReg stored into DestReg => Load from the stack where SrcReg is stored, and save to DestReg

11037 ↗

(On Diff #190779)

using the two -> use two

llvm/lib/Target/PowerPC/PPCInstrInfo.td

1313 ↗

(On Diff #190779)

round mode => rounding mode

[ZhangKang]

Modify the comments to follow reviewer's comments..

[ZhangKang]

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

10996 ↗

(On Diff #190779)

Yes, it a typo, I have fixed it.

llvm/trunk/include/llvm/IR/IntrinsicsPowerPC.td

def int_ppc_tresume : GCCBuiltin<"__builtin_tresume">,
       Intrinsic<[llvm_i32_ty], [], []>;
 def int_ppc_tsuspend : GCCBuiltin<"__builtin_tsuspend">,
       Intrinsic<[llvm_i32_ty], [], []>;
 
 def int_ppc_ttest : GCCBuiltin<"__builtin_ttest">,
       Intrinsic<[llvm_i64_ty], [], []>;
 
 def int_ppc_cfence : Intrinsic<[], [llvm_anyint_ty], []>;
+
+// PowerPC set FPSCR Intrinsic Definitions.
+def int_ppc_setrnd : GCCBuiltin<"__builtin_setrnd">,
+      Intrinsic<[llvm_double_ty], [llvm_i32_ty], []>;
+
 }

llvm/trunk/lib/Target/PowerPC/PPCISelLowering.cpp

BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY),
             MI.getOperand(0).getReg())
         .addReg(isEQ ? PPC::CR0EQ : PPC::CR0GT);
   } else if (MI.getOpcode() == PPC::TCHECK_RET) {
     DebugLoc Dl = MI.getDebugLoc();
     MachineRegisterInfo &RegInfo = F->getRegInfo();
     unsigned CRReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
     BuildMI(*BB, MI, Dl, TII->get(PPC::TCHECK), CRReg);
     return BB;
+  } else if (MI.getOpcode() == PPC::SETRNDi) {
+    DebugLoc dl = MI.getDebugLoc();
+    unsigned OldFPSCRReg = MI.getOperand(0).getReg();
+
+    // Save FPSCR value.
+    BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), OldFPSCRReg);
+
+    // The floating point rounding mode is in the bits 62:63 of FPCSR, and has 
+    // the following settings:
+    //   00 Round to nearest
+    //   01 Round to 0
+    //   10 Round to +inf
+    //   11 Round to -inf
+
+    // When the operand is immediate, using the two least significant bits of
+    // the immediate to set the bits 62:63 of FPSCR.
+    unsigned Mode = MI.getOperand(1).getImm();
+    BuildMI(*BB, MI, dl, TII->get(Mode & 1 ? PPC::MTFSB1 : PPC::MTFSB0))
+      .addImm(31);
+
+    BuildMI(*BB, MI, dl, TII->get(Mode & 2 ? PPC::MTFSB1 : PPC::MTFSB0))
+      .addImm(30);
+  } else if (MI.getOpcode() == PPC::SETRND) {
+    DebugLoc dl = MI.getDebugLoc();
+
+    // Copy register from F8RCRegClass::SrcReg to G8RCRegClass::DestReg
+    // or copy register from G8RCRegClass::SrcReg to F8RCRegClass::DestReg.
+    // If the target doesn't have DirectMove, we should use stack to do the 
+    // conversion, because the target doesn't have the instructions like mtvsrd
+    // or mfvsrd to do this conversion directly.
+    auto copyRegFromG8RCOrF8RC = [&] (unsigned DestReg, unsigned SrcReg) {
+      if (Subtarget.hasDirectMove()) {
+        BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), DestReg)
+          .addReg(SrcReg);
+      } else {
+        // Use stack to do the register copy.
+        unsigned StoreOp = PPC::STD, LoadOp = PPC::LFD;
+        MachineRegisterInfo &RegInfo = F->getRegInfo();
+        const TargetRegisterClass *RC = RegInfo.getRegClass(SrcReg);
+        if (RC == &PPC::F8RCRegClass) {
+          // Copy register from F8RCRegClass to G8RCRegclass.
+          assert((RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) &&
+                 "Unsupported RegClass.");
+
+          StoreOp = PPC::STFD;
+          LoadOp = PPC::LD;
+        } else {
+          // Copy register from G8RCRegClass to F8RCRegclass.
+          assert((RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) &&
+                 (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) &&
+                 "Unsupported RegClass.");
+        }
+
+        MachineFrameInfo &MFI = F->getFrameInfo();
+        int FrameIdx = MFI.CreateStackObject(8, 8, false);
+
+        MachineMemOperand *MMOStore = F->getMachineMemOperand(
+          MachinePointerInfo::getFixedStack(*F, FrameIdx, 0),
+          MachineMemOperand::MOStore, MFI.getObjectSize(FrameIdx),
+          MFI.getObjectAlignment(FrameIdx));
+
+        // Store the SrcReg into the stack.
+        BuildMI(*BB, MI, dl, TII->get(StoreOp))
+          .addReg(SrcReg)
+          .addImm(0)
+          .addFrameIndex(FrameIdx)
+          .addMemOperand(MMOStore);
+
+        MachineMemOperand *MMOLoad = F->getMachineMemOperand(
+          MachinePointerInfo::getFixedStack(*F, FrameIdx, 0),
+          MachineMemOperand::MOLoad, MFI.getObjectSize(FrameIdx),
+          MFI.getObjectAlignment(FrameIdx));
+
+        // Load from the stack where SrcReg is stored, and save to DestReg, 
+        // so we have done the RegClass conversion from RegClass::SrcReg to 
+        // RegClass::DestReg.
+        BuildMI(*BB, MI, dl, TII->get(LoadOp), DestReg)
+          .addImm(0)
+          .addFrameIndex(FrameIdx)
+          .addMemOperand(MMOLoad);
+      }
+    };
+
+    unsigned OldFPSCRReg = MI.getOperand(0).getReg();
+    
+    // Save FPSCR value.
+    BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), OldFPSCRReg);
+
+    // When the operand is gprc register, use two least significant bits of the
+    // register and mtfsf instruction to set the bits 62:63 of FPSCR. 
+    // 
+    // copy OldFPSCRTmpReg, OldFPSCRReg 
+    // (INSERT_SUBREG ExtSrcReg, (IMPLICIT_DEF ImDefReg), SrcOp, 1)
+    // rldimi NewFPSCRTmpReg, ExtSrcReg, OldFPSCRReg, 0, 62
+    // copy NewFPSCRReg, NewFPSCRTmpReg
+    // mtfsf 255, NewFPSCRReg
+    MachineOperand SrcOp = MI.getOperand(1);
+    MachineRegisterInfo &RegInfo = F->getRegInfo();
+    unsigned OldFPSCRTmpReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
+
+    copyRegFromG8RCOrF8RC(OldFPSCRTmpReg, OldFPSCRReg);
+    
+    unsigned ImDefReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
+    unsigned ExtSrcReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
+
+    // The first operand of INSERT_SUBREG should be a register which has
+    // subregisters, we only care about its RegClass, so we should use an
+    // IMPLICIT_DEF register.
+    BuildMI(*BB, MI, dl, TII->get(TargetOpcode::IMPLICIT_DEF), ImDefReg);
+    BuildMI(*BB, MI, dl, TII->get(PPC::INSERT_SUBREG), ExtSrcReg)
+      .addReg(ImDefReg)
+      .add(SrcOp)
+      .addImm(1);
+
+    unsigned NewFPSCRTmpReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
+    BuildMI(*BB, MI, dl, TII->get(PPC::RLDIMI), NewFPSCRTmpReg)
+      .addReg(OldFPSCRTmpReg)
+      .addReg(ExtSrcReg)
+      .addImm(0)
+      .addImm(62);
+
+    unsigned NewFPSCRReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
+    copyRegFromG8RCOrF8RC(NewFPSCRReg, NewFPSCRTmpReg);
+
+    // The mask 255 means that put the 32:63 bits of NewFPSCRReg to the 32:63
+    // bits of FPSCR.
+    BuildMI(*BB, MI, dl, TII->get(PPC::MTFSF))
+      .addImm(255)
+      .addReg(NewFPSCRReg)
+      .addImm(0)
+      .addImm(0);
   } else {
     llvm_unreachable("Unexpected instr type to insert");
   }
 
   MI.eraseFromParent(); // The pseudo instruction is gone now.
   return BB;
 }
 

llvm/trunk/lib/Target/PowerPC/PPCInstrInfo.td

let isCodeGenOnly = 1 in {
       def BCCTR :  XLForm_2_br2<19, 528, 12, 0, (outs), (ins crbitrc:$bi),
                                 "bcctr 12, $bi, 0", IIC_BrB, []>;
       def BCCTRn : XLForm_2_br2<19, 528, 4, 0, (outs), (ins crbitrc:$bi),
                                 "bcctr 4, $bi, 0", IIC_BrB, []>;
     }
   }
 }
 
+// Set the float rounding mode.
+let Uses = [RM], Defs = [RM] in { 
+def SETRNDi : PPCCustomInserterPseudo<(outs f8rc:$FRT), (ins u2imm:$RND),
+                    "#SETRNDi", [(set f64:$FRT, (int_ppc_setrnd (i32 imm:$RND)))]>;
+
+def SETRND : PPCCustomInserterPseudo<(outs f8rc:$FRT), (ins gprc:$in),
+                    "#SETRND", [(set f64:$FRT, (int_ppc_setrnd gprc :$in))]>;
+}
+
 let Defs = [LR] in
   def MovePCtoLR : PPCEmitTimePseudo<(outs), (ins), "#MovePCtoLR", []>,
                    PPC970_Unit_BRU;
 let Defs = [LR] in
   def MoveGOTtoLR : PPCEmitTimePseudo<(outs), (ins), "#MoveGOTtoLR", []>,
                     PPC970_Unit_BRU;
 
 let isBranch = 1, isTerminator = 1, hasCtrlDep = 1, PPC970_Unit = 7 in {

llvm/trunk/test/CodeGen/PowerPC/setrnd.ll

+; RUN: llc -mcpu=pwr9 -mtriple=powerpc64le-unknown-linux-gnu < %s \
+; RUN:   -verify-machineinstrs | FileCheck %s
+; RUN: llc -mcpu=pwr7 -mtriple=powerpc64le-unknown-linux-gnu < %s \
+; RUN:   -verify-machineinstrs | FileCheck -check-prefix=CHECK-PWR7 %s
+
+define double @test_setrndi() {
+entry:
+  %0 = tail call double @llvm.ppc.setrnd(i32 2)
+  ret double %0
+
+; CHECK-LABEL: @test_setrndi
+; CHECK:      # %bb.0:
+; CHECK-DAG:   mffs 1 
+; CHECK-DAG:   mtfsb0 31
+; CHECK-DAG:   mtfsb1 30
+; CHECK:       blr
+}
+
+define double @test_setrnd(i32 signext %x) {
+entry:
+  %0 = tail call double @llvm.ppc.setrnd(i32 %x)
+  ret double %0
+
+; CHECK-LABEL: @test_setrnd
+; CHECK:      # %bb.0:
+; CHECK-DAG:   mffs 1 
+; CHECK-DAG:   mffprd [[REG1:[0-9]+]], 1 
+; CHECK-DAG:   rldimi [[REG1]], 3, 0, 62
+; CHECK-DAG:   mtvsrd [[REG2:[0-9]+]], [[REG1]]
+; CHECK-DAG:   mtfsf 255, [[REG2]] 
+; CHECK:       blr
+
+; CHECK-PWR7-LABEL: @test_setrnd
+; CHECK-PWR7:      # %bb.0:
+; CHECK-PWR7-DAG:   mffs 1 
+; CHECK-PWR7-DAG:   stfd 1, -8(1) 
+; CHECK-PWR7-DAG:   ld [[REG1:[0-9]+]], -8(1) 
+; CHECK-PWR7-DAG:   rldimi [[REG1]], 3, 0, 62
+; CHECK-PWR7-DAG:   std [[REG1]], -16(1)
+; CHECK-PWR7-DAG:   lfd [[REG2:[0-9]+]], -16(1)
+; CHECK-PWR7-DAG:   mtfsf 255, [[REG2]] 
+; CHECK-PWR7:       blr
+}
+
+declare double @llvm.ppc.setrnd(i32) 
+