Fix FP_TO_INT**_IN_MEM lowering.
llvm-svn: 25368
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 22194aa..bd53e61 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -13,6 +13,7 @@
//===----------------------------------------------------------------------===//
#include "X86.h"
+#include "X86InstrBuilder.h"
#include "X86ISelLowering.h"
#include "X86TargetMachine.h"
#include "llvm/CallingConv.h"
@@ -1261,54 +1262,117 @@
MachineBasicBlock *
X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
MachineBasicBlock *BB) {
- assert((MI->getOpcode() == X86::CMOV_FR32 ||
- MI->getOpcode() == X86::CMOV_FR64) &&
- "Unexpected instr type to insert");
+ switch (MI->getOpcode()) {
+ default: assert(false && "Unexpected instr type to insert");
+ case X86::CMOV_FR32:
+ case X86::CMOV_FR64: {
+ // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
+ // control-flow pattern. The incoming instruction knows the destination vreg
+ // to set, the condition code register to branch on, the true/false values to
+ // select between, and a branch opcode to use.
+ const BasicBlock *LLVM_BB = BB->getBasicBlock();
+ ilist<MachineBasicBlock>::iterator It = BB;
+ ++It;
+
+ // thisMBB:
+ // ...
+ // TrueVal = ...
+ // cmpTY ccX, r1, r2
+ // bCC copy1MBB
+ // fallthrough --> copy0MBB
+ MachineBasicBlock *thisMBB = BB;
+ MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
+ MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
+ unsigned Opc = getCondBrOpcodeForX86CC(MI->getOperand(3).getImmedValue());
+ BuildMI(BB, Opc, 1).addMBB(sinkMBB);
+ MachineFunction *F = BB->getParent();
+ F->getBasicBlockList().insert(It, copy0MBB);
+ F->getBasicBlockList().insert(It, sinkMBB);
+ // Update machine-CFG edges
+ BB->addSuccessor(copy0MBB);
+ BB->addSuccessor(sinkMBB);
+
+ // copy0MBB:
+ // %FalseValue = ...
+ // # fallthrough to sinkMBB
+ BB = copy0MBB;
+
+ // Update machine-CFG edges
+ BB->addSuccessor(sinkMBB);
+
+ // sinkMBB:
+ // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
+ // ...
+ BB = sinkMBB;
+ BuildMI(BB, X86::PHI, 4, MI->getOperand(0).getReg())
+ .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
+ .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
- // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
- // control-flow pattern. The incoming instruction knows the destination vreg
- // to set, the condition code register to branch on, the true/false values to
- // select between, and a branch opcode to use.
- const BasicBlock *LLVM_BB = BB->getBasicBlock();
- ilist<MachineBasicBlock>::iterator It = BB;
- ++It;
-
- // thisMBB:
- // ...
- // TrueVal = ...
- // cmpTY ccX, r1, r2
- // bCC copy1MBB
- // fallthrough --> copy0MBB
- MachineBasicBlock *thisMBB = BB;
- MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
- MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
- unsigned Opc = getCondBrOpcodeForX86CC(MI->getOperand(3).getImmedValue());
- BuildMI(BB, Opc, 1).addMBB(sinkMBB);
- MachineFunction *F = BB->getParent();
- F->getBasicBlockList().insert(It, copy0MBB);
- F->getBasicBlockList().insert(It, sinkMBB);
- // Update machine-CFG edges
- BB->addSuccessor(copy0MBB);
- BB->addSuccessor(sinkMBB);
-
- // copy0MBB:
- // %FalseValue = ...
- // # fallthrough to sinkMBB
- BB = copy0MBB;
-
- // Update machine-CFG edges
- BB->addSuccessor(sinkMBB);
-
- // sinkMBB:
- // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
- // ...
- BB = sinkMBB;
- BuildMI(BB, X86::PHI, 4, MI->getOperand(0).getReg())
- .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
- .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
+ delete MI; // The pseudo instruction is gone now.
+ return BB;
+ }
- delete MI; // The pseudo instruction is gone now.
- return BB;
+ case X86::FP_TO_INT16_IN_MEM:
+ case X86::FP_TO_INT32_IN_MEM:
+ case X86::FP_TO_INT64_IN_MEM: {
+ // Change the floating point control register to use "round towards zero"
+ // mode when truncating to an integer value.
+ MachineFunction *F = BB->getParent();
+ int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
+ addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
+
+ // Load the old value of the high byte of the control word...
+ unsigned OldCW =
+ F->getSSARegMap()->createVirtualRegister(X86::R16RegisterClass);
+ addFrameReference(BuildMI(BB, X86::MOV16rm, 4, OldCW), CWFrameIdx);
+
+ // Set the high part to be round to zero...
+ addFrameReference(BuildMI(BB, X86::MOV16mi, 5), CWFrameIdx).addImm(0xC7F);
+
+ // Reload the modified control word now...
+ addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
+
+ // Restore the memory image of control word to original value
+ addFrameReference(BuildMI(BB, X86::MOV16mr, 5), CWFrameIdx).addReg(OldCW);
+
+ // Get the X86 opcode to use.
+ unsigned Opc;
+ switch (MI->getOpcode()) {
+ case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
+ case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
+ case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
+ }
+
+ X86AddressMode AM;
+ MachineOperand &Op = MI->getOperand(0);
+ if (Op.isRegister()) {
+ AM.BaseType = X86AddressMode::RegBase;
+ AM.Base.Reg = Op.getReg();
+ } else {
+ AM.BaseType = X86AddressMode::FrameIndexBase;
+ AM.Base.FrameIndex = Op.getFrameIndex();
+ }
+ Op = MI->getOperand(1);
+ if (Op.isImmediate())
+ AM.Scale = Op.getImmedValue();
+ Op = MI->getOperand(2);
+ if (Op.isImmediate())
+ AM.IndexReg = Op.getImmedValue();
+ Op = MI->getOperand(3);
+ if (Op.isGlobalAddress()) {
+ AM.GV = Op.getGlobal();
+ } else {
+ AM.Disp = Op.getImmedValue();
+ }
+ addFullAddress(BuildMI(BB, Opc, 5), AM).addReg(MI->getOperand(4).getReg());
+
+ // Reload the original control word now.
+ addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
+
+ delete MI; // The pseudo instruction is gone now.
+ return BB;
+ }
+ }
}