The VPERM2F128 is a AVX instruction which permutes between two 256-bit

[oota-llvm.git] / lib / Target / X86 / X86FloatingPoint.cpp

diff --git a/lib/Target/X86/X86FloatingPoint.cpp b/lib/Target/X86/X86FloatingPoint.cpp
index 463cde076a3b94570474e8ad4b31cc5925e36e9b..e3461c82c7a6cc11fb0a447d336190148a9b99bc 100644 (file)
--- a/lib/Target/X86/X86FloatingPoint.cpp
+++ b/lib/Target/X86/X86FloatingPoint.cpp
@@ -260,6 +260,21 @@ namespace {
       BuildMI(*MBB, I, dl, TII->get(X86::LD_Frr)).addReg(STReg);
     }
 
+    /// duplicatePendingSTBeforeKill - The instruction at I is about to kill
+    /// RegNo. If any PendingST registers still need the RegNo value, duplicate
+    /// them to new scratch registers.
+    void duplicatePendingSTBeforeKill(unsigned RegNo, MachineInstr *I) {
+      for (unsigned i = 0; i != NumPendingSTs; ++i) {
+        if (PendingST[i] != RegNo)
+          continue;
+        unsigned SR = getScratchReg();
+        DEBUG(dbgs() << "Duplicating pending ST" << i
+                     << " in FP" << RegNo << " to FP" << SR << '\n');
+        duplicateToTop(RegNo, SR, I);
+        PendingST[i] = SR;
+      }
+    }
+
     /// popStackAfter - Pop the current value off of the top of the FP stack
     /// after the specified instruction.
     void popStackAfter(MachineBasicBlock::iterator &I);
@@ -406,6 +421,10 @@ bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) {
     if (MI->isCopy() && isFPCopy(MI))
       FPInstClass = X86II::SpecialFP;
 
+    if (MI->isImplicitDef() &&
+        X86::RFP80RegClass.contains(MI->getOperand(0).getReg()))
+      FPInstClass = X86II::SpecialFP;
+
     if (FPInstClass == X86II::NotFP)
       continue;  // Efficiently ignore non-fp insts!
 
@@ -461,6 +480,7 @@ bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) {
       }
       dumpStack();
     );
+    (void)PrevMI;
 
     Changed = true;
   }
@@ -884,7 +904,7 @@ void FPS::adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I) {
   // Kill registers by popping.
   if (Kills && I != MBB->begin()) {
     MachineBasicBlock::iterator I2 = llvm::prior(I);
-    for (;;) {
+    while (StackTop) {
       unsigned KReg = getStackEntry(0);
       if (!(Kills & (1 << KReg)))
         break;
@@ -969,6 +989,9 @@ void FPS::handleOneArgFP(MachineBasicBlock::iterator &I) {
   unsigned Reg = getFPReg(MI->getOperand(NumOps-1));
   bool KillsSrc = MI->killsRegister(X86::FP0+Reg);
 
+  if (KillsSrc)
+    duplicatePendingSTBeforeKill(Reg, I);
+
   // FISTP64m is strange because there isn't a non-popping versions.
   // If we have one _and_ we don't want to pop the operand, duplicate the value
   // on the stack instead of moving it.  This ensure that popping the value is
@@ -1032,6 +1055,7 @@ void FPS::handleOneArgFPRW(MachineBasicBlock::iterator &I) {
   bool KillsSrc = MI->killsRegister(X86::FP0+Reg);
 
   if (KillsSrc) {
+    duplicatePendingSTBeforeKill(Reg, I);
     // If this is the last use of the source register, just make sure it's on
     // the top of the stack.
     moveToTop(Reg, I);
@@ -1318,6 +1342,7 @@ void FPS::handleSpecialFP(MachineBasicBlock::iterator &I) {
 
       // When the source is killed, allocate a scratch FP register.
       if (KillsSrc) {
+        duplicatePendingSTBeforeKill(SrcFP, I);
         unsigned Slot = getSlot(SrcFP);
         unsigned SR = getScratchReg();
         PendingST[DstST] = SR;
@@ -1369,6 +1394,15 @@ void FPS::handleSpecialFP(MachineBasicBlock::iterator &I) {
     break;
   }
 
+  case TargetOpcode::IMPLICIT_DEF: {
+    // All FP registers must be explicitly defined, so load a 0 instead.
+    unsigned Reg = MI->getOperand(0).getReg() - X86::FP0;
+    DEBUG(dbgs() << "Emitting LD_F0 for implicit FP" << Reg << '\n');
+    BuildMI(*MBB, I, MI->getDebugLoc(), TII->get(X86::LD_F0));
+    pushReg(Reg);
+    break;
+  }
+
   case X86::FpPOP_RETVAL: {
     // The FpPOP_RETVAL instruction is used after calls that return a value on
     // the floating point stack. We cannot model this with ST defs since CALL
@@ -1467,25 +1501,24 @@ void FPS::handleSpecialFP(MachineBasicBlock::iterator &I) {
     }
 
     if (STUses && !isMask_32(STUses))
-      report_fatal_error("Inline asm fixed input regs"
-                         " must be last on the x87 stack");
+      MI->emitError("fixed input regs must be last on the x87 stack");
     unsigned NumSTUses = CountTrailingOnes_32(STUses);
 
     // Defs must be contiguous from the stack top. ST0-STn.
-    if (STDefs && !isMask_32(STDefs))
-      report_fatal_error("Inline asm output regs"
-                         " must be last on the x87 stack");
+    if (STDefs && !isMask_32(STDefs)) {
+      MI->emitError("output regs must be last on the x87 stack");
+      STDefs = NextPowerOf2(STDefs) - 1;
+    }
     unsigned NumSTDefs = CountTrailingOnes_32(STDefs);
 
     // So must the clobbered stack slots. ST0-STm, m >= n.
     if (STClobbers && !isMask_32(STDefs | STClobbers))
-      report_fatal_error("Inline asm clobbers must be last on the x87 stack");
+      MI->emitError("clobbers must be last on the x87 stack");
 
     // Popped inputs are the ones that are also clobbered or defined.
     unsigned STPopped = STUses & (STDefs | STClobbers);
     if (STPopped && !isMask_32(STPopped))
-      report_fatal_error("Inline asm implicitly popped regs"
-                         " must be last on the x87 stack");
+      MI->emitError("implicitly popped regs must be last on the x87 stack");
     unsigned NumSTPopped = CountTrailingOnes_32(STPopped);
 
     DEBUG(dbgs() << "Asm uses " << NumSTUses << " fixed regs, pops "
@@ -1501,7 +1534,7 @@ void FPS::handleSpecialFP(MachineBasicBlock::iterator &I) {
       if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
         continue;
       if (!Op.isUse())
-        report_fatal_error("Illegal \"f\" output constraint in inline asm");
+        MI->emitError("illegal \"f\" output constraint");
       unsigned FPReg = getFPReg(Op);
       FPUsed |= 1U << FPReg;