X86 fast-isel: Avoid explicit AH subreg reference for [SU]Rem.

Explicit references to %AH for an i8 remainder instruction can lead to
references to %AH in a REX prefixed instruction, which causes things to
blow up. Do the same thing in FastISel as we do for DAG isel and instead
shift %AX right by 8 bits and then extract the 8-bit subreg from that
result.

rdar://14203849
http://llvm.org/bugs/show_bug.cgi?id=16105

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@185899 91177308-0d34-0410-b5e6-96231b3b80d8

Author: Jim Grosbach

lib/Target/X86/X86FastISel.cpp

@@ -1376,10 +1376,37 @@ bool X86FastISel::X86SelectDivRem(const Instruction *I) {
   // Generate the DIV/IDIV instruction.
   BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
           TII.get(OpEntry.OpDivRem)).addReg(Op1Reg);
-  // Copy output register into result register.
-  unsigned ResultReg = createResultReg(TypeEntry.RC);
-  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
-          TII.get(Copy), ResultReg).addReg(OpEntry.DivRemResultReg);
+  // For i8 remainder, we can't reference AH directly, as we'll end
+  // up with bogus copies like %R9B = COPY %AH. Reference AX
+  // instead to prevent AH references in a REX instruction.
+  //
+  // The current assumption of the fast register allocator is that isel
+  // won't generate explicit references to the GPR8_NOREX registers. If
+  // the allocator and/or the backend get enhanced to be more robust in
+  // that regard, this can be, and should be, removed.
+  unsigned ResultReg = 0;
+  if ((I->getOpcode() == Instruction::SRem ||
+       I->getOpcode() == Instruction::URem) &&
+      OpEntry.DivRemResultReg == X86::AH && Subtarget->is64Bit()) {
+    unsigned SourceSuperReg = createResultReg(&X86::GR16RegClass);
+    unsigned ResultSuperReg = createResultReg(&X86::GR16RegClass);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+            TII.get(Copy), SourceSuperReg).addReg(X86::AX);
+
+    // Shift AX right by 8 bits instead of using AH.
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SHR16ri),
+            ResultSuperReg).addReg(SourceSuperReg).addImm(8);
+
+    // Now reference the 8-bit subreg of the result.
+    ResultReg = FastEmitInst_extractsubreg(MVT::i8, ResultSuperReg,
+                                           /*Kill=*/true, X86::sub_8bit);
+  }
+  // Copy the result out of the physreg if we haven't already.
+  if (!ResultReg) {
+    ResultReg = createResultReg(TypeEntry.RC);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Copy), ResultReg)
+        .addReg(OpEntry.DivRemResultReg);
+  }
   UpdateValueMap(I, ResultReg);
 
   return true;