Bug 13662: Enable GPRPair for all i64 operands of inline asm on ARM

This patch assigns paired GPRs  for inline asm with
64-bit data on ARM. It's enabled for both ARM and Thumb to support modifiers
like %H, %Q, %R.

llvm-svn: 185169

lib/Target/ARM/ARMAsmPrinter.cpp

@@ -464,8 +464,14 @@ bool ARMAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNum,
       // This takes advantage of the 2 operand-ness of ldm/stm and that we've
       // already got the operands in registers that are operands to the
       // inline asm statement.
-
-      O << "{" << ARMInstPrinter::getRegisterName(RegBegin);
+      O << "{";
+      if (ARM::GPRPairRegClass.contains(RegBegin)) {
+        const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
+        unsigned Reg0 = TRI->getSubReg(RegBegin, ARM::gsub_0);
+        O << ARMInstPrinter::getRegisterName(Reg0) << ", ";;
+        RegBegin = TRI->getSubReg(RegBegin, ARM::gsub_1);
+      }
+      O << ARMInstPrinter::getRegisterName(RegBegin);
 
       // FIXME: The register allocator not only may not have given us the
       // registers in sequence, but may not be in ascending registers. This
@@ -491,6 +497,20 @@ bool ARMAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNum,
         return true;
       unsigned Flags = FlagsOP.getImm();
       unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
+      unsigned RC;
+      InlineAsm::hasRegClassConstraint(Flags, RC);
+      if (RC == ARM::GPRPairRegClassID) {
+        if (NumVals != 1)
+          return true;
+        const MachineOperand &MO = MI->getOperand(OpNum);
+        if (!MO.isReg())
+          return true;
+        const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
+        unsigned Reg = TRI->getSubReg(MO.getReg(), ExtraCode[0] == 'Q' ?
+            ARM::gsub_0 : ARM::gsub_1);
+        O << ARMInstPrinter::getRegisterName(Reg);
+        return false;
+      }
       if (NumVals != 2)
         return true;
       unsigned RegOp = ExtraCode[0] == 'Q' ? OpNum : OpNum + 1;

lib/Target/ARM/ARMISelDAGToDAG.cpp

@@ -3472,16 +3472,16 @@ SDNode *ARMDAGToDAGISel::SelectInlineAsm(SDNode *N){
   // However, some instrstions (e.g. ldrexd/strexd in ARM mode) require
   // (even/even+1) GPRs and use %n and %Hn to refer to the individual regs
   // respectively. Since there is no constraint to explicitly specify a
-  // reg pair, we search %H operand inside the asm string. If it is found, the
-  // transformation below enforces a GPRPair reg class for "%r" for 64-bit data.
-  if (AsmString.find(":H}") == StringRef::npos)
-    return NULL;
+  // reg pair, we use GPRPair reg class for "%r" for 64-bit data. For Thumb,
+  // the 64-bit data may be referred by H, Q, R modifiers, so we still pack
+  // them into a GPRPair.
 
   SDLoc dl(N);
-  SDValue Glue = N->getOperand(NumOps-1);
+  SDValue Glue = N->getGluedNode() ? N->getOperand(NumOps-1) : SDValue(0,0);
 
+  SmallVector<bool, 8> OpChanged;
   // Glue node will be appended late.
-  for(unsigned i = 0; i < NumOps -1; ++i) {
+  for(unsigned i = 0, e = N->getGluedNode() ? NumOps - 1 : NumOps; i < e; ++i) {
     SDValue op = N->getOperand(i);
     AsmNodeOperands.push_back(op);
 
@@ -3495,17 +3495,28 @@ SDNode *ARMDAGToDAGISel::SelectInlineAsm(SDNode *N){
     else
       continue;
 
+    unsigned NumRegs = InlineAsm::getNumOperandRegisters(Flag);
+    if (NumRegs)
+      OpChanged.push_back(false);
+
+    unsigned DefIdx = 0;
+    bool IsTiedToChangedOp = false;
+    // If it's a use that is tied with a previous def, it has no
+    // reg class constraint.
+    if (Changed && InlineAsm::isUseOperandTiedToDef(Flag, DefIdx))
+      IsTiedToChangedOp = OpChanged[DefIdx];
+
     if (Kind != InlineAsm::Kind_RegUse && Kind != InlineAsm::Kind_RegDef
         && Kind != InlineAsm::Kind_RegDefEarlyClobber)
       continue;
 
-    unsigned RegNum = InlineAsm::getNumOperandRegisters(Flag);
     unsigned RC;
     bool HasRC = InlineAsm::hasRegClassConstraint(Flag, RC);
-    if (!HasRC || RC != ARM::GPRRegClassID || RegNum != 2)
+    if ((!IsTiedToChangedOp && (!HasRC || RC != ARM::GPRRegClassID))
+        || NumRegs != 2)
       continue;
 
-    assert((i+2 < NumOps-1) && "Invalid number of operands in inline asm");
+    assert((i+2 < NumOps) && "Invalid number of operands in inline asm");
     SDValue V0 = N->getOperand(i+1);
     SDValue V1 = N->getOperand(i+2);
     unsigned Reg0 = cast<RegisterSDNode>(V0)->getReg();
@@ -3566,6 +3577,7 @@ SDNode *ARMDAGToDAGISel::SelectInlineAsm(SDNode *N){
     Changed = true;
 
     if(PairedReg.getNode()) {
+      OpChanged[OpChanged.size() -1 ] = true;
       Flag = InlineAsm::getFlagWord(Kind, 1 /* RegNum*/);
       Flag = InlineAsm::getFlagWordForRegClass(Flag, ARM::GPRPairRegClassID);
       // Replace the current flag.
@@ -3578,7 +3590,8 @@ SDNode *ARMDAGToDAGISel::SelectInlineAsm(SDNode *N){
     }
   }
 
-  AsmNodeOperands.push_back(Glue);
+  if (Glue.getNode())
+    AsmNodeOperands.push_back(Glue);
   if (!Changed)
     return NULL;
 

test/CodeGen/ARM/inlineasm-64bit.ll

@@ -1,5 +1,5 @@
 ; RUN: llc < %s -O3  -mtriple=arm-linux-gnueabi | FileCheck %s
-
+; RUN: llc -mtriple=thumbv7-none-linux-gnueabi -verify-machineinstrs < %s | FileCheck %s
 ; check if regs are passing correctly
 define void @i64_write(i64* %p, i64 %val) nounwind {
 ; CHECK: i64_write:
@@ -45,10 +45,43 @@ entry:
 ; check if callee-saved registers used by inline asm are saved/restored
 define void @foo(i64* %p, i64 %i) nounwind {
 ; CHECK:foo:
-; CHECK: push {{{r[4-9]|r10|r11}}
+; CHECK: {{push|push.w}} {{{r[4-9]|r10|r11}}
 ; CHECK: ldrexd [[REG1:(r[0-9]?[02468])]], {{r[0-9]?[13579]}}, [r{{[0-9]+}}]
 ; CHECK: strexd [[REG1]], {{r[0-9]?[02468]}}, {{r[0-9]?[13579]}}
-; CHECK: pop {{{r[4-9]|r10|r11}}
+; CHECK: {{pop|pop.w}} {{{r[4-9]|r10|r11}}
   %1 = tail call { i64, i64 } asm sideeffect "@ atomic64_set\0A1: ldrexd $0, ${0:H}, [$3]\0Aldrexd $1, ${1:H}, [$3]\0A strexd $0, $4, ${4:H}, [$3]\0A teq $0, #0\0A bne 1b", "=&r,=&r,=*Qo,r,r,~{cc}"(i64* %p, i64* %p, i64 %i) nounwind
   ret void
 }
+
+; return *p;
+define i64 @ldrd_test(i64* %p) nounwind {
+; CHECK: ldrd_test:
+  %1 = tail call i64 asm "ldrd $0, ${0:H}, [$1]", "=r,r"(i64* %p) nounwind
+  ret i64 %1
+}
+
+define i64 @QR_test(i64* %p) nounwind {
+; CHECK: QR_test:
+; CHECK: ldrd {{r[0-9]?[02468]}}, {{r[0-9]?[13579]}}
+  %1 = tail call i64 asm "ldrd ${0:Q}, ${0:R}, [$1]", "=r,r"(i64* %p) nounwind
+  ret i64 %1
+}
+
+define i64 @defuse_test(i64 %p) nounwind {
+; CHECK: defuse_test:
+; CHECK: add {{r[0-9]?[02468]}}, {{r[0-9]?[13579]}}, #1
+  %1 = tail call i64 asm "add $0, ${0:H}, #1", "=r,0"(i64 %p) nounwind
+  ret i64 %1
+}
+
+; *p = (hi << 32) | lo;
+define void @strd_test(i64* %p, i32 %lo, i32 %hi) nounwind {
+; CHECK: strd_test:
+; CHECK: strd {{r[0-9]?[02468]}}, {{r[0-9]?[13579]}}
+  %1 = zext i32 %hi to i64
+  %2 = shl nuw i64 %1, 32
+  %3 = sext i32 %lo to i64
+  %4 = or i64 %2, %3
+  tail call void asm sideeffect "strd $0, ${0:H}, [$1]", "r,r"(i64 %4, i64* %p) nounwind
+  ret void
+}