Fix support for inline asm input / output operand tying when operand spans across...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@74053 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/MachineInstr.cpp b/lib/CodeGen/MachineInstr.cpp
index c351593..8dad76b 100644 (file)
--- a/lib/CodeGen/MachineInstr.cpp
+++ b/lib/CodeGen/MachineInstr.cpp
@@ -716,31 +716,37 @@ isRegTiedToUseOperand(unsigned DefOpIdx, unsigned *UseOpIdx) const {
     const MachineOperand &MO = getOperand(DefOpIdx);
     if (!MO.isReg() || !MO.isDef() || MO.getReg() == 0)
       return false;
-    // Determine the actual operand no corresponding to this index.
+    // Determine the actual operand index that corresponds to this index.
     unsigned DefNo = 0;
+    unsigned DefPart = 0;
     for (unsigned i = 1, e = getNumOperands(); i < e; ) {
       const MachineOperand &FMO = getOperand(i);
       assert(FMO.isImm());
       // Skip over this def.
-      i += InlineAsm::getNumOperandRegisters(FMO.getImm()) + 1;
-      if (i > DefOpIdx)
+      unsigned NumOps = InlineAsm::getNumOperandRegisters(FMO.getImm());
+      unsigned PrevDef = i + 1;
+      i = PrevDef + NumOps;
+      if (i > DefOpIdx) {
+        DefPart = DefOpIdx - PrevDef;
         break;
+      }
       ++DefNo;
     }
-    for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
+    for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
       const MachineOperand &FMO = getOperand(i);
       if (!FMO.isImm())
         continue;
       if (i+1 >= e || !getOperand(i+1).isReg() || !getOperand(i+1).isUse())
         continue;
       unsigned Idx;
-      if (InlineAsm::isUseOperandTiedToDef(FMO.getImm(), Idx) && 
+      if (InlineAsm::isUseOperandTiedToDef(FMO.getImm(), Idx) &&
           Idx == DefNo) {
         if (UseOpIdx)
-          *UseOpIdx = (unsigned)i + 1;
+          *UseOpIdx = (unsigned)i + 1 + DefPart;
         return true;
       }
     }
+    return false;
   }
 
   assert(getOperand(DefOpIdx).isDef() && "DefOpIdx is not a def!");
@@ -766,10 +772,16 @@ isRegTiedToDefOperand(unsigned UseOpIdx, unsigned *DefOpIdx) const {
     const MachineOperand &MO = getOperand(UseOpIdx);
     if (!MO.isReg() || !MO.isUse() || MO.getReg() == 0)
       return false;
-    assert(UseOpIdx > 0);
-    const MachineOperand &UFMO = getOperand(UseOpIdx-1);
-    if (!UFMO.isImm())
-      return false;  // Must be physreg uses.
+    int FlagIdx = UseOpIdx - 1;
+    if (FlagIdx < 1)
+      return false;
+    while (!getOperand(FlagIdx).isImm()) {
+      if (--FlagIdx == 0)
+        return false;
+    }
+    const MachineOperand &UFMO = getOperand(FlagIdx);
+    if (FlagIdx + InlineAsm::getNumOperandRegisters(UFMO.getImm()) < UseOpIdx)
+      return false;
     unsigned DefNo;
     if (InlineAsm::isUseOperandTiedToDef(UFMO.getImm(), DefNo)) {
       if (!DefOpIdx)
@@ -785,7 +797,7 @@ isRegTiedToDefOperand(unsigned UseOpIdx, unsigned *DefOpIdx) const {
         DefIdx += InlineAsm::getNumOperandRegisters(FMO.getImm()) + 1;
         --DefNo;
       }
-      *DefOpIdx = DefIdx+1;
+      *DefOpIdx = DefIdx + UseOpIdx - FlagIdx;
       return true;
     }
     return false;

diff --git a/test/CodeGen/X86/inline-asm-tied.ll b/test/CodeGen/X86/inline-asm-tied.ll
new file mode 100644 (file)
index 0000000..6df2c48
--- /dev/null
+++ b/test/CodeGen/X86/inline-asm-tied.ll
@@ -0,0 +1,19 @@
+; RUN: llvm-as < %s | llc -mtriple=i386-apple-darwin9 -O0 | grep {movl %edx, 4(%esp)} | count 2
+; rdar://6992609
+
+target triple = "i386-apple-darwin9.0"
+@llvm.used = appending global [1 x i8*] [i8* bitcast (i64 (i64)* @_OSSwapInt64 to i8*)], section "llvm.metadata"               ; <[1 x i8*]*> [#uses=0]
+
+define i64 @_OSSwapInt64(i64 %_data) nounwind {
+entry:
+       %retval = alloca i64            ; <i64*> [#uses=2]
+       %_data.addr = alloca i64                ; <i64*> [#uses=4]
+       store i64 %_data, i64* %_data.addr
+       %tmp = load i64* %_data.addr            ; <i64> [#uses=1]
+       %0 = call i64 asm "bswap   %eax\0A\09bswap   %edx\0A\09xchgl   %eax, %edx", "=A,0,~{dirflag},~{fpsr},~{flags}"(i64 %tmp) nounwind               ; <i64> [#uses=1]
+       store i64 %0, i64* %_data.addr
+       %tmp1 = load i64* %_data.addr           ; <i64> [#uses=1]
+       store i64 %tmp1, i64* %retval
+       %1 = load i64* %retval          ; <i64> [#uses=1]
+       ret i64 %1
+}