PowerPC inline asm was emitting two output operands

for a single "m" constraint; this is wrong because the
opcode of a load or store would have to change in parallel.
This patch makes it always compute addresses into a register,
which is correct but not as efficient as possible.  7144566.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@79292 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/PowerPC/AsmPrinter/PPCAsmPrinter.cpp b/lib/Target/PowerPC/AsmPrinter/PPCAsmPrinter.cpp
index efb9e5c62cdcdcfabe2a24d4e290c95fe09bd72c..5dca6e6ad24b421fdf397207d5d63c6463b90f37 100644 (file)
--- a/lib/Target/PowerPC/AsmPrinter/PPCAsmPrinter.cpp
+++ b/lib/Target/PowerPC/AsmPrinter/PPCAsmPrinter.cpp
@@ -507,15 +507,17 @@ bool PPCAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
   return false;
 }
 
+// At the moment, all inline asm memory operands are a single register.
+// In any case, the output of this routine should always be just one
+// assembler operand.
+
 bool PPCAsmPrinter::PrintAsmMemoryOperand(const MachineInstr *MI, unsigned OpNo,
                                           unsigned AsmVariant,
                                           const char *ExtraCode) {
   if (ExtraCode && ExtraCode[0])
     return true; // Unknown modifier.
-  if (MI->getOperand(OpNo).isReg())
-    printMemRegReg(MI, OpNo);
-  else
-    printMemRegImm(MI, OpNo);
+  assert (MI->getOperand(OpNo).isReg());
+  printOperand(MI, OpNo);
   return false;
 }
 

diff --git a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
index d9a4daefb13a36e623864ee9907989aff895c12b..1a068f09810c6862f08de7d0baf3803dff81fca5 100644 (file)
--- a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
+++ b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
@@ -143,30 +143,14 @@ namespace {
     }
       
     /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
-    /// inline asm expressions.
-    virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
+    /// inline asm expressions.  It is always correct to compute the value into
+    /// a register.  The case of adding a (possibly relocatable) constant to a
+    /// register can be improved, but it is wrong to substitute Reg+Reg for
+    /// Reg in an asm, because the load or store opcode would have to change.
+   virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
                                               char ConstraintCode,
                                               std::vector<SDValue> &OutOps) {
-      SDValue Op0, Op1;
-      switch (ConstraintCode) {
-      default: return true;
-      case 'm':   // memory
-        if (!SelectAddrIdx(Op, Op, Op0, Op1))
-          SelectAddrImm(Op, Op, Op0, Op1);
-        break;
-      case 'o':   // offsetable
-        if (!SelectAddrImm(Op, Op, Op0, Op1)) {
-          Op0 = Op;
-          Op1 = getSmallIPtrImm(0);
-        }
-        break;
-      case 'v':   // not offsetable
-        SelectAddrIdxOnly(Op, Op, Op0, Op1);
-        break;
-      }
-      
-      OutOps.push_back(Op0);
-      OutOps.push_back(Op1);
+      OutOps.push_back(Op);
       return false;
     }
     

diff --git a/test/CodeGen/PowerPC/2009-08-17-inline-asm-addr-mode-breakage.ll b/test/CodeGen/PowerPC/2009-08-17-inline-asm-addr-mode-breakage.ll
new file mode 100644 (file)
index 0000000..9e0c5de
--- /dev/null
+++ b/test/CodeGen/PowerPC/2009-08-17-inline-asm-addr-mode-breakage.ll
@@ -0,0 +1,22 @@
+; RUN: llvm-as < %s | llc -march=ppc32 | grep add
+; ModuleID = '<stdin>'
+target datalayout = "E-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f128:64:128"
+target triple = "powerpc-apple-darwin10.0"
+; It is wrong on powerpc to substitute reg+reg for $0; the stw opcode
+; would have to change.
+
+@x = external global [0 x i32]                    ; <[0 x i32]*> [#uses=1]
+
+define void @foo(i32 %y) nounwind ssp {
+entry:
+  %y_addr = alloca i32                            ; <i32*> [#uses=2]
+  %"alloca point" = bitcast i32 0 to i32          ; <i32> [#uses=0]
+  store i32 %y, i32* %y_addr
+  %0 = load i32* %y_addr, align 4                 ; <i32> [#uses=1]
+  %1 = getelementptr inbounds [0 x i32]* @x, i32 0, i32 %0 ; <i32*> [#uses=1]
+  call void asm sideeffect "isync\0A\09eieio\0A\09stw $1, $0", "=*o,r,~{memory}"(i32* %1, i32 0) nounwind
+  br label %return
+
+return:                                           ; preds = %entry
+  ret void
+}