[PowerPC] Fix inline asm memory operands not to use r0

On PowerPC, inline asm memory operands might be expanded as 0($r), where $r is
a register containing the address. As a result, this register cannot be r0, and
we need to enforce this register subclass constraint to prevent miscompiling
the code (we'd get this constraint for free with the usual instruction
definitions, but that scheme has no knowledge of how we end up printing inline
asm memory operands, and so here we need to do it 'by hand'). We can accomplish
this within the current address-mode selection framework by introducing an
explicit COPY_TO_REGCLASS node.

Fixes PR21443.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223318 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
index cb7d73a04c2f1791535868caa0ccc710c92c3b9d..e04d7e1bc359bb09c42b869d68acc4a6b95cc9ba 100644 (file)
--- a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
+++ b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
@@ -173,10 +173,20 @@ namespace {
     /// 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.
-   bool SelectInlineAsmMemoryOperand(const SDValue &Op,
+    bool SelectInlineAsmMemoryOperand(const SDValue &Op,
                                       char ConstraintCode,
                                       std::vector<SDValue> &OutOps) override {
-      OutOps.push_back(Op);
+      // We need to make sure that this one operand does not end up in r0
+      // (because we might end up lowering this as 0(%op)).
+      const TargetRegisterInfo *TRI = TM.getSubtargetImpl()->getRegisterInfo();
+      const TargetRegisterClass *TRC = TRI->getPointerRegClass(*MF, /*Kind=*/1);
+      SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
+      SDValue NewOp =
+        SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
+                                       SDLoc(Op), Op.getValueType(),
+                                       Op, RC), 0);
+
+      OutOps.push_back(NewOp);
       return false;
     }
 

diff --git a/test/CodeGen/PowerPC/ia-mem-r0.ll b/test/CodeGen/PowerPC/ia-mem-r0.ll
new file mode 100644 (file)
index 0000000..4ab17ed
--- /dev/null
+++ b/test/CodeGen/PowerPC/ia-mem-r0.ll
@@ -0,0 +1,94 @@
+; RUN: llc -mcpu=pwr7 < %s | FileCheck %s
+target datalayout = "e-m:e-i64:64-n32:64"
+target triple = "powerpc64le-unknown-linux-gnu"
+
+; Make sure that we don't generate a std r, 0(0) -- the memory address cannot
+; be stored in r0.
+; CHECK-LABEL: @test1
+; CHECK-NOT: std {{[0-9]+}}, 0(0) 
+; CHECK: blr
+
+define void @test1({ i8*, void (i8*, i8*)* } %fn_arg) {
+  %fn = alloca { i8*, void (i8*, i8*)* }
+  %sp = alloca i8*, align 8
+  %regs = alloca [18 x i64], align 8
+  store { i8*, void (i8*, i8*)* } %fn_arg, { i8*, void (i8*, i8*)* }* %fn
+  %1 = bitcast [18 x i64]* %regs to i64*
+  call void asm sideeffect "std  14, $0", "=*m"(i64* %1)
+  %2 = bitcast [18 x i64]* %regs to i8*
+  %3 = getelementptr i8* %2, i32 8
+  %4 = bitcast i8* %3 to i64*
+  call void asm sideeffect "std  15, $0", "=*m"(i64* %4)
+  %5 = bitcast [18 x i64]* %regs to i8*
+  %6 = getelementptr i8* %5, i32 16
+  %7 = bitcast i8* %6 to i64*
+  call void asm sideeffect "std  16, $0", "=*m"(i64* %7)
+  %8 = bitcast [18 x i64]* %regs to i8*
+  %9 = getelementptr i8* %8, i32 24
+  %10 = bitcast i8* %9 to i64*
+  call void asm sideeffect "std  17, $0", "=*m"(i64* %10)
+  %11 = bitcast [18 x i64]* %regs to i8*
+  %12 = getelementptr i8* %11, i32 32
+  %13 = bitcast i8* %12 to i64*
+  call void asm sideeffect "std  18, $0", "=*m"(i64* %13)
+  %14 = bitcast [18 x i64]* %regs to i8*
+  %15 = getelementptr i8* %14, i32 40
+  %16 = bitcast i8* %15 to i64*
+  call void asm sideeffect "std  19, $0", "=*m"(i64* %16)
+  %17 = bitcast [18 x i64]* %regs to i8*
+  %18 = getelementptr i8* %17, i32 48
+  %19 = bitcast i8* %18 to i64*
+  call void asm sideeffect "std  20, $0", "=*m"(i64* %19)
+  %20 = bitcast [18 x i64]* %regs to i8*
+  %21 = getelementptr i8* %20, i32 56
+  %22 = bitcast i8* %21 to i64*
+  call void asm sideeffect "std  21, $0", "=*m"(i64* %22)
+  %23 = bitcast [18 x i64]* %regs to i8*
+  %24 = getelementptr i8* %23, i32 64
+  %25 = bitcast i8* %24 to i64*
+  call void asm sideeffect "std  22, $0", "=*m"(i64* %25)
+  %26 = bitcast [18 x i64]* %regs to i8*
+  %27 = getelementptr i8* %26, i32 72
+  %28 = bitcast i8* %27 to i64*
+  call void asm sideeffect "std  23, $0", "=*m"(i64* %28)
+  %29 = bitcast [18 x i64]* %regs to i8*
+  %30 = getelementptr i8* %29, i32 80
+  %31 = bitcast i8* %30 to i64*
+  call void asm sideeffect "std  24, $0", "=*m"(i64* %31)
+  %32 = bitcast [18 x i64]* %regs to i8*
+  %33 = getelementptr i8* %32, i32 88
+  %34 = bitcast i8* %33 to i64*
+  call void asm sideeffect "std  25, $0", "=*m"(i64* %34)
+  %35 = bitcast [18 x i64]* %regs to i8*
+  %36 = getelementptr i8* %35, i32 96
+  %37 = bitcast i8* %36 to i64*
+  call void asm sideeffect "std  26, $0", "=*m"(i64* %37)
+  %38 = bitcast [18 x i64]* %regs to i8*
+  %39 = getelementptr i8* %38, i32 104
+  %40 = bitcast i8* %39 to i64*
+  call void asm sideeffect "std  27, $0", "=*m"(i64* %40)
+  %41 = bitcast [18 x i64]* %regs to i8*
+  %42 = getelementptr i8* %41, i32 112
+  %43 = bitcast i8* %42 to i64*
+  call void asm sideeffect "std  28, $0", "=*m"(i64* %43)
+  %44 = bitcast [18 x i64]* %regs to i8*
+  %45 = getelementptr i8* %44, i32 120
+  %46 = bitcast i8* %45 to i64*
+  call void asm sideeffect "std  29, $0", "=*m"(i64* %46)
+  %47 = bitcast [18 x i64]* %regs to i8*
+  %48 = getelementptr i8* %47, i32 128
+  %49 = bitcast i8* %48 to i64*
+  call void asm sideeffect "std  30, $0", "=*m"(i64* %49)
+  %50 = bitcast [18 x i64]* %regs to i8*
+  %51 = getelementptr i8* %50, i32 136
+  %52 = bitcast i8* %51 to i64*
+  call void asm sideeffect "std  31, $0", "=*m"(i64* %52)
+  %53 = getelementptr { i8*, void (i8*, i8*)* }* %fn, i32 0, i32 1
+  %.funcptr = load void (i8*, i8*)** %53
+  %54 = getelementptr { i8*, void (i8*, i8*)* }* %fn, i32 0, i32 0
+  %.ptr = load i8** %54
+  %55 = load i8** %sp
+  call void %.funcptr(i8* %.ptr, i8* %55)
+  ret void
+}
+