[ARM] Do not perform a tail call when the caller returns several values.

The fix is slightly different then x86 (see r216117) because the number of values
attached to a return can vary even for a single returned value (e.g., f64 yields
two returned values).

<rdar://problem/18352998>

llvm-svn: 218076

2 files changed, 95 insertions, 1 deletions

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 14cd7d7bd85..ca6b614ce2e 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -2313,9 +2313,15 @@ bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
       if (Copies.count(UseChain.getNode()))
         // Second CopyToReg
         Copy = *UI;
-      else
+      else {
+        // We are at the top of this chain.
+        // If the copy has a glue operand, we conservatively assume it
+        // isn't safe to perform a tail call.
+        if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
+          return false;
         // First CopyToReg
         TCChain = UseChain;
+      }
     }
   } else if (Copy->getOpcode() == ISD::BITCAST) {
     // f32 returned in a single GPR.
@@ -2324,6 +2330,10 @@ bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
     Copy = *Copy->use_begin();
     if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
       return false;
+    // If the copy has a glue operand, we conservatively assume it isn't safe to
+    // perform a tail call.
+    if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
+      return false;
     TCChain = Copy->getOperand(0);
   } else {
     return false;
diff --git a/llvm/test/CodeGen/ARM/no-tail-call.ll b/llvm/test/CodeGen/ARM/no-tail-call.ll
new file mode 100644
index 00000000000..3a8cb21bee9
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/no-tail-call.ll
@@ -0,0 +1,84 @@
+; RUN: llc < %s -O0 -o - | FileCheck %s
+target datalayout = "e-m:o-p:32:32-f64:32:64-v64:32:64-v128:32:128-a:0:32-n32-S32"
+target triple = "armv7s-apple-ios7"
+
+%foo = type <{ %Sf }>
+%Sf = type <{ float }>
+
+declare float @llvm.ceil.f32(float) 
+
+; Check that we are not emitting a tail call for the last call to ceil.
+; This function returns three different results.
+; CHECK-LABEL: func1:
+; CHECK-NOT: b _ceilf
+; CHECK: pop
+define { float, float, float } @func1() {
+entry:
+  %0 = alloca %foo, align 4
+  %1 = alloca %foo, align 4
+  %2 = alloca %foo, align 4
+  %.native = getelementptr inbounds %foo* %0, i32 0, i32 0
+  %.native.value = getelementptr inbounds %Sf* %.native, i32 0, i32 0
+  store float 0.000000e+00, float* %.native.value, align 4
+  %.native1 = getelementptr inbounds %foo* %1, i32 0, i32 0
+  %.native1.value = getelementptr inbounds %Sf* %.native1, i32 0, i32 0
+  store float 1.000000e+00, float* %.native1.value, align 4
+  %.native2 = getelementptr inbounds %foo* %2, i32 0, i32 0
+  %.native2.value = getelementptr inbounds %Sf* %.native2, i32 0, i32 0
+  store float 5.000000e+00, float* %.native2.value, align 4
+  br i1 true, label %3, label %4
+
+; <label>:3                                       ; preds = %entry
+  %.native4 = getelementptr inbounds %foo* %1, i32 0, i32 0
+  %.native4.value = getelementptr inbounds %Sf* %.native4, i32 0, i32 0
+  store float 2.000000e+00, float* %.native4.value, align 4
+  br label %4
+
+; <label>:4                                       ; preds = %3, %entry
+  %5 = call float @llvm.ceil.f32(float 5.000000e+00)
+  %.native3 = getelementptr inbounds %foo* %1, i32 0, i32 0
+  %.native3.value = getelementptr inbounds %Sf* %.native3, i32 0, i32 0
+  %6 = load float* %.native3.value, align 4
+  %7 = call float @llvm.ceil.f32(float %6)
+  %8 = insertvalue { float, float, float } { float 0.000000e+00, float undef, float undef }, float %5, 1
+  %9 = insertvalue { float, float, float } %8, float %7, 2
+  ret { float, float, float } %9
+}
+
+; Check that we are not emitting a tail call for the last call to ceil.
+; This function returns two different results.
+; CHECK-LABEL: func2:
+; CHECK-NOT: b _ceilf
+; CHECK: pop
+define { float, float } @func2() {
+entry:
+  %0 = alloca %foo, align 4
+  %1 = alloca %foo, align 4
+  %2 = alloca %foo, align 4
+  %.native = getelementptr inbounds %foo* %0, i32 0, i32 0
+  %.native.value = getelementptr inbounds %Sf* %.native, i32 0, i32 0
+  store float 0.000000e+00, float* %.native.value, align 4
+  %.native1 = getelementptr inbounds %foo* %1, i32 0, i32 0
+  %.native1.value = getelementptr inbounds %Sf* %.native1, i32 0, i32 0
+  store float 1.000000e+00, float* %.native1.value, align 4
+  %.native2 = getelementptr inbounds %foo* %2, i32 0, i32 0
+  %.native2.value = getelementptr inbounds %Sf* %.native2, i32 0, i32 0
+  store float 5.000000e+00, float* %.native2.value, align 4
+  br i1 true, label %3, label %4
+
+; <label>:3                                       ; preds = %entry
+  %.native4 = getelementptr inbounds %foo* %1, i32 0, i32 0
+  %.native4.value = getelementptr inbounds %Sf* %.native4, i32 0, i32 0
+  store float 2.000000e+00, float* %.native4.value, align 4
+  br label %4
+
+; <label>:4                                       ; preds = %3, %entry
+  %5 = call float @llvm.ceil.f32(float 5.000000e+00)
+  %.native3 = getelementptr inbounds %foo* %1, i32 0, i32 0
+  %.native3.value = getelementptr inbounds %Sf* %.native3, i32 0, i32 0
+  %6 = load float* %.native3.value, align 4
+  %7 = call float @llvm.ceil.f32(float %6)
+  %8 = insertvalue { float, float } { float 0.000000e+00, float undef }, float %7, 1
+  ret { float, float } %8
+}
+