[arm-fast-isel] Unaligned stores of floats require special care.

rdar://10510150

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145742 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMFastISel.cpp b/lib/Target/ARM/ARMFastISel.cpp
index f43f084c03a706f68a4fa5bd0bce4d326cfc9812..63f3d4548704ded3f4c68d8fef20db0d1bfe7e42 100644 (file)
--- a/lib/Target/ARM/ARMFastISel.cpp
+++ b/lib/Target/ARM/ARMFastISel.cpp
@@ -181,7 +181,7 @@ class ARMFastISel : public FastISel {
     bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr, bool isZExt,
                      bool allocReg);
                      
-    bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr);
+    bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr, unsigned Alignment = 0);
     bool ARMComputeAddress(const Value *Obj, Address &Addr);
     void ARMSimplifyAddress(Address &Addr, EVT VT, bool useAM3);
     bool ARMIsMemCpySmall(uint64_t Len);
@@ -1053,7 +1053,7 @@ bool ARMFastISel::SelectLoad(const Instruction *I) {
   return true;
 }
 
-bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
+bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr, unsigned Alignment) {
   unsigned StrOpc;
   bool useAM3 = false;
   switch (VT.getSimpleVT().SimpleTy) {
@@ -1102,9 +1102,23 @@ bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
     case MVT::f32:
       if (!Subtarget->hasVFP2()) return false;
       StrOpc = ARM::VSTRS;
+      // Unaligned stores need special handling.
+      if (Alignment && Alignment < 4) {
+        unsigned MoveReg = createResultReg(TLI.getRegClassFor(MVT::i32));
+        AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+                                TII.get(ARM::VMOVRS), MoveReg)
+                        .addReg(SrcReg));
+        SrcReg = MoveReg;
+        VT = MVT::i32;
+        StrOpc = isThumb2 ? ARM::t2STRi12 : ARM::STRi12;
+      }
       break;
     case MVT::f64:
       if (!Subtarget->hasVFP2()) return false;
+      // FIXME: Unaligned stores need special handling.
+      if (Alignment && Alignment < 8) {
+          return false;
+      }
       StrOpc = ARM::VSTRD;
       break;
   }
@@ -1141,7 +1155,8 @@ bool ARMFastISel::SelectStore(const Instruction *I) {
   if (!ARMComputeAddress(I->getOperand(1), Addr))
     return false;
 
-  if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
+  if (!ARMEmitStore(VT, SrcReg, Addr, cast<StoreInst>(I)->getAlignment()))
+    return false;
   return true;
 }
 

diff --git a/test/CodeGen/ARM/fast-isel.ll b/test/CodeGen/ARM/fast-isel.ll
index 465e85f9a832f7ddfe6762b69c5c7ce19d8bdc40..f9c1a9d7d354c7201a2833ce2b3f54149392147d 100644 (file)
--- a/test/CodeGen/ARM/fast-isel.ll
+++ b/test/CodeGen/ARM/fast-isel.ll
@@ -158,3 +158,25 @@ define void @test4() {
 ; ARM: ldr r1, [r1]
 ; ARM: str r0, [r1]
 }
+
+; Check unaligned stores
+%struct.anon = type <{ float }>
+
+@a = common global %struct.anon* null, align 4
+
+define void @unaligned_store(float %x, float %y) nounwind {
+entry:
+; ARM: @unaligned_store
+; ARM: vmov r1, s0
+; ARM: str r1, [r0]
+
+; THUMB: @unaligned_store
+; THUMB: vmov r1, s0
+; THUMB: str r1, [r0]
+
+  %add = fadd float %x, %y
+  %0 = load %struct.anon** @a, align 4
+  %x1 = getelementptr inbounds %struct.anon* %0, i32 0, i32 0
+  store float %add, float* %x1, align 1
+  ret void
+}