Preserve alignment information when promoting integer loads and stores.

2 files changed, 67 insertions, 5 deletions

diff --git a/lib/Transforms/NaCl/PromoteIntegers.cpp b/lib/Transforms/NaCl/PromoteIntegers.cpp
index d48bdfc37b..b8050b5ba2 100644
--- a/lib/Transforms/NaCl/PromoteIntegers.cpp
+++ b/lib/Transforms/NaCl/PromoteIntegers.cpp
@@ -43,13 +43,25 @@
 using namespace llvm;
 
 namespace {
+class ConversionState;
+
 class PromoteIntegers : public FunctionPass {
+  DataLayout *DL;
+
+  Value *splitLoad(LoadInst *Inst, ConversionState &State);
+  Value *splitStore(StoreInst *Inst, ConversionState &State);
+  void convertInstruction(Instruction *Inst, ConversionState &State);
+
  public:
   static char ID;
   PromoteIntegers() : FunctionPass(ID) {
     initializePromoteIntegersPass(*PassRegistry::getPassRegistry());
   }
   virtual bool runOnFunction(Function &F);
+  virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+    AU.addRequired<DataLayout>();
+    return FunctionPass::getAnalysisUsage(AU);
+  }
 };
 }
 
@@ -135,6 +147,8 @@ static Value *convertConstant(Constant *C, bool SignExt=false) {
   }
 }
 
+namespace {
+
 // Holds the state for converting/replacing values. Conversion is done in one
 // pass, with each value requiring conversion possibly having two stages. When
 // an instruction needs to be replaced (i.e. it has illegal operands or result)
@@ -212,9 +226,11 @@ class ConversionState {
   SmallVector<Instruction *, 8> ToErase;
 };
 
+} // anonymous namespace
+
 // Split an illegal load into multiple legal loads and return the resulting
 // promoted value. The size of the load is assumed to be a multiple of 8.
-static Value *splitLoad(LoadInst *Inst, ConversionState &State) {
+Value *PromoteIntegers::splitLoad(LoadInst *Inst, ConversionState &State) {
   if (Inst->isVolatile() || Inst->isAtomic())
     report_fatal_error("Can't split volatile/atomic loads");
   if (cast<IntegerType>(Inst->getType())->getBitWidth() % 8 != 0)
@@ -246,7 +262,15 @@ static Value *splitLoad(LoadInst *Inst, ConversionState &State) {
         HiType->getPointerTo(),
         OrigPtr->getName() + ".hity");
 
-  Value *LoadHi = IRB.CreateAlignedLoad(BCHi, 1, Inst->getName() + ".hi"); // XXX EMSCRIPTEN: worst-case alignment assumption
+#if 0 // XXX EMSCRIPTEN: We want the full-strength alignment.
+  Value *LoadHi = IRB.CreateAlignedLoad(BCHi, 1, Inst->getName() + ".hi");
+#else
+  unsigned HiAlign = MinAlign(Inst->getAlignment() == 0 ?
+                                  DL->getABITypeAlignment(Inst->getType()) :
+                                  Inst->getAlignment(),
+                              LoWidth / 8);
+  Value *LoadHi = IRB.CreateAlignedLoad(BCHi, HiAlign, Inst->getName() + ".hi");
+#endif
   if (!isLegalSize(Width - LoWidth)) {
     LoadHi = splitLoad(cast<LoadInst>(LoadHi), State);
 #if 0 /// XXX EMSCRIPTEN: We don't need to convert pointers.
@@ -266,7 +290,7 @@ static Value *splitLoad(LoadInst *Inst, ConversionState &State) {
   return Result;
 }
 
-static Value *splitStore(StoreInst *Inst, ConversionState &State) {
+Value *PromoteIntegers::splitStore(StoreInst *Inst, ConversionState &State) {
   if (Inst->isVolatile() || Inst->isAtomic())
     report_fatal_error("Can't split volatile/atomic stores");
   if (cast<IntegerType>(Inst->getValueOperand()->getType())->getBitWidth() % 8
@@ -305,7 +329,15 @@ static Value *splitStore(StoreInst *Inst, ConversionState &State) {
         HiType->getPointerTo(),
         OrigPtr->getName() + ".hity");
 
-  Value *StoreHi = IRB.CreateAlignedStore(HiTrunc, BCHi, 1); // XXX EMSCRIPTEN: worst-case alignment assumption
+#if 0 // XXX EMSCRIPTEN: We want the full-strength alignment.
+  Value *StoreHi = IRB.CreateAlignedStore(HiTrunc, BCHi, 1);
+#else
+  unsigned HiAlign = MinAlign(Inst->getAlignment() == 0 ?
+                                  DL->getABITypeAlignment(Inst->getValueOperand()->getType()) :
+                                  Inst->getAlignment(),
+                              LoWidth / 8);
+  Value *StoreHi = IRB.CreateAlignedStore(HiTrunc, BCHi, HiAlign);
+#endif
 
   if (!isLegalSize(Width - LoWidth)) {
     // HiTrunc is still illegal, and is redundant with the truncate in the
@@ -389,7 +421,7 @@ static Value *getSignExtend(Value *Operand, Value *OrigOperand,
       InsertPt), Shl);
 }
 
-static void convertInstruction(Instruction *Inst, ConversionState &State) {
+void PromoteIntegers::convertInstruction(Instruction *Inst, ConversionState &State) {
   if (SExtInst *Sext = dyn_cast<SExtInst>(Inst)) {
     Value *Op = Sext->getOperand(0);
     Value *NewInst = NULL;
@@ -642,6 +674,7 @@ static void convertInstruction(Instruction *Inst, ConversionState &State) {
 }
 
 bool PromoteIntegers::runOnFunction(Function &F) {
+  DL = &getAnalysis<DataLayout>();
   // Don't support changing the function arguments. This should not be
   // generated by clang.
   for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
diff --git a/test/Transforms/NaCl/promote-integer-align.ll b/test/Transforms/NaCl/promote-integer-align.ll
new file mode 100644
index 0000000000..0866d99a72
--- /dev/null
+++ b/test/Transforms/NaCl/promote-integer-align.ll
@@ -0,0 +1,29 @@
+; RUN: opt -S -nacl-promote-ints < %s | FileCheck %s
+
+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-f80:128:128-n8:16:32"
+
+; CHECK:     define void @aligned_copy(i24* %p, i24* %q) {
+; CHECK-NEXT:  %p.loty = bitcast i24* %p to i16*
+; CHECK-NEXT:  %t.lo = load i16* %p.loty, align 64
+; CHECK-NEXT:  %t.lo.ext = zext i16 %t.lo to i32
+; CHECK-NEXT:  %p.hi = getelementptr i16* %p.loty, i32 1
+; CHECK-NEXT:  %p.hity = bitcast i16* %p.hi to i8*
+; CHECK-NEXT:  %t.hi = load i8* %p.hity, align 1
+; CHECK-NEXT:  %t.hi.ext = zext i8 %t.hi to i32
+; CHECK-NEXT:  %t.hi.ext.sh = shl i32 %t.hi.ext, 16
+; CHECK-NEXT:  %t = or i32 %t.lo.ext, %t.hi.ext.sh
+; CHECK-NEXT:  %q.loty = bitcast i24* %q to i16*
+; CHECK-NEXT:  %t.lo1 = trunc i32 %t to i16
+; CHECK-NEXT:  store i16 %t.lo1, i16* %q.loty, align 64
+; CHECK-NEXT:  %t.hi.sh = lshr i32 %t, 16
+; CHECK-NEXT:  %q.hi = getelementptr i16* %q.loty, i32 1
+; CHECK-NEXT:  %t.hi2 = trunc i32 %t.hi.sh to i8
+; CHECK-NEXT:  %q.hity = bitcast i16* %q.hi to i8*
+; CHECK-NEXT:  store i8 %t.hi2, i8* %q.hity, align 1
+; CHECK-NEXT:  ret void
+; CHECK-NEXT:}
+define void @aligned_copy(i24* %p, i24* %q) {
+  %t = load i24* %p, align 64
+  store i24 %t, i24* %q, align 64
+  ret void
+}