Subzero: Align spill locations to natural alignment.

src/IceTargetLowering.h

@@ -134,7 +134,7 @@ public:
 
   virtual bool hasFramePointer() const { return false; }
   virtual SizeT getFrameOrStackReg() const = 0;
-  virtual size_t typeWidthInBytesOnStack(Type Ty) = 0;
+  virtual size_t typeWidthInBytesOnStack(Type Ty) const = 0;
   bool hasComputedFrame() const { return HasComputedFrame; }
   int32_t getStackAdjustment() const { return StackAdjustment; }
   void updateStackAdjustment(int32_t Offset) { StackAdjustment += Offset; }

src/IceTargetLoweringX8632.h

@@ -40,7 +40,7 @@ public:
   virtual SizeT getFrameOrStackReg() const {
     return IsEbpBasedFrame ? Reg_ebp : Reg_esp;
   }
-  virtual size_t typeWidthInBytesOnStack(Type Ty) {
+  virtual size_t typeWidthInBytesOnStack(Type Ty) const {
     // Round up to the next multiple of 4 bytes.  In particular, i1,
     // i8, and i16 are rounded up to 4 bytes.
     return (typeWidthInBytes(Ty) + 3) & ~3;
@@ -125,6 +125,8 @@ protected:
   void scalarizeArithmetic(InstArithmetic::OpKind K, Variable *Dest,
                            Operand *Src0, Operand *Src1);
 
+  void sortByAlignment(VarList &Dest, const VarList &Source) const;
+
   // Operand legalization helpers.  To deal with address mode
   // constraints, the helpers will create a new Operand and emit
   // instructions that guarantee that the Operand kind is one of those
@@ -458,7 +460,7 @@ protected:
   bool IsEbpBasedFrame;
   bool NeedsStackAlignment;
   size_t FrameSizeLocals;
-  size_t LocalsSizeBytes;
+  size_t SpillAreaSizeBytes;
   llvm::SmallBitVector TypeToRegisterSet[IceType_NUM];
   llvm::SmallBitVector ScratchRegs;
   llvm::SmallBitVector RegsUsed;

tests_lit/llvm2ice_tests/align-spill-locations.ll

+; This checks to ensure that Subzero aligns spill slots.
+
+; RUN: %llvm2ice --verbose none %s | FileCheck  %s
+; RUN: %llvm2ice -O2 --verbose none %s | FileCheck  %s
+; RUN: %llvm2ice --verbose none %s | FileCheck --check-prefix=ERRORS %s
+
+; The location of the stack slot for a variable is inferred from the
+; return sequence.
+
+; In this file, "global" refers to a variable with a live range across
+; multiple basic blocks (not an LLVM global variable) and "local"
+; refers to a variable that is live in only a single basic block.
+
+define <4 x i32> @align_global_vector(i32 %arg) {
+entry:
+  %vec.global = insertelement <4 x i32> undef, i32 %arg, i32 0
+  br label %block
+block:
+  call void @ForceXmmSpills()
+  ret <4 x i32> %vec.global
+; CHECK-LABEL: align_global_vector:
+; CHECK: movups xmm0, xmmword ptr [esp]
+; CHECK-NEXT: add esp, 28
+; CHECK-NEXT: ret
+}
+
+define <4 x i32> @align_local_vector(i32 %arg) {
+entry:
+  br label %block
+block:
+  %vec.local = insertelement <4 x i32> undef, i32 %arg, i32 0
+  call void @ForceXmmSpills()
+  ret <4 x i32> %vec.local
+; CHECK-LABEL: align_local_vector:
+; CHECK: movups xmm0, xmmword ptr [esp]
+; CHECK-NEXT: add esp, 28
+; CHECK-NEXT: ret
+}
+
+declare void @ForceXmmSpills()
+
+define <4 x i32> @align_global_vector_ebp_based(i32 %arg) {
+entry:
+  %alloc = alloca i8, i32 1, align 1
+  %vec.global = insertelement <4 x i32> undef, i32 %arg, i32 0
+  br label %block
+block:
+  call void @ForceXmmSpillsAndUseAlloca(i8* %alloc)
+  ret <4 x i32> %vec.global
+; CHECK-LABEL: align_global_vector_ebp_based:
+; CHECK: movups xmm0, xmmword ptr [ebp-24]
+; CHECK-NEXT: mov esp, ebp
+; CHECK-NEXT: pop ebp
+; CHECK: ret
+}
+
+define <4 x i32> @align_local_vector_ebp_based(i32 %arg) {
+entry:
+  %alloc = alloca i8, i32 1, align 1
+  %vec.local = insertelement <4 x i32> undef, i32 %arg, i32 0
+  call void @ForceXmmSpillsAndUseAlloca(i8* %alloc)
+  ret <4 x i32> %vec.local
+; CHECK-LABEL: align_local_vector_ebp_based:
+; CHECK: movups xmm0, xmmword ptr [ebp-24]
+; CHECK-NEXT: mov esp, ebp
+; CHECK-NEXT: pop ebp
+; CHECK: ret
+}
+
+define <4 x i32> @align_local_vector_and_global_float(i32 %arg) {
+entry:
+  %float.global = sitofp i32 %arg to float
+  call void @ForceXmmSpillsAndUseFloat(float %float.global)
+  br label %block
+block:
+  %vec.local = insertelement <4 x i32> undef, i32 undef, i32 0
+  call void @ForceXmmSpillsAndUseFloat(float %float.global)
+  ret <4 x i32> %vec.local
+; CHECK-LABEL: align_local_vector_and_global_float:
+; CHECK: cvtsi2ss xmm0, eax
+; CHECK-NEXT: movss dword ptr [esp+28], xmm0
+; CHECK: movups xmm0, xmmword ptr [esp]
+; CHECK-NEXT: add esp, 44
+; CHECK-NEXT: ret
+}
+
+declare void @ForceXmmSpillsAndUseAlloca(i8*)
+declare void @ForceXmmSpillsAndUseFloat(float)
+
+; ERRORS-NOT: ICE translation error