[X86] Align f128 and i128 to 16 bytes when passing on x86-32

llvm/docs/ReleaseNotes.md

@@ -233,6 +233,8 @@ Changes to the X86 Backend
 --------------------------
 
 * `fp128` will now use `*f128` libcalls on 32-bit GNU targets as well.
+* On x86-32, `fp128` and `i128` are now passed with the expected 16-byte stack
+  alignment.
 
 Changes to the OCaml bindings
 -----------------------------

llvm/lib/Target/X86/X86CallingConv.cpp

@@ -374,5 +374,37 @@ static bool CC_X86_64_I128(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
   return true;
 }
 
+/// Special handling for i128 and fp128: on x86-32, i128 and fp128 get legalized
+/// as four i32s, but fp128 must be passed on the stack with 16-byte alignment.
+/// Technically only fp128 has a specified ABI, but it makes sense to handle
+/// i128 the same until we hear differently.
+static bool CC_X86_32_I128_FP128(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
+                                 CCValAssign::LocInfo &LocInfo,
+                                 ISD::ArgFlagsTy &ArgFlags, CCState &State) {
+  assert(ValVT == MVT::i32 && "Should have i32 parts");
+  SmallVectorImpl<CCValAssign> &PendingMembers = State.getPendingLocs();
+  PendingMembers.push_back(
+      CCValAssign::getPending(ValNo, ValVT, LocVT, LocInfo));
+
+  if (!ArgFlags.isInConsecutiveRegsLast())
+    return true;
+
+  unsigned NumRegs = PendingMembers.size();
+  assert(NumRegs == 4 && "Should have two parts");
+
+  int64_t Offset = State.AllocateStack(16, Align(16));
+  PendingMembers[0].convertToMem(Offset);
+  PendingMembers[1].convertToMem(Offset + 4);
+  PendingMembers[2].convertToMem(Offset + 8);
+  PendingMembers[3].convertToMem(Offset + 12);
+
+  State.addLoc(PendingMembers[0]);
+  State.addLoc(PendingMembers[1]);
+  State.addLoc(PendingMembers[2]);
+  State.addLoc(PendingMembers[3]);
+  PendingMembers.clear();
+  return true;
+}
+
 // Provides entry points of CC_X86 and RetCC_X86.
 #include "X86GenCallingConv.inc"

llvm/lib/Target/X86/X86CallingConv.td

@@ -859,6 +859,11 @@ def CC_X86_32_C : CallingConv<[
   // The 'nest' parameter, if any, is passed in ECX.
   CCIfNest<CCAssignToReg<[ECX]>>,
 
+  // i128 and fp128 need to be passed on the stack with a higher alignment than
+  // their legal types. Handle this with a custom function.
+  CCIfType<[i32],
+           CCIfConsecutiveRegs<CCCustom<"CC_X86_32_I128_FP128">>>,
+
   // On swifttailcc pass swiftself in ECX.
   CCIfCC<"CallingConv::SwiftTail",
          CCIfSwiftSelf<CCIfType<[i32], CCAssignToReg<[ECX]>>>>,

llvm/lib/Target/X86/X86ISelLoweringCall.cpp

@@ -237,9 +237,18 @@ EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL,
 bool X86TargetLowering::functionArgumentNeedsConsecutiveRegisters(
     Type *Ty, CallingConv::ID CallConv, bool isVarArg,
     const DataLayout &DL) const {
-  // i128 split into i64 needs to be allocated to two consecutive registers,
-  // or spilled to the stack as a whole.
-  return Ty->isIntegerTy(128);
+  // On x86-64 i128 is split into two i64s and needs to be allocated to two
+  // consecutive registers, or spilled to the stack as a whole. On x86-32 i128
+  // is split to four i32s and never actually passed in registers, but we use
+  // the consecutive register mark to match it in TableGen.
+  if (Ty->isIntegerTy(128))
+    return true;
+
+  // On x86-32, fp128 acts the same as i128.
+  if (Subtarget.is32Bit() && Ty->isFP128Ty())
+    return true;
+
+  return false;
 }
 
 /// Helper for getByValTypeAlignment to determine