AMDGPU: Correct memory size for image intrinsics

This was incorrectly rounding up to the next power of 2. v4f32 was
rounding up to v8f32, which was just wrong. There are also v3i16/v3f16
available in MVT, so we don't even need to round the f16 cases
anymore. Additionally, this field is really an EVT so we don't even
need to consider this.

Also switch some asserts to return invalid. We should have an IR
verifier for these intrinsic return types, but for now it's better to
not assert on IR that passes the verifier.

This should also probably be fixed to consider that dmask is really
eliminating some of the loaded components.

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index c3a17c1..98654c1 100644 (file)
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -878,45 +878,20 @@ unsigned SITargetLowering::getVectorTypeBreakdownForCallingConv(
     Context, CC, VT, IntermediateVT, NumIntermediates, RegisterVT);
 }
 
-static MVT memVTFromAggregate(Type *Ty) {
+// Peek through TFE struct returns to only use the data size.
+static EVT memVTFromImageReturn(Type *Ty) {
+  auto *ST = dyn_cast<StructType>(Ty);
+  if (!ST)
+    return EVT::getEVT(Ty, true);
+
+  // Some intrinsics return an aggregate type - special case to work out the
+  // correct memVT.
+  //
   // Only limited forms of aggregate type currently expected.
-  assert(Ty->isStructTy() && "Expected struct type");
-
-
-  Type *ElementType = nullptr;
-  unsigned NumElts;
-  if (Ty->getContainedType(0)->isVectorTy()) {
-    VectorType *VecComponent = cast<VectorType>(Ty->getContainedType(0));
-    ElementType = VecComponent->getElementType();
-    NumElts = VecComponent->getNumElements();
-  } else {
-    ElementType = Ty->getContainedType(0);
-    NumElts = 1;
-  }
-
-  assert((Ty->getContainedType(1) && Ty->getContainedType(1)->isIntegerTy(32)) && "Expected int32 type");
-
-  // Calculate the size of the memVT type from the aggregate
-  unsigned Pow2Elts = 0;
-  unsigned ElementSize;
-  switch (ElementType->getTypeID()) {
-    default:
-      llvm_unreachable("Unknown type!");
-    case Type::IntegerTyID:
-      ElementSize = cast<IntegerType>(ElementType)->getBitWidth();
-      break;
-    case Type::HalfTyID:
-      ElementSize = 16;
-      break;
-    case Type::FloatTyID:
-      ElementSize = 32;
-      break;
-  }
-  unsigned AdditionalElts = ElementSize == 16 ? 2 : 1;
-  Pow2Elts = 1 << Log2_32_Ceil(NumElts + AdditionalElts);
-
-  return MVT::getVectorVT(MVT::getVT(ElementType, false),
-                          Pow2Elts);
+  if (ST->getNumContainedTypes() != 2 ||
+      !ST->getContainedType(1)->isIntegerTy(32))
+    return EVT();
+  return EVT::getEVT(ST->getContainedType(0));
 }
 
 bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
@@ -946,12 +921,8 @@ bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
     Info.flags = MachineMemOperand::MODereferenceable;
     if (Attr.hasFnAttribute(Attribute::ReadOnly)) {
       Info.opc = ISD::INTRINSIC_W_CHAIN;
-      Info.memVT = MVT::getVT(CI.getType(), true);
-      if (Info.memVT == MVT::Other) {
-        // Some intrinsics return an aggregate type - special case to work out
-        // the correct memVT
-        Info.memVT = memVTFromAggregate(CI.getType());
-      }
+      // TODO: Account for dmask reducing loaded size.
+      Info.memVT = memVTFromImageReturn(CI.getType());
       Info.flags |= MachineMemOperand::MOLoad;
     } else if (Attr.hasFnAttribute(Attribute::WriteOnly)) {
       Info.opc = ISD::INTRINSIC_VOID;