[SubZero] Vector types support for MIPS
This patch implements vector operations on MIPS32 using VariableVecOn32 method (on the lines of Variable64On32).
Vector operations are scalarized prior to lowering. Each vector variable is split into 4 containers to hold a variable of vector type.
For MIPS32, four GP/FP registers are used to hold a vector variable. Arguments are passed in GP registers irrespective of the type of the vector variable.
Lit test vector-mips.ll has been added to test this implementation.
R=stichnot@chromium.org
Review URL: https://codereview.chromium.org/2380023002 .
Patch from Jaydeep Patil <jaydeep.patil@imgtec.com>.
diff --git a/src/IceTargetLowering.h b/src/IceTargetLowering.h
index 278b7a8..82ee5c7 100644
--- a/src/IceTargetLowering.h
+++ b/src/IceTargetLowering.h
@@ -256,6 +256,12 @@
/// Return whether a 64-bit Variable should be split into a Variable64On32.
virtual bool shouldSplitToVariable64On32(Type Ty) const = 0;
+ /// Return whether a Vector Variable should be split into a VariableVecOn32.
+ virtual bool shouldSplitToVariableVecOn32(Type Ty) const {
+ (void)Ty;
+ return false;
+ }
+
bool hasComputedFrame() const { return HasComputedFrame; }
/// Returns true if this function calls a function that has the "returns
/// twice" attribute.
@@ -503,6 +509,9 @@
const SizeT NumElements = typeNumElements(DestTy);
Variable *T = Func->makeVariable(DestTy);
+ if (auto *VarVecOn32 = llvm::dyn_cast<VariableVecOn32>(T)) {
+ VarVecOn32->initVecElement(Func);
+ }
Context.insert<InstFakeDef>(T);
for (SizeT I = 0; I < NumElements; ++I) {