Allow __fp16 as a function arg or return type for AArch64

[olista01]

Summary

ACLE 2.0 [1] allows __fp16 to be used as a function argument or return type. This enables this for AArch64.

I have not enabled this for 32-bit ARM targets yet, as we are expecting to release an updated version of the 32-bit AAPCS soon, which changes the handling of __fp16 in a non backwards-compatible way.

[1] http://infocenter.arm.com/help/topic/com.arm.doc.ihi0053c/IHI0053C_acle_2_0.pdf

[t.p.northover]

Hi Oliver,

I think this looks reasonable.

Cheers.

Tim.

[t.p.northover]

Actually, I've just noticed that this only applies to AAPCS. Could you add the CallingConvention lines to DarwinPCS too? We'll want to support ACLE too.

Cheers.

Tim.

[olista01]

Thanks, committed revision 212812.

Are you also going to review the dependent patch, D4456?

[t.p.northover]

Are you also going to review the dependent patch, D4456?

I am thinking hard about it, but that's an ugly hack to put into Clang.

It would be better if we could use "half" unconditionally and fptrunc/fpext instead of the conversion intrinsics, but that brings its own problems (like InstCombine folding things down to real, non-trivial operations on "half").

I'm still not quite sure what the best solution is or how much work would be involved in the different approaches.

Cheers.

Tim.

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
   if (!UseCP) {
     assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
     return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
                            (VT == MVT::f64) ? MVT::i64 : MVT::i32);
   }
 
   EVT OrigVT = VT;
   EVT SVT = VT;
-  while (SVT != MVT::f32) {
+  while (SVT != MVT::f32 && SVT != MVT::f16) {
     SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
     if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
         // Only do this if the target has a native EXTLOAD instruction from
         // smaller type.
         TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
         TLI.ShouldShrinkFPConstant(OrigVT)) {
       Type *SType = SVT.getTypeForEVT(*DAG.getContext());
       LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));

lib/Target/AArch64/AArch64CallingConvention.td

def CC_AArch64_AAPCS : CallingConv<[
   CCIfType<[i64], CCIfSplit<CCAssignToRegWithShadow<[X0, X2, X4, X6],
                                                     [X0, X1, X3, X5]>>>,
 
   // i128 is split to two i64s, and its stack alignment is 16 bytes.
   CCIfType<[i64], CCIfSplit<CCAssignToStackWithShadow<8, 16, [X7]>>>,
 
   CCIfType<[i64], CCAssignToRegWithShadow<[X0, X1, X2, X3, X4, X5, X6, X7],
                                           [W0, W1, W2, W3, W4, W5, W6, W7]>>,
+  CCIfType<[f16], CCAssignToRegWithShadow<[H0, H1, H2, H3, H4, H5, H6, H7],
+                                          [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]>>,
   CCIfType<[f32], CCAssignToRegWithShadow<[S0, S1, S2, S3, S4, S5, S6, S7],
                                           [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]>>,
   CCIfType<[f64], CCAssignToRegWithShadow<[D0, D1, D2, D3, D4, D5, D6, D7],
                                           [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]>>,
   CCIfType<[v1i64, v2i32, v4i16, v8i8, v1f64, v2f32],
            CCAssignToRegWithShadow<[D0, D1, D2, D3, D4, D5, D6, D7],
                                    [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]>>,
   CCIfType<[f128, v2i64, v4i32, v8i16, v16i8, v4f32, v2f64],
            CCAssignToReg<[Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]>>,
 
   // If more than will fit in registers, pass them on the stack instead.
-  CCIfType<[i1, i8, i16], CCAssignToStack<8, 8>>,
+  CCIfType<[i1, i8, i16, f16], CCAssignToStack<8, 8>>,
   CCIfType<[i32, f32], CCAssignToStack<8, 8>>,
   CCIfType<[i64, f64, v1f64, v2f32, v1i64, v2i32, v4i16, v8i8],
            CCAssignToStack<8, 8>>,
   CCIfType<[f128, v2i64, v4i32, v8i16, v16i8, v4f32, v2f64],
            CCAssignToStack<16, 16>>
 ]>;
 
 def RetCC_AArch64_AAPCS : CallingConv<[
   CCIfType<[v2f32], CCBitConvertToType<v2i32>>,
   CCIfType<[v2f64, v4f32], CCBitConvertToType<v2i64>>,
 
   // Big endian vectors must be passed as if they were 1-element vectors so that
   // their lanes are in a consistent order.
   CCIfBigEndian<CCIfType<[v2i32, v2f32, v4i16, v4f16, v8i8],
                          CCBitConvertToType<f64>>>,
   CCIfBigEndian<CCIfType<[v2i64, v2f64, v4i32, v4f32, v8i16, v8f16, v16i8],
                          CCBitConvertToType<f128>>>,
 
   CCIfType<[i32], CCAssignToRegWithShadow<[W0, W1, W2, W3, W4, W5, W6, W7],
                                           [X0, X1, X2, X3, X4, X5, X6, X7]>>,
   CCIfType<[i64], CCAssignToRegWithShadow<[X0, X1, X2, X3, X4, X5, X6, X7],
                                           [W0, W1, W2, W3, W4, W5, W6, W7]>>,
+  CCIfType<[f16], CCAssignToRegWithShadow<[H0, H1, H2, H3, H4, H5, H6, H7],
+                                          [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]>>,
   CCIfType<[f32], CCAssignToRegWithShadow<[S0, S1, S2, S3, S4, S5, S6, S7],
                                           [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]>>,
   CCIfType<[f64], CCAssignToRegWithShadow<[D0, D1, D2, D3, D4, D5, D6, D7],
                                           [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]>>,
   CCIfType<[v1i64, v2i32, v4i16, v8i8, v1f64, v2f32],
       CCAssignToRegWithShadow<[D0, D1, D2, D3, D4, D5, D6, D7],
                               [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7]>>,
   CCIfType<[f128, v2i64, v4i32, v8i16, v16i8, v4f32, v2f64],

lib/Target/AArch64/AArch64ISelLowering.cpp

if (VA.isRegLoc()) {
 
       SDValue ArgValue;
       const TargetRegisterClass *RC;
 
       if (RegVT == MVT::i32)
         RC = &AArch64::GPR32RegClass;
       else if (RegVT == MVT::i64)
         RC = &AArch64::GPR64RegClass;
+      else if (RegVT == MVT::f16)
+        RC = &AArch64::FPR16RegClass;
       else if (RegVT == MVT::f32)
         RC = &AArch64::FPR32RegClass;
       else if (RegVT == MVT::f64 || RegVT.is64BitVector())
         RC = &AArch64::FPR64RegClass;
       else if (RegVT == MVT::f128 || RegVT.is128BitVector())
         RC = &AArch64::FPR128RegClass;
       else
         llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");

test/CodeGen/AArch64/arm64-aapcs.ll

 
 ; Check if VPR can be correctly pass by stack.
 define <2 x double> @test_vreg_stack([8 x <2 x double>], <2 x double> %varg_stack) {
 entry:
 ; CHECK-LABEL: test_vreg_stack:
 ; CHECK: ldr {{q[0-9]+}}, [sp]
   ret <2 x double> %varg_stack;
 }
+
+; Check that f16 can be passed and returned (ACLE 2.0 extension)
+define half @test_half(float, half %arg) {
+; CHECK-LABEL: test_half:
+; CHECK: mov v0.16b, v{{[0-9]+}}.16b
+  ret half %arg;
+}
+
+; Check that f16 constants are materialized correctly
+define half @test_half_const() {
+; CHECK-LABEL: test_half_const:
+; CHECK: ldr h0, [x{{[0-9]+}}, :lo12:{{.*}}]
+  ret half 0xH4248
+}