Fix fptosi, fptoui from f16 vectors to i8, i16 vectors

[pirama]

Summary

Convert v4f16 and v8f16 to corresponding f32 vectors before doing the
conversion to int.

Add tests for v4f16, v8f16.

[pirama]

Ping...

[jmolloy]

Hi Pirama,

Why are we only doing this for v4f16 and v8f16? What about oversized vectors - is there any specific reason you're excluding these? I'm not even sure oversized/undersized vectors can get this far, but it seems strange to whitelist.

I also think this code should be sunk down into LowerVectorFP_TO_INT.

Cheers,

James

[pirama]

Removed vector length checks and refactored as suggested. Larger vectors indeed don't reach here. But it is still better to do the promotion unconditionally.

[pirama]

Remove unnecessary braces.

[jmolloy]

This now looks good to me. Thanks for making those changes!

llvm/trunk/lib/Target/AArch64/AArch64ISelLowering.cpp

 }
 
 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
   // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
   // Any additional optimization in this function should be recorded
   // in the cost tables.
   EVT InVT = Op.getOperand(0).getValueType();
   EVT VT = Op.getValueType();
+  unsigned NumElts = InVT.getVectorNumElements();
+
+  // f16 vectors are promoted to f32 before a conversion.
+  if (InVT.getVectorElementType() == MVT::f16) {
+    MVT NewVT = MVT::getVectorVT(MVT::f32, NumElts);
+    SDLoc dl(Op);
+    return DAG.getNode(
+        Op.getOpcode(), dl, Op.getValueType(),
+        DAG.getNode(ISD::FP_EXTEND, dl, NewVT, Op.getOperand(0)));
+  }
 
   if (VT.getSizeInBits() < InVT.getSizeInBits()) {
     SDLoc dl(Op);
     SDValue Cv =
         DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
                     Op.getOperand(0));
     return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
   }

llvm/trunk/test/CodeGen/AArch64/fp16-v4-instructions.ll

 
 define <4 x i16> @bitcast_h_to_i(float, <4 x half> %a) {
 ; CHECK-LABEL: bitcast_h_to_i:
 ; CHECK: mov v0.16b, v1.16b
   %2 = bitcast <4 x half> %a to <4 x i16>
   ret <4 x i16> %2
 }
 
-
 define <4 x half> @sitofp_i8(<4 x i8> %a) #0 {
 ; CHECK-LABEL: sitofp_i8:
 ; CHECK-NEXT: shl [[OP1:v[0-9]+\.4h]], v0.4h, #8
 ; CHECK-NEXT: sshr [[OP2:v[0-9]+\.4h]], [[OP1]], #8
 ; CHECK-NEXT: sshll [[OP3:v[0-9]+\.4s]], [[OP2]], #0
 ; CHECK-NEXT: scvtf [[OP4:v[0-9]+\.4s]], [[OP3]]
 ; CHECK-NEXT: fcvtn v0.4h, [[OP4]]
 ; CHECK-NEXT: ret

 ; CHECK-LABEL: test_insert_at_zero:
 ; CHECK-NEXT: str d0, [x0]
 ; CHECK-NEXT: ret
   %1 = insertelement <4 x half> undef, half %a, i64 0
   store <4 x half> %1, <4 x half>* %b, align 4
   ret void
 }
 
+define <4 x i8> @fptosi_i8(<4 x half> %a) #0 {
+; CHECK-LABEL: fptosi_i8:
+; CHECK-NEXT: fcvtl  [[REG1:v[0-9]+\.4s]], v0.4h
+; CHECK-NEXT: fcvtzs [[REG2:v[0-9]+\.4s]], [[REG1]]
+; CHECK-NEXT: xtn    v0.4h, [[REG2]]
+; CHECK-NEXT: ret
+  %1 = fptosi<4 x half> %a to <4 x i8>
+  ret <4 x i8> %1
+}
+
+define <4 x i16> @fptosi_i16(<4 x half> %a) #0 {
+; CHECK-LABEL: fptosi_i16:
+; CHECK-NEXT: fcvtl  [[REG1:v[0-9]+\.4s]], v0.4h
+; CHECK-NEXT: fcvtzs [[REG2:v[0-9]+\.4s]], [[REG1]]
+; CHECK-NEXT: xtn    v0.4h, [[REG2]]
+; CHECK-NEXT: ret
+  %1 = fptosi<4 x half> %a to <4 x i16>
+  ret <4 x i16> %1
+}
+
+define <4 x i8> @fptoui_i8(<4 x half> %a) #0 {
+; CHECK-LABEL: fptoui_i8:
+; CHECK-NEXT: fcvtl  [[REG1:v[0-9]+\.4s]], v0.4h
+; NOTE: fcvtzs selected here because the xtn shaves the sign bit
+; CHECK-NEXT: fcvtzs [[REG2:v[0-9]+\.4s]], [[REG1]]
+; CHECK-NEXT: xtn    v0.4h, [[REG2]]
+; CHECK-NEXT: ret
+  %1 = fptoui<4 x half> %a to <4 x i8>
+  ret <4 x i8> %1
+}
+
+define <4 x i16> @fptoui_i16(<4 x half> %a) #0 {
+; CHECK-LABEL: fptoui_i16:
+; CHECK-NEXT: fcvtl  [[REG1:v[0-9]+\.4s]], v0.4h
+; CHECK-NEXT: fcvtzu [[REG2:v[0-9]+\.4s]], [[REG1]]
+; CHECK-NEXT: xtn    v0.4h, [[REG2]]
+; CHECK-NEXT: ret
+  %1 = fptoui<4 x half> %a to <4 x i16>
+  ret <4 x i16> %1
+}
+
 attributes #0 = { nounwind }

llvm/trunk/test/CodeGen/AArch64/fp16-v8-instructions.ll

 ; CHECK-LABEL: test_insert_at_zero:
 ; CHECK-NEXT: str q0, [x0]
 ; CHECK-NEXT: ret
   %1 = insertelement <8 x half> undef, half %a, i64 0
   store <8 x half> %1, <8 x half>* %b, align 4
   ret void
 }
 
+define <8 x i8> @fptosi_i8(<8 x half> %a) #0 {
+; CHECK-LABEL: fptosi_i8:
+; CHECK-DAG: fcvtl   [[LO:v[0-9]+\.4s]], v0.4h
+; CHECK-DAG: fcvtl2  [[HI:v[0-9]+\.4s]], v0.8h
+; CHECK-DAG: fcvtzs  [[LOF32:v[0-9]+\.4s]], [[LO]]
+; CHECK-DAG: xtn     [[I16:v[0-9]+]].4h, [[LOF32]]
+; CHECK-DAG: fcvtzs  [[HIF32:v[0-9]+\.4s]], [[HI]]
+; CHECK-DAG: xtn2    [[I16]].8h, [[HIF32]]
+; CHECK-NEXT: xtn     v0.8b, [[I16]].8h
+; CHECK-NEXT: ret
+  %1 = fptosi<8 x half> %a to <8 x i8>
+  ret <8 x i8> %1
+}
+
+define <8 x i16> @fptosi_i16(<8 x half> %a) #0 {
+; CHECK-LABEL: fptosi_i16:
+; CHECK-DAG: fcvtl   [[LO:v[0-9]+\.4s]], v0.4h
+; CHECK-DAG: fcvtl2  [[HI:v[0-9]+\.4s]], v0.8h
+; CHECK-DAG: fcvtzs  [[LOF32:v[0-9]+\.4s]], [[LO]]
+; CHECK-DAG: xtn     [[I16:v[0-9]+]].4h, [[LOF32]]
+; CHECK-DAG: fcvtzs  [[HIF32:v[0-9]+\.4s]], [[HI]]
+; CHECK-NEXT: xtn2    [[I16]].8h, [[HIF32]]
+; CHECK-NEXT: ret
+  %1 = fptosi<8 x half> %a to <8 x i16>
+  ret <8 x i16> %1
+}
+
+define <8 x i8> @fptoui_i8(<8 x half> %a) #0 {
+; CHECK-LABEL: fptoui_i8:
+; CHECK-DAG: fcvtl   [[LO:v[0-9]+\.4s]], v0.4h
+; CHECK-DAG: fcvtl2  [[HI:v[0-9]+\.4s]], v0.8h
+; CHECK-DAG: fcvtzu  [[LOF32:v[0-9]+\.4s]], [[LO]]
+; CHECK-DAG: xtn     [[I16:v[0-9]+]].4h, [[LOF32]]
+; CHECK-DAG: fcvtzu  [[HIF32:v[0-9]+\.4s]], [[HI]]
+; CHECK-DAG: xtn2    [[I16]].8h, [[HIF32]]
+; CHECK-NEXT: xtn     v0.8b, [[I16]].8h
+; CHECK-NEXT: ret
+  %1 = fptoui<8 x half> %a to <8 x i8>
+  ret <8 x i8> %1
+}
+
+define <8 x i16> @fptoui_i16(<8 x half> %a) #0 {
+; CHECK-LABEL: fptoui_i16:
+; CHECK-DAG: fcvtl   [[LO:v[0-9]+\.4s]], v0.4h
+; CHECK-DAG: fcvtl2  [[HI:v[0-9]+\.4s]], v0.8h
+; CHECK-DAG: fcvtzu  [[LOF32:v[0-9]+\.4s]], [[LO]]
+; CHECK-DAG: xtn     [[I16:v[0-9]+]].4h, [[LOF32]]
+; CHECK-DAG: fcvtzu  [[HIF32:v[0-9]+\.4s]], [[HI]]
+; CHECK-NEXT: xtn2    [[I16]].8h, [[HIF32]]
+; CHECK-NEXT: ret
+  %1 = fptoui<8 x half> %a to <8 x i16>
+  ret <8 x i16> %1
+}
+
 attributes #0 = { nounwind }