[llvm] [WebAssembly] Support BUILD_VECTOR with F16x8. (PR #108117)

[Brendan Dahl via llvm-commits]

https://github.com/brendandahl created https://github.com/llvm/llvm-project/pull/108117

Convert BUILD_VECTORS with FP16x8 to I16x8 since there's no FP16 scalar value to intialize v128.const.

>From 9791930d57ca81d30d1bf33d762ccdd9053bbc0f Mon Sep 17 00:00:00 2001
From: Brendan Dahl <brendan.dahl at gmail.com>
Date: Tue, 10 Sep 2024 23:52:22 +0000
Subject: [PATCH] [WebAssembly] Support BUILD_VECTOR with F16x8.

Convert BUILD_VECTORS with FP16x8 to I16x8 since there's no FP16
scalar value to intialize v128.const.
---
 .../WebAssembly/WebAssemblyISelLowering.cpp       | 15 +++++++++++++++
 llvm/test/CodeGen/WebAssembly/half-precision.ll   |  7 +++++++
 2 files changed, 22 insertions(+)

diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp b/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
index 5cc084f3ab1387..3d7b121c182fab 100644
--- a/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
@@ -208,6 +208,9 @@ WebAssemblyTargetLowering::WebAssemblyTargetLowering(
                    MVT::v2f64})
       setOperationAction(ISD::BUILD_VECTOR, T, Custom);
 
+    if (Subtarget->hasFP16())
+      setOperationAction(ISD::BUILD_VECTOR, MVT::f16, Custom);
+
     // We have custom shuffle lowering to expose the shuffle mask
     for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
                    MVT::v2f64})
@@ -2055,6 +2058,18 @@ static SDValue LowerConvertLow(SDValue Op, SelectionDAG &DAG) {
 
 SDValue WebAssemblyTargetLowering::LowerBUILD_VECTOR(SDValue Op,
                                                      SelectionDAG &DAG) const {
+  MVT VT = Op.getSimpleValueType();
+  if (VT == MVT::v8f16) {
+    // BUILD_VECTOR can't handle FP16 operands since Wasm doesn't have a scaler
+    // FP16 type, so cast them to I16s.
+    MVT IVT = VT.changeVectorElementType(MVT::i16);
+    SmallVector<SDValue, 8> NewOps;
+    for (unsigned I = 0, E = Op.getNumOperands(); I < E; ++I)
+      NewOps.push_back(DAG.getBitcast(MVT::i16, Op.getOperand(I)));
+    SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(), IVT, NewOps);
+    return DAG.getBitcast(VT, Res);
+  }
+
   if (auto ConvertLow = LowerConvertLow(Op, DAG))
     return ConvertLow;
 
diff --git a/llvm/test/CodeGen/WebAssembly/half-precision.ll b/llvm/test/CodeGen/WebAssembly/half-precision.ll
index c0b14d2064d5eb..5a0bc56886956a 100644
--- a/llvm/test/CodeGen/WebAssembly/half-precision.ll
+++ b/llvm/test/CodeGen/WebAssembly/half-precision.ll
@@ -28,6 +28,13 @@ define <8 x half> @splat_v8f16(float %x) {
   ret <8 x half> %v
 }
 
+; CHECK-LABEL: const_splat_v8f16:
+; CHECK:       v128.const      $push0=, 20800, 0, 0, 0, 0, 0, 0, 20800
+; CHECK-NEXT:  return $pop0
+define <8 x half> @const_splat_v8f16() {
+  ret <8 x half> <half 42., half 0., half 0., half 0., half 0., half 0., half 0., half 42.>
+}
+
 ; CHECK-LABEL: extract_lane_v8f16:
 ; CHECK:       f16x8.extract_lane $push0=, $0, 1
 ; CHECK-NEXT:  return $pop0