[WebAssembly] Support promoting lower lanes of f16x8 to f32x4.

Author: brendandahl

clang/lib/Headers/wasm_simd128.h

@@ -45,6 +45,7 @@ typedef int __i32x2 __attribute__((__vector_size__(8), __aligned__(8)));
 typedef unsigned int __u32x2
     __attribute__((__vector_size__(8), __aligned__(8)));
 typedef float __f32x2 __attribute__((__vector_size__(8), __aligned__(8)));
+typedef __fp16 __f16x4 __attribute__((__vector_size__(8), __aligned__(8)));
 
 #define __DEFAULT_FN_ATTRS                                                     \
   __attribute__((__always_inline__, __nodebug__, __target__("simd128"),        \
@@ -2010,6 +2011,13 @@ static __inline__ v128_t __FP16_FN_ATTRS wasm_f16x8_convert_u16x8(v128_t __a) {
   return (v128_t) __builtin_convertvector((__u16x8)__a, __f16x8);
 }
 
+static __inline__ v128_t __FP16_FN_ATTRS
+wasm_f32x4_promote_low_f16x8(v128_t __a) {
+  return (v128_t) __builtin_convertvector(
+      (__f16x4){((__f16x8)__a)[0], ((__f16x8)__a)[1],
+                ((__f16x8)__a)[2], ((__f16x8)__a)[3]}, __f32x4);
+}
+
 static __inline__ v128_t __FP16_FN_ATTRS wasm_f16x8_relaxed_madd(v128_t __a,
                                                                  v128_t __b,
                                                                  v128_t __c) {

cross-project-tests/intrinsic-header-tests/wasm_simd128.c

@@ -1033,6 +1033,12 @@ v128_t test_f64x2_promote_low_f32x4(v128_t a) {
   return wasm_f64x2_promote_low_f32x4(a);
 }
 
+// CHECK-LABEL: test_f32x4_promote_low_f16x8:
+// CHECK: f32x4.promote_low_f16x8{{$}}
+v128_t test_f32x4_promote_low_f16x8(v128_t a) {
+  return wasm_f32x4_promote_low_f16x8(a);
+}
+
 // CHECK-LABEL: test_i8x16_shuffle:
 // CHECK: i8x16.shuffle 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1,
 // 0{{$}}

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

@@ -2341,7 +2341,7 @@ WebAssemblyTargetLowering::LowerEXTEND_VECTOR_INREG(SDValue Op,
 
 static SDValue LowerConvertLow(SDValue Op, SelectionDAG &DAG) {
   SDLoc DL(Op);
-  if (Op.getValueType() != MVT::v2f64)
+  if (Op.getValueType() != MVT::v2f64 && Op.getValueType() != MVT::v4f32)
     return SDValue();
 
   auto GetConvertedLane = [](SDValue Op, unsigned &Opcode, SDValue &SrcVec,
@@ -2354,6 +2354,7 @@ static SDValue LowerConvertLow(SDValue Op, SelectionDAG &DAG) {
       Opcode = WebAssemblyISD::CONVERT_LOW_U;
       break;
     case ISD::FP_EXTEND:
+    case ISD::FP16_TO_FP:
       Opcode = WebAssemblyISD::PROMOTE_LOW;
       break;
     default:
@@ -2372,36 +2373,52 @@ static SDValue LowerConvertLow(SDValue Op, SelectionDAG &DAG) {
     return true;
   };
 
-  unsigned LHSOpcode, RHSOpcode, LHSIndex, RHSIndex;
-  SDValue LHSSrcVec, RHSSrcVec;
-  if (!GetConvertedLane(Op.getOperand(0), LHSOpcode, LHSSrcVec, LHSIndex) ||
-      !GetConvertedLane(Op.getOperand(1), RHSOpcode, RHSSrcVec, RHSIndex))
+  unsigned NumLanes = Op.getValueType() == MVT::v2f64 ? 2 : 4;
+  unsigned FirstOpcode = 0, SecondOpcode = 0, ThirdOpcode = 0, FourthOpcode = 0;
+  unsigned FirstIndex = 0, SecondIndex = 0, ThirdIndex = 0, FourthIndex = 0;
+  SDValue FirstSrcVec, SecondSrcVec, ThirdSrcVec, FourthSrcVec;
+
+  if (!GetConvertedLane(Op.getOperand(0), FirstOpcode, FirstSrcVec, FirstIndex) ||
+      !GetConvertedLane(Op.getOperand(1), SecondOpcode, SecondSrcVec, SecondIndex))
+    return SDValue();
+
+  // If we're converting to v4f32, check the third and fourth lanes, too.
+  if (NumLanes == 4 && (!GetConvertedLane(Op.getOperand(2), ThirdOpcode, ThirdSrcVec, ThirdIndex) ||
+                        !GetConvertedLane(Op.getOperand(3), FourthOpcode, FourthSrcVec, FourthIndex)))
+    return SDValue();
+
+  if (FirstOpcode != SecondOpcode)
     return SDValue();
 
-  if (LHSOpcode != RHSOpcode)
+  // TODO Add an optimization similar to the v2f64 below for shuffling the
+  // vectors when the lanes are in the wrong order or come from different src
+  // vectors.
+  if (NumLanes == 4 && (FirstOpcode != ThirdOpcode || FirstOpcode != FourthOpcode ||
+                        FirstSrcVec != SecondSrcVec || FirstSrcVec != ThirdSrcVec || FirstSrcVec != FourthSrcVec ||
+                        FirstIndex != 0 || SecondIndex != 1 || ThirdIndex != 2 || FourthIndex != 3))
     return SDValue();
 
   MVT ExpectedSrcVT;
-  switch (LHSOpcode) {
+  switch (FirstOpcode) {
   case WebAssemblyISD::CONVERT_LOW_S:
   case WebAssemblyISD::CONVERT_LOW_U:
     ExpectedSrcVT = MVT::v4i32;
     break;
   case WebAssemblyISD::PROMOTE_LOW:
-    ExpectedSrcVT = MVT::v4f32;
+    ExpectedSrcVT = NumLanes == 2 ? MVT::v4f32 : MVT::v8i16;
     break;
   }
-  if (LHSSrcVec.getValueType() != ExpectedSrcVT)
+  if (FirstSrcVec.getValueType() != ExpectedSrcVT)
     return SDValue();
 
-  auto Src = LHSSrcVec;
-  if (LHSIndex != 0 || RHSIndex != 1 || LHSSrcVec != RHSSrcVec) {
+  auto Src = FirstSrcVec;
+  if (NumLanes == 2 && (FirstIndex != 0 || SecondIndex != 1 || FirstSrcVec != SecondSrcVec)) {
     // Shuffle the source vector so that the converted lanes are the low lanes.
     Src = DAG.getVectorShuffle(
-        ExpectedSrcVT, DL, LHSSrcVec, RHSSrcVec,
-        {static_cast<int>(LHSIndex), static_cast<int>(RHSIndex) + 4, -1, -1});
+        ExpectedSrcVT, DL, FirstSrcVec, SecondSrcVec,
+        {static_cast<int>(FirstIndex), static_cast<int>(SecondIndex) + 4, -1, -1});
   }
-  return DAG.getNode(LHSOpcode, DL, MVT::v2f64, Src);
+  return DAG.getNode(FirstOpcode, DL, NumLanes == 2 ? MVT::v2f64 : MVT::v4f32, Src);
 }
 
 SDValue WebAssemblyTargetLowering::LowerBUILD_VECTOR(SDValue Op,

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

@@ -1468,6 +1468,7 @@ defm "" : SIMDConvert<F32x4, F64x2, demote_zero,
 def promote_t : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>]>;
 def promote_low : SDNode<"WebAssemblyISD::PROMOTE_LOW", promote_t>;
 defm "" : SIMDConvert<F64x2, F32x4, promote_low, "promote_low_f32x4", 0x5f>;
+defm "" : HalfPrecisionConvert<F32x4, I16x8, promote_low, "promote_low_f16x8", 0x14b>;
 
 // Lower extending loads to load64_zero + promote_low
 def extloadv2f32 : PatFrag<(ops node:$ptr), (extload node:$ptr)> {

llvm/test/CodeGen/WebAssembly/half-precision.ll

@@ -369,3 +369,23 @@ define <8 x half> @shuffle_poison_v8f16(<8 x half> %x, <8 x half> %y) {
                i32 poison, i32 poison, i32 poison, i32 poison>
   ret <8 x half> %res
 }
+
+define <4 x float> @promote_low_v4f32(<8 x half> %x) {
+; CHECK-LABEL: promote_low_v4f32:
+; CHECK: .functype promote_low_v4f32 (v128) -> (v128){{$}}
+; CHECK-NEXT: f32x4.promote_low_f16x8 $push[[R:[0-9]+]]=, $0
+; CHECK-NEXT: return $pop[[R]]
+  %v = shufflevector <8 x half> %x, <8 x half> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %a = fpext <4 x half> %v to <4 x float>
+  ret <4 x float> %a
+}
+
+define <4 x float> @promote_low_v4f32_2(<8 x half> %x) {
+; CHECK-LABEL: promote_low_v4f32_2:
+; CHECK:         .functype promote_low_v4f32_2 (v128) -> (v128)
+; CHECK-NEXT:    f32x4.promote_low_f16x8 $push[[R:[0-9]+]]=, $0
+; CHECK-NEXT: return $pop[[R]]
+  %v = fpext <8 x half> %x to <8 x float>
+  %a = shufflevector <8 x float> %v, <8 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  ret <4 x float> %a
+}

llvm/test/MC/WebAssembly/simd-encodings.s

@@ -935,4 +935,7 @@ main:
     # CHECK: f16x8.convert_i16x8_u # encoding: [0xfd,0xc8,0x02]
     f16x8.convert_i16x8_u
 
+    # CHECK: f32x4.promote_low_f16x8 # encoding: [0xfd,0xcb,0x02]
+    f32x4.promote_low_f16x8
+
     end_function