[WebAssembly] Explicitly add {z,s}ext so extends are selected

[luke]

Summary

During DAG legalization, {u,s}itofp instructions on v2i8, v2i16, v4i8
and v4i16 types ended up being legalized into scalar instructions, when
they could just be extended to v2i32/v4i32 instead.

Fixes https://github.com/llvm/llvm-project/issues/57182

[tlively]

Nice!

[luke]

Remove accidental changes

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

performVECTOR_SHUFFLECombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
   if (!SrcType.is128BitVector() ||
       SrcType.getVectorNumElements() != DstType.getVectorNumElements())
     return SDValue();
   SDValue NewShuffle = DAG.getVectorShuffle(
       SrcType, SDLoc(N), CastOp, DAG.getUNDEF(SrcType), Shuffle->getMask());
   return DAG.getBitcast(DstType, NewShuffle);
 }
 
+/// Convert ({u,s}itofp vec) --> ({u,s}itofp ({s,z}ext vec)) so it doesn't get
+/// split up into scalar instructions during legalization, and the vector
+/// extending instructions are selected in performVectorExtendCombine below.
+static SDValue
+performVectorExtendToFPCombine(SDNode *N,
+                               TargetLowering::DAGCombinerInfo &DCI) {
+  auto &DAG = DCI.DAG;
+  assert(N->getOpcode() == ISD::UINT_TO_FP ||
+         N->getOpcode() == ISD::SINT_TO_FP);
+
+  EVT InVT = N->getOperand(0)->getValueType(0);
+  EVT ResVT = N->getValueType(0);
+  MVT ExtVT;
+  if (ResVT == MVT::v4f32 && (InVT == MVT::v4i16 || InVT == MVT::v4i8))
+    ExtVT = MVT::v4i32;
+  else if (ResVT == MVT::v2f64 && (InVT == MVT::v2i16 || InVT == MVT::v2i8))
+    ExtVT = MVT::v2i32;
+  else
+    return SDValue();
+
+  unsigned Op =
+      N->getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
+  SDValue Conv = DAG.getNode(Op, SDLoc(N), ExtVT, N->getOperand(0));
+  return DAG.getNode(N->getOpcode(), SDLoc(N), ResVT, Conv);
+}
+
 static SDValue
 performVectorExtendCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
   auto &DAG = DCI.DAG;
   assert(N->getOpcode() == ISD::SIGN_EXTEND ||
          N->getOpcode() == ISD::ZERO_EXTEND);
 
   // Combine ({s,z}ext (extract_subvector src, i)) into a widening operation if
   // possible before the extract_subvector can be expanded.
WebAssemblyTargetLowering::PerformDAGCombine(SDNode *N,
   switch (N->getOpcode()) {
   default:
     return SDValue();
   case ISD::VECTOR_SHUFFLE:
     return performVECTOR_SHUFFLECombine(N, DCI);
   case ISD::SIGN_EXTEND:
   case ISD::ZERO_EXTEND:
     return performVectorExtendCombine(N, DCI);
+  case ISD::UINT_TO_FP:
+  case ISD::SINT_TO_FP:
+    return performVectorExtendToFPCombine(N, DCI);
   case ISD::FP_TO_SINT_SAT:
   case ISD::FP_TO_UINT_SAT:
   case ISD::FP_ROUND:
   case ISD::CONCAT_VECTORS:
     return performVectorTruncZeroCombine(N, DCI);
   case ISD::TRUNCATE:
     return performTruncateCombine(N, DCI);
   }
 }

llvm/test/CodeGen/WebAssembly/simd-extending-convert.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mattr=+simd128 | FileCheck %s
 
 ; TODO: These tests should check that floating point conversions select
 ; extending instructions where possible
 
 target triple = "wasm32-unknown-unknown"
 
 define <4 x float> @extend_to_float_low_i16x8_u(<8 x i16> %x) {
 ; CHECK-LABEL: extend_to_float_low_i16x8_u:
 ; CHECK:         .functype extend_to_float_low_i16x8_u (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_u 0
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.splat
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_u 1
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 1
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_u 2
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 2
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_u 3
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 3
+; CHECK-NEXT:    i32x4.extend_low_i16x8_u
+; CHECK-NEXT:    f32x4.convert_i32x4_u
 ; CHECK-NEXT:    # fallthrough-return
   %low = shufflevector <8 x i16> %x, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %extended = uitofp <4 x i16> %low to <4 x float>
   ret <4 x float> %extended
 }
 
 define <4 x float> @extend_to_float_high_i16x8_u(<8 x i16> %x) {
 ; CHECK-LABEL: extend_to_float_high_i16x8_u:
 ; CHECK:         .functype extend_to_float_high_i16x8_u (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_u 4
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.splat
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_u 5
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 1
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_u 6
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 2
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_u 7
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 3
+; CHECK-NEXT:    i32x4.extend_high_i16x8_u
+; CHECK-NEXT:    f32x4.convert_i32x4_u
 ; CHECK-NEXT:    # fallthrough-return
   %high = shufflevector <8 x i16> %x, <8 x i16> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
   %extended = uitofp <4 x i16> %high to <4 x float>
   ret <4 x float> %extended
 }
 
 define <4 x float> @extend_to_float_low_i8x16_u(<8 x i8> %x) {
 ; CHECK-LABEL: extend_to_float_low_i8x16_u:
 ; CHECK:         .functype extend_to_float_low_i8x16_u (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    v128.const 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_u 0
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.splat
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_u 1
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 1
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_u 2
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 2
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_u 3
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 3
+; CHECK-NEXT:    i8x16.shuffle 16, 1, 2, 3, 17, 5, 6, 7, 18, 9, 10, 11, 19, 13, 14, 15
+; CHECK-NEXT:    f32x4.convert_i32x4_u
 ; CHECK-NEXT:    # fallthrough-return
   %low = shufflevector <8 x i8> %x, <8 x i8> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %extended = uitofp <4 x i8> %low to <4 x float>
   ret <4 x float> %extended
 }
 
 define <4 x float> @extend_to_float_high_i8x16_u(<8 x i8> %x) {
 ; CHECK-LABEL: extend_to_float_high_i8x16_u:
 ; CHECK:         .functype extend_to_float_high_i8x16_u (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_u 4
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.splat
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_u 5
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 1
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_u 6
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 2
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_u 7
-; CHECK-NEXT:    f32.convert_i32_u
-; CHECK-NEXT:    f32x4.replace_lane 3
+; CHECK-NEXT:    v128.const 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
+; CHECK-NEXT:    i8x16.shuffle 4, 17, 18, 19, 5, 21, 22, 23, 6, 25, 26, 27, 7, 29, 30, 31
+; CHECK-NEXT:    f32x4.convert_i32x4_u
 ; CHECK-NEXT:    # fallthrough-return
   %high = shufflevector <8 x i8> %x, <8 x i8> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
   %extended = uitofp <4 x i8> %high to <4 x float>
   ret <4 x float> %extended
 }
 
 define <4 x float> @extend_to_float_low_i16x8_s(<8 x i16> %x) {
 ; CHECK-LABEL: extend_to_float_low_i16x8_s:
 ; CHECK:         .functype extend_to_float_low_i16x8_s (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_s 0
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.splat
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_s 1
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 1
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_s 2
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 2
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_s 3
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 3
+; CHECK-NEXT:    i32x4.extend_low_i16x8_s
+; CHECK-NEXT:    f32x4.convert_i32x4_s
 ; CHECK-NEXT:    # fallthrough-return
   %low = shufflevector <8 x i16> %x, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %extended = sitofp <4 x i16> %low to <4 x float>
   ret <4 x float> %extended
 }
 
 define <4 x float> @extend_to_float_high_i16x8_s(<8 x i16> %x) {
 ; CHECK-LABEL: extend_to_float_high_i16x8_s:
 ; CHECK:         .functype extend_to_float_high_i16x8_s (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_s 4
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.splat
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_s 5
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 1
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_s 6
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 2
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_s 7
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 3
+; CHECK-NEXT:    i32x4.extend_high_i16x8_s
+; CHECK-NEXT:    f32x4.convert_i32x4_s
 ; CHECK-NEXT:    # fallthrough-return
   %high = shufflevector <8 x i16> %x, <8 x i16> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
   %extended = sitofp <4 x i16> %high to <4 x float>
   ret <4 x float> %extended
 }
 
 define <4 x float> @extend_to_float_low_i8x16_s(<8 x i8> %x) {
 ; CHECK-LABEL: extend_to_float_low_i8x16_s:
 ; CHECK:         .functype extend_to_float_low_i8x16_s (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_s 0
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.splat
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_s 1
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 1
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_s 2
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 2
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_s 3
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 3
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i8x16.shuffle 0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 3, 0, 0, 0
+; CHECK-NEXT:    i32.const 24
+; CHECK-NEXT:    i32x4.shl
+; CHECK-NEXT:    i32.const 24
+; CHECK-NEXT:    i32x4.shr_s
+; CHECK-NEXT:    f32x4.convert_i32x4_s
 ; CHECK-NEXT:    # fallthrough-return
   %low = shufflevector <8 x i8> %x, <8 x i8> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %extended = sitofp <4 x i8> %low to <4 x float>
   ret <4 x float> %extended
 }
 
 define <4 x float> @extend_to_float_high_i8x16_s(<8 x i8> %x) {
 ; CHECK-LABEL: extend_to_float_high_i8x16_s:
 ; CHECK:         .functype extend_to_float_high_i8x16_s (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_s 4
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.splat
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_s 5
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 1
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_s 6
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 2
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i8x16.extract_lane_s 7
-; CHECK-NEXT:    f32.convert_i32_s
-; CHECK-NEXT:    f32x4.replace_lane 3
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i8x16.shuffle 4, 0, 0, 0, 5, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0
+; CHECK-NEXT:    i32.const 24
+; CHECK-NEXT:    i32x4.shl
+; CHECK-NEXT:    i32.const 24
+; CHECK-NEXT:    i32x4.shr_s
+; CHECK-NEXT:    f32x4.convert_i32x4_s
 ; CHECK-NEXT:    # fallthrough-return
   %high = shufflevector <8 x i8> %x, <8 x i8> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
   %extended = sitofp <4 x i8> %high to <4 x float>
   ret <4 x float> %extended
 }
 
 define <2 x double> @extend_to_double_low_i32x4_u(<4 x i32> %x) {
 ; CHECK-LABEL: extend_to_double_low_i32x4_u:
 ; CHECK:         .functype extend_to_double_low_i32x4_u (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    f64x2.convert_low_i32x4_u
 ; CHECK-NEXT:    # fallthrough-return
   %low = shufflevector <4 x i32> %x, <4 x i32> undef, <2 x i32> <i32 0, i32 1>
   %extended = uitofp <2 x i32> %low to <2 x double>
   ret <2 x double> %extended
 }
 
 define <2 x double> @extend_to_double_low_i16x4_u(<4 x i16> %x) {
 ; CHECK-LABEL: extend_to_double_low_i16x4_u:
 ; CHECK:         .functype extend_to_double_low_i16x4_u (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    v128.const 0, 0, 0, 0, 0, 0, 0, 0
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_u 0
-; CHECK-NEXT:    f64.convert_i32_u
-; CHECK-NEXT:    f64x2.splat
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i16x8.extract_lane_u 1
-; CHECK-NEXT:    f64.convert_i32_u
-; CHECK-NEXT:    f64x2.replace_lane 1
+; CHECK-NEXT:    i8x16.shuffle 16, 17, 2, 3, 18, 19, 6, 7, 20, 21, 10, 11, 22, 23, 14, 15
+; CHECK-NEXT:    f64x2.convert_low_i32x4_u
 ; CHECK-NEXT:    # fallthrough-return
   %low = shufflevector <4 x i16> %x, <4 x i16> undef, <2 x i32> <i32 0, i32 1>
   %extended = uitofp <2 x i16> %low to <2 x double>
   ret <2 x double> %extended
 }