[WebAssembly] Implement prototype v128.load{32,64}_zero instructions

Specified in WebAssembly/simd#237, these
instructions load the first vector lane from memory and zero the other
lanes. Since these instructions are not officially part of the SIMD
proposal, they are only available on an opt-in basis via LLVM
intrinsics and clang builtin functions. If these instructions are
merged to the proposal, this implementation will change so that the
instructions will be generated from normal IR. At that point the
intrinsics and builtin functions would be removed.

This PR also changes the opcodes for the experimental f32x4.qfm{a,s}
instructions because their opcodes conflicted with those of the
v128.load{32,64}_zero instructions. The new opcodes were chosen to
match those used in V8.

Differential Revision: https://reviews.llvm.org/D84820

Author: tlively

clang/include/clang/Basic/BuiltinsWebAssembly.def

@@ -169,5 +169,8 @@ TARGET_BUILTIN(__builtin_wasm_narrow_u_i8x16_i16x8, "V16cV8sV8s", "nc", "simd128
 TARGET_BUILTIN(__builtin_wasm_narrow_s_i16x8_i32x4, "V8sV4iV4i", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_narrow_u_i16x8_i32x4, "V8sV4iV4i", "nc", "simd128")
 
+TARGET_BUILTIN(__builtin_wasm_load32_zero, "V4ii*", "nU", "simd128")
+TARGET_BUILTIN(__builtin_wasm_load64_zero, "V2LLiLLi*", "nU", "simd128")
+
 #undef BUILTIN
 #undef TARGET_BUILTIN

clang/lib/CodeGen/CGBuiltin.cpp

@@ -16852,6 +16852,16 @@ Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
         CGM.getIntrinsic(IntNo, {ConvertType(E->getType()), Low->getType()});
     return Builder.CreateCall(Callee, {Low, High});
   }
+  case WebAssembly::BI__builtin_wasm_load32_zero: {
+    Value *Ptr = EmitScalarExpr(E->getArg(0));
+    Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_load32_zero);
+    return Builder.CreateCall(Callee, {Ptr});
+  }
+  case WebAssembly::BI__builtin_wasm_load64_zero: {
+    Value *Ptr = EmitScalarExpr(E->getArg(0));
+    Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_load64_zero);
+    return Builder.CreateCall(Callee, {Ptr});
+  }
   case WebAssembly::BI__builtin_wasm_shuffle_v8x16: {
     Value *Ops[18];
     size_t OpIdx = 0;

clang/test/CodeGen/builtins-wasm.c

@@ -737,6 +737,18 @@ i16x8 narrow_u_i16x8_i32x4(i32x4 low, i32x4 high) {
   // WEBASSEMBLY: ret
 }
 
+i32x4 load32_zero(int *p) {
+  return __builtin_wasm_load32_zero(p);
+  // WEBASSEMBLY: call <4 x i32> @llvm.wasm.load32.zero(i32* %p)
+  // WEBASSEMBLY: ret
+}
+
+i64x2 load64_zero(long long *p) {
+  return __builtin_wasm_load64_zero(p);
+  // WEBASSEMBLY: call <2 x i64> @llvm.wasm.load64.zero(i64* %p)
+  // WEBASSEMBLY: ret
+}
+
 i8x16 swizzle_v8x16(i8x16 x, i8x16 y) {
   return __builtin_wasm_swizzle_v8x16(x, y);
   // WEBASSEMBLY: call <16 x i8> @llvm.wasm.swizzle(<16 x i8> %x, <16 x i8> %y)

llvm/include/llvm/IR/IntrinsicsWebAssembly.td

@@ -190,6 +190,20 @@ def int_wasm_nearest :
             [LLVMMatchType<0>],
             [IntrNoMem, IntrSpeculatable]>;
 
+// TODO: Replace these intrinsic with normal ISel patterns once the
+// load_zero instructions are merged to the proposal.
+def int_wasm_load32_zero :
+  Intrinsic<[llvm_v4i32_ty],
+            [LLVMPointerType<llvm_i32_ty>],
+            [IntrReadMem, IntrArgMemOnly],
+             "", [SDNPMemOperand]>;
+
+def int_wasm_load64_zero :
+  Intrinsic<[llvm_v2i64_ty],
+            [LLVMPointerType<llvm_i64_ty>],
+            [IntrReadMem, IntrArgMemOnly],
+             "", [SDNPMemOperand]>;
+
 //===----------------------------------------------------------------------===//
 // Thread-local storage intrinsics
 //===----------------------------------------------------------------------===//

llvm/lib/Target/WebAssembly/MCTargetDesc/WebAssemblyMCTargetDesc.h

@@ -232,6 +232,7 @@ inline unsigned GetDefaultP2AlignAny(unsigned Opc) {
   WASM_LOAD_STORE(ATOMIC_NOTIFY)
   WASM_LOAD_STORE(ATOMIC_WAIT_I32)
   WASM_LOAD_STORE(LOAD_SPLAT_v32x4)
+  WASM_LOAD_STORE(LOAD_ZERO_v4i32)
     return 2;
   WASM_LOAD_STORE(LOAD_I64)
   WASM_LOAD_STORE(LOAD_F64)
@@ -254,6 +255,7 @@ inline unsigned GetDefaultP2AlignAny(unsigned Opc) {
   WASM_LOAD_STORE(LOAD_EXTEND_U_v4i32)
   WASM_LOAD_STORE(LOAD_EXTEND_S_v2i64)
   WASM_LOAD_STORE(LOAD_EXTEND_U_v2i64)
+  WASM_LOAD_STORE(LOAD_ZERO_v2i64)
     return 3;
   WASM_LOAD_STORE(LOAD_V128)
   WASM_LOAD_STORE(STORE_V128)

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

@@ -675,6 +675,15 @@ bool WebAssemblyTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
     Info.align = Align(8);
     Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
     return true;
+  case Intrinsic::wasm_load32_zero:
+  case Intrinsic::wasm_load64_zero:
+    Info.opc = ISD::INTRINSIC_W_CHAIN;
+    Info.memVT = Intrinsic == Intrinsic::wasm_load32_zero ? MVT::i32 : MVT::i64;
+    Info.ptrVal = I.getArgOperand(0);
+    Info.offset = 0;
+    Info.align = Info.memVT == MVT::i32 ? Align(4) : Align(8);
+    Info.flags = MachineMemOperand::MOLoad;
+    return true;
   default:
     return false;
   }

llvm/lib/Target/WebAssembly/WebAssemblyInstrMemory.td

@@ -70,7 +70,7 @@ defm LOAD_F64 : WebAssemblyLoad<F64, "f64.load", 0x2b, []>;
 multiclass LoadPatNoOffset<ValueType ty, PatFrag kind, string inst> {
   def : Pat<(ty (kind I32:$addr)), (!cast<NI>(inst # "_A32") 0, 0, I32:$addr)>,
         Requires<[HasAddr32]>;
-  def : Pat<(ty (kind I64:$addr)), (!cast<NI>(inst # "_A64") 0, 0, I64:$addr)>,
+  def : Pat<(ty (kind (i64 I64:$addr))), (!cast<NI>(inst # "_A64") 0, 0, I64:$addr)>,
         Requires<[HasAddr64]>;
 }
 

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

@@ -163,6 +163,43 @@ defm : LoadPatGlobalAddrOffOnly<types[0], !cast<PatFrag>(exts[0]#types[1]),
                                 "LOAD_EXTEND"#exts[1]#"_"#types[0]>;
 }
 
+// Load lane into zero vector
+multiclass SIMDLoadZero<ValueType vec_t, string name, bits<32> simdop> {
+  let mayLoad = 1, UseNamedOperandTable = 1 in {
+  defm LOAD_ZERO_#vec_t#_A32 :
+    SIMD_I<(outs V128:$dst),
+           (ins P2Align:$p2align, offset32_op:$off, I32:$addr),
+           (outs), (ins P2Align:$p2align, offset32_op:$off), [],
+           name#"\t$dst, ${off}(${addr})$p2align",
+           name#"\t$off$p2align", simdop>;
+  defm LOAD_ZERO_#vec_t#_A64 :
+    SIMD_I<(outs V128:$dst),
+           (ins P2Align:$p2align, offset64_op:$off, I64:$addr),
+           (outs), (ins P2Align:$p2align, offset64_op:$off), [],
+           name#"\t$dst, ${off}(${addr})$p2align",
+           name#"\t$off$p2align", simdop>;
+  } // mayLoad = 1, UseNamedOperandTable = 1
+}
+
+// TODO: Also support v4f32 and v2f64 once the instructions are merged
+// to the proposal
+defm "" : SIMDLoadZero<v4i32, "v128.load32_zero", 252>;
+defm "" : SIMDLoadZero<v2i64, "v128.load64_zero", 253>;
+
+defm : LoadPatNoOffset<v4i32, int_wasm_load32_zero, "LOAD_ZERO_v4i32">;
+defm : LoadPatNoOffset<v2i64, int_wasm_load64_zero, "LOAD_ZERO_v2i64">;
+
+defm : LoadPatImmOff<v4i32, int_wasm_load32_zero, regPlusImm, "LOAD_ZERO_v4i32">;
+defm : LoadPatImmOff<v2i64, int_wasm_load64_zero, regPlusImm, "LOAD_ZERO_v2i64">;
+
+defm : LoadPatImmOff<v4i32, int_wasm_load32_zero, or_is_add, "LOAD_ZERO_v4i32">;
+defm : LoadPatImmOff<v2i64, int_wasm_load64_zero, or_is_add, "LOAD_ZERO_v2i64">;
+
+defm : LoadPatOffsetOnly<v4i32, int_wasm_load32_zero, "LOAD_ZERO_v4i32">;
+defm : LoadPatOffsetOnly<v2i64, int_wasm_load64_zero, "LOAD_ZERO_v2i64">;
+
+defm : LoadPatGlobalAddrOffOnly<v4i32, int_wasm_load32_zero, "LOAD_ZERO_v4i32">;
+defm : LoadPatGlobalAddrOffOnly<v2i64, int_wasm_load64_zero, "LOAD_ZERO_v2i64">;
 
 // Store: v128.store
 let mayStore = 1, UseNamedOperandTable = 1 in {
@@ -800,7 +837,7 @@ let isCommutable = 1 in
 defm DOT : SIMD_I<(outs V128:$dst), (ins V128:$lhs, V128:$rhs), (outs), (ins),
                   [(set V128:$dst, (int_wasm_dot V128:$lhs, V128:$rhs))],
                   "i32x4.dot_i16x8_s\t$dst, $lhs, $rhs", "i32x4.dot_i16x8_s",
-                  180>;
+                  186>;
 
 //===----------------------------------------------------------------------===//
 // Floating-point unary arithmetic
@@ -1038,20 +1075,21 @@ def : Pat<(t1 (bitconvert (t2 V128:$v))), (t1 V128:$v)>;
 // Quasi-Fused Multiply- Add and Subtract (QFMA/QFMS)
 //===----------------------------------------------------------------------===//
 
-multiclass SIMDQFM<ValueType vec_t, string vec, bits<32> baseInst> {
+multiclass SIMDQFM<ValueType vec_t, string vec, bits<32> simdopA,
+                   bits<32> simdopS> {
   defm QFMA_#vec_t :
     SIMD_I<(outs V128:$dst), (ins V128:$a, V128:$b, V128:$c),
            (outs), (ins),
            [(set (vec_t V128:$dst),
              (int_wasm_qfma (vec_t V128:$a), (vec_t V128:$b), (vec_t V128:$c)))],
-           vec#".qfma\t$dst, $a, $b, $c", vec#".qfma", baseInst>;
+           vec#".qfma\t$dst, $a, $b, $c", vec#".qfma", simdopA>;
   defm QFMS_#vec_t :
     SIMD_I<(outs V128:$dst), (ins V128:$a, V128:$b, V128:$c),
            (outs), (ins),
            [(set (vec_t V128:$dst),
              (int_wasm_qfms (vec_t V128:$a), (vec_t V128:$b), (vec_t V128:$c)))],
-           vec#".qfms\t$dst, $a, $b, $c", vec#".qfms", !add(baseInst, 1)>;
+           vec#".qfms\t$dst, $a, $b, $c", vec#".qfms", simdopS>;
 }
 
-defm "" : SIMDQFM<v4f32, "f32x4", 252>;
-defm "" : SIMDQFM<v2f64, "f64x2", 254>;
+defm "" : SIMDQFM<v4f32, "f32x4", 180, 212>;
+defm "" : SIMDQFM<v2f64, "f64x2", 254, 255>;