[WebAssembly] Narrowing and widening SIMD ops

Summary:
Implements target-specific LLVM intrinsics and clang builtins for
these new SIMD operations, as described at https://github.com/WebAssembly/simd/blob/master/proposals/simd/SIMD.md#integer-to-integer-narrowing.

Reviewers: aheejin

Subscribers: dschuff, sbc100, jgravelle-google, hiraditya, sunfish, cfe-commits, llvm-commits

Tags: #clang, #llvm

Differential Revision: https://reviews.llvm.org/D67425

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@371906 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/IR/IntrinsicsWebAssembly.td b/include/llvm/IR/IntrinsicsWebAssembly.td
index 73d190b0b550ebbeb05e539f58ed022cab2688c0..6b5d0cbf799abb794a53189e0594a9b6baa74cc1 100644 (file)
--- a/include/llvm/IR/IntrinsicsWebAssembly.td
+++ b/include/llvm/IR/IntrinsicsWebAssembly.td
@@ -117,6 +117,31 @@ def int_wasm_qfms :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],
             [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_narrow_signed :
+  Intrinsic<[llvm_anyvector_ty],
+            [llvm_anyvector_ty, LLVMMatchType<1>],
+            [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_narrow_unsigned :
+  Intrinsic<[llvm_anyvector_ty],
+            [llvm_anyvector_ty, LLVMMatchType<1>],
+            [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_widen_low_signed :
+  Intrinsic<[llvm_anyvector_ty],
+            [llvm_anyvector_ty],
+            [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_widen_high_signed :
+  Intrinsic<[llvm_anyvector_ty],
+            [llvm_anyvector_ty],
+            [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_widen_low_unsigned :
+  Intrinsic<[llvm_anyvector_ty],
+            [llvm_anyvector_ty],
+            [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_widen_high_unsigned :
+  Intrinsic<[llvm_anyvector_ty],
+            [llvm_anyvector_ty],
+            [IntrNoMem, IntrSpeculatable]>;
+
 
 //===----------------------------------------------------------------------===//
 // Bulk memory intrinsics

diff --git a/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td b/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
index ff03167595900642593c48c164697fbc9b586051..64fb77136cd07624a0ea7448caa9a816c74d54d8 100644 (file)
--- a/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
+++ b/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
@@ -712,6 +712,42 @@ defm "" : SIMDConvert<v4i32, v4f32, fp_to_uint, "i32x4.trunc_sat_f32x4_u", 172>;
 defm "" : SIMDConvert<v2i64, v2f64, fp_to_sint, "i64x2.trunc_sat_f64x2_s", 173>;
 defm "" : SIMDConvert<v2i64, v2f64, fp_to_uint, "i64x2.trunc_sat_f64x2_u", 174>;
 
+// Widening operations
+multiclass SIMDWiden<ValueType vec_t, string vec, ValueType arg_t, string arg,
+                     bits<32> baseInst> {
+  defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_low_signed,
+                        vec#".widen_low_"#arg#"_s", baseInst>;
+  defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_high_signed,
+                        vec#".widen_high_"#arg#"_s", !add(baseInst, 1)>;
+  defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_low_unsigned,
+                        vec#".widen_low_"#arg#"_u", !add(baseInst, 2)>;
+  defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_high_unsigned,
+                        vec#".widen_high_"#arg#"_u", !add(baseInst, 3)>;
+}
+
+defm "" : SIMDWiden<v8i16, "i16x8", v16i8, "i8x16", 202>;
+defm "" : SIMDWiden<v4i32, "i32x4", v8i16, "i16x8", 206>;
+
+// Narrowing operations
+multiclass SIMDNarrow<ValueType vec_t, string vec, ValueType arg_t, string arg,
+                      bits<32> baseInst> {
+  defm NARROW_S_#vec_t :
+    SIMD_I<(outs V128:$dst), (ins V128:$low, V128:$high), (outs), (ins),
+           [(set (vec_t V128:$dst), (vec_t (int_wasm_narrow_signed
+             (arg_t V128:$low), (arg_t V128:$high))))],
+           vec#".narrow_"#arg#"_s\t$dst, $low, $high", vec#".narrow_"#arg#"_s",
+           baseInst>;
+  defm NARROW_U_#vec_t :
+    SIMD_I<(outs V128:$dst), (ins V128:$low, V128:$high), (outs), (ins),
+           [(set (vec_t V128:$dst), (vec_t (int_wasm_narrow_unsigned
+             (arg_t V128:$low), (arg_t V128:$high))))],
+           vec#".narrow_"#arg#"_u\t$dst, $low, $high", vec#".narrow_"#arg#"_u",
+           !add(baseInst, 1)>;
+}
+
+defm "" : SIMDNarrow<v16i8, "i8x16", v8i16, "i16x8", 198>;
+defm "" : SIMDNarrow<v8i16, "i16x8", v4i32, "i32x4", 200>;
+
 // Lower llvm.wasm.trunc.saturate.* to saturating instructions
 def : Pat<(v4i32 (int_wasm_trunc_saturate_signed (v4f32 V128:$src))),
           (fp_to_sint_v4i32_v4f32 (v4f32 V128:$src))>;

diff --git a/test/CodeGen/WebAssembly/simd-intrinsics.ll b/test/CodeGen/WebAssembly/simd-intrinsics.ll
index 2077cf8c70e69f2d4977a08f6c366c5fb1b81f23..f88b8b454268c61cdd87bb8d8030f694704a481e 100644 (file)
--- a/test/CodeGen/WebAssembly/simd-intrinsics.ll
+++ b/test/CodeGen/WebAssembly/simd-intrinsics.ll
@@ -87,6 +87,30 @@ define <16 x i8> @bitselect_v16i8(<16 x i8> %v1, <16 x i8> %v2, <16 x i8> %c) {
   ret <16 x i8> %a
 }
 
+; CHECK-LABEL: narrow_signed_v16i8:
+; SIMD128-NEXT: .functype narrow_signed_v16i8 (v128, v128) -> (v128){{$}}
+; SIMD128-NEXT: i8x16.narrow_i16x8_s $push[[R:[0-9]+]]=, $0, $1{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <16 x i8> @llvm.wasm.narrow.signed.v16i8.v8i16(<8 x i16>, <8 x i16>)
+define <16 x i8> @narrow_signed_v16i8(<8 x i16> %low, <8 x i16> %high) {
+  %a = call <16 x i8> @llvm.wasm.narrow.signed.v16i8.v8i16(
+    <8 x i16> %low, <8 x i16> %high
+  )
+  ret <16 x i8> %a
+}
+
+; CHECK-LABEL: narrow_unsigned_v16i8:
+; SIMD128-NEXT: .functype narrow_unsigned_v16i8 (v128, v128) -> (v128){{$}}
+; SIMD128-NEXT: i8x16.narrow_i16x8_u $push[[R:[0-9]+]]=, $0, $1{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <16 x i8> @llvm.wasm.narrow.unsigned.v16i8.v8i16(<8 x i16>, <8 x i16>)
+define <16 x i8> @narrow_unsigned_v16i8(<8 x i16> %low, <8 x i16> %high) {
+  %a = call <16 x i8> @llvm.wasm.narrow.unsigned.v16i8.v8i16(
+    <8 x i16> %low, <8 x i16> %high
+  )
+  ret <16 x i8> %a
+}
+
 ; ==============================================================================
 ; 8 x i16
 ; ==============================================================================
@@ -166,6 +190,70 @@ define <8 x i16> @bitselect_v8i16(<8 x i16> %v1, <8 x i16> %v2, <8 x i16> %c) {
   ret <8 x i16> %a
 }
 
+; CHECK-LABEL: narrow_signed_v8i16:
+; SIMD128-NEXT: .functype narrow_signed_v8i16 (v128, v128) -> (v128){{$}}
+; SIMD128-NEXT: i16x8.narrow_i32x4_s $push[[R:[0-9]+]]=, $0, $1{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <8 x i16> @llvm.wasm.narrow.signed.v8i16.v4i32(<4 x i32>, <4 x i32>)
+define <8 x i16> @narrow_signed_v8i16(<4 x i32> %low, <4 x i32> %high) {
+  %a = call <8 x i16> @llvm.wasm.narrow.signed.v8i16.v4i32(
+    <4 x i32> %low, <4 x i32> %high
+  )
+  ret <8 x i16> %a
+}
+
+; CHECK-LABEL: narrow_unsigned_v8i16:
+; SIMD128-NEXT: .functype narrow_unsigned_v8i16 (v128, v128) -> (v128){{$}}
+; SIMD128-NEXT: i16x8.narrow_i32x4_u $push[[R:[0-9]+]]=, $0, $1{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <8 x i16> @llvm.wasm.narrow.unsigned.v8i16.v4i32(<4 x i32>, <4 x i32>)
+define <8 x i16> @narrow_unsigned_v8i16(<4 x i32> %low, <4 x i32> %high) {
+  %a = call <8 x i16> @llvm.wasm.narrow.unsigned.v8i16.v4i32(
+    <4 x i32> %low, <4 x i32> %high
+  )
+  ret <8 x i16> %a
+}
+
+; CHECK-LABEL: widen_low_signed_v8i16:
+; SIMD128-NEXT: .functype widen_low_signed_v8i16 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i16x8.widen_low_i8x16_s $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <8 x i16> @llvm.wasm.widen.low.signed.v8i16.v16i8(<16 x i8>)
+define <8 x i16> @widen_low_signed_v8i16(<16 x i8> %v) {
+  %a = call <8 x i16> @llvm.wasm.widen.low.signed.v8i16.v16i8(<16 x i8> %v)
+  ret <8 x i16> %a
+}
+
+; CHECK-LABEL: widen_high_signed_v8i16:
+; SIMD128-NEXT: .functype widen_high_signed_v8i16 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i16x8.widen_high_i8x16_s $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <8 x i16> @llvm.wasm.widen.high.signed.v8i16.v16i8(<16 x i8>)
+define <8 x i16> @widen_high_signed_v8i16(<16 x i8> %v) {
+  %a = call <8 x i16> @llvm.wasm.widen.high.signed.v8i16.v16i8(<16 x i8> %v)
+  ret <8 x i16> %a
+}
+
+; CHECK-LABEL: widen_low_unsigned_v8i16:
+; SIMD128-NEXT: .functype widen_low_unsigned_v8i16 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i16x8.widen_low_i8x16_u $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <8 x i16> @llvm.wasm.widen.low.unsigned.v8i16.v16i8(<16 x i8>)
+define <8 x i16> @widen_low_unsigned_v8i16(<16 x i8> %v) {
+  %a = call <8 x i16> @llvm.wasm.widen.low.unsigned.v8i16.v16i8(<16 x i8> %v)
+  ret <8 x i16> %a
+}
+
+; CHECK-LABEL: widen_high_unsigned_v8i16:
+; SIMD128-NEXT: .functype widen_high_unsigned_v8i16 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i16x8.widen_high_i8x16_u $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <8 x i16> @llvm.wasm.widen.high.unsigned.v8i16.v16i8(<16 x i8>)
+define <8 x i16> @widen_high_unsigned_v8i16(<16 x i8> %v) {
+  %a = call <8 x i16> @llvm.wasm.widen.high.unsigned.v8i16.v16i8(<16 x i8> %v)
+  ret <8 x i16> %a
+}
+
 ; ==============================================================================
 ; 4 x i32
 ; ==============================================================================
@@ -223,6 +311,46 @@ define <4 x i32> @trunc_sat_u_v4i32(<4 x float> %x) {
   ret <4 x i32> %a
 }
 
+; CHECK-LABEL: widen_low_signed_v4i32:
+; SIMD128-NEXT: .functype widen_low_signed_v4i32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i32x4.widen_low_i16x8_s $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x i32> @llvm.wasm.widen.low.signed.v4i32.v8i16(<8 x i16>)
+define <4 x i32> @widen_low_signed_v4i32(<8 x i16> %v) {
+  %a = call <4 x i32> @llvm.wasm.widen.low.signed.v4i32.v8i16(<8 x i16> %v)
+  ret <4 x i32> %a
+}
+
+; CHECK-LABEL: widen_high_signed_v4i32:
+; SIMD128-NEXT: .functype widen_high_signed_v4i32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i32x4.widen_high_i16x8_s $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x i32> @llvm.wasm.widen.high.signed.v4i32.v8i16(<8 x i16>)
+define <4 x i32> @widen_high_signed_v4i32(<8 x i16> %v) {
+  %a = call <4 x i32> @llvm.wasm.widen.high.signed.v4i32.v8i16(<8 x i16> %v)
+  ret <4 x i32> %a
+}
+
+; CHECK-LABEL: widen_low_unsigned_v4i32:
+; SIMD128-NEXT: .functype widen_low_unsigned_v4i32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i32x4.widen_low_i16x8_u $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x i32> @llvm.wasm.widen.low.unsigned.v4i32.v8i16(<8 x i16>)
+define <4 x i32> @widen_low_unsigned_v4i32(<8 x i16> %v) {
+  %a = call <4 x i32> @llvm.wasm.widen.low.unsigned.v4i32.v8i16(<8 x i16> %v)
+  ret <4 x i32> %a
+}
+
+; CHECK-LABEL: widen_high_unsigned_v4i32:
+; SIMD128-NEXT: .functype widen_high_unsigned_v4i32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i32x4.widen_high_i16x8_u $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x i32> @llvm.wasm.widen.high.unsigned.v4i32.v8i16(<8 x i16>)
+define <4 x i32> @widen_high_unsigned_v4i32(<8 x i16> %v) {
+  %a = call <4 x i32> @llvm.wasm.widen.high.unsigned.v4i32.v8i16(<8 x i16> %v)
+  ret <4 x i32> %a
+}
+
 ; ==============================================================================
 ; 2 x i64
 ; ==============================================================================

diff --git a/test/MC/WebAssembly/simd-encodings.s b/test/MC/WebAssembly/simd-encodings.s
index 491b4844d7fa53593ce5c2594decad7dbbe236bc..c1c5243a6cabf46ea5a7a3c553f8ea36470072e5 100644 (file)
--- a/test/MC/WebAssembly/simd-encodings.s
+++ b/test/MC/WebAssembly/simd-encodings.s
@@ -463,4 +463,40 @@ main:
     # CHECK: f64x2.convert_i64x2_u # encoding: [0xfd,0xb2,0x01]
     f64x2.convert_i64x2_u
 
+    # CHECK: i8x16.narrow_i16x8_s # encoding: [0xfd,0xc6,0x01]
+    i8x16.narrow_i16x8_s
+
+    # CHECK: i8x16.narrow_i16x8_u # encoding: [0xfd,0xc7,0x01]
+    i8x16.narrow_i16x8_u
+
+    # CHECK: i16x8.narrow_i32x4_s # encoding: [0xfd,0xc8,0x01]
+    i16x8.narrow_i32x4_s
+
+    # CHECK: i16x8.narrow_i32x4_u # encoding: [0xfd,0xc9,0x01]
+    i16x8.narrow_i32x4_u
+
+    # CHECK: i16x8.widen_low_i8x16_s # encoding: [0xfd,0xca,0x01]
+    i16x8.widen_low_i8x16_s
+
+    # CHECK: i16x8.widen_high_i8x16_s # encoding: [0xfd,0xcb,0x01]
+    i16x8.widen_high_i8x16_s
+
+    # CHECK: i16x8.widen_low_i8x16_u # encoding: [0xfd,0xcc,0x01]
+    i16x8.widen_low_i8x16_u
+
+    # CHECK: i16x8.widen_high_i8x16_u # encoding: [0xfd,0xcd,0x01]
+    i16x8.widen_high_i8x16_u
+
+    # CHECK: i32x4.widen_low_i16x8_s # encoding: [0xfd,0xce,0x01]
+    i32x4.widen_low_i16x8_s
+
+    # CHECK: i32x4.widen_high_i16x8_s # encoding: [0xfd,0xcf,0x01]
+    i32x4.widen_high_i16x8_s
+
+    # CHECK: i32x4.widen_low_i16x8_u # encoding: [0xfd,0xd0,0x01]
+    i32x4.widen_low_i16x8_u
+
+    # CHECK: i32x4.widen_high_i16x8_u # encoding: [0xfd,0xd1,0x01]
+    i32x4.widen_high_i16x8_u
+
     end_function