S390 [liftoff]: Implement simd replace lane ops

Change-Id: I00da20528553e4135681790998c03126931bca9a
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3042719
Commit-Queue: Milad Fa <mfarazma@redhat.com>
Reviewed-by: Michael Achenbach <machenbach@chromium.org>
Reviewed-by: Junliang Yan <junyan@redhat.com>
Cr-Commit-Position: refs/heads/master@{#75883}

src/codegen/s390/macro-assembler-s390.cc

@@ -5158,6 +5158,58 @@ void TurboAssembler::I8x16ExtractLaneS(Register dst, Simd128Register src,
   lgbr(dst, r0);
 }
 
+void TurboAssembler::F64x2ReplaceLane(Simd128Register dst, Simd128Register src1,
+                                      DoubleRegister src2,
+                                      uint8_t imm_lane_idx) {
+  vlgv(r0, src2, MemOperand(r0, 0), Condition(3));
+  if (src1 != dst) {
+    vlr(dst, src1, Condition(0), Condition(0), Condition(0));
+  }
+  vlvg(dst, r0, MemOperand(r0, 1 - imm_lane_idx), Condition(3));
+}
+
+void TurboAssembler::F32x4ReplaceLane(Simd128Register dst, Simd128Register src1,
+                                      DoubleRegister src2,
+                                      uint8_t imm_lane_idx) {
+  vlgv(r0, src2, MemOperand(r0, 0), Condition(2));
+  if (src1 != dst) {
+    vlr(dst, src1, Condition(0), Condition(0), Condition(0));
+  }
+  vlvg(dst, r0, MemOperand(r0, 3 - imm_lane_idx), Condition(2));
+}
+
+void TurboAssembler::I64x2ReplaceLane(Simd128Register dst, Simd128Register src1,
+                                      Register src2, uint8_t imm_lane_idx) {
+  if (src1 != dst) {
+    vlr(dst, src1, Condition(0), Condition(0), Condition(0));
+  }
+  vlvg(dst, src2, MemOperand(r0, 1 - imm_lane_idx), Condition(3));
+}
+
+void TurboAssembler::I32x4ReplaceLane(Simd128Register dst, Simd128Register src1,
+                                      Register src2, uint8_t imm_lane_idx) {
+  if (src1 != dst) {
+    vlr(dst, src1, Condition(0), Condition(0), Condition(0));
+  }
+  vlvg(dst, src2, MemOperand(r0, 3 - imm_lane_idx), Condition(2));
+}
+
+void TurboAssembler::I16x8ReplaceLane(Simd128Register dst, Simd128Register src1,
+                                      Register src2, uint8_t imm_lane_idx) {
+  if (src1 != dst) {
+    vlr(dst, src1, Condition(0), Condition(0), Condition(0));
+  }
+  vlvg(dst, src2, MemOperand(r0, 7 - imm_lane_idx), Condition(1));
+}
+
+void TurboAssembler::I8x16ReplaceLane(Simd128Register dst, Simd128Register src1,
+                                      Register src2, uint8_t imm_lane_idx) {
+  if (src1 != dst) {
+    vlr(dst, src1, Condition(0), Condition(0), Condition(0));
+  }
+  vlvg(dst, src2, MemOperand(r0, 15 - imm_lane_idx), Condition(0));
+}
+
 }  // namespace internal
 }  // namespace v8
 

src/codegen/s390/macro-assembler-s390.h

@@ -1054,6 +1054,18 @@ class V8_EXPORT_PRIVATE TurboAssembler : public TurboAssemblerBase {
                          uint8_t imm_lane_idx);
   void I8x16ExtractLaneS(Register dst, Simd128Register src,
                          uint8_t imm_lane_idx);
+  void F64x2ReplaceLane(Simd128Register dst, Simd128Register src1,
+                        DoubleRegister src2, uint8_t imm_lane_idx);
+  void F32x4ReplaceLane(Simd128Register dst, Simd128Register src1,
+                        DoubleRegister src2, uint8_t imm_lane_idx);
+  void I64x2ReplaceLane(Simd128Register dst, Simd128Register src1,
+                        Register src2, uint8_t imm_lane_idx);
+  void I32x4ReplaceLane(Simd128Register dst, Simd128Register src1,
+                        Register src2, uint8_t imm_lane_idx);
+  void I16x8ReplaceLane(Simd128Register dst, Simd128Register src1,
+                        Register src2, uint8_t imm_lane_idx);
+  void I8x16ReplaceLane(Simd128Register dst, Simd128Register src1,
+                        Register src2, uint8_t imm_lane_idx);
 
   // ---------------------------------------------------------------------------
   // Pointer compression Support

src/compiler/backend/s390/code-generator-s390.cc

@@ -2517,72 +2517,25 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       SIMD_EXTRACT_LANE_LIST(EMIT_SIMD_EXTRACT_LANE)
 #undef EMIT_SIMD_EXTRACT_LANE
 #undef SIMD_EXTRACT_LANE_LIST
-      // vector replace element
-    case kS390_F64x2ReplaceLane: {
-      Simd128Register src = i.InputSimd128Register(0);
-      Simd128Register dst = i.OutputSimd128Register();
-      __ vlgv(kScratchReg, i.InputDoubleRegister(2), MemOperand(r0, 0),
-              Condition(3));
-      if (src != dst) {
-        __ vlr(dst, src, Condition(0), Condition(0), Condition(0));
-      }
-      __ vlvg(dst, kScratchReg, MemOperand(r0, 1 - i.InputInt8(1)),
-              Condition(3));
-      break;
-    }
-    case kS390_F32x4ReplaceLane: {
-      Simd128Register src = i.InputSimd128Register(0);
-      Simd128Register dst = i.OutputSimd128Register();
-      __ vlgv(kScratchReg, i.InputDoubleRegister(2), MemOperand(r0, 0),
-              Condition(2));
-      if (src != dst) {
-        __ vlr(dst, src, Condition(0), Condition(0), Condition(0));
-      }
-      __ vlvg(dst, kScratchReg, MemOperand(r0, 3 - i.InputInt8(1)),
-              Condition(2));
-      break;
-    }
-    case kS390_I64x2ReplaceLane: {
-      Simd128Register src = i.InputSimd128Register(0);
-      Simd128Register dst = i.OutputSimd128Register();
-      if (src != dst) {
-        __ vlr(dst, src, Condition(0), Condition(0), Condition(0));
-      }
-      __ vlvg(dst, i.InputRegister(2), MemOperand(r0, 1 - i.InputInt8(1)),
-              Condition(3));
-      break;
-    }
-    case kS390_I32x4ReplaceLane: {
-      Simd128Register src = i.InputSimd128Register(0);
-      Simd128Register dst = i.OutputSimd128Register();
-      if (src != dst) {
-        __ vlr(dst, src, Condition(0), Condition(0), Condition(0));
-      }
-      __ vlvg(dst, i.InputRegister(2), MemOperand(r0, 3 - i.InputInt8(1)),
-              Condition(2));
-      break;
-    }
-    case kS390_I16x8ReplaceLane: {
-      Simd128Register src = i.InputSimd128Register(0);
-      Simd128Register dst = i.OutputSimd128Register();
-      if (src != dst) {
-        __ vlr(dst, src, Condition(0), Condition(0), Condition(0));
-      }
-      __ vlvg(dst, i.InputRegister(2), MemOperand(r0, 7 - i.InputInt8(1)),
-              Condition(1));
-      break;
-    }
-    case kS390_I8x16ReplaceLane: {
-      Simd128Register src = i.InputSimd128Register(0);
-      Simd128Register dst = i.OutputSimd128Register();
-      if (src != dst) {
-        __ vlr(dst, src, Condition(0), Condition(0), Condition(0));
-      }
-      __ vlvg(dst, i.InputRegister(2), MemOperand(r0, 15 - i.InputInt8(1)),
-              Condition(0));
-      break;
-    }
-    // vector binops
+
+#define SIMD_REPLACE_LANE_LIST(V)     \
+  V(F64x2ReplaceLane, DoubleRegister) \
+  V(F32x4ReplaceLane, DoubleRegister) \
+  V(I64x2ReplaceLane, Register)       \
+  V(I32x4ReplaceLane, Register)       \
+  V(I16x8ReplaceLane, Register)       \
+  V(I8x16ReplaceLane, Register)
+
+#define EMIT_SIMD_REPLACE_LANE(name, stype)                       \
+  case kS390_##name: {                                            \
+    __ name(i.OutputSimd128Register(), i.InputSimd128Register(0), \
+            i.Input##stype(2), i.InputInt8(1));                   \
+    break;                                                        \
+  }
+      SIMD_REPLACE_LANE_LIST(EMIT_SIMD_REPLACE_LANE)
+#undef EMIT_SIMD_REPLACE_LANE
+#undef SIMD_REPLACE_LANE_LIST
+      // vector binops
     case kS390_F64x2Add: {
       __ vfa(i.OutputSimd128Register(), i.InputSimd128Register(0),
              i.InputSimd128Register(1), Condition(0), Condition(0),

src/wasm/baseline/s390/liftoff-assembler-s390.h

@@ -2178,6 +2178,24 @@ SIMD_EXTRACT_LANE_LIST(EMIT_SIMD_EXTRACT_LANE)
 #undef EMIT_SIMD_EXTRACT_LANE
 #undef SIMD_EXTRACT_LANE_LIST
 
+#define SIMD_REPLACE_LANE_LIST(V)             \
+  V(f64x2_replace_lane, F64x2ReplaceLane, fp) \
+  V(f32x4_replace_lane, F32x4ReplaceLane, fp) \
+  V(i64x2_replace_lane, I64x2ReplaceLane, gp) \
+  V(i32x4_replace_lane, I32x4ReplaceLane, gp) \
+  V(i16x8_replace_lane, I16x8ReplaceLane, gp) \
+  V(i8x16_replace_lane, I8x16ReplaceLane, gp)
+
+#define EMIT_SIMD_REPLACE_LANE(name, op, stype)                        \
+  void LiftoffAssembler::emit_##name(                                  \
+      LiftoffRegister dst, LiftoffRegister src1, LiftoffRegister src2, \
+      uint8_t imm_lane_idx) {                                          \
+    op(dst.fp(), src1.fp(), src2.stype(), imm_lane_idx);               \
+  }
+SIMD_REPLACE_LANE_LIST(EMIT_SIMD_REPLACE_LANE)
+#undef EMIT_SIMD_REPLACE_LANE
+#undef SIMD_REPLACE_LANE_LIST
+
 void LiftoffAssembler::LoadTransform(LiftoffRegister dst, Register src_addr,
                                      Register offset_reg, uintptr_t offset_imm,
                                      LoadType type,
@@ -2206,13 +2224,6 @@ void LiftoffAssembler::emit_i8x16_swizzle(LiftoffRegister dst,
   bailout(kUnsupportedArchitecture, "emit_i8x16_swizzle");
 }
 
-void LiftoffAssembler::emit_f64x2_replace_lane(LiftoffRegister dst,
-                                               LiftoffRegister src1,
-                                               LiftoffRegister src2,
-                                               uint8_t imm_lane_idx) {
-  bailout(kUnsupportedArchitecture, "emit_f64x2replacelane");
-}
-
 void LiftoffAssembler::emit_f64x2_abs(LiftoffRegister dst,
                                       LiftoffRegister src) {
   bailout(kUnsupportedArchitecture, "emit_f64x2_abs");
@@ -2307,13 +2318,6 @@ void LiftoffAssembler::emit_f64x2_promote_low_f32x4(LiftoffRegister dst,
   bailout(kSimd, "f64x2.promote_low_f32x4");
 }
 
-void LiftoffAssembler::emit_f32x4_replace_lane(LiftoffRegister dst,
-                                               LiftoffRegister src1,
-                                               LiftoffRegister src2,
-                                               uint8_t imm_lane_idx) {
-  bailout(kUnsupportedArchitecture, "emit_f32x4replacelane");
-}
-
 void LiftoffAssembler::emit_f32x4_abs(LiftoffRegister dst,
                                       LiftoffRegister src) {
   bailout(kUnsupportedArchitecture, "emit_f32x4_abs");
@@ -2393,13 +2397,6 @@ void LiftoffAssembler::emit_f32x4_pmax(LiftoffRegister dst, LiftoffRegister lhs,
   bailout(kSimd, "pmax unimplemented");
 }
 
-void LiftoffAssembler::emit_i64x2_replace_lane(LiftoffRegister dst,
-                                               LiftoffRegister src1,
-                                               LiftoffRegister src2,
-                                               uint8_t imm_lane_idx) {
-  bailout(kUnsupportedArchitecture, "emit_i64x2replacelane");
-}
-
 void LiftoffAssembler::emit_i64x2_neg(LiftoffRegister dst,
                                       LiftoffRegister src) {
   bailout(kUnsupportedArchitecture, "emit_i64x2neg");
@@ -2506,13 +2503,6 @@ void LiftoffAssembler::emit_i64x2_extmul_high_i32x4_u(LiftoffRegister dst,
   bailout(kSimd, "i64x2_extmul_high_i32x4_u unsupported");
 }
 
-void LiftoffAssembler::emit_i32x4_replace_lane(LiftoffRegister dst,
-                                               LiftoffRegister src1,
-                                               LiftoffRegister src2,
-                                               uint8_t imm_lane_idx) {
-  bailout(kUnsupportedArchitecture, "emit_i32x4replacelane");
-}
-
 void LiftoffAssembler::emit_i32x4_neg(LiftoffRegister dst,
                                       LiftoffRegister src) {
   bailout(kUnsupportedArchitecture, "emit_i32x4neg");
@@ -2749,13 +2739,6 @@ void LiftoffAssembler::emit_i16x8_max_u(LiftoffRegister dst,
   bailout(kUnsupportedArchitecture, "emit_i16x8_max_u");
 }
 
-void LiftoffAssembler::emit_i16x8_replace_lane(LiftoffRegister dst,
-                                               LiftoffRegister src1,
-                                               LiftoffRegister src2,
-                                               uint8_t imm_lane_idx) {
-  bailout(kUnsupportedArchitecture, "emit_i16x8replacelane");
-}
-
 void LiftoffAssembler::emit_i16x8_extadd_pairwise_i8x16_s(LiftoffRegister dst,
                                                           LiftoffRegister src) {
   bailout(kSimd, "i16x8.extadd_pairwise_i8x16_s");
@@ -2809,13 +2792,6 @@ void LiftoffAssembler::emit_i8x16_popcnt(LiftoffRegister dst,
   bailout(kSimd, "i8x16.popcnt");
 }
 
-void LiftoffAssembler::emit_i8x16_replace_lane(LiftoffRegister dst,
-                                               LiftoffRegister src1,
-                                               LiftoffRegister src2,
-                                               uint8_t imm_lane_idx) {
-  bailout(kUnsupportedArchitecture, "emit_i8x16replacelane");
-}
-
 void LiftoffAssembler::emit_i8x16_neg(LiftoffRegister dst,
                                       LiftoffRegister src) {
   bailout(kUnsupportedArchitecture, "emit_i8x16neg");