[wasm-simd][x64] Optimize i8x16.popcnt

Instead of loading the same mask twice, we load from an external
reference twice. This saves some some binary size and a bunch of
instructions.

Bug: v8:11002
Change-Id: Ice80bd10694dcca920e18b8043390d7631c65805
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2643404Reviewed-by: Deepti Gandluri <gdeepti@chromium.org>
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#72242}

src/codegen/external-reference.cc

@@ -75,6 +75,12 @@ constexpr struct alignas(16) {
 } double_negate_constant = {uint64_t{0x8000000000000000},
                             uint64_t{0x8000000000000000}};
 
+constexpr struct alignas(16) {
+  uint64_t a;
+  uint64_t b;
+} wasm_i8x16_popcnt_mask = {uint64_t{0x03020201'02010100},
+                            uint64_t{0x04030302'03020201}};
+
 // Implementation of ExternalReference
 
 static ExternalReference::Type BuiltinCallTypeForResultSize(int result_size) {
@@ -480,6 +486,10 @@ ExternalReference ExternalReference::address_of_double_neg_constant() {
   return ExternalReference(reinterpret_cast<Address>(&double_negate_constant));
 }
 
+ExternalReference ExternalReference::address_of_wasm_i8x16_popcnt_mask() {
+  return ExternalReference(reinterpret_cast<Address>(&wasm_i8x16_popcnt_mask));
+}
+
 ExternalReference
 ExternalReference::address_of_enable_experimental_regexp_engine() {
   return ExternalReference(&FLAG_enable_experimental_regexp_engine);

src/codegen/external-reference.h

@@ -115,6 +115,7 @@ class StatsCounter;
   V(address_of_runtime_stats_flag, "TracingFlags::runtime_stats")              \
   V(address_of_the_hole_nan, "the_hole_nan")                                   \
   V(address_of_uint32_bias, "uint32_bias")                                     \
+  V(address_of_wasm_i8x16_popcnt_mask, "wasm_i8x16_popcnt_mask")               \
   V(bytecode_size_table_address, "Bytecodes::bytecode_size_table_address")     \
   V(check_object_type, "check_object_type")                                    \
   V(compute_integer_hash, "ComputeSeededHash")                                 \

src/compiler/backend/x64/code-generator-x64.cc

@@ -3808,16 +3808,18 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
         __ vpshufb(kScratchDoubleReg, tmp, kScratchDoubleReg);
         __ vpaddb(dst, dst, kScratchDoubleReg);
       } else {
+        Operand mask = __ ExternalReferenceAsOperand(
+            ExternalReference::address_of_wasm_i8x16_popcnt_mask());
         __ shufps(tmp, tmp, 0);
-        __ Move(kScratchDoubleReg, src);
-        __ andps(kScratchDoubleReg, tmp);
-        __ andnps(tmp, src);
-        __ Move(dst, 0x04030302'03020201, 0x03020201'02010100);
-        __ psrlw(tmp, 4);
-        __ pshufb(dst, kScratchDoubleReg);
-        __ Move(kScratchDoubleReg, 0x04030302'03020201, 0x03020201'02010100);
-        __ pshufb(kScratchDoubleReg, tmp);
-        __ paddb(dst, kScratchDoubleReg);
+        __ Move(kScratchDoubleReg, tmp);
+        __ andps(tmp, src);
+        __ andnps(kScratchDoubleReg, src);
+        __ psrlw(kScratchDoubleReg, 4);
+        __ movups(dst, mask);
+        __ pshufb(dst, tmp);
+        __ movups(tmp, mask);
+        __ pshufb(tmp, kScratchDoubleReg);
+        __ paddb(dst, tmp);
       }
       break;
     }