PR #1662: Replace shift with addition in crc multiply

Imported from GitHub PR https://github.com/abseil/abseil-cpp/pull/1662

Merge 4b2c6c909b573d31a1cccba7cb72d4d8badeef8b into cba31a956209e68e4d4049e8a9bc03b1fd67320a

Merging this change closes #1662

COPYBARA_INTEGRATE_REVIEW=https://github.com/abseil/abseil-cpp/pull/1662 from pps83:crc-add 4b2c6c909b573d31a1cccba7cb72d4d8badeef8b
PiperOrigin-RevId: 631470883
Change-Id: I4a72be643ed341ddf0e0007418ab4a613a03db4b

2 files changed, 12 insertions, 10 deletions

diff --git a/absl/crc/internal/crc32_x86_arm_combined_simd.h b/absl/crc/internal/crc32_x86_arm_combined_simd.h
index aa6a6595..0f6e3479 100644
--- a/absl/crc/internal/crc32_x86_arm_combined_simd.h
+++ b/absl/crc/internal/crc32_x86_arm_combined_simd.h
@@ -123,8 +123,8 @@ uint64_t V128_Extract64(const V128 l);
 // Extracts the low 64 bits from V128.
 int64_t V128_Low64(const V128 l);
 
-// Left-shifts packed 64-bit integers in l by r.
-V128 V128_ShiftLeft64(const V128 l, const V128 r);
+// Add packed 64-bit integers in |l| and |r|.
+V128 V128_Add64(const V128 l, const V128 r);
 
 #endif
 
@@ -193,8 +193,8 @@ inline uint64_t V128_Extract64(const V128 l) {
 
 inline int64_t V128_Low64(const V128 l) { return _mm_cvtsi128_si64(l); }
 
-inline V128 V128_ShiftLeft64(const V128 l, const V128 r) {
-  return _mm_sll_epi64(l, r);
+inline V128 V128_Add64(const V128 l, const V128 r) {
+  return _mm_add_epi64(l, r);
 }
 
 #elif defined(ABSL_CRC_INTERNAL_HAVE_ARM_SIMD)
@@ -289,9 +289,7 @@ inline int64_t V128_Low64(const V128 l) {
   return vgetq_lane_s64(vreinterpretq_s64_u64(l), 0);
 }
 
-inline V128 V128_ShiftLeft64(const V128 l, const V128 r) {
-  return vshlq_u64(l, vreinterpretq_s64_u64(r));
-}
+inline V128 V128_Add64(const V128 l, const V128 r) { return vaddq_u64(l, r); }
 
 #endif
 
diff --git a/absl/crc/internal/crc_x86_arm_combined.cc b/absl/crc/internal/crc_x86_arm_combined.cc
index 20dd3e01..79dace34 100644
--- a/absl/crc/internal/crc_x86_arm_combined.cc
+++ b/absl/crc/internal/crc_x86_arm_combined.cc
@@ -101,13 +101,17 @@ constexpr size_t kMediumCutoff = 2048;
 namespace {
 
 uint32_t multiply(uint32_t a, uint32_t b) {
-  V128 shifts = V128_From64WithZeroFill(1);
   V128 power = V128_From64WithZeroFill(a);
   V128 crc = V128_From64WithZeroFill(b);
   V128 res = V128_PMulLow(power, crc);
 
-  // Combine crc values
-  res = V128_ShiftLeft64(res, shifts);
+  // Combine crc values.
+  //
+  // Adding res to itself is equivalent to multiplying by 2,
+  // or shifting left by 1. Addition is used as not all compilers
+  // are able to generate optimal code without this hint.
+  // https://godbolt.org/z/rr3fMnf39
+  res = V128_Add64(res, res);
   return static_cast<uint32_t>(V128_Extract32<1>(res)) ^
          CRC32_u32(0, static_cast<uint32_t>(V128_Low64(res)));
 }