add intdiv imp

Signed-off-by: Abhishek Kumar <abhishek.r.kumar@fujitsu.com>

Author: abhishek-iitmadras

BUILD

@@ -539,6 +539,22 @@ cc_library(
     local_defines = ["HWY_HEADER_ONLY"],
 )
 
+cc_library(
+    name = "intdiv",
+    compatible_with = [],
+    copts = COPTS,
+    hdrs = [
+        "hwy/contrib/intdiv/intdiv.h",
+    ],
+    textual_hdrs = [
+        "hwy/contrib/intdiv/intdiv-inl.h",
+    ],
+    deps = [
+        ":hwy",
+    ],
+)
+
+
 cc_test(
     name = "list_targets",
     size = "small",
@@ -644,6 +660,17 @@ cc_test(
     ],
 )
 
+cc_test(
+    name = "intdiv_test",
+    size = "medium",
+    timeout = "long",
+    srcs = ["hwy/contrib/intdiv/intdiv_test.cc"],
+    copts = COPTS + HWY_TEST_COPTS,
+    local_defines = ["HWY_IS_TEST"],
+    tags = ["hwy_ops_test"],
+    deps = HWY_TEST_DEPS + [":intdiv"],
+)
+
 # For manually building the tests we define here (:all does not work in --config=msvc)
 test_suite(
     name = "hwy_ops_tests",

CMakeLists.txt

@@ -222,6 +222,8 @@ list(APPEND HWY_CONTRIB_SOURCES
     hwy/contrib/image/image.cc
     hwy/contrib/image/image.h
     hwy/contrib/math/fast_math-inl.h
+    hwy/contrib/intdiv/intdiv.h
+    hwy/contrib/intdiv/intdiv-inl.h
     hwy/contrib/math/math-inl.h
     hwy/contrib/matvec/matvec-inl.h
     hwy/contrib/random/random-inl.h
@@ -915,6 +917,7 @@ list(APPEND HWY_TEST_FILES
   hwy/contrib/bit_pack/bit_pack_test.cc
   hwy/contrib/dot/dot_test.cc
   hwy/contrib/matvec/matvec_test.cc
+  hwy/contrib/intdiv/intdiv_test.cc
   hwy/contrib/image/image_test.cc
   # Disabled due to SIGILL in clang7 debug build during gtest discovery phase,
   # not reproducible locally. Still tested via bazel build.

hwy.gni

@@ -52,6 +52,8 @@ hwy_contrib_public = [
   "$_hwy/contrib/dot/dot-inl.h",
   "$_hwy/contrib/image/image.h",
   "$_hwy/contrib/math/fast_math-inl.h",
+  "$_hwy/contrib/intdiv/intdiv.h",
+  "$_hwy/contrib/intdiv/intdiv-inl.h",
   "$_hwy/contrib/math/math-inl.h",
 ]
 

hwy/base.h

@@ -3049,10 +3049,11 @@ class Divisor64 {
   // Returns n % divisor_.
   uint64_t Remainder(uint64_t n) const { return n - (Divide(n) * divisor_); }
 
- private:
-  uint64_t divisor_;
+//  private:
+//   uint64_t divisor_;
 
   static uint64_t Div128(uint64_t hi, uint64_t div) {
+    HWY_DASSERT(div != 0);
 #if HWY_COMPILER_MSVC >= 1920 && HWY_ARCH_X86_64
     unsigned __int64 remainder;  // unused
     return _udiv128(hi, uint64_t{0}, div, &remainder);
@@ -3063,7 +3064,9 @@ class Divisor64 {
 #endif
   }
 
-  static uint64_t MulHigh(uint64_t a, uint64_t b) {
+  private:
+    uint64_t divisor_;
+    static uint64_t MulHigh(uint64_t a, uint64_t b) {
 #if HWY_COMPILER_MSVC >= 1920 && HWY_ARCH_X86_64
     return __umulh(a, b);
 #else
@@ -3074,9 +3077,9 @@ class Divisor64 {
 #endif
   }
 
-  uint64_t mul_ = 1;
-  uint64_t shift1_ = 0;
-  uint64_t shift2_ = 0;
+    uint64_t mul_ = 1;
+    uint64_t shift1_ = 0;
+    uint64_t shift2_ = 0;
 };
 #else
 // No Div128 available, use built-in 64-bit division on each call.