commit 1b0e0d75f59db59edf95cfb17d5060626d6ade38
author Cindy Liu <hcindyl@google.com> Thu Aug 17 10:24:42 2023 -0700
committer Cindy Liu <hcindyl@google.com> Thu Aug 17 13:40:06 2023 -0700
tree cd4d1ea06997cc741f0f3cb9a85a4af45136eeac
parent be9f3c6fa413145a0673f655adb1f985fd76f7b1

Fix `vrev` implementation

Update the maximum manipulated bits based on data type to match HW behavior

PiperOrigin-RevId: 557864112

sim/kelvin_vector_instructions.cc

diff --git a/sim/kelvin_vector_instructions.cc b/sim/kelvin_vector_instructions.cc
index 432d91d..1afa457 100644
--- a/sim/kelvin_vector_instructions.cc
+++ b/sim/kelvin_vector_instructions.cc
@@ -623,13 +623,13 @@
       inst, true /* scalar */, strip_mine,
       std::function<T(T, T)>([](T vs1, T vs2) -> T {
         T r = vs1;
-        // TODO(leonidl): revisit after spec clarification.
-        // For now it's set to always use 5 lower bits, regardless of type.
         T count = vs2 & 0b11111;
         if (count & 1) r = ((r & 0x55555555) << 1) | ((r & 0xAAAAAAAA) >> 1);
         if (count & 2) r = ((r & 0x33333333) << 2) | ((r & 0xCCCCCCCC) >> 2);
         if (count & 4) r = ((r & 0x0F0F0F0F) << 4) | ((r & 0xF0F0F0F0) >> 4);
+        if (sizeof(T) == 1) return r;
         if (count & 8) r = ((r & 0x00FF00FF) << 8) | ((r & 0xFF00FF00) >> 8);
+        if (sizeof(T) == 2) return r;
         if (count & 16) r = ((r & 0x0000FFFF) << 16) | ((r & 0xFFFF0000) >> 16);
         return r;
       }));

sim/test/kelvin_vector_instructions_test.cc

diff --git a/sim/test/kelvin_vector_instructions_test.cc b/sim/test/kelvin_vector_instructions_test.cc
index 5e92867..91cc4c4 100644
--- a/sim/test/kelvin_vector_instructions_test.cc
+++ b/sim/test/kelvin_vector_instructions_test.cc
@@ -879,7 +879,9 @@
     if (count & 1) r = ((r & 0x55555555) << 1) | ((r & 0xAAAAAAAA) >> 1);
     if (count & 2) r = ((r & 0x33333333) << 2) | ((r & 0xCCCCCCCC) >> 2);
     if (count & 4) r = ((r & 0x0F0F0F0F) << 4) | ((r & 0xF0F0F0F0) >> 4);
+    if (sizeof(Vs1) == 1) return r;
     if (count & 8) r = ((r & 0x00FF00FF) << 8) | ((r & 0xFF00FF00) >> 8);
+    if (sizeof(Vs1) == 2) return r;
     if (count & 16) r = ((r & 0x0000FFFF) << 16) | ((r & 0xFFFF0000) >> 16);
     return r;
   }