Rewrite non-temporal instructions

src/trusted/validator_ragel/dfa_validate_common.c

Index: src/trusted/validator_ragel/dfa_validate_common.c
diff --git a/src/trusted/validator_ragel/dfa_validate_common.c b/src/trusted/validator_ragel/dfa_validate_common.c
index 4972cbd654b3fcd0d884dec0099db63750991802..4f20121f79ea976cf6f79d54fb1a7a67b64aed3c 100644
--- a/src/trusted/validator_ragel/dfa_validate_common.c
+++ b/src/trusted/validator_ragel/dfa_validate_common.c
@@ -12,6 +12,7 @@
 #include "native_client/src/shared/platform/nacl_check.h"
 #include "native_client/src/trusted/service_runtime/nacl_config.h"
 #include "native_client/src/trusted/validator_ragel/validator.h"
+#include "native_client/src/include/build_config.h"
 
 /* Used as an argument to copy_func when unsupported instruction must be
    replaced with HLTs.  */
@@ -33,6 +34,58 @@ Bool NaClDfaProcessValidationError(const uint8_t *begin, const uint8_t *end,
   return FALSE;
 }
 
+#if NACL_BUILD_SUBARCH == 64
+static Bool IsREX(uint8_t byte) {
+  return byte >= 0x40 && byte <= 0x4f;
+}
+#endif
+
+static Bool RewriteUnsupportedInstruction(const uint8_t *begin,
+                                          const uint8_t *end) {
+  uint8_t *ptr = (uint8_t *) begin;
+  /* We usually only check and rewrite the first few bytes without examining
+   * further because this function is only called when the validator tells us
+   * that it is an 'unsupported instruction' and there are no other validation
+   * failures.
+   */
+#if NACL_BUILD_SUBARCH == 32
+  UNREFERENCED_PARAMETER(end);
+  if (memcmp(begin, "\x0f\xe7", 2) == 0) {
+    /* movntq => movq */
+    ptr[1] = 0x7f;
+    return TRUE;
+  } else if (memcmp(begin, "\x66\x0f\xe7", 3) == 0) {
+    /* movntdq => movdqa */
+    ptr[2] = 0x7f;
+    return TRUE;
+  }
+#elif NACL_BUILD_SUBARCH == 64
+  if (IsREX(begin[0])) {
+    if (memcmp(begin + 1, "\x0f\x2b", 2) == 0) {

[Petr Hosek, 2015/08/05 21:59:56]

Couldn't you rewrite this using a switch statement rather than using the
complicated nested if statements?

[ruiq, 2015/08/06 04:53:18]

Done.

+      /* movntps => movaps */
+      ptr[2] = 0x29;
+      return TRUE;
+    } else if (memcmp(begin + 1, "\x0f\xc3", 2) == 0) {
+      /* movnti => mov */
+      ptr[1] = 0x89;
+      memmove(ptr + 2, ptr + 3, end - begin - 3);
+      ptr[end - begin - 1] = 0x90;
+      return TRUE;
+    } else if (memcmp(begin + 1, "\x0f\x18", 2) == 0) {
+      /* prefetchnta => nop */
+      memset(ptr, 0x90, end - begin);
+      return TRUE;
+    }
+  } else if (begin[0] == 0x66 && IsREX(begin[1]) &&
+      memcmp(begin + 2, "\x0f\xe7", 2) == 0) {
+    /* movntdq => movdqa */
+    ptr[3] = 0x7f;
+    return TRUE;
+  }
+#endif
+  return FALSE;
+}
+
 Bool NaClDfaStubOutUnsupportedInstruction(const uint8_t *begin,
                                           const uint8_t *end,
                                           uint32_t info,
@@ -47,11 +100,13 @@ Bool NaClDfaStubOutUnsupportedInstruction(const uint8_t *begin,
     if (data->flags & NACL_DISABLE_NONTEMPORALS_X86) {
       return FALSE;
     } else {
-      /* TODO(ruiq): rewrite instruction. For now, we keep the original
-       * instruction and indicate validation success, which is consistent
-       * with current validation results. */
-      data->did_rewrite = 0;
-      return TRUE;
+      if (RewriteUnsupportedInstruction(begin, end)) {

[Petr Hosek, 2015/08/05 21:59:56]

You could merge the nested if statement with the outer one to make the code
simpler, i.e.:

if (!(data->flags & NACL_DISABLE_NONTEMPORALS_X86) &&
RewriteUnsupportedInstruction(begin, end)) {
  data->did_rewrite = 1;
  return TRUE;
}
return FALSE;

In fact, you could go even further and merge the `return FALSE;` with the
outermost loop and have just a single `return FALSE;` statement.

[ruiq, 2015/08/06 04:53:18]

Done.

+        data->did_rewrite = 1;
+        return TRUE;
+      } else {
+        data->did_rewrite = 0;
+        return FALSE;
+      }
     }
   } else {
     return FALSE;