Scope the `GroupSse2Impl` `movzx` inline-asm workaround from `MoveMask` to `Match`.

`GroupSse2Impl::MoveMask` previously applied `asm("" : "+r"(mask))` on Clang to prevent an unnecessary `movzx` when iterating over `BitMask` in `Match(h2)`. However, `MoveMask` is also called by `MaskEmpty()`, `MaskFull()`, `MaskNonFull()`, `MaskEmptyOrDeleted()`, and `MaskFullOrSentinel()`.

In non-`Match` callers, the `asm("" : "+r"(mask))` tied-register inline assembly had two negative effects:

1. In `find_or_prepare_insert_large` (such as `IntPolicy::apply<EmplaceDecomposable>`), the tied `+r` constraint on `MoveMask` inside `MaskEmpty()` caused LLVM's `RegisterCoalescer` to coalesce the `vpmovmskb` destination into `%ecx` (which holds `seq.index()` in the probe loop), requiring two extra register-to-register copies (`movq %rcx, %rbx` and `movq %rbx, %rcx`) on every probe loop iteration.

2. In LLVM's `InlineCost` analysis (`visitCallBase`), every inline `asm` statement adds `InlineConstants::CallPenalty` (`+10` cost) per inlined callsite, inflating the perceived inlining cost of functions that call `MaskEmpty()`, `MaskFull()`, or `MaskEmptyOrDeleted()`.

Moving the `asm("" : "+r"(mask))` barrier into `GroupSse2Impl::Match` preserves the `movzx` workaround for `Match(h2)` while eliminating the register coalescing copies and inline-asm call penalty in all other `MoveMask` callers.

PiperOrigin-RevId: 1007541334

absl/container/internal/hashtable_control_bytes.h

diff --git a/absl/container/internal/hashtable_control_bytes.h b/absl/container/internal/hashtable_control_bytes.h
index 34df1d4..9a83c3b 100644
--- a/absl/container/internal/hashtable_control_bytes.h
+++ b/absl/container/internal/hashtable_control_bytes.h
@@ -289,7 +289,15 @@
   // Returns a bitmask representing the positions of slots that match hash.
   BitMaskType Match(h2_t hash) const {
     auto match = _mm_set1_epi8(static_cast<char>(hash));
-    return BitMaskType(MoveMask(_mm_cmpeq_epi8(match, ctrl)));
+    auto mask = MoveMask(_mm_cmpeq_epi8(match, ctrl));
+#ifdef __clang__
+    // TODO(b/472522597): Without the inline asm, clang ends up generating an
+    // unnecessary movzx to zero the upper bits of the output when iterating
+    // over matches, even though those bits are already zero. See
+    // https://godbolt.org/z/G6xW1Ecbx.
+    asm("" : "+r"(mask));  // NOLINT
+#endif
+    return BitMaskType(mask);
   }
 
   // Returns a bitmask representing the positions of empty slots.
@@ -343,14 +351,7 @@
   }
 
   static MaskInt MoveMask(__m128i xmm) {
-    auto mask = static_cast<MaskInt>(_mm_movemask_epi8(xmm));
-#ifdef __clang__
-    // TODO(b/472522597): Without the inline asm, clang ends up generating an
-    // unnecessary movzx to zero the upper bits of the output, but those bits
-    // are already zero. See https://godbolt.org/z/G6xW1Ecbx.
-    asm("" : "+r"(mask));  // NOLINT
-#endif
-    return mask;
+    return static_cast<MaskInt>(_mm_movemask_epi8(xmm));
   }
 
   __m128i ctrl;