diff --git a/hwy/contrib/thread_pool/thread_pool.h b/hwy/contrib/thread_pool/thread_pool.h
index 1f19ab4a8f..41abd7dd51 100644
--- a/hwy/contrib/thread_pool/thread_pool.h
+++ b/hwy/contrib/thread_pool/thread_pool.h
@@ -458,7 +458,7 @@ class Stats {
     printf(
         "%3zu: %5d x %.2f/%5d x %4.1f tasks, %.2f steals; "
         "wake %7.3f ns, latency %6.3f < %7.3f us, barrier %7.3f us; "
-        "wait %.1f us (%5.0f reps, %4.1f%%), balance %4.1f%%-%5.1f%%, "
+        "wait %.1f us (%6.0f reps, %4.1f%%), balance %4.1f%%-%5.1f%%, "
         "func: %6.3f + %7.3f, "
         "%.1f%% of thread time %7.3f s; main:worker %5.1f%%\n",
         num_threads, num_run_static_, avg_tasks_static, num_run_dynamic_,
@@ -589,7 +589,7 @@ class CallerAccumulator {
     const double task_len = avg_elapsed / avg_tasks_per_worker;
     printf(
         "%40s: %7.0f x (%3.0f%%) %2zu clusters, %4.1f workers @ "
-        "%5.1f tasks (%u-%u), "
+        "%5.1f tasks (%5u-%5u), "
         "%5.0f us wait, %6.1E us run (task len %6.1E us), total %6.2f s\n",
         caller, static_cast<double>(calls_), pc_root, active_clusters,
         avg_workers, avg_tasks_per_worker, static_cast<uint32_t>(min_tasks_),
@@ -1605,7 +1605,7 @@ class alignas(HWY_ALIGNMENT) ThreadPool {
   // Returns whether threads were used. If not, there is no need to update
   // the autotuner config.
   template <class Closure>
-  bool RunWithoutAutotune(uint64_t begin, uint64_t end, HWY_MAYBE_UNUSED pool::Caller caller,
+  bool RunWithoutAutotune(uint64_t begin, uint64_t end, pool::Caller caller,
                           const Closure& closure) {
     pool::Worker& main = workers_[0];
 
@@ -1649,6 +1649,8 @@ class alignas(HWY_ALIGNMENT) ThreadPool {
       PROFILER_END_ROOT_RUN();
       shared_.LastRootEnd().Reset();
     }
+#else
+    (void)caller;
 #endif
 
     busy_.Clear();
diff --git a/hwy/ops/set_macros-inl.h b/hwy/ops/set_macros-inl.h
index 2a5e1dc1ca..bf604d96da 100644
--- a/hwy/ops/set_macros-inl.h
+++ b/hwy/ops/set_macros-inl.h
@@ -141,7 +141,7 @@
 #define HWY_TARGET_STR_AVX2 \
   HWY_TARGET_STR_SSE4 ",avx,avx2" HWY_TARGET_STR_BMI2_FMA HWY_TARGET_STR_F16C
 
-#ifndef HWY_HAVE_EVEX512
+#ifndef HWY_HAVE_EVEX512  // allow override
 // evex512 has been removed from clang 22, see
 // https://github.com/llvm/llvm-project/pull/157034
 #if (1400 <= HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1600) || \
@@ -167,15 +167,21 @@
   ",vpclmulqdq,avx512vbmi,avx512vbmi2,vaes,avx512vnni,avx512bitalg," \
   "avx512vpopcntdq,gfni"
 
-// Force-disable for compilers that do not properly support avx512bf16.
-#if !defined(HWY_AVX3_DISABLE_AVX512BF16) &&                        \
+// Opt-out for compilers that do not properly support avx512bf16.
+#ifndef HWY_AVX3_ENABLE_AVX512BF16  // allow override
+// Default is to disable if the DISABLE macro is defined, or if old compiler.
+// clang-cl 21.1.4 reportedly works; feel free to define this to 1 there.
+#if defined(HWY_AVX3_DISABLE_AVX512BF16) ||                         \
     (HWY_COMPILER_CLANGCL ||                                        \
      (HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1000) || \
      (HWY_COMPILER_CLANG && HWY_COMPILER_CLANG < 900))
-#define HWY_AVX3_DISABLE_AVX512BF16
+#define HWY_AVX3_ENABLE_AVX512BF16 0
+#else
+#define HWY_AVX3_ENABLE_AVX512BF16 1
 #endif
+#endif  // HWY_AVX3_ENABLE_AVX512BF16
 
-#if !defined(HWY_AVX3_DISABLE_AVX512BF16)
+#if HWY_AVX3_ENABLE_AVX512BF16
 #define HWY_TARGET_STR_AVX3_ZEN4 HWY_TARGET_STR_AVX3_DL ",avx512bf16"
 #else
 #define HWY_TARGET_STR_AVX3_ZEN4 HWY_TARGET_STR_AVX3_DL
@@ -338,7 +344,7 @@
 #define HWY_HAVE_FLOAT64 1
 #define HWY_MEM_OPS_MIGHT_FAULT 0
 #define HWY_NATIVE_FMA 1
-#if (HWY_TARGET <= HWY_AVX3_ZEN4) && !defined(HWY_AVX3_DISABLE_AVX512BF16)
+#if (HWY_TARGET <= HWY_AVX3_ZEN4) && HWY_AVX3_ENABLE_AVX512BF16
 #define HWY_NATIVE_DOT_BF16 1
 #else
 #define HWY_NATIVE_DOT_BF16 0
diff --git a/hwy/ops/x86_128-inl.h b/hwy/ops/x86_128-inl.h
index e496e3e9e2..5e1573831f 100644
--- a/hwy/ops/x86_128-inl.h
+++ b/hwy/ops/x86_128-inl.h
@@ -57,7 +57,7 @@ namespace detail {
 #undef HWY_AVX3_HAVE_F32_TO_BF16C
 #if HWY_TARGET <= HWY_AVX3_ZEN4 && !HWY_COMPILER_CLANGCL &&           \
     (HWY_COMPILER_GCC_ACTUAL >= 1000 || HWY_COMPILER_CLANG >= 900) && \
-    !defined(HWY_AVX3_DISABLE_AVX512BF16)
+    HWY_AVX3_ENABLE_AVX512BF16
 #define HWY_AVX3_HAVE_F32_TO_BF16C 1
 #else
 #define HWY_AVX3_HAVE_F32_TO_BF16C 0