diff --git a/src/amd/compiler/tests/main.cpp b/src/amd/compiler/tests/main.cpp
index cb646e2dd30..eac0a244adf 100644
--- a/src/amd/compiler/tests/main.cpp
+++ b/src/amd/compiler/tests/main.cpp
@@ -34,6 +34,8 @@
 #include "aco_ir.h"
 #include "framework.h"
 
+#include "util/u_cpu_detect.h"
+
 static const char *help_message =
    "Usage: %s [-h] [-l --list] [--no-check] [TEST [TEST ...]]\n"
    "\n"
@@ -227,6 +229,8 @@ int main(int argc, char **argv)
       return 99;
    }
 
+   util_cpu_detect();
+
    if (do_list) {
       for (auto test : tests)
          printf("%s\n", test.first.c_str());
diff --git a/src/compiler/glsl/standalone.cpp b/src/compiler/glsl/standalone.cpp
index ca187001186..2714d8b95ed 100644
--- a/src/compiler/glsl/standalone.cpp
+++ b/src/compiler/glsl/standalone.cpp
@@ -401,6 +401,8 @@ standalone_compile_shader(const struct standalone_options *_options,
    int status = EXIT_SUCCESS;
    bool glsl_es = false;
 
+   util_cpu_detect();
+
    options = _options;
 
    switch (options->glsl_version) {
diff --git a/src/compiler/nir/tests/negative_equal_tests.cpp b/src/compiler/nir/tests/negative_equal_tests.cpp
index f83041a4fbf..76472e48309 100644
--- a/src/compiler/nir/tests/negative_equal_tests.cpp
+++ b/src/compiler/nir/tests/negative_equal_tests.cpp
@@ -36,6 +36,7 @@ protected:
    const_value_negative_equal_test()
    {
       glsl_type_singleton_init_or_ref();
+      util_cpu_detect();
 
       memset(c1, 0, sizeof(c1));
       memset(c2, 0, sizeof(c2));
@@ -55,6 +56,7 @@ protected:
    alu_srcs_negative_equal_test()
    {
       glsl_type_singleton_init_or_ref();
+      util_cpu_detect();
 
       static const nir_shader_compiler_options options = { };
       nir_builder_init_simple_shader(&bld, NULL, MESA_SHADER_VERTEX, &options);
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.c b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
index 165d73d94fc..33269e528fe 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_arit.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
@@ -104,13 +104,13 @@ lp_build_min_simple(struct lp_build_context *bld,
 
    /* TODO: optimize the constant case */
 
-   if (type.floating && util_cpu_caps.has_sse) {
+   if (type.floating && util_get_cpu_caps()->has_sse) {
       if (type.width == 32) {
          if (type.length == 1) {
             intrinsic = "llvm.x86.sse.min.ss";
             intr_size = 128;
          }
-         else if (type.length <= 4 || !util_cpu_caps.has_avx) {
+         else if (type.length <= 4 || !util_get_cpu_caps()->has_avx) {
             intrinsic = "llvm.x86.sse.min.ps";
             intr_size = 128;
          }
@@ -119,12 +119,12 @@ lp_build_min_simple(struct lp_build_context *bld,
             intr_size = 256;
          }
       }
-      if (type.width == 64 && util_cpu_caps.has_sse2) {
+      if (type.width == 64 && util_get_cpu_caps()->has_sse2) {
          if (type.length == 1) {
             intrinsic = "llvm.x86.sse2.min.sd";
             intr_size = 128;
          }
-         else if (type.length == 2 || !util_cpu_caps.has_avx) {
+         else if (type.length == 2 || !util_get_cpu_caps()->has_avx) {
             intrinsic = "llvm.x86.sse2.min.pd";
             intr_size = 128;
          }
@@ -134,7 +134,7 @@ lp_build_min_simple(struct lp_build_context *bld,
          }
       }
    }
-   else if (type.floating && util_cpu_caps.has_altivec) {
+   else if (type.floating && util_get_cpu_caps()->has_altivec) {
       if (nan_behavior == GALLIVM_NAN_RETURN_NAN ||
           nan_behavior == GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN) {
          debug_printf("%s: altivec doesn't support nan return nan behavior\n",
@@ -144,7 +144,7 @@ lp_build_min_simple(struct lp_build_context *bld,
          intrinsic = "llvm.ppc.altivec.vminfp";
          intr_size = 128;
       }
-   } else if (util_cpu_caps.has_altivec) {
+   } else if (util_get_cpu_caps()->has_altivec) {
       intr_size = 128;
       if (type.width == 8) {
          if (!type.sign) {
@@ -174,7 +174,7 @@ lp_build_min_simple(struct lp_build_context *bld,
        * The sse intrinsics return the second operator in case of nan by
        * default so we need to special code to handle those.
        */
-      if (util_cpu_caps.has_sse && type.floating &&
+      if (util_get_cpu_caps()->has_sse && type.floating &&
           nan_behavior != GALLIVM_NAN_BEHAVIOR_UNDEFINED &&
           nan_behavior != GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN &&
           nan_behavior != GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN) {
@@ -274,13 +274,13 @@ lp_build_max_simple(struct lp_build_context *bld,
 
    /* TODO: optimize the constant case */
 
-   if (type.floating && util_cpu_caps.has_sse) {
+   if (type.floating && util_get_cpu_caps()->has_sse) {
       if (type.width == 32) {
          if (type.length == 1) {
             intrinsic = "llvm.x86.sse.max.ss";
             intr_size = 128;
          }
-         else if (type.length <= 4 || !util_cpu_caps.has_avx) {
+         else if (type.length <= 4 || !util_get_cpu_caps()->has_avx) {
             intrinsic = "llvm.x86.sse.max.ps";
             intr_size = 128;
          }
@@ -289,12 +289,12 @@ lp_build_max_simple(struct lp_build_context *bld,
             intr_size = 256;
          }
       }
-      if (type.width == 64 && util_cpu_caps.has_sse2) {
+      if (type.width == 64 && util_get_cpu_caps()->has_sse2) {
          if (type.length == 1) {
             intrinsic = "llvm.x86.sse2.max.sd";
             intr_size = 128;
          }
-         else if (type.length == 2 || !util_cpu_caps.has_avx) {
+         else if (type.length == 2 || !util_get_cpu_caps()->has_avx) {
             intrinsic = "llvm.x86.sse2.max.pd";
             intr_size = 128;
          }
@@ -304,7 +304,7 @@ lp_build_max_simple(struct lp_build_context *bld,
          }
       }
    }
-   else if (type.floating && util_cpu_caps.has_altivec) {
+   else if (type.floating && util_get_cpu_caps()->has_altivec) {
       if (nan_behavior == GALLIVM_NAN_RETURN_NAN ||
           nan_behavior == GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN) {
          debug_printf("%s: altivec doesn't support nan return nan behavior\n",
@@ -314,7 +314,7 @@ lp_build_max_simple(struct lp_build_context *bld,
          intrinsic = "llvm.ppc.altivec.vmaxfp";
          intr_size = 128;
       }
-   } else if (util_cpu_caps.has_altivec) {
+   } else if (util_get_cpu_caps()->has_altivec) {
      intr_size = 128;
      if (type.width == 8) {
        if (!type.sign) {
@@ -338,7 +338,7 @@ lp_build_max_simple(struct lp_build_context *bld,
    }
 
    if (intrinsic) {
-      if (util_cpu_caps.has_sse && type.floating &&
+      if (util_get_cpu_caps()->has_sse && type.floating &&
           nan_behavior != GALLIVM_NAN_BEHAVIOR_UNDEFINED &&
           nan_behavior != GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN &&
           nan_behavior != GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN) {
@@ -472,12 +472,12 @@ lp_build_add(struct lp_build_context *bld,
             return lp_build_intrinsic_binary(builder, intrin, bld->vec_type, a, b);
          }
          if (type.width * type.length == 128) {
-            if (util_cpu_caps.has_sse2) {
+            if (util_get_cpu_caps()->has_sse2) {
                if (type.width == 8)
                  intrinsic = type.sign ? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b";
                if (type.width == 16)
                  intrinsic = type.sign ? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w";
-            } else if (util_cpu_caps.has_altivec) {
+            } else if (util_get_cpu_caps()->has_altivec) {
                if (type.width == 8)
                   intrinsic = type.sign ? "llvm.ppc.altivec.vaddsbs" : "llvm.ppc.altivec.vaddubs";
                if (type.width == 16)
@@ -485,7 +485,7 @@ lp_build_add(struct lp_build_context *bld,
             }
          }
          if (type.width * type.length == 256) {
-            if (util_cpu_caps.has_avx2) {
+            if (util_get_cpu_caps()->has_avx2) {
                if (type.width == 8)
                   intrinsic = type.sign ? "llvm.x86.avx2.padds.b" : "llvm.x86.avx2.paddus.b";
                if (type.width == 16)
@@ -713,11 +713,11 @@ lp_build_hadd_partial4(struct lp_build_context *bld,
    tmp[2] = num_vecs > 2 ? vectors[2] : vectors[0];
    tmp[3] = num_vecs > 3 ? vectors[3] : vectors[0];
 
-   if (util_cpu_caps.has_sse3 && bld->type.width == 32 &&
+   if (util_get_cpu_caps()->has_sse3 && bld->type.width == 32 &&
        bld->type.length == 4) {
       intrinsic = "llvm.x86.sse3.hadd.ps";
    }
-   else if (util_cpu_caps.has_avx && bld->type.width == 32 &&
+   else if (util_get_cpu_caps()->has_avx && bld->type.width == 32 &&
             bld->type.length == 8) {
       intrinsic = "llvm.x86.avx.hadd.ps.256";
    }
@@ -796,12 +796,12 @@ lp_build_sub(struct lp_build_context *bld,
             return lp_build_intrinsic_binary(builder, intrin, bld->vec_type, a, b);
          }
          if (type.width * type.length == 128) {
-            if (util_cpu_caps.has_sse2) {
+            if (util_get_cpu_caps()->has_sse2) {
                if (type.width == 8)
                   intrinsic = type.sign ? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b";
                if (type.width == 16)
                   intrinsic = type.sign ? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w";
-            } else if (util_cpu_caps.has_altivec) {
+            } else if (util_get_cpu_caps()->has_altivec) {
                if (type.width == 8)
                   intrinsic = type.sign ? "llvm.ppc.altivec.vsubsbs" : "llvm.ppc.altivec.vsububs";
                if (type.width == 16)
@@ -809,7 +809,7 @@ lp_build_sub(struct lp_build_context *bld,
             }
          }
          if (type.width * type.length == 256) {
-            if (util_cpu_caps.has_avx2) {
+            if (util_get_cpu_caps()->has_avx2) {
                if (type.width == 8)
                   intrinsic = type.sign ? "llvm.x86.avx2.psubs.b" : "llvm.x86.avx2.psubus.b";
                if (type.width == 16)
@@ -1078,8 +1078,8 @@ lp_build_mul_32_lohi_cpu(struct lp_build_context *bld,
     */
    if (LLVM_VERSION_MAJOR < 7 &&
        (bld->type.length == 4 || bld->type.length == 8) &&
-       ((util_cpu_caps.has_sse2 && (bld->type.sign == 0)) ||
-        util_cpu_caps.has_sse4_1)) {
+       ((util_get_cpu_caps()->has_sse2 && (bld->type.sign == 0)) ||
+        util_get_cpu_caps()->has_sse4_1)) {
       const char *intrinsic = NULL;
       LLVMValueRef aeven, aodd, beven, bodd, muleven, mulodd;
       LLVMValueRef shuf[LP_MAX_VECTOR_WIDTH / 32], shuf_vec;
@@ -1096,7 +1096,7 @@ lp_build_mul_32_lohi_cpu(struct lp_build_context *bld,
       aodd = LLVMBuildShuffleVector(builder, aeven, bld->undef, shuf_vec, "");
       bodd = LLVMBuildShuffleVector(builder, beven, bld->undef, shuf_vec, "");
 
-      if (util_cpu_caps.has_avx2 && bld->type.length == 8) {
+      if (util_get_cpu_caps()->has_avx2 && bld->type.length == 8) {
          if (bld->type.sign) {
             intrinsic = "llvm.x86.avx2.pmul.dq";
          } else {
@@ -1331,8 +1331,8 @@ lp_build_div(struct lp_build_context *bld,
 
    /* fast rcp is disabled (just uses div), so makes no sense to try that */
    if(FALSE &&
-      ((util_cpu_caps.has_sse && type.width == 32 && type.length == 4) ||
-       (util_cpu_caps.has_avx && type.width == 32 && type.length == 8)) &&
+      ((util_get_cpu_caps()->has_sse && type.width == 32 && type.length == 4) ||
+       (util_get_cpu_caps()->has_avx && type.width == 32 && type.length == 8)) &&
       type.floating)
       return lp_build_mul(bld, a, lp_build_rcp(bld, b));
 
@@ -1745,7 +1745,7 @@ lp_build_abs(struct lp_build_context *bld,
       return lp_build_intrinsic_unary(builder, intrinsic, vec_type, a);
    }
 
-   if(type.width*type.length == 128 && util_cpu_caps.has_ssse3 && LLVM_VERSION_MAJOR < 6) {
+   if(type.width*type.length == 128 && util_get_cpu_caps()->has_ssse3 && LLVM_VERSION_MAJOR < 6) {
       switch(type.width) {
       case 8:
          return lp_build_intrinsic_unary(builder, "llvm.x86.ssse3.pabs.b.128", vec_type, a);
@@ -1755,7 +1755,7 @@ lp_build_abs(struct lp_build_context *bld,
          return lp_build_intrinsic_unary(builder, "llvm.x86.ssse3.pabs.d.128", vec_type, a);
       }
    }
-   else if (type.width*type.length == 256 && util_cpu_caps.has_avx2 && LLVM_VERSION_MAJOR < 6) {
+   else if (type.width*type.length == 256 && util_get_cpu_caps()->has_avx2 && LLVM_VERSION_MAJOR < 6) {
       switch(type.width) {
       case 8:
          return lp_build_intrinsic_unary(builder, "llvm.x86.avx2.pabs.b", vec_type, a);
@@ -1897,15 +1897,15 @@ lp_build_int_to_float(struct lp_build_context *bld,
 static boolean
 arch_rounding_available(const struct lp_type type)
 {
-   if ((util_cpu_caps.has_sse4_1 &&
+   if ((util_get_cpu_caps()->has_sse4_1 &&
        (type.length == 1 || type.width*type.length == 128)) ||
-       (util_cpu_caps.has_avx && type.width*type.length == 256) ||
-       (util_cpu_caps.has_avx512f && type.width*type.length == 512))
+       (util_get_cpu_caps()->has_avx && type.width*type.length == 256) ||
+       (util_get_cpu_caps()->has_avx512f && type.width*type.length == 512))
       return TRUE;
-   else if ((util_cpu_caps.has_altivec &&
+   else if ((util_get_cpu_caps()->has_altivec &&
             (type.width == 32 && type.length == 4)))
       return TRUE;
-   else if (util_cpu_caps.has_neon)
+   else if (util_get_cpu_caps()->has_neon)
       return TRUE;
 
    return FALSE;
@@ -1935,7 +1935,7 @@ lp_build_iround_nearest_sse2(struct lp_build_context *bld,
    assert(type.width == 32);
 
    assert(lp_check_value(type, a));
-   assert(util_cpu_caps.has_sse2);
+   assert(util_get_cpu_caps()->has_sse2);
 
    /* This is relying on MXCSR rounding mode, which should always be nearest. */
    if (type.length == 1) {
@@ -1961,7 +1961,7 @@ lp_build_iround_nearest_sse2(struct lp_build_context *bld,
       }
       else {
          assert(type.width*type.length == 256);
-         assert(util_cpu_caps.has_avx);
+         assert(util_get_cpu_caps()->has_avx);
 
          intrinsic = "llvm.x86.avx.cvt.ps2dq.256";
       }
@@ -1987,7 +1987,7 @@ lp_build_round_altivec(struct lp_build_context *bld,
    assert(type.floating);
 
    assert(lp_check_value(type, a));
-   assert(util_cpu_caps.has_altivec);
+   assert(util_get_cpu_caps()->has_altivec);
 
    (void)type;
 
@@ -2014,7 +2014,7 @@ lp_build_round_arch(struct lp_build_context *bld,
                     LLVMValueRef a,
                     enum lp_build_round_mode mode)
 {
-   if (util_cpu_caps.has_sse4_1 || util_cpu_caps.has_neon) {
+   if (util_get_cpu_caps()->has_sse4_1 || util_get_cpu_caps()->has_neon) {
       LLVMBuilderRef builder = bld->gallivm->builder;
       const struct lp_type type = bld->type;
       const char *intrinsic_root;
@@ -2042,7 +2042,7 @@ lp_build_round_arch(struct lp_build_context *bld,
       lp_format_intrinsic(intrinsic, sizeof intrinsic, intrinsic_root, bld->vec_type);
       return lp_build_intrinsic_unary(builder, intrinsic, bld->vec_type, a);
    }
-   else /* (util_cpu_caps.has_altivec) */
+   else /* (util_get_cpu_caps()->has_altivec) */
      return lp_build_round_altivec(bld, a, mode);
 }
 
@@ -2377,9 +2377,9 @@ lp_build_iround(struct lp_build_context *bld,
 
    assert(lp_check_value(type, a));
 
-   if ((util_cpu_caps.has_sse2 &&
+   if ((util_get_cpu_caps()->has_sse2 &&
        ((type.width == 32) && (type.length == 1 || type.length == 4))) ||
-       (util_cpu_caps.has_avx && type.width == 32 && type.length == 8)) {
+       (util_get_cpu_caps()->has_avx && type.width == 32 && type.length == 8)) {
       return lp_build_iround_nearest_sse2(bld, a);
    }
    if (arch_rounding_available(type)) {
@@ -2664,8 +2664,8 @@ lp_build_rcp(struct lp_build_context *bld,
     * particular uses that require less workarounds.
     */
 
-   if (FALSE && ((util_cpu_caps.has_sse && type.width == 32 && type.length == 4) ||
-         (util_cpu_caps.has_avx && type.width == 32 && type.length == 8))){
+   if (FALSE && ((util_get_cpu_caps()->has_sse && type.width == 32 && type.length == 4) ||
+         (util_get_cpu_caps()->has_avx && type.width == 32 && type.length == 8))){
       const unsigned num_iterations = 0;
       LLVMValueRef res;
       unsigned i;
@@ -2784,8 +2784,8 @@ lp_build_fast_rsqrt_available(struct lp_type type)
 {
    assert(type.floating);
 
-   if ((util_cpu_caps.has_sse && type.width == 32 && type.length == 4) ||
-       (util_cpu_caps.has_avx && type.width == 32 && type.length == 8)) {
+   if ((util_get_cpu_caps()->has_sse && type.width == 32 && type.length == 4) ||
+       (util_get_cpu_caps()->has_avx && type.width == 32 && type.length == 8)) {
       return true;
    }
    return false;
@@ -3694,7 +3694,7 @@ lp_build_is_inf_or_nan(struct gallivm_state *gallivm,
 LLVMValueRef
 lp_build_fpstate_get(struct gallivm_state *gallivm)
 {
-   if (util_cpu_caps.has_sse) {
+   if (util_get_cpu_caps()->has_sse) {
       LLVMBuilderRef builder = gallivm->builder;
       LLVMValueRef mxcsr_ptr = lp_build_alloca(
          gallivm,
@@ -3715,7 +3715,7 @@ void
 lp_build_fpstate_set_denorms_zero(struct gallivm_state *gallivm,
                                   boolean zero)
 {
-   if (util_cpu_caps.has_sse) {
+   if (util_get_cpu_caps()->has_sse) {
       /* turn on DAZ (64) | FTZ (32768) = 32832 if available */
       int daz_ftz = _MM_FLUSH_ZERO_MASK;
 
@@ -3724,7 +3724,7 @@ lp_build_fpstate_set_denorms_zero(struct gallivm_state *gallivm,
       LLVMValueRef mxcsr =
          LLVMBuildLoad(builder, mxcsr_ptr, "mxcsr");
 
-      if (util_cpu_caps.has_daz) {
+      if (util_get_cpu_caps()->has_daz) {
          /* Enable denormals are zero mode */
          daz_ftz |= _MM_DENORMALS_ZERO_MASK;
       }
@@ -3745,7 +3745,7 @@ void
 lp_build_fpstate_set(struct gallivm_state *gallivm,
                      LLVMValueRef mxcsr_ptr)
 {
-   if (util_cpu_caps.has_sse) {
+   if (util_get_cpu_caps()->has_sse) {
       LLVMBuilderRef builder = gallivm->builder;
       mxcsr_ptr = LLVMBuildPointerCast(builder, mxcsr_ptr,
                      LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0), "");
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_conv.c b/src/gallium/auxiliary/gallivm/lp_bld_conv.c
index c68b8850473..af445b00c1a 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_conv.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_conv.c
@@ -101,7 +101,7 @@ lp_build_half_to_float(struct gallivm_state *gallivm,
    LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
    LLVMValueRef h;
 
-   if (util_cpu_caps.has_f16c &&
+   if (util_get_cpu_caps()->has_f16c &&
        (src_length == 4 || src_length == 8)) {
       if (LLVM_VERSION_MAJOR < 11) {
          const char *intrinsic = NULL;
@@ -167,7 +167,7 @@ lp_build_float_to_half(struct gallivm_state *gallivm,
     * useless.
     */
 
-   if (util_cpu_caps.has_f16c &&
+   if (util_get_cpu_caps()->has_f16c &&
        (length == 4 || length == 8)) {
       struct lp_type i168_type = lp_type_int_vec(16, 16 * 8);
       unsigned mode = 3; /* same as LP_BUILD_ROUND_TRUNCATE */
@@ -489,7 +489,7 @@ int lp_build_conv_auto(struct gallivm_state *gallivm,
 
       /* Special case 4x4x32 --> 1x16x8 */
       if (src_type.length == 4 &&
-            (util_cpu_caps.has_sse2 || util_cpu_caps.has_altivec))
+            (util_get_cpu_caps()->has_sse2 || util_get_cpu_caps()->has_altivec))
       {
          num_dsts = (num_srcs + 3) / 4;
          dst_type->length = num_srcs * 4 >= 16 ? 16 : num_srcs * 4;
@@ -500,7 +500,7 @@ int lp_build_conv_auto(struct gallivm_state *gallivm,
 
       /* Special case 2x8x32 --> 1x16x8 */
       if (src_type.length == 8 &&
-          util_cpu_caps.has_avx)
+          util_get_cpu_caps()->has_avx)
       {
          num_dsts = (num_srcs + 1) / 2;
          dst_type->length = num_srcs * 8 >= 16 ? 16 : num_srcs * 8;
@@ -597,7 +597,7 @@ lp_build_conv(struct gallivm_state *gallivm,
        ((dst_type.length == 16 && 4 * num_dsts == num_srcs) ||
         (num_dsts == 1 && dst_type.length * num_srcs == 16 && num_srcs != 3)) &&
 
-       (util_cpu_caps.has_sse2 || util_cpu_caps.has_altivec))
+       (util_get_cpu_caps()->has_sse2 || util_get_cpu_caps()->has_altivec))
    {
       struct lp_build_context bld;
       struct lp_type int16_type, int32_type;
@@ -710,7 +710,7 @@ lp_build_conv(struct gallivm_state *gallivm,
       ((dst_type.length == 16 && 2 * num_dsts == num_srcs) ||
        (num_dsts == 1 && dst_type.length * num_srcs == 8)) &&
 
-      util_cpu_caps.has_avx) {
+      util_get_cpu_caps()->has_avx) {
 
       struct lp_build_context bld;
       struct lp_type int16_type, int32_type;
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_s3tc.c b/src/gallium/auxiliary/gallivm/lp_bld_format_s3tc.c
index 174857e06d9..e17c7881e7d 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_format_s3tc.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format_s3tc.c
@@ -642,8 +642,8 @@ s3tc_dxt1_full_to_rgba_aos(struct gallivm_state *gallivm,
        * XXX with sse2 and 16x8 vectors, should use pavgb even when n == 1.
        * Much cheaper (but we don't care that much if n == 1).
        */
-      if ((util_cpu_caps.has_sse2 && n == 4) ||
-          (util_cpu_caps.has_avx2 && n == 8)) {
+      if ((util_get_cpu_caps()->has_sse2 && n == 4) ||
+          (util_get_cpu_caps()->has_avx2 && n == 8)) {
          color2_2 = lp_build_pavgb(&bld8, colors0, colors1);
          color2_2 = LLVMBuildBitCast(builder, color2_2, bld32.vec_type, "");
       }
@@ -1350,7 +1350,7 @@ s3tc_decode_block_dxt1(struct gallivm_state *gallivm,
    if (is_dxt1_variant) {
       LLVMValueRef color23_2, color2_2;
 
-      if (util_cpu_caps.has_sse2) {
+      if (util_get_cpu_caps()->has_sse2) {
          LLVMValueRef intrargs[2];
          intrargs[0] = LLVMBuildBitCast(builder, color01, bld8.vec_type, "");
          /* same interleave as for lerp23 - correct result in 2nd element */
@@ -1389,7 +1389,7 @@ s3tc_decode_block_dxt1(struct gallivm_state *gallivm,
       color23 = lp_build_select(&bld32, sel_mask, color23, color23_2);
    }
 
-   if (util_cpu_caps.has_ssse3) {
+   if (util_get_cpu_caps()->has_ssse3) {
       /*
        * Use pshufb as mini-lut. (Only doable with intrinsics as the
        * final shuffles are non-constant. pshufb is awesome!)
@@ -1689,7 +1689,7 @@ s3tc_decode_block_dxt5(struct gallivm_state *gallivm,
    type16.sign = FALSE;
    sel_mask = LLVMBuildBitCast(builder, sel_mask, bld8.vec_type, "");
 
-   if (!util_cpu_caps.has_ssse3) {
+   if (!util_get_cpu_caps()->has_ssse3) {
       LLVMValueRef acodeg, mask1, acode0, acode1;
 
       /* extraction of the 3 bit values into something more useful is HARD */
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_yuv.c b/src/gallium/auxiliary/gallivm/lp_bld_format_yuv.c
index 121452d7596..97deffe1de0 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_format_yuv.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format_yuv.c
@@ -90,7 +90,7 @@ uyvy_to_yuv_soa(struct gallivm_state *gallivm,
     * per element. Didn't measure performance but cuts shader size
     * by quite a bit (less difference if cpu has no sse4.1 support).
     */
-   if (util_cpu_caps.has_sse2 && n > 1) {
+   if (util_get_cpu_caps()->has_sse2 && n > 1) {
       LLVMValueRef sel, tmp, tmp2;
       struct lp_build_context bld32;
 
@@ -174,7 +174,7 @@ yuyv_to_yuv_soa(struct gallivm_state *gallivm,
     * per element. Didn't measure performance but cuts shader size
     * by quite a bit (less difference if cpu has no sse4.1 support).
     */
-   if (util_cpu_caps.has_sse2 && n > 1) {
+   if (util_get_cpu_caps()->has_sse2 && n > 1) {
       LLVMValueRef sel, tmp;
       struct lp_build_context bld32;
 
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_gather.c b/src/gallium/auxiliary/gallivm/lp_bld_gather.c
index e991b0dc375..42cc17371a0 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_gather.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_gather.c
@@ -488,7 +488,7 @@ lp_build_gather(struct gallivm_state *gallivm,
        * 32bit/64bit fetches you're doing it wrong (this is gather, not
        * conversion) and it would be awkward for floats.
        */
-   } else if (util_cpu_caps.has_avx2 && !need_expansion &&
+   } else if (util_get_cpu_caps()->has_avx2 && !need_expansion &&
               src_width == 32 && (length == 4 || length == 8)) {
       return lp_build_gather_avx2(gallivm, length, src_width, dst_type,
                                   base_ptr, offsets);
@@ -500,7 +500,7 @@ lp_build_gather(struct gallivm_state *gallivm,
     * (In general, should be more of a win if the fetch is 256bit wide -
     * this is true for the 32bit case above too.)
     */
-   } else if (0 && util_cpu_caps.has_avx2 && !need_expansion &&
+   } else if (0 && util_get_cpu_caps()->has_avx2 && !need_expansion &&
               src_width == 64 && (length == 2 || length == 4)) {
       return lp_build_gather_avx2(gallivm, length, src_width, dst_type,
                                   base_ptr, offsets);
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_init.c b/src/gallium/auxiliary/gallivm/lp_bld_init.c
index 685ed0e58aa..dd428242cb9 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_init.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_init.c
@@ -433,6 +433,7 @@ lp_build_init(void)
    /* For simulating less capable machines */
 #ifdef DEBUG
    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
+      extern struct util_cpu_caps_t util_cpu_caps;
       assert(util_cpu_caps.has_sse2);
       util_cpu_caps.has_sse3 = 0;
       util_cpu_caps.has_ssse3 = 0;
@@ -445,7 +446,7 @@ lp_build_init(void)
    }
 #endif
 
-   if (util_cpu_caps.has_avx2 || util_cpu_caps.has_avx) {
+   if (util_get_cpu_caps()->has_avx2 || util_get_cpu_caps()->has_avx) {
       lp_native_vector_width = 256;
    } else {
       /* Leave it at 128, even when no SIMD extensions are available.
@@ -460,16 +461,16 @@ lp_build_init(void)
 #if LLVM_VERSION_MAJOR < 4
    if (lp_native_vector_width <= 128) {
       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
-       * "util_cpu_caps.has_avx" predicate, and lack the
+       * "util_get_cpu_caps()->has_avx" predicate, and lack the
        * "lp_native_vector_width > 128" predicate. And also to ensure a more
        * consistent behavior, allowing one to test SSE2 on AVX machines.
        * XXX: should not play games with util_cpu_caps directly as it might
        * get used for other things outside llvm too.
        */
-      util_cpu_caps.has_avx = 0;
-      util_cpu_caps.has_avx2 = 0;
-      util_cpu_caps.has_f16c = 0;
-      util_cpu_caps.has_fma = 0;
+      util_get_cpu_caps()->has_avx = 0;
+      util_get_cpu_caps()->has_avx2 = 0;
+      util_get_cpu_caps()->has_f16c = 0;
+      util_get_cpu_caps()->has_fma = 0;
    }
 #endif
 
@@ -482,7 +483,7 @@ lp_build_init(void)
     * Right now denorms get explicitly disabled (but elsewhere) for x86,
     * whereas ppc64 explicitly enables them...
     */
-   if (util_cpu_caps.has_altivec) {
+   if (util_get_cpu_caps()->has_altivec) {
       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
       __asm (
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_logic.c b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
index 315977ae745..3ed3b5a74b1 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_logic.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
@@ -196,7 +196,7 @@ lp_build_compare(struct gallivm_state *gallivm,
 
    if (!type.floating && !type.sign &&
        type.width * type.length == 128 &&
-       util_cpu_caps.has_sse2 &&
+       util_get_cpu_caps()->has_sse2 &&
        (func == PIPE_FUNC_LESS ||
         func == PIPE_FUNC_LEQUAL ||
         func == PIPE_FUNC_GREATER ||
@@ -348,11 +348,11 @@ lp_build_select(struct lp_build_context *bld,
 
       res = LLVMBuildSelect(builder, mask, a, b, "");
    }
-   else if (((util_cpu_caps.has_sse4_1 &&
+   else if (((util_get_cpu_caps()->has_sse4_1 &&
               type.width * type.length == 128) ||
-             (util_cpu_caps.has_avx &&
+             (util_get_cpu_caps()->has_avx &&
               type.width * type.length == 256 && type.width >= 32) ||
-             (util_cpu_caps.has_avx2 &&
+             (util_get_cpu_caps()->has_avx2 &&
               type.width * type.length == 256)) &&
             !LLVMIsConstant(a) &&
             !LLVMIsConstant(b) &&
@@ -379,7 +379,7 @@ lp_build_select(struct lp_build_context *bld,
             intrinsic = "llvm.x86.avx.blendv.ps.256";
             arg_type = LLVMVectorType(LLVMFloatTypeInContext(lc), 8);
          } else {
-            assert(util_cpu_caps.has_avx2);
+            assert(util_get_cpu_caps()->has_avx2);
             intrinsic = "llvm.x86.avx2.pblendvb";
             arg_type = LLVMVectorType(LLVMInt8TypeInContext(lc), 32);
          }
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp b/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
index 9b75676a4e2..4f3e696816c 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
+++ b/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
@@ -400,22 +400,22 @@ lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
     * http://llvm.org/PR19429
     * http://llvm.org/PR16721
     */
-   MAttrs.push_back(util_cpu_caps.has_sse    ? "+sse"    : "-sse"   );
-   MAttrs.push_back(util_cpu_caps.has_sse2   ? "+sse2"   : "-sse2"  );
-   MAttrs.push_back(util_cpu_caps.has_sse3   ? "+sse3"   : "-sse3"  );
-   MAttrs.push_back(util_cpu_caps.has_ssse3  ? "+ssse3"  : "-ssse3" );
-   MAttrs.push_back(util_cpu_caps.has_sse4_1 ? "+sse4.1" : "-sse4.1");
-   MAttrs.push_back(util_cpu_caps.has_sse4_2 ? "+sse4.2" : "-sse4.2");
+   MAttrs.push_back(util_get_cpu_caps()->has_sse    ? "+sse"    : "-sse"   );
+   MAttrs.push_back(util_get_cpu_caps()->has_sse2   ? "+sse2"   : "-sse2"  );
+   MAttrs.push_back(util_get_cpu_caps()->has_sse3   ? "+sse3"   : "-sse3"  );
+   MAttrs.push_back(util_get_cpu_caps()->has_ssse3  ? "+ssse3"  : "-ssse3" );
+   MAttrs.push_back(util_get_cpu_caps()->has_sse4_1 ? "+sse4.1" : "-sse4.1");
+   MAttrs.push_back(util_get_cpu_caps()->has_sse4_2 ? "+sse4.2" : "-sse4.2");
    /*
     * AVX feature is not automatically detected from CPUID by the X86 target
     * yet, because the old (yet default) JIT engine is not capable of
     * emitting the opcodes. On newer llvm versions it is and at least some
     * versions (tested with 3.3) will emit avx opcodes without this anyway.
     */
-   MAttrs.push_back(util_cpu_caps.has_avx  ? "+avx"  : "-avx");
-   MAttrs.push_back(util_cpu_caps.has_f16c ? "+f16c" : "-f16c");
-   MAttrs.push_back(util_cpu_caps.has_fma  ? "+fma"  : "-fma");
-   MAttrs.push_back(util_cpu_caps.has_avx2 ? "+avx2" : "-avx2");
+   MAttrs.push_back(util_get_cpu_caps()->has_avx  ? "+avx"  : "-avx");
+   MAttrs.push_back(util_get_cpu_caps()->has_f16c ? "+f16c" : "-f16c");
+   MAttrs.push_back(util_get_cpu_caps()->has_fma  ? "+fma"  : "-fma");
+   MAttrs.push_back(util_get_cpu_caps()->has_avx2 ? "+avx2" : "-avx2");
    /* disable avx512 and all subvariants */
    MAttrs.push_back("-avx512cd");
    MAttrs.push_back("-avx512er");
@@ -426,7 +426,7 @@ lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
    MAttrs.push_back("-avx512vl");
 #endif
 #if defined(PIPE_ARCH_ARM)
-   if (!util_cpu_caps.has_neon) {
+   if (!util_get_cpu_caps()->has_neon) {
       MAttrs.push_back("-neon");
       MAttrs.push_back("-crypto");
       MAttrs.push_back("-vfp2");
@@ -434,7 +434,7 @@ lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
 #endif
 
 #if defined(PIPE_ARCH_PPC)
-   MAttrs.push_back(util_cpu_caps.has_altivec ? "+altivec" : "-altivec");
+   MAttrs.push_back(util_get_cpu_caps()->has_altivec ? "+altivec" : "-altivec");
 #if (LLVM_VERSION_MAJOR < 4)
    /*
     * Make sure VSX instructions are disabled
@@ -444,7 +444,7 @@ lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
     * https://llvm.org/bugs/show_bug.cgi?id=33531 (fixed in 4.0)
     * https://llvm.org/bugs/show_bug.cgi?id=34647 (llc performance on certain unusual shader IR; intro'd in 4.0, pending as of 5.0)
     */
-   if (util_cpu_caps.has_altivec) {
+   if (util_get_cpu_caps()->has_altivec) {
       MAttrs.push_back("-vsx");
    }
 #else
@@ -458,8 +458,8 @@ lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
     * Make sure VSX instructions are ENABLED (if supported), unless
     * VSX instructions are explicitly enabled/disabled via GALLIVM_VSX=1 or 0.
     */
-   if (util_cpu_caps.has_altivec) {
-      MAttrs.push_back(util_cpu_caps.has_vsx ? "+vsx" : "-vsx");
+   if (util_get_cpu_caps()->has_altivec) {
+      MAttrs.push_back(util_get_cpu_caps()->has_vsx ? "+vsx" : "-vsx");
    }
 #endif
 #endif
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_pack.c b/src/gallium/auxiliary/gallivm/lp_bld_pack.c
index e1f652a9342..76e57c52f80 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_pack.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_pack.c
@@ -322,7 +322,7 @@ lp_build_interleave2(struct gallivm_state *gallivm,
 {
    LLVMValueRef shuffle;
 
-   if (type.length == 2 && type.width == 128 && util_cpu_caps.has_avx) {
+   if (type.length == 2 && type.width == 128 && util_get_cpu_caps()->has_avx) {
       /*
        * XXX: This is a workaround for llvm code generation deficiency. Strangely
        * enough, while this needs vinsertf128/vextractf128 instructions (hence
@@ -484,7 +484,7 @@ lp_build_unpack2_native(struct gallivm_state *gallivm,
 
    /* Interleave bits */
 #if UTIL_ARCH_LITTLE_ENDIAN
-   if (src_type.length * src_type.width == 256 && util_cpu_caps.has_avx2) {
+   if (src_type.length * src_type.width == 256 && util_get_cpu_caps()->has_avx2) {
       *dst_lo = lp_build_interleave2_half(gallivm, src_type, src, msb, 0);
       *dst_hi = lp_build_interleave2_half(gallivm, src_type, src, msb, 1);
    } else {
@@ -585,22 +585,22 @@ lp_build_pack2(struct gallivm_state *gallivm,
    assert(src_type.length * 2 == dst_type.length);
 
    /* Check for special cases first */
-   if ((util_cpu_caps.has_sse2 || util_cpu_caps.has_altivec) &&
+   if ((util_get_cpu_caps()->has_sse2 || util_get_cpu_caps()->has_altivec) &&
         src_type.width * src_type.length >= 128) {
       const char *intrinsic = NULL;
       boolean swap_intrinsic_operands = FALSE;
 
       switch(src_type.width) {
       case 32:
-         if (util_cpu_caps.has_sse2) {
+         if (util_get_cpu_caps()->has_sse2) {
            if (dst_type.sign) {
               intrinsic = "llvm.x86.sse2.packssdw.128";
            } else {
-              if (util_cpu_caps.has_sse4_1) {
+              if (util_get_cpu_caps()->has_sse4_1) {
                  intrinsic = "llvm.x86.sse41.packusdw";
               }
            }
-         } else if (util_cpu_caps.has_altivec) {
+         } else if (util_get_cpu_caps()->has_altivec) {
             if (dst_type.sign) {
                intrinsic = "llvm.ppc.altivec.vpkswss";
             } else {
@@ -613,18 +613,18 @@ lp_build_pack2(struct gallivm_state *gallivm,
          break;
       case 16:
          if (dst_type.sign) {
-            if (util_cpu_caps.has_sse2) {
+            if (util_get_cpu_caps()->has_sse2) {
                intrinsic = "llvm.x86.sse2.packsswb.128";
-            } else if (util_cpu_caps.has_altivec) {
+            } else if (util_get_cpu_caps()->has_altivec) {
                intrinsic = "llvm.ppc.altivec.vpkshss";
 #if UTIL_ARCH_LITTLE_ENDIAN
                swap_intrinsic_operands = TRUE;
 #endif
             }
          } else {
-            if (util_cpu_caps.has_sse2) {
+            if (util_get_cpu_caps()->has_sse2) {
                intrinsic = "llvm.x86.sse2.packuswb.128";
-            } else if (util_cpu_caps.has_altivec) {
+            } else if (util_get_cpu_caps()->has_altivec) {
                intrinsic = "llvm.ppc.altivec.vpkshus";
 #if UTIL_ARCH_LITTLE_ENDIAN
                swap_intrinsic_operands = TRUE;
@@ -740,7 +740,7 @@ lp_build_pack2_native(struct gallivm_state *gallivm,
 
    /* At this point only have special case for avx2 */
    if (src_type.length * src_type.width == 256 &&
-       util_cpu_caps.has_avx2) {
+       util_get_cpu_caps()->has_avx2) {
       switch(src_type.width) {
       case 32:
          if (dst_type.sign) {
@@ -793,7 +793,7 @@ lp_build_packs2(struct gallivm_state *gallivm,
 
    /* All X86 SSE non-interleaved pack instructions take signed inputs and
     * saturate them, so no need to clamp for those cases. */
-   if(util_cpu_caps.has_sse2 &&
+   if(util_get_cpu_caps()->has_sse2 &&
       src_type.width * src_type.length >= 128 &&
       src_type.sign &&
       (src_type.width == 32 || src_type.width == 16))
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample.c b/src/gallium/auxiliary/gallivm/lp_bld_sample.c
index 686abc08620..98dcde912b5 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample.c
@@ -1152,7 +1152,7 @@ lp_build_minify(struct lp_build_context *bld,
       LLVMValueRef size;
       assert(bld->type.sign);
       if (lod_scalar ||
-         (util_cpu_caps.has_avx2 || !util_cpu_caps.has_sse)) {
+         (util_get_cpu_caps()->has_avx2 || !util_get_cpu_caps()->has_sse)) {
          size = LLVMBuildLShr(builder, base_size, level, "minify");
          size = lp_build_max(bld, size, bld->one);
       }
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
index 2b91edd37c7..6e47640e70d 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
@@ -3234,7 +3234,7 @@ lp_build_sample_soa_code(struct gallivm_state *gallivm,
        * as it appears to be a loss with just AVX)
        */
       if (num_quads == 1 || !use_aos ||
-          (util_cpu_caps.has_avx2 &&
+          (util_get_cpu_caps()->has_avx2 &&
            (bld.num_lods == 1 ||
             derived_sampler_state.min_img_filter == derived_sampler_state.mag_img_filter))) {
          if (use_aos) {
diff --git a/src/gallium/auxiliary/rtasm/rtasm_cpu.c b/src/gallium/auxiliary/rtasm/rtasm_cpu.c
index b1c8b990ef1..03b11f914b4 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_cpu.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_cpu.c
@@ -35,10 +35,10 @@
 
 DEBUG_GET_ONCE_BOOL_OPTION(nosse, "GALLIUM_NOSSE", false);
 
-static struct util_cpu_caps *get_cpu_caps(void)
+static const struct util_cpu_caps_t *get_cpu_caps(void)
 {
    util_cpu_detect();
-   return &util_cpu_caps;
+   return util_get_cpu_caps();
 }
 
 int rtasm_cpu_has_sse(void)
diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c
index ad687f32853..ddd65fb6a08 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c
@@ -2152,17 +2152,17 @@ static void x86_init_func_common( struct x86_function *p )
 {
    util_cpu_detect();
    p->caps = 0;
-   if(util_cpu_caps.has_mmx)
+   if(util_get_cpu_caps()->has_mmx)
       p->caps |= X86_MMX;
-   if(util_cpu_caps.has_mmx2)
+   if(util_get_cpu_caps()->has_mmx2)
       p->caps |= X86_MMX2;
-   if(util_cpu_caps.has_sse)
+   if(util_get_cpu_caps()->has_sse)
       p->caps |= X86_SSE;
-   if(util_cpu_caps.has_sse2)
+   if(util_get_cpu_caps()->has_sse2)
       p->caps |= X86_SSE2;
-   if(util_cpu_caps.has_sse3)
+   if(util_get_cpu_caps()->has_sse3)
       p->caps |= X86_SSE3;
-   if(util_cpu_caps.has_sse4_1)
+   if(util_get_cpu_caps()->has_sse4_1)
       p->caps |= X86_SSE4_1;
    p->csr = p->store;
 #if defined(PIPE_ARCH_X86)
diff --git a/src/gallium/auxiliary/util/u_threaded_context.c b/src/gallium/auxiliary/util/u_threaded_context.c
index 1eaff77724e..bf56993db09 100644
--- a/src/gallium/auxiliary/util/u_threaded_context.c
+++ b/src/gallium/auxiliary/util/u_threaded_context.c
@@ -2071,8 +2071,8 @@ tc_set_context_param(struct pipe_context *_pipe,
    if (param == PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE) {
       /* Pin the gallium thread as requested. */
       util_set_thread_affinity(tc->queue.threads[0],
-                               util_cpu_caps.L3_affinity_mask[value],
-                               NULL, UTIL_MAX_CPUS);
+                               util_get_cpu_caps()->L3_affinity_mask[value],
+                               NULL, util_get_cpu_caps()->num_cpu_mask_bits);
 
       /* Execute this immediately (without enqueuing).
        * It's required to be thread-safe.
@@ -2720,7 +2720,7 @@ threaded_context_create(struct pipe_context *pipe,
 
    util_cpu_detect();
 
-   if (!debug_get_bool_option("GALLIUM_THREAD", util_cpu_caps.nr_cpus > 1))
+   if (!debug_get_bool_option("GALLIUM_THREAD", util_get_cpu_caps()->nr_cpus > 1))
       return pipe;
 
    tc = os_malloc_aligned(sizeof(struct threaded_context), 16);
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_depth.c b/src/gallium/drivers/llvmpipe/lp_bld_depth.c
index 64cf72ae101..913c1bd2462 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_depth.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_depth.c
@@ -435,7 +435,7 @@ lp_build_occlusion_count(struct gallivm_state *gallivm,
    assert(type.length <= 16);
    assert(type.floating);
 
-   if(util_cpu_caps.has_sse && type.length == 4) {
+   if(util_get_cpu_caps()->has_sse && type.length == 4) {
       const char *movmskintr = "llvm.x86.sse.movmsk.ps";
       const char *popcntintr = "llvm.ctpop.i32";
       LLVMValueRef bits = LLVMBuildBitCast(builder, maskvalue,
@@ -446,7 +446,7 @@ lp_build_occlusion_count(struct gallivm_state *gallivm,
                                        LLVMInt32TypeInContext(context), bits);
       count = LLVMBuildZExt(builder, count, LLVMIntTypeInContext(context, 64), "");
    }
-   else if(util_cpu_caps.has_avx && type.length == 8) {
+   else if(util_get_cpu_caps()->has_avx && type.length == 8) {
       const char *movmskintr = "llvm.x86.avx.movmsk.ps.256";
       const char *popcntintr = "llvm.ctpop.i32";
       LLVMValueRef bits = LLVMBuildBitCast(builder, maskvalue,
diff --git a/src/gallium/drivers/llvmpipe/lp_screen.c b/src/gallium/drivers/llvmpipe/lp_screen.c
index f133bbf8a4d..628a4338c1e 100644
--- a/src/gallium/drivers/llvmpipe/lp_screen.c
+++ b/src/gallium/drivers/llvmpipe/lp_screen.c
@@ -915,7 +915,7 @@ llvmpipe_create_screen(struct sw_winsys *winsys)
 
    screen->allow_cl = !!getenv("LP_CL");
    screen->use_tgsi = (LP_DEBUG & DEBUG_TGSI_IR);
-   screen->num_threads = util_cpu_caps.nr_cpus > 1 ? util_cpu_caps.nr_cpus : 0;
+   screen->num_threads = util_get_cpu_caps()->nr_cpus > 1 ? util_get_cpu_caps()->nr_cpus : 0;
 #ifdef EMBEDDED_DEVICE
    screen->num_threads = 0;
 #endif
diff --git a/src/gallium/drivers/llvmpipe/lp_test_arit.c b/src/gallium/drivers/llvmpipe/lp_test_arit.c
index 873dcf37fac..725854cc25c 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_arit.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_arit.c
@@ -382,7 +382,7 @@ flush_denorm_to_zero(float val)
    fi_val.f = val;
 
 #if defined(PIPE_ARCH_SSE)
-   if (util_cpu_caps.has_sse) {
+   if (util_get_cpu_caps()->has_sse) {
       if ((fi_val.ui & 0x7f800000) == 0) {
          fi_val.ui &= 0xff800000;
       }
@@ -458,7 +458,7 @@ test_unary(unsigned verbose, FILE *fp, const struct unary_test_t *test, unsigned
             continue;
          }
 
-         if (!util_cpu_caps.has_neon &&
+         if (!util_get_cpu_caps()->has_neon &&
              test->ref == &nearbyintf && length == 2 &&
              ref != roundf(testval)) {
             /* FIXME: The generic (non SSE) path in lp_build_iround, which is
diff --git a/src/gallium/drivers/llvmpipe/lp_texture.c b/src/gallium/drivers/llvmpipe/lp_texture.c
index 2bf223d66f9..815736166d5 100644
--- a/src/gallium/drivers/llvmpipe/lp_texture.c
+++ b/src/gallium/drivers/llvmpipe/lp_texture.c
@@ -85,7 +85,7 @@ llvmpipe_texture_layout(struct llvmpipe_screen *screen,
     * of a block for all formats) though this should not be strictly necessary
     * neither. In any case it can only affect compressed or 1d textures.
     */
-   unsigned mip_align = MAX2(64, util_cpu_caps.cacheline);
+   unsigned mip_align = MAX2(64, util_get_cpu_caps()->cacheline);
 
    assert(LP_MAX_TEXTURE_2D_LEVELS <= LP_MAX_TEXTURE_LEVELS);
    assert(LP_MAX_TEXTURE_3D_LEVELS <= LP_MAX_TEXTURE_LEVELS);
@@ -123,7 +123,7 @@ llvmpipe_texture_layout(struct llvmpipe_screen *screen,
       if (util_format_is_compressed(pt->format))
          lpr->row_stride[level] = nblocksx * block_size;
       else
-         lpr->row_stride[level] = align(nblocksx * block_size, util_cpu_caps.cacheline);
+         lpr->row_stride[level] = align(nblocksx * block_size, util_get_cpu_caps()->cacheline);
 
       /* if row_stride * height > LP_MAX_TEXTURE_SIZE */
       if ((uint64_t)lpr->row_stride[level] * nblocksy > LP_MAX_TEXTURE_SIZE) {
diff --git a/src/gallium/drivers/swr/swr_loader.cpp b/src/gallium/drivers/swr/swr_loader.cpp
index 97db7ca3e8b..d891b6b14e8 100644
--- a/src/gallium/drivers/swr/swr_loader.cpp
+++ b/src/gallium/drivers/swr/swr_loader.cpp
@@ -91,7 +91,7 @@ swr_create_screen(struct sw_winsys *winsys)
 
    util_cpu_detect();
 
-   if (util_cpu_caps.has_avx512f && util_cpu_caps.has_avx512er) {
+   if (util_get_cpu_caps()->has_avx512f && util_get_cpu_caps()->has_avx512er) {
       swr_print_info("SWR detected KNL instruction support ");
 #ifndef HAVE_SWR_KNL
       swr_print_info("(skipping: not built).\n");
@@ -103,7 +103,7 @@ swr_create_screen(struct sw_winsys *winsys)
 #endif
    }
 
-   if (util_cpu_caps.has_avx512f && util_cpu_caps.has_avx512bw) {
+   if (util_get_cpu_caps()->has_avx512f && util_get_cpu_caps()->has_avx512bw) {
       swr_print_info("SWR detected SKX instruction support ");
 #ifndef HAVE_SWR_SKX
       swr_print_info("(skipping not built).\n");
@@ -113,7 +113,7 @@ swr_create_screen(struct sw_winsys *winsys)
 #endif
    }
 
-   if (util_cpu_caps.has_avx2) {
+   if (util_get_cpu_caps()->has_avx2) {
       swr_print_info("SWR detected AVX2 instruction support ");
 #ifndef HAVE_SWR_AVX2
       swr_print_info("(skipping not built).\n");
@@ -123,7 +123,7 @@ swr_create_screen(struct sw_winsys *winsys)
 #endif
    }
 
-   if (util_cpu_caps.has_avx) {
+   if (util_get_cpu_caps()->has_avx) {
       swr_print_info("SWR detected AVX instruction support ");
 #ifndef HAVE_SWR_AVX
       swr_print_info("(skipping not built).\n");
diff --git a/src/gallium/drivers/vc4/vc4_tiling.h b/src/gallium/drivers/vc4/vc4_tiling.h
index 66767e7f1f8..5afe32939a8 100644
--- a/src/gallium/drivers/vc4/vc4_tiling.h
+++ b/src/gallium/drivers/vc4/vc4_tiling.h
@@ -90,7 +90,7 @@ vc4_load_lt_image(void *dst, uint32_t dst_stride,
                   int cpp, const struct pipe_box *box)
 {
 #ifdef USE_ARM_ASM
-        if (util_cpu_caps.has_neon) {
+        if (util_get_cpu_caps()->has_neon) {
                 vc4_load_lt_image_neon(dst, dst_stride, src, src_stride,
                                        cpp, box);
                 return;
@@ -106,7 +106,7 @@ vc4_store_lt_image(void *dst, uint32_t dst_stride,
                    int cpp, const struct pipe_box *box)
 {
 #ifdef USE_ARM_ASM
-        if (util_cpu_caps.has_neon) {
+        if (util_get_cpu_caps()->has_neon) {
                 vc4_store_lt_image_neon(dst, dst_stride, src, src_stride,
                                         cpp, box);
                 return;
diff --git a/src/gallium/tests/unit/translate_test.c b/src/gallium/tests/unit/translate_test.c
index 4d9c4e27ebf..782f16e7f78 100644
--- a/src/gallium/tests/unit/translate_test.c
+++ b/src/gallium/tests/unit/translate_test.c
@@ -50,6 +50,7 @@ int main(int argc, char** argv)
 {
    struct translate *(*create_fn)(const struct translate_key *key) = 0;
 
+   extern struct util_cpu_caps_t util_cpu_caps;
    struct translate_key key;
    unsigned output_format;
    unsigned input_format;
@@ -87,7 +88,7 @@ int main(int argc, char** argv)
    }
    else if (!strcmp(argv[1], "sse"))
    {
-      if(!util_cpu_caps.has_sse || !rtasm_cpu_has_sse())
+      if(!util_get_cpu_caps()->has_sse || !rtasm_cpu_has_sse())
       {
          printf("Error: CPU doesn't support SSE (test with qemu)\n");
          return 2;
@@ -99,7 +100,7 @@ int main(int argc, char** argv)
    }
    else if (!strcmp(argv[1], "sse2"))
    {
-      if(!util_cpu_caps.has_sse2 || !rtasm_cpu_has_sse())
+      if(!util_get_cpu_caps()->has_sse2 || !rtasm_cpu_has_sse())
       {
          printf("Error: CPU doesn't support SSE2 (test with qemu)\n");
          return 2;
@@ -110,7 +111,7 @@ int main(int argc, char** argv)
    }
    else if (!strcmp(argv[1], "sse3"))
    {
-      if(!util_cpu_caps.has_sse3 || !rtasm_cpu_has_sse())
+      if(!util_get_cpu_caps()->has_sse3 || !rtasm_cpu_has_sse())
       {
          printf("Error: CPU doesn't support SSE3 (test with qemu)\n");
          return 2;
@@ -120,7 +121,7 @@ int main(int argc, char** argv)
    }
    else if (!strcmp(argv[1], "sse4.1"))
    {
-      if(!util_cpu_caps.has_sse4_1 || !rtasm_cpu_has_sse())
+      if(!util_get_cpu_caps()->has_sse4_1 || !rtasm_cpu_has_sse())
       {
          printf("Error: CPU doesn't support SSE4.1 (test with qemu)\n");
          return 2;
diff --git a/src/gallium/tests/unit/u_half_test.c b/src/gallium/tests/unit/u_half_test.c
index 7f2eba9382b..4474cfb82b0 100644
--- a/src/gallium/tests/unit/u_half_test.c
+++ b/src/gallium/tests/unit/u_half_test.c
@@ -36,13 +36,14 @@ test(void)
 int
 main(int argc, char **argv)
 {
-   assert(!util_cpu_caps.has_f16c);
+   util_cpu_detect();
    test();
 
-   /* Test f16c. */
-   util_cpu_detect();
-   if (util_cpu_caps.has_f16c)
+   /* Test non-f16c. */
+   if (util_get_cpu_caps()->has_f16c) {
+      ((struct util_cpu_caps_t *)util_get_cpu_caps())->has_f16c = false;
       test();
+   }
 
    printf("Success!\n");
    return 0;
diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c b/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c
index 8a0aedfed64..a18362ce6ea 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c
@@ -312,8 +312,8 @@ static void amdgpu_pin_threads_to_L3_cache(struct radeon_winsys *rws,
    struct amdgpu_winsys *ws = amdgpu_winsys(rws);
 
    util_set_thread_affinity(ws->cs_queue.threads[0],
-                            util_cpu_caps.L3_affinity_mask[cache],
-                            NULL, UTIL_MAX_CPUS);
+                            util_get_cpu_caps()->L3_affinity_mask[cache],
+                            NULL, util_get_cpu_caps()->num_cpu_mask_bits);
 }
 
 static uint32_t kms_handle_hash(const void *key)
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
index f0e1b9f7df3..4430ce50466 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
@@ -801,8 +801,8 @@ static void radeon_pin_threads_to_L3_cache(struct radeon_winsys *ws,
 
    if (util_queue_is_initialized(&rws->cs_queue)) {
       util_set_thread_affinity(rws->cs_queue.threads[0],
-                               util_cpu_caps.L3_affinity_mask[cache],
-                               NULL, UTIL_MAX_CPUS);
+                               util_get_cpu_caps()->L3_affinity_mask[cache],
+                               NULL, util_get_cpu_caps()->num_cpu_mask_bits);
    }
 }
 
diff --git a/src/mesa/main/glthread.c b/src/mesa/main/glthread.c
index eb8eb30cabc..c9dfef541fc 100644
--- a/src/mesa/main/glthread.c
+++ b/src/mesa/main/glthread.c
@@ -199,19 +199,20 @@ _mesa_glthread_flush_batch(struct gl_context *ctx)
    /* Pin threads regularly to the same Zen CCX that the main thread is
     * running on. The main thread can move between CCXs.
     */
-   if (util_cpu_caps.nr_cpus != util_cpu_caps.cores_per_L3 &&
+   if (util_get_cpu_caps()->nr_cpus != util_get_cpu_caps()->cores_per_L3 &&
        /* driver support */
        ctx->Driver.PinDriverToL3Cache &&
        ++glthread->pin_thread_counter % 128 == 0) {
       int cpu = util_get_current_cpu();
 
       if (cpu >= 0) {
-         unsigned L3_cache = util_cpu_caps.cpu_to_L3[cpu];
-
-         util_set_thread_affinity(glthread->queue.threads[0],
-                                  util_cpu_caps.L3_affinity_mask[L3_cache],
-                                  NULL, UTIL_MAX_CPUS);
-         ctx->Driver.PinDriverToL3Cache(ctx, L3_cache);
+         uint16_t L3_cache = util_get_cpu_caps()->cpu_to_L3[cpu];
+         if (L3_cache != U_CPU_INVALID_L3) {
+            util_set_thread_affinity(glthread->queue.threads[0],
+                                     util_get_cpu_caps()->L3_affinity_mask[L3_cache],
+                                     NULL, util_get_cpu_caps()->num_cpu_mask_bits);
+            ctx->Driver.PinDriverToL3Cache(ctx, L3_cache);
+         }
       }
    }
 
diff --git a/src/mesa/state_tracker/st_context.c b/src/mesa/state_tracker/st_context.c
index 40364296664..f27fa7ff29c 100644
--- a/src/mesa/state_tracker/st_context.c
+++ b/src/mesa/state_tracker/st_context.c
@@ -815,6 +815,10 @@ st_create_context_priv(struct gl_context *ctx, struct pipe_context *pipe,
          !st->lower_ucp;
    st->shader_has_one_variant[MESA_SHADER_COMPUTE] = st->has_shareable_shaders;
 
+   if (util_get_cpu_caps()->cores_per_L3 == util_get_cpu_caps()->nr_cpus ||
+       !st->pipe->set_context_param)
+      st->pin_thread_counter = ST_L3_PINNING_DISABLED;
+
    st->bitmap.cache.empty = true;
 
    if (ctx->Const.ForceGLNamesReuse && ctx->Shared->RefCount == 1) {
diff --git a/src/mesa/state_tracker/st_context.h b/src/mesa/state_tracker/st_context.h
index b1fda06ff3e..9ab6969de62 100644
--- a/src/mesa/state_tracker/st_context.h
+++ b/src/mesa/state_tracker/st_context.h
@@ -55,6 +55,7 @@ struct st_program;
 struct st_perf_monitor_group;
 struct u_upload_mgr;
 
+#define ST_L3_PINNING_DISABLED 0xffffffff
 
 struct st_bitmap_cache
 {
@@ -130,6 +131,9 @@ struct st_context
    struct draw_stage *feedback_stage;  /**< For GL_FEEDBACK rendermode */
    struct draw_stage *selection_stage;  /**< For GL_SELECT rendermode */
    struct draw_stage *rastpos_stage;  /**< For glRasterPos */
+
+   unsigned pin_thread_counter; /* for L3 thread pinning on AMD Zen */
+
    GLboolean clamp_frag_color_in_shader;
    GLboolean clamp_vert_color_in_shader;
    boolean clamp_frag_depth_in_shader;
@@ -235,8 +239,6 @@ struct st_context
    /** This masks out unused shader resources. Only valid in draw calls. */
    uint64_t active_states;
 
-   unsigned pin_thread_counter; /* for L3 thread pinning on AMD Zen */
-
    /* If true, further analysis of states is required to know if something
     * has changed. Used mainly for shaders.
     */
diff --git a/src/mesa/state_tracker/st_draw.c b/src/mesa/state_tracker/st_draw.c
index 996d985510c..159d7017b07 100644
--- a/src/mesa/state_tracker/st_draw.c
+++ b/src/mesa/state_tracker/st_draw.c
@@ -124,26 +124,26 @@ prepare_draw(struct st_context *st, struct gl_context *ctx)
       st_validate_state(st, ST_PIPELINE_RENDER);
    }
 
-   struct pipe_context *pipe = st->pipe;
-
    /* Pin threads regularly to the same Zen CCX that the main thread is
     * running on. The main thread can move between CCXs.
     */
-   if (unlikely(/* AMD Zen */
-                util_cpu_caps.nr_cpus != util_cpu_caps.cores_per_L3 &&
+   if (unlikely(st->pin_thread_counter != ST_L3_PINNING_DISABLED &&
                 /* no glthread */
                 ctx->CurrentClientDispatch != ctx->MarshalExec &&
-                /* driver support */
-                pipe->set_context_param &&
                 /* do it occasionally */
                 ++st->pin_thread_counter % 512 == 0)) {
+      st->pin_thread_counter = 0;
+
       int cpu = util_get_current_cpu();
       if (cpu >= 0) {
-         unsigned L3_cache = util_cpu_caps.cpu_to_L3[cpu];
-
-         pipe->set_context_param(pipe,
-                                 PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE,
-                                 L3_cache);
+         struct pipe_context *pipe = st->pipe;
+         uint16_t L3_cache = util_get_cpu_caps()->cpu_to_L3[cpu];
+
+         if (L3_cache != U_CPU_INVALID_L3) {
+            pipe->set_context_param(pipe,
+                                    PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE,
+                                    L3_cache);
+         }
       }
    }
 }
diff --git a/src/util/half_float.h b/src/util/half_float.h
index c52bccf8d1e..8f1a1dbf11d 100644
--- a/src/util/half_float.h
+++ b/src/util/half_float.h
@@ -59,7 +59,7 @@ static inline uint16_t
 _mesa_float_to_half(float val)
 {
 #if defined(USE_X86_64_ASM)
-   if (util_cpu_caps.has_f16c) {
+   if (util_get_cpu_caps()->has_f16c) {
       __m128 in = {val};
       __m128i out;
 
@@ -75,7 +75,7 @@ static inline float
 _mesa_half_to_float(uint16_t val)
 {
 #if defined(USE_X86_64_ASM)
-   if (util_cpu_caps.has_f16c) {
+   if (util_get_cpu_caps()->has_f16c) {
       __m128i in = {val};
       __m128 out;
 
@@ -90,7 +90,7 @@ static inline uint16_t
 _mesa_float_to_float16_rtz(float val)
 {
 #if defined(USE_X86_64_ASM)
-   if (util_cpu_caps.has_f16c) {
+   if (util_get_cpu_caps()->has_f16c) {
       __m128 in = {val};
       __m128i out;
 
diff --git a/src/util/tests/format/u_format_test.c b/src/util/tests/format/u_format_test.c
index f4a62a5c6a8..e6473c2bf6d 100644
--- a/src/util/tests/format/u_format_test.c
+++ b/src/util/tests/format/u_format_test.c
@@ -850,6 +850,8 @@ int main(int argc, char **argv)
 {
    boolean success;
 
+   util_cpu_detect();
+
    success = test_all();
 
    return success ? 0 : 1;
diff --git a/src/util/u_cpu_detect.c b/src/util/u_cpu_detect.c
index 025f2f30156..4a4b06e1bc6 100644
--- a/src/util/u_cpu_detect.c
+++ b/src/util/u_cpu_detect.c
@@ -90,7 +90,7 @@
 DEBUG_GET_ONCE_BOOL_OPTION(dump_cpu, "GALLIUM_DUMP_CPU", false)
 
 
-struct util_cpu_caps util_cpu_caps;
+struct util_cpu_caps_t util_cpu_caps;
 
 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
 static int has_cpuid(void);
@@ -438,26 +438,22 @@ get_cpu_topology(void)
    util_cpu_caps.cores_per_L3 = util_cpu_caps.nr_cpus;
    util_cpu_caps.num_L3_caches = 1;
 
+   memset(util_cpu_caps.cpu_to_L3, 0xff, sizeof(util_cpu_caps.cpu_to_L3));
+
 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
    /* AMD Zen */
    if (util_cpu_caps.family >= CPU_AMD_ZEN1_ZEN2 &&
        util_cpu_caps.family < CPU_AMD_LAST) {
       uint32_t regs[4];
 
-      /* Query the L3 cache count. */
-      cpuid_count(0x8000001D, 3, regs);
-      unsigned cache_level = (regs[0] >> 5) & 0x7;
-      unsigned cores_per_L3 = ((regs[0] >> 14) & 0xfff) + 1;
-
-      if (cache_level != 3 || cores_per_L3 == util_cpu_caps.nr_cpus)
-         return;
-
       uint32_t saved_mask[UTIL_MAX_CPUS / 32] = {0};
       uint32_t mask[UTIL_MAX_CPUS / 32] = {0};
-      uint32_t allowed_mask[UTIL_MAX_CPUS / 32] = {0};
-      uint32_t apic_id[UTIL_MAX_CPUS];
       bool saved = false;
 
+      uint32_t L3_found[UTIL_MAX_CPUS] = {0};
+      uint32_t num_L3_caches = 0;
+      util_affinity_mask *L3_affinity_masks = NULL;
+
       /* Query APIC IDs from each CPU core.
        *
        * An APIC ID is a logical ID of the CPU with respect to the cache
@@ -482,41 +478,60 @@ get_cpu_topology(void)
 
          if (util_set_current_thread_affinity(mask,
                                               !saved ? saved_mask : NULL,
-                                              UTIL_MAX_CPUS)) {
+                                              util_cpu_caps.num_cpu_mask_bits)) {
             saved = true;
-            allowed_mask[i / 32] |= cpu_bit;
 
             /* Query the APIC ID of the current core. */
             cpuid(0x00000001, regs);
-            apic_id[i] = regs[1] >> 24;
+            unsigned apic_id = regs[1] >> 24;
+
+            /* Query the total core count for the CPU */
+            uint32_t core_count = 1;
+            if (regs[3] & (1 << 28))
+               core_count = (regs[1] >> 16) & 0xff;
+
+            core_count = util_next_power_of_two(core_count);
+
+            /* Query the L3 cache count. */
+            cpuid_count(0x8000001D, 3, regs);
+            unsigned cache_level = (regs[0] >> 5) & 0x7;
+            unsigned cores_per_L3 = ((regs[0] >> 14) & 0xfff) + 1;
+
+            if (cache_level != 3)
+               continue;
+
+            unsigned local_core_id = apic_id & (core_count - 1);
+            unsigned phys_id = (apic_id & ~(core_count - 1)) >> util_logbase2(core_count);
+            unsigned local_l3_cache_index = local_core_id / util_next_power_of_two(cores_per_L3);
+#define L3_ID(p, i) (p << 16 | i << 1 | 1);
+
+            unsigned l3_id = L3_ID(phys_id, local_l3_cache_index);
+            int idx = -1;
+            for (unsigned c = 0; c < num_L3_caches; c++) {
+               if (L3_found[c] == l3_id) {
+                  idx = c;
+                  break;
+               }
+            }
+            if (idx == -1) {
+               idx = num_L3_caches;
+               L3_found[num_L3_caches++] = l3_id;
+               L3_affinity_masks = realloc(L3_affinity_masks, sizeof(util_affinity_mask) * num_L3_caches);
+               if (!L3_affinity_masks)
+                  return;
+               memset(&L3_affinity_masks[num_L3_caches - 1], 0, sizeof(util_affinity_mask));
+            }
+            util_cpu_caps.cpu_to_L3[i] = idx;
+            L3_affinity_masks[idx][i / 32] |= cpu_bit;
+
          }
          mask[i / 32] = 0;
       }
 
-      if (saved) {
-
-         /* We succeeded in using at least one CPU. */
-         util_cpu_caps.num_L3_caches = util_cpu_caps.nr_cpus / cores_per_L3;
-         util_cpu_caps.cores_per_L3 = cores_per_L3;
-         util_cpu_caps.L3_affinity_mask = calloc(sizeof(util_affinity_mask),
-                                                 util_cpu_caps.num_L3_caches);
-
-         for (unsigned i = 0; i < util_cpu_caps.nr_cpus && i < UTIL_MAX_CPUS;
-              i++) {
-            uint32_t cpu_bit = 1u << (i % 32);
-
-            if (allowed_mask[i / 32] & cpu_bit) {
-               /* Each APIC ID bit represents a topology level, so we need
-                * to round up to the next power of two.
-                */
-               unsigned L3_index = apic_id[i] /
-                                   util_next_power_of_two(cores_per_L3);
-
-               util_cpu_caps.L3_affinity_mask[L3_index][i / 32] |= cpu_bit;
-               util_cpu_caps.cpu_to_L3[i] = L3_index;
-            }
-         }
+      util_cpu_caps.num_L3_caches = num_L3_caches;
+      util_cpu_caps.L3_affinity_mask = L3_affinity_masks;
 
+      if (saved) {
          if (debug_get_option_dump_cpu()) {
             fprintf(stderr, "CPU <-> L3 cache mapping:\n");
             for (unsigned i = 0; i < util_cpu_caps.num_L3_caches; i++) {
@@ -528,7 +543,8 @@ get_cpu_topology(void)
          }
 
          /* Restore the original affinity mask. */
-         util_set_current_thread_affinity(saved_mask, NULL, UTIL_MAX_CPUS);
+         util_set_current_thread_affinity(saved_mask, NULL,
+                                          util_cpu_caps.num_cpu_mask_bits);
       } else {
          if (debug_get_option_dump_cpu())
             fprintf(stderr, "Cannot set thread affinity for any thread.\n");
@@ -547,7 +563,7 @@ util_cpu_detect_once(void)
    {
       SYSTEM_INFO system_info;
       GetSystemInfo(&system_info);
-      util_cpu_caps.nr_cpus = system_info.dwNumberOfProcessors;
+      util_cpu_caps.nr_cpus = MAX2(1, system_info.dwNumberOfProcessors);
    }
 #elif defined(PIPE_OS_UNIX) && defined(_SC_NPROCESSORS_ONLN)
    util_cpu_caps.nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
@@ -569,6 +585,8 @@ util_cpu_detect_once(void)
    util_cpu_caps.nr_cpus = 1;
 #endif
 
+   util_cpu_caps.num_cpu_mask_bits = align(util_cpu_caps.nr_cpus, 32);
+
    /* Make the fallback cacheline size nonzero so that it can be
     * safely passed to align().
     */
diff --git a/src/util/u_cpu_detect.h b/src/util/u_cpu_detect.h
index a76fd912910..1c7239b2ec7 100644
--- a/src/util/u_cpu_detect.h
+++ b/src/util/u_cpu_detect.h
@@ -55,7 +55,7 @@ enum cpu_family {
 
 typedef uint32_t util_affinity_mask[UTIL_MAX_CPUS / 32];
 
-struct util_cpu_caps {
+struct util_cpu_caps_t {
    int nr_cpus;
    enum cpu_family family;
 
@@ -98,14 +98,27 @@ struct util_cpu_caps {
 
    unsigned num_L3_caches;
    unsigned cores_per_L3;
+   unsigned num_cpu_mask_bits;
 
    uint16_t cpu_to_L3[UTIL_MAX_CPUS];
    /* Affinity masks for each L3 cache. */
    util_affinity_mask *L3_affinity_mask;
 };
 
-extern struct util_cpu_caps
-util_cpu_caps;
+#define U_CPU_INVALID_L3 0xffff
+
+static inline const struct util_cpu_caps_t *
+util_get_cpu_caps(void)
+{
+	extern struct util_cpu_caps_t util_cpu_caps;
+
+	/* If you hit this assert, it means that something is using the
+	 * cpu-caps without having first called util_cpu_detect()
+	 */
+	assert(util_cpu_caps.nr_cpus >= 1);
+
+	return &util_cpu_caps;
+}
 
 void util_cpu_detect(void);
 
diff --git a/src/util/u_math.c b/src/util/u_math.c
index 9a8a9ecbbde..41e7f599eb0 100644
--- a/src/util/u_math.c
+++ b/src/util/u_math.c
@@ -92,7 +92,7 @@ util_fpstate_get(void)
    unsigned mxcsr = 0;
 
 #if defined(PIPE_ARCH_SSE)
-   if (util_cpu_caps.has_sse) {
+   if (util_get_cpu_caps()->has_sse) {
       mxcsr = _mm_getcsr();
    }
 #endif
@@ -110,10 +110,10 @@ unsigned
 util_fpstate_set_denorms_to_zero(unsigned current_mxcsr)
 {
 #if defined(PIPE_ARCH_SSE)
-   if (util_cpu_caps.has_sse) {
+   if (util_get_cpu_caps()->has_sse) {
       /* Enable flush to zero mode */
       current_mxcsr |= _MM_FLUSH_ZERO_MASK;
-      if (util_cpu_caps.has_daz) {
+      if (util_get_cpu_caps()->has_daz) {
          /* Enable denormals are zero mode */
          current_mxcsr |= _MM_DENORMALS_ZERO_MASK;
       }
@@ -132,7 +132,7 @@ void
 util_fpstate_set(unsigned mxcsr)
 {
 #if defined(PIPE_ARCH_SSE)
-   if (util_cpu_caps.has_sse) {
+   if (util_get_cpu_caps()->has_sse) {
       _mm_setcsr(mxcsr);
    }
 #endif
diff --git a/src/util/u_queue.c b/src/util/u_queue.c
index b11b297a45c..8f21f0667c6 100644
--- a/src/util/u_queue.c
+++ b/src/util/u_queue.c
@@ -27,7 +27,7 @@
 #include "u_queue.h"
 
 #include "c11/threads.h"
-
+#include "util/u_cpu_detect.h"
 #include "util/os_time.h"
 #include "util/u_string.h"
 #include "util/u_thread.h"
@@ -258,7 +258,8 @@ util_queue_thread_func(void *input)
       uint32_t mask[UTIL_MAX_CPUS / 32];
 
       memset(mask, 0xff, sizeof(mask));
-      util_set_current_thread_affinity(mask, NULL, UTIL_MAX_CPUS);
+      util_set_current_thread_affinity(mask, NULL,
+                                       util_get_cpu_caps()->num_cpu_mask_bits);
    }
 
 #if defined(__linux__)