Insert match::opaque into deep matchers to fix large symbol names on windows and debug gcc builds

.github/workflows/ci.yaml

@@ -454,7 +454,7 @@ jobs:
       env:
         CMAKE_PREFIX_PATH: ${{ github.workspace }}/cget
         CCACHE_LOGFILE: /tmp/ccache.log
-        CXXFLAGS: -Werror -pthread -fdebug-prefix-map=$PWD=. -fdebug-types-section -DMIGRAPHX_USE_TYPE_ERASED_MATCHERS=1 --param ggc-min-expand=5 --param ggc-min-heapsize=8192
+        CXXFLAGS: -Werror -pthread -fdebug-prefix-map=$PWD=. -fdebug-types-section --param ggc-min-expand=5 --param ggc-min-heapsize=8192
       run: |
         echo "leak:dnnl::impl::malloc" > suppressions.txt
         export LSAN_OPTIONS="suppressions=$(pwd)/suppressions.txt"

src/CMakeLists.txt

@@ -159,11 +159,6 @@ enable_static_init(migraphx)
 file(GLOB BUILDER_SRCS CONFIGURE_DEPENDS op/builder/*.cpp op/builder/torch/*.cpp)
 target_sources(migraphx PRIVATE ${BUILDER_SRCS})
 
-if(WIN32)
-    # Due to compilation crashing, we need to use type-erased matchers on Windows.
-    target_compile_definitions(migraphx PUBLIC MIGRAPHX_USE_TYPE_ERASED_MATCHERS=1)
-endif()
-
 configure_file(config.h.in include/migraphx/config.h)
 
 configure_file(version.h.in include/migraphx/version.h)

src/fuse_attention.cpp

@@ -141,13 +141,14 @@ struct find_quant_attention
 {
     auto matcher() const
     {
-        auto gemm1 =
+        auto gemm1 = match::opaque(
             match::name("dequantizelinear")(match::arg(0)(match::name("quant_dot").bind("qgemm1")))
-                .bind("deq1");
-        auto softmax = match::softmax_input(match::skip(match::name("convert"))(gemm1));
-        auto probs   = match::name("quantizelinear")(
-            match::arg(0)(match::skip(match::name("convert"))(softmax)));
-        auto gemm2 = match::name("quant_dot")(match::arg(0)(probs)).bind("qgemm2");
+                .bind("deq1"));
+        auto softmax =
+            match::opaque(match::softmax_input(match::skip(match::name("convert"))(gemm1)));
+        auto softmax_cvt = match::opaque(match::skip(match::name("convert"))(softmax));
+        auto probs       = match::opaque(match::name("quantizelinear")(match::arg(0)(softmax_cvt)));
+        auto gemm2 = match::opaque(match::name("quant_dot")(match::arg(0)(probs)).bind("qgemm2"));
         return match::name("dequantizelinear")(match::arg(0)(gemm2)).bind("deq2");
     }
 
@@ -193,8 +194,10 @@ struct find_transposed_attention
 {
     auto matcher() const
     {
-        auto gemm1         = match::any_of[pointwise_inputs()](match::name("dot").bind("dot1"));
-        auto softmax       = match::skip(match::name("convert"))(match::softmax_input(gemm1));
+        auto gemm1 =
+            match::opaque(match::any_of[pointwise_inputs()](match::name("dot").bind("dot1")));
+        auto softmax =
+            match::opaque(match::skip(match::name("convert"))(match::softmax_input(gemm1)));
         auto swap_last_two = match::make_basic_pred_matcher([](instruction_ref ins) {
             auto perm = ins->get_operator().to_value()["permutation"].to_vector<int64_t>();
             if(perm.size() < 2)
@@ -204,8 +207,9 @@ struct find_transposed_attention
                    perm[perm.size() - 2] == static_cast<int64_t>(perm.size() - 1) and
                    perm[perm.size() - 1] == static_cast<int64_t>(perm.size() - 2);
         });
-        auto transposed_softmax = match::name("transpose")(swap_last_two, match::arg(0)(softmax))
-                                      .bind("transposed_softmax");
+        auto transposed_softmax =
+            match::opaque(match::name("transpose")(swap_last_two, match::arg(0)(softmax))
+                              .bind("transposed_softmax"));
         auto input_of_dot2 = match::any().bind("input_of_dot2");
         return match::name("dot")(match::arg(0)(input_of_dot2), match::arg(1)(transposed_softmax));
     }
@@ -248,8 +252,10 @@ struct find_attention
 
     auto matcher() const
     {
-        auto gemm1   = match::any_of[pointwise_inputs()](match::name("dot").bind("dot1"));
-        auto softmax = match::skip(match::name("convert"))(match::softmax_input(gemm1));
+        auto gemm1 =
+            match::opaque(match::any_of[pointwise_inputs()](match::name("dot").bind("dot1")));
+        auto softmax =
+            match::opaque(match::skip(match::name("convert"))(match::softmax_input(gemm1)));
         return match::name("dot")(match::arg(0)(softmax));
     }
 
@@ -888,36 +894,45 @@ struct find_kv_cache_attention
         static const std::unordered_set<std::string> skip_set = {
             "multibroadcast", "broadcast", "reshape", "unsqueeze", "squeeze"};
 
-        auto keys =
-            match::skip(match::name(skip_set))(match::name("concat_past_present")).bind("pres_k");
-        auto k_transpose =
-            match::skip(match::name(skip_set))(match::name("transpose")(match::arg(0)(keys)));
+        auto keys = match::opaque(
+            match::skip(match::name(skip_set))(match::name("concat_past_present")).bind("pres_k"));
+        auto keys_transpose = match::opaque(match::name("transpose")(match::arg(0)(keys)));
+        auto k_transpose    = match::opaque(match::skip(match::name(skip_set))(keys_transpose));
         auto queries = match::name("slice");
-        auto gemm1   = match::name("dot")(match::arg(0)(queries), match::arg(1)(k_transpose));
-        auto gemm1_maybe_cvt = match::skip(match::name("convert"))(gemm1);
-        auto scale           = match::name("mul")(match::any_arg(0, 1)(gemm1_maybe_cvt));
-        auto broadcasted_const = match::name("multibroadcast")(match::arg(0)(match::is_constant()));
-        auto attn_scores       = match::any_of(scale, gemm1_maybe_cvt);
-        auto causal_mask =
-            match::name("where")(match::arg(0)(broadcasted_const), match::arg(2)(attn_scores));
-        auto conv_grtr         = match::name("convert")(match::arg(0)(match::name("greater")));
-        auto local_window_comp = match::skip(match::name(skip_set))(conv_grtr);
-        auto local_window_mask =
-            match::name("where")(match::arg(0)(match::any_of(local_window_comp, broadcasted_const)),
-                                 match::arg(2)(match::any_of(causal_mask, scale, gemm1_maybe_cvt)));
-        auto greater = match::name("greater")(match::arg(1)(match::any().bind("total_sl")));
-        auto conv_greater =
-            match::skip(match::name("unsqueeze"))(match::name("convert")(match::arg(0)(greater)));
-        auto bc_greater = match::name("multibroadcast")(match::arg(0)(conv_greater));
-        auto mask       = match::name("where")(
-            match::arg(0)(bc_greater),
-            match::arg(2)(match::any_of(local_window_mask, causal_mask, scale, gemm1_maybe_cvt)));
-        auto attn_probabilities = match::skip(match::name("convert"))(
-            match::softmax_input(match::skip(match::name("convert"))(mask)));
-        auto values =
-            match::skip(match::name(skip_set))(match::name("concat_past_present")).bind("pres_v");
-        auto gemm2 = match::name("dot")(match::arg(0)(attn_probabilities), match::arg(1)(values));
-        auto transpose_out = match::name("transpose")(match::arg(0)(gemm2));
+        auto gemm1 =
+            match::opaque(match::name("dot")(match::arg(0)(queries), match::arg(1)(k_transpose)));
+        auto gemm1_maybe_cvt = match::opaque(match::skip(match::name("convert"))(gemm1));
+        auto scale = match::opaque(match::name("mul")(match::any_arg(0, 1)(gemm1_maybe_cvt)));
+        auto constant = match::opaque(match::is_constant());
+        auto broadcasted_const =
+            match::opaque(match::name("multibroadcast")(match::arg(0)(constant)));
+        auto attn_scores = match::opaque(match::any_of(scale, gemm1_maybe_cvt));
+        auto causal_mask = match::opaque(
+            match::name("where")(match::arg(0)(broadcasted_const), match::arg(2)(attn_scores)));
+        auto grtr              = match::opaque(match::name("greater"));
+        auto conv_grtr         = match::opaque(match::name("convert")(match::arg(0)(grtr)));
+        auto local_window_comp = match::opaque(match::skip(match::name(skip_set))(conv_grtr));
+        auto local_window_cond = match::opaque(match::any_of(local_window_comp, broadcasted_const));
+        auto local_window_val  = match::opaque(match::any_of(causal_mask, scale, gemm1_maybe_cvt));
+        auto local_window_mask = match::opaque(match::name("where")(
+            match::arg(0)(local_window_cond), match::arg(2)(local_window_val)));
+        auto greater =
+            match::opaque(match::name("greater")(match::arg(1)(match::any().bind("total_sl"))));
+        auto greater_cvt  = match::opaque(match::name("convert")(match::arg(0)(greater)));
+        auto conv_greater = match::opaque(match::skip(match::name("unsqueeze"))(greater_cvt));
+        auto bc_greater = match::opaque(match::name("multibroadcast")(match::arg(0)(conv_greater)));
+        auto mask_val =
+            match::opaque(match::any_of(local_window_mask, causal_mask, scale, gemm1_maybe_cvt));
+        auto mask =
+            match::opaque(match::name("where")(match::arg(0)(bc_greater), match::arg(2)(mask_val)));
+        auto mask_cvt = match::opaque(match::skip(match::name("convert"))(mask));
+        auto attn_probabilities =
+            match::opaque(match::skip(match::name("convert"))(match::softmax_input(mask_cvt)));
+        auto values             = match::opaque(
+            match::skip(match::name(skip_set))(match::name("concat_past_present")).bind("pres_v"));
+        auto gemm2 = match::opaque(
+            match::name("dot")(match::arg(0)(attn_probabilities), match::arg(1)(values)));
+        auto transpose_out = match::opaque(match::name("transpose")(match::arg(0)(gemm2)));
         return match::name("reshape")(match::arg(0)(transpose_out));
     }
 

src/include/migraphx/match/gelu_erf.hpp

@@ -1,7 +1,7 @@
 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2026 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -38,15 +38,16 @@ struct gelu_erf_matcher
     F f;
     auto erf_fn() const
     {
-        auto mul_1_sqrt_2 = f("mul")(
-            either_arg(0, 1)(none_of(has_value(M_SQRT1_2)).bind("x"), has_value(M_SQRT1_2)));
-        auto div_sqrt_2 = f("div")(args(none_of(has_value(M_SQRT2)).bind("x"), has_value(M_SQRT2)));
-        return f("erf")(used_once(), arg(0)(used_once(), any_of(mul_1_sqrt_2, div_sqrt_2)));
+        auto mul_1_sqrt_2 = opaque(f("mul")(
+            either_arg(0, 1)(none_of(has_value(M_SQRT1_2)).bind("x"), has_value(M_SQRT1_2))));
+        auto div_sqrt_2 =
+            opaque(f("div")(args(none_of(has_value(M_SQRT2)).bind("x"), has_value(M_SQRT2))));
+        return opaque(f("erf")(used_once(), arg(0)(used_once(), any_of(mul_1_sqrt_2, div_sqrt_2))));
     }
 
     auto add_erf() const
     {
-        return f("add")(used_once(), either_arg(0, 1)(erf_fn(), has_value(1.0)));
+        return opaque(f("add")(used_once(), either_arg(0, 1)(erf_fn(), has_value(1.0))));
     }
 
     auto one_half() const { return has_value(0.5); }

src/include/migraphx/match/gelu_tanh.hpp

@@ -1,7 +1,7 @@
 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2026 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -38,41 +38,44 @@ struct gelu_tanh_matcher
     F f;
 
     /// x ^ 3
-    auto pow_fn() const { return f("pow")(used_once(), arg(1)(has_value(3.0))); }
+    auto pow_fn() const { return opaque(f("pow")(used_once(), arg(1)(has_value(3.0)))); }
 
     auto tanh_fn() const
     {
         /// Gelu tanh approximation
         /// tanh( sqrt(2/M_PI) * (x + 0.044715 * x ^ 3 )
-        auto mul_const_pow1   = f("mul")(either_arg(0, 1)(has_value(0.044715), pow_fn()));
-        auto add_any_mul      = f("add")(any_arg(0, 1)(mul_const_pow1));
-        auto mul_sqrt2rpi_add = f("mul")(either_arg(0, 1)(has_value(sqrt(M_2_PI)), add_any_mul));
+        auto mul_const_pow1 = opaque(f("mul")(either_arg(0, 1)(has_value(0.044715), pow_fn())));
+        auto add_any_mul    = opaque(f("add")(any_arg(0, 1)(mul_const_pow1)));
+        auto mul_sqrt2rpi_add =
+            opaque(f("mul")(either_arg(0, 1)(has_value(sqrt(M_2_PI)), add_any_mul)));
 
         /// FastGelu tanh approximation
         /// tanh( 0.797885 * x + 0.035677 * x ^ 3 )
-        auto mul_const_pow2    = f("mul")(either_arg(0, 1)(has_value(0.035677), pow_fn()));
-        auto mul_const_x       = f("mul")(any_arg(0, 1)(has_value(0.797885)));
-        auto add_mul_x_mul_pow = f("add")(either_arg(0, 1)(mul_const_pow2, mul_const_x));
-        return f("tanh")(used_once(), arg(0)(any_of(add_mul_x_mul_pow, mul_sqrt2rpi_add)));
+        auto mul_const_pow2    = opaque(f("mul")(either_arg(0, 1)(has_value(0.035677), pow_fn())));
+        auto mul_const_x       = opaque(f("mul")(any_arg(0, 1)(has_value(0.797885))));
+        auto add_mul_x_mul_pow = opaque(f("add")(either_arg(0, 1)(mul_const_pow2, mul_const_x)));
+        return opaque(f("tanh")(used_once(), arg(0)(any_of(add_mul_x_mul_pow, mul_sqrt2rpi_add))));
     }
 
     /// x * (0.5? + 0.5 * tanh( sqrt(2/M_PI) * (x? + 0.044715 * x? ^ 3) ) ) or
     /// x * (0.5? + 0.5 * tanh( 0.797885 * x? + 0.035677 * x? ^ 3 ) )
     /// <item>? question mark means it doesn't explicitly match that item (anything will work)
     auto matcher_v0() const
     {
-        auto mul_half_tanh = f("mul")(either_arg(0, 1)(has_value(0.5), tanh_fn()));
-        auto add_any_mul   = f("add")(any_arg(0, 1)(mul_half_tanh));
-        return f("mul")(either_arg(0, 1)(any().bind("x"), add_any_mul));
+        auto mul_half_tanh = opaque(f("mul")(either_arg(0, 1)(has_value(0.5), tanh_fn())));
+        auto add_any_mul   = opaque(f("add")(any_arg(0, 1)(mul_half_tanh)));
+        return opaque(f("mul")(either_arg(0, 1)(any().bind("x"), add_any_mul)));
     }
 
     /// x * 0.5 * (1.0 + tanh( sqrt(2/M_PI) * (x + 0.044715 * x ^ 3) ) ) or
     /// x * 0.5 * (1.0 + tanh( 0.797885 * x + 0.035677 * x ^ 3 ) ) )
     auto matcher_v1() const
     {
-        auto add_one_tanh = f("add")(used_once(), either_arg(0, 1)(has_value(1.0), tanh_fn()));
-        auto mul_half_x = f("mul")(used_once(), either_arg(0, 1)(has_value(0.5), any().bind("x")));
-        return f("mul")(either_arg(0, 1)(mul_half_x, add_one_tanh));
+        auto add_one_tanh =
+            opaque(f("add")(used_once(), either_arg(0, 1)(has_value(1.0), tanh_fn())));
+        auto mul_half_x =
+            opaque(f("mul")(used_once(), either_arg(0, 1)(has_value(0.5), any().bind("x"))));
+        return opaque(f("mul")(either_arg(0, 1)(mul_half_x, add_one_tanh)));
     }
 };
 } // namespace detail

src/include/migraphx/match/layernorm.hpp

@@ -1,7 +1,7 @@
 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2025 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2026 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -51,31 +51,31 @@ struct layernorm_matcher
         });
     }
 
-    auto reduce_mean() const { return f("reduce_mean")(last_axis()); }
+    auto reduce_mean() const { return opaque(f("reduce_mean")(last_axis())); }
 
     auto x_minus_mean() const
     {
-        return f("sub")(arg(0)(any().bind("x")), arg(1)(skip_broadcasts(reduce_mean())));
+        return opaque(f("sub")(arg(0)(any().bind("x")), arg(1)(skip_broadcasts(reduce_mean()))));
     }
 
     auto variance() const
     {
-        return reduce_mean()(arg(0)(any_of(
+        return opaque(reduce_mean()(arg(0)(any_of(
             f("pow")(arg(0)(x_minus_mean()), arg(1)(has_value(2.0f))),
             f("mul")(arg(0)(x_minus_mean()), arg(1)(x_minus_mean())),
-            f("sqdiff")(either_arg(0, 1)(any().bind("x"), skip_broadcasts(reduce_mean()))))));
+            f("sqdiff")(either_arg(0, 1)(any().bind("x"), skip_broadcasts(reduce_mean())))))));
     }
 
     auto sqrt_add_eps(const std::string& name) const
     {
-        auto add_eps = f("add")(either_arg(0, 1)(variance(), is_constant().bind("eps")));
-        return skip_broadcasts(f(name)(arg(0)(any_of(add_eps, variance()))));
+        auto add_eps = opaque(f("add")(either_arg(0, 1)(variance(), is_constant().bind("eps"))));
+        return opaque(skip_broadcasts(f(name)(arg(0)(any_of(add_eps, variance())))));
     }
 
     auto layernorm_onnx() const
     {
-        auto div_sqrt  = f("div")(arg(0)(x_minus_mean()), arg(1)(sqrt_add_eps("sqrt")));
-        auto mul_rsqrt = f("mul")(either_arg(0, 1)(x_minus_mean(), sqrt_add_eps("rsqrt")));
+        auto div_sqrt  = opaque(f("div")(arg(0)(x_minus_mean()), arg(1)(sqrt_add_eps("sqrt"))));
+        auto mul_rsqrt = opaque(f("mul")(either_arg(0, 1)(x_minus_mean(), sqrt_add_eps("rsqrt"))));
         return any(any_of(div_sqrt, mul_rsqrt));
     }
 

src/include/migraphx/matcher.hpp

@@ -46,6 +46,17 @@
 #define MIGRAPHX_USE_TYPE_ERASED_MATCHERS 0
 #endif
 
+#ifndef MIGRAPHX_USE_TYPE_ERASED_OPAQUE_MATCHER
+#if MIGRAPHX_USE_TYPE_ERASED_MATCHERS
+#define MIGRAPHX_USE_TYPE_ERASED_OPAQUE_MATCHER 0
+// Windows and gcc use an excessive amount of memory to compile deeply nested matcher types
+#elif defined(_WIN32) || (defined(__GNUC__) && not defined(__clang__))
+#define MIGRAPHX_USE_TYPE_ERASED_OPAQUE_MATCHER 1
+#else
+#define MIGRAPHX_USE_TYPE_ERASED_OPAQUE_MATCHER 0
+#endif
+#endif
+
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
@@ -269,6 +280,16 @@ struct any_matcher : any_matcher_base
     }
 };
 
+template <class M>
+auto opaque(M m)
+{
+#if MIGRAPHX_USE_TYPE_ERASED_OPAQUE_MATCHER
+    return any_matcher{m};
+#else
+    return m;
+#endif
+}
+
 /// Create a basic matcher from a matcher
 template <class M>
 typename type_erased_matcher<M>::type make_basic_matcher(M m)

src/rewrite_gelu.cpp

@@ -1,7 +1,7 @@
 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2025 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2026 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -66,7 +66,10 @@ static void replace_with_tanh_exp_gelu(module& m, const match::matcher_result& r
  */
 struct find_gelu_erf
 {
-    auto matcher() const { return match::any_of(match::gelu_erf(), match::gelu_tanh()); }
+    auto matcher() const
+    {
+        return match::any_of(match::opaque(match::gelu_erf()), match::opaque(match::gelu_tanh()));
+    }
 
     void apply(module& m, const match::matcher_result& r) const
     {
@@ -86,7 +89,7 @@ struct find_gelu_erf
  */
 struct find_tanh_fast_gelu
 {
-    auto matcher() const { return match::gelu_tanh(); }
+    auto matcher() const { return match::opaque(match::gelu_tanh()); }
 
     void apply(module& m, const match::matcher_result& r) const
     {