Only do complex return style autodetection on non-Windows x86_64

Other architectures we don't care about as much (since we won't be using
MKL on them) and I don't have time to figure out the assembly syntax to
probe without smashing the stack on `i686`.

Windows prefers argument retstyles, so just let it be.

Author: staticfloat

src/Make.inc

@@ -89,6 +89,15 @@ ifneq (,$(filter $(ARCH), x86_64 i686))
   CBLAS_DIVERGENCE_AUTODETECTION := 1
 endif
 
+# If we're on x86_64 (and NOT on windows), we can do complex return style autodetection
+# We don't do this on Windows because Windows actually prefers arugment retstyle
+ifneq ($(OS),WINNT)
+ifneq (,$(filter $(ARCH), x86_64))
+  LBT_CFLAGS += -DCOMPLEX_RETSTYLE_AUTODETECTION
+  COMPLEX_RETSTYLE_AUTODETECTION := 1
+endif
+endif
+
 
 ifeq ($(VERBOSE),0)
 ENDCOLOR := "\033[0m"

src/Makefile

@@ -7,7 +7,6 @@ all: $(builddir)/libblastrampoline.$(SHLIB_EXT)
 # Objects we'll build
 MAIN_OBJS := libblastrampoline.o dl_utils.o config.o \
 			 autodetection.o \
-			 complex_return_style_adapters.o \
 			 threading.o deepbindless.o trampolines/trampolines_$(ARCH).o
 
 # Include win_utils.c on windws
@@ -24,6 +23,10 @@ ifeq ($(CBLAS_DIVERGENCE_AUTODETECTION),1)
 MAIN_OBJS += cblas_adapters.o
 endif
 
+ifeq ($(COMPLEX_RETSTYLE_AUTODETECTION),1)
+MAIN_OBJS += complex_return_style_adapters.o
+endif
+
 # Place the `.o` files into `$(builddir)`
 MAIN_OBJS := $(addprefix $(builddir)/,$(MAIN_OBJS))
 

src/autodetection.c

@@ -154,6 +154,7 @@ int32_t autodetect_interface(void * handle, const char * suffix) {
     return LBT_INTERFACE_UNKNOWN;
 }
 
+#ifdef COMPLEX_RETSTYLE_AUTODETECTION
 int32_t autodetect_complex_return_style(void * handle, const char * suffix) {
     char symbol_name[MAX_SYMBOL_LEN];
 
@@ -196,6 +197,7 @@ int32_t autodetect_complex_return_style(void * handle, const char * suffix) {
     // If that was not reset either, we have no idea what's going on.
     return LBT_COMPLEX_RETSTYLE_UNKNOWN;
 }
+#endif // COMPLEX_RETSTYLE_AUTODETECTION
 
 #ifdef F2C_AUTODETECTION
 int32_t autodetect_f2c(void * handle, const char * suffix) {
@@ -232,7 +234,7 @@ int32_t autodetect_f2c(void * handle, const char * suffix) {
     // We have no idea what happened; nothing works and everything is broken
     return LBT_F2C_UNKNOWN;
 }
-#endif
+#endif // F2C_AUTODETECTION
 
 #ifdef CBLAS_DIVERGENCE_AUTODETECTION
 int32_t autodetect_cblas_divergence(void * handle, const char * suffix) {
@@ -261,4 +263,4 @@ int32_t autodetect_cblas_divergence(void * handle, const char * suffix) {
     // If we can't even find `zdotc_64`, we don't know what this is.
     return LBT_CBLAS_UNKNOWN;
 }
-#endif
+#endif // CBLAS_DIVERGENCE_AUTODETECTION

src/libblastrampoline.c

@@ -1,7 +1,9 @@
 #include "libblastrampoline_internal.h"
 #include "libblastrampoline_trampdata.h"
-#include "libblastrampoline_complex_retdata.h"
 
+#ifdef COMPLEX_RETSTYLE_AUTODETECTION
+#include "libblastrampoline_complex_retdata.h"
+#endif
 #ifdef F2C_AUTODETECTION
 #include "libblastrampoline_f2cdata.h"
 #endif
@@ -67,6 +69,7 @@ int32_t set_forward_by_index(int32_t symbol_idx, const void * addr, int32_t inte
         }
     }
 
+#ifdef COMPLEX_RETSTYLE_AUTODETECTION
     if (complex_retstyle == LBT_COMPLEX_RETSTYLE_ARGUMENT) {
         // Check to see if this symbol is one of the complex-returning functions
         for (int complex_symbol_idx=0; cmplxret_func_idxs[complex_symbol_idx] != -1; ++complex_symbol_idx) {
@@ -90,6 +93,7 @@ int32_t set_forward_by_index(int32_t symbol_idx, const void * addr, int32_t inte
             }
         }
     }
+#endif // COMPLEX_RETSTYLE_AUTODETECTION
 
 #ifdef F2C_AUTODETECTION
     if (f2c == LBT_F2C_REQUIRED) {
@@ -216,7 +220,9 @@ LBT_DLLEXPORT int32_t lbt_forward(const char * libname, int32_t clear, int32_t v
     }
 
     // Next, let's figure out what the complex return style is:
-    int complex_retstyle = autodetect_complex_return_style(handle, lib_suffix);
+    int complex_retstyle = LBT_COMPLEX_RETSTYLE_UNKNOWN;
+#ifdef COMPLEX_RETSTYLE_AUTODETECTION
+    complex_retstyle = autodetect_complex_return_style(handle, lib_suffix);
     if (complex_retstyle == LBT_COMPLEX_RETSTYLE_UNKNOWN) {
         fprintf(stderr, "Unable to autodetect complex return style of \"%s\"\n", libname);
         return 0;
@@ -229,6 +235,7 @@ LBT_DLLEXPORT int32_t lbt_forward(const char * libname, int32_t clear, int32_t v
             printf(" -> Autodetected argument-passing complex return style\n");
         }
     }
+#endif // COMPLEX_RETSTYLE_AUTODETECTION
 
     int f2c = LBT_F2C_PLAIN;
 #ifdef F2C_AUTODETECTION
@@ -247,7 +254,7 @@ LBT_DLLEXPORT int32_t lbt_forward(const char * libname, int32_t clear, int32_t v
             printf(" -> Autodetected gfortran calling convention\n");
         }
     }
-#endif
+#endif // F2C_AUTODETECTION
 
     int cblas = LBT_CBLAS_UNKNOWN;
 #ifdef CBLAS_DIVERGENCE_AUTODETECTION
@@ -273,7 +280,7 @@ LBT_DLLEXPORT int32_t lbt_forward(const char * libname, int32_t clear, int32_t v
             }
         }
     }
-#endif
+#endif // CBLAS_DIVERGENCE_AUTODETECTION
 
     /*
      * Now, if we are opening a 64-bit library with 32-bit names (e.g. suffix == ""),
@@ -378,7 +385,7 @@ LBT_DLLEXPORT int32_t lbt_forward(const char * libname, int32_t clear, int32_t v
             }
         }
     }
-#endif
+#endif // CBLAS_DIVERGENCE_AUTODETECTION
 
     record_library_load(libname, handle, lib_suffix, &forwards[0], interface, complex_retstyle, f2c, cblas);
     if (verbose) {

src/libblastrampoline_internal.h

@@ -80,7 +80,9 @@ const char * autodetect_symbol_suffix(void * handle, const char * suffix_hint);
 int32_t autodetect_blas_interface(void * isamax_addr);
 int32_t autodetect_lapack_interface(void * dpotrf_addr);
 int32_t autodetect_interface(void * handle, const char * suffix);
+#ifdef COMPLEX_RETSTYLE_AUTODETECTION
 int32_t autodetect_complex_return_style(void * handle, const char * suffix);
+#endif
 
 #ifdef F2C_AUTODETECTION
 int32_t autodetect_f2c(void * handle, const char * suffix);

test/direct.jl

@@ -71,10 +71,18 @@ lbt_handle = dlopen("$(lbt_prefix)/$(binlib)/lib$(lbt_link_name).$(shlib_ext)",
     @test libs[1].f2c == LBT_F2C_PLAIN
     if Sys.ARCH == :x86_64
         @test libs[1].cblas == LBT_CBLAS_CONFORMANT
+        if Sys.iswindows()
+            # Technically, this should be "argument", but we disable complex return style
+            # autodetection on windows since the default compilers seem to prefer argument
+            # style, so we'd rather just not touch things on that platform.
+            @test libs[1].complex_retstyle == LBT_COMPLEX_RETSTYLE_UNKNOWN
+        else
+            @test libs[1].complex_retstyle == LBT_COMPLEX_RETSTYLE_NORMAL
+        end
     else
         @test libs[1].cblas == LBT_CBLAS_UNKNOWN
+        @test libs[1].complex_retstyle == LBT_COMPLEX_RETSTYLE_UNKNOWN
     end
-    @test libs[1].complex_retstyle == LBT_COMLPEX_RETSTYLE_NORMAL
 
     @test bitfield_get(libs[1].active_forwards, dgemm_idx) != 0
 
@@ -83,7 +91,6 @@ lbt_handle = dlopen("$(lbt_prefix)/$(binlib)/lib$(lbt_link_name).$(shlib_ext)",
     @test libs[2].suffix == ""
     @test libs[2].interface == LBT_INTERFACE_LP64
     @test libs[2].f2c == LBT_F2C_PLAIN
-    @test libs[2].complex_retstyle == LBT_COMLPEX_RETSTYLE_NORMAL
 
     # If OpenBLAS32 and OpenBLAS are the same interface (e.g. i686)
     # then libs[2].active_forwards should be all zero!
@@ -189,7 +196,7 @@ end
     @test length(self_traces) == 3
 end
 
-if MKL_jll.is_available() && Sys.WORD_SIZE == 64
+if MKL_jll.is_available() && Sys.ARCH == :x86_64
     # Since MKL v2022, we can explicitly link against ILP64-suffixed symbols
     @testset "MKL v2022 ILP64 loading" begin
         # Load the ILP64 interface library.  Remember, you must load the `core`
@@ -237,9 +244,13 @@ if MKL_jll.is_available() && Sys.WORD_SIZE == 64
         @test length(libs) == 1
         @test libs[1].interface == LBT_INTERFACE_ILP64
         @test libs[1].cblas == LBT_CBLAS_DIVERGENT
-        @test libs[1].complex_retstyle == LBT_COMLPEX_RETSTYLE_ARGUMENT
+        if Sys.iswindows()
+            @test libs[1].complex_retstyle == LBT_COMPLEX_RETSTYLE_UNKNOWN
+        else
+            @test libs[1].complex_retstyle == LBT_COMPLEX_RETSTYLE_ARGUMENT
+        end
 
-        # Call cblas_zdotu_sub, showcasing that it doesn't work
+        # Call cblas_zdotu_sub, asserting that it does not try to call a forwardless-symbol
         empty!(stacktraces)
         A = ComplexF64[3.1 + 1.4im, -1.0 +  1.2im]
         B = ComplexF64[1.3 + 0.3im, -1.1 + -3.4im]

test/runtests.jl

@@ -104,7 +104,8 @@ end
 
 # Build version that links against vanilla OpenBLAS32
 @testset "Vanilla OpenBLAS32_jll (LP64)" begin
-    run_all_tests("openblas", OpenBLAS32_jll.LIBPATH_list, :LP64, "")
+    # Reverse OpenBLAS32_jll's LIBPATH_list so that we get the right openblas.so
+    run_all_tests("openblas", reverse(OpenBLAS32_jll.LIBPATH_list), :LP64, "")
 end
 
 # Next, build a version that links against `libblastrampoline`, and tell

test/utils.jl

@@ -117,9 +117,9 @@ end
 const LBT_INTERFACE_LP64 = 32
 const LBT_INTERFACE_ILP64 = 64
 const LBT_F2C_PLAIN = 0
-const LBT_COMLPEX_RETSTYLE_NORMAL = 0
-const LBT_COMLPEX_RETSTYLE_ARGUMENT = 1
-const LBT_COMLPEX_RETSTYLE_UNKNOWN = -1
+const LBT_COMPLEX_RETSTYLE_NORMAL = 0
+const LBT_COMPLEX_RETSTYLE_ARGUMENT = 1
+const LBT_COMPLEX_RETSTYLE_UNKNOWN = -1
 const LBT_CBLAS_CONFORMANT = 0
 const LBT_CBLAS_DIVERGENT = 1
 const LBT_CBLAS_UNKNOWN = -1
@@ -152,7 +152,7 @@ function lbt_get_forward(handle, symbol_name, interface, f2c = LBT_F2C_PLAIN)
     return ccall(dlsym(handle, :lbt_get_forward), Ptr{Cvoid}, (Cstring, Int32, Int32), symbol_name, interface, f2c)
 end
 
-function lbt_set_forward(handle, symbol_name, addr, interface, complex_retstyle = LBT_COMLPEX_RETSTYLE_NORMAL, f2c = LBT_F2C_PLAIN; verbose::Bool = false)
+function lbt_set_forward(handle, symbol_name, addr, interface, complex_retstyle = LBT_COMPLEX_RETSTYLE_NORMAL, f2c = LBT_F2C_PLAIN; verbose::Bool = false)
     return ccall(dlsym(handle, :lbt_set_forward), Int32, (Cstring, Ptr{Cvoid}, Int32, Int32, Int32, Int32), symbol_name, addr, interface, complex_retstyle, f2c, verbose ? 1 : 0)
 end
 

test/zdotc_test/zdotc_test.c

@@ -10,7 +10,7 @@ typedef int64_t blasint;
 typedef int32_t blasint;
 #endif
 
-extern double complex MANGLE(cblas_zdotc_sub)(blasint, double complex *, blasint, double complex *, blasint, double complex *);
+extern void MANGLE(cblas_zdotc_sub)(blasint, double complex *, blasint, double complex *, blasint, double complex *);
 extern double complex MANGLE(zdotc_)(blasint *, double complex *, blasint *, double complex *, blasint *);
 
 #define N 2