[cpu] Backport fix for rbx clobbering on Windows with Clang

.github/workflows/dev-short-tests.yml

@@ -272,14 +272,15 @@ jobs:
     - uses: actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11 # tag=v4.1.1
     - name: Add MSBuild to PATH
       uses: microsoft/setup-msbuild@6fb02220983dee41ce7ae257b6f4d8f9bf5ed4ce # tag=v2.0.0
-    - name: Build
+    - name: Build & Test
       working-directory: ${{env.GITHUB_WORKSPACE}}
       run: |
         cd build\cmake
         mkdir build
         cd build
-        cmake.exe -G "${{matrix.generator}}" ${{matrix.flags}} ..
+        cmake.exe -G "${{matrix.generator}}" ${{matrix.flags}} -DCMAKE_BUILD_TYPE=Debug -DZSTD_BUILD_TESTS:BOOL=ON -DZSTD_ZSTREAM_FLAGS=-T30s -DZSTD_FUZZER_FLAGS=-T30s -DZSTD_FULLBENCH_FLAGS=-i0 ..
         cmake.exe --build .
+        ctest.exe -V -C Debug
 
   msbuild-visual-studio:
     strategy:

build/cmake/tests/CMakeLists.txt

@@ -61,7 +61,7 @@ if (NOT MSVC)
     target_compile_options(fullbench PRIVATE "-Wno-deprecated-declarations")
 endif()
 target_link_libraries(fullbench libzstd_static)
-add_test(NAME fullbench COMMAND fullbench ${ZSTD_FULLBENCH_FLAGS})
+add_test(NAME fullbench COMMAND "$<TARGET_FILE:fullbench>" ${ZSTD_FULLBENCH_FLAGS})
 
 #
 # fuzzer
@@ -73,7 +73,7 @@ endif()
 target_link_libraries(fuzzer libzstd_static)
 AddTestFlagsOption(ZSTD_FUZZER_FLAGS "$ENV{FUZZERTEST} $ENV{FUZZER_FLAGS}"
     "Semicolon-separated list of flags to pass to the fuzzer test (see `fuzzer -h` for usage)")
-add_test(NAME fuzzer COMMAND fuzzer ${ZSTD_FUZZER_FLAGS})
+add_test(NAME fuzzer COMMAND "$<TARGET_FILE:fuzzer>" ${ZSTD_FUZZER_FLAGS})
 # Disable the timeout since the run time is too long for the default timeout of
 # 1500 seconds and varies considerably between low-end and high-end CPUs.
 # set_tests_properties(fuzzer PROPERTIES TIMEOUT 0)
@@ -88,7 +88,7 @@ endif()
 target_link_libraries(zstreamtest libzstd_static)
 AddTestFlagsOption(ZSTD_ZSTREAM_FLAGS "$ENV{ZSTREAM_TESTTIME} $ENV{FUZZER_FLAGS}"
     "Semicolon-separated list of flags to pass to the zstreamtest test (see `zstreamtest -h` for usage)")
-add_test(NAME zstreamtest COMMAND zstreamtest ${ZSTD_ZSTREAM_FLAGS})
+add_test(NAME zstreamtest COMMAND "$<TARGET_FILE:zstreamtest>" ${ZSTD_ZSTREAM_FLAGS})
 
 #
 # playTests.sh

lib/common/cpu.h

@@ -35,6 +35,7 @@ MEM_STATIC ZSTD_cpuid_t ZSTD_cpuid(void) {
     U32 f7b = 0;
     U32 f7c = 0;
 #if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))
+#if !defined(__clang__)
     int reg[4];
     __cpuid((int*)reg, 0);
     {
@@ -50,6 +51,41 @@ MEM_STATIC ZSTD_cpuid_t ZSTD_cpuid(void) {
             f7c = (U32)reg[2];
         }
     }
+#else
+    /* Clang compiler has a bug (fixed in https://reviews.llvm.org/D101338) in
+     * which the `__cpuid` intrinsic does not save and restore `rbx` as it needs
+     * to due to being a reserved register. So in that case, do the `cpuid`
+     * ourselves. Clang supports inline assembly anyway.
+     */
+    U32 n;
+    __asm__(
+        "pushq %%rbx\n\t"
+        "cpuid\n\t"
+        "popq %%rbx\n\t"
+        : "=a"(n)
+        : "a"(0)
+        : "rcx", "rdx");
+    if (n >= 1) {
+      U32 f1a;
+      __asm__(
+          "pushq %%rbx\n\t"
+          "cpuid\n\t"
+          "popq %%rbx\n\t"
+          : "=a"(f1a), "=c"(f1c), "=d"(f1d)
+          : "a"(1)
+          :);
+    }
+    if (n >= 7) {
+      __asm__(
+          "pushq %%rbx\n\t"
+          "cpuid\n\t"
+          "movq %%rbx, %%rax\n\t"
+          "popq %%rbx"
+          : "=a"(f7b), "=c"(f7c)
+          : "a"(7), "c"(0)
+          : "rdx");
+    }
+#endif
 #elif defined(__i386__) && defined(__PIC__) && !defined(__clang__) && defined(__GNUC__)
     /* The following block like the normal cpuid branch below, but gcc
      * reserves ebx for use of its pic register so we must specially