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

Backport folly fix for rbx clobbering: https://github.com/facebook/folly/commit/f22f88b8b9d70160388f0f149bc9abaeb82c250b

This supercedes PR #3646.

diff --git a/lib/common/cpu.h b/lib/common/cpu.h
index 8bc34a36da25fa4c67ec0f7e47b12fb4b066d9c5..0e684d9ad8ed02a18f9ea35dcdb28b9bb3ee4768 100644 (file)
--- a/lib/common/cpu.h
+++ b/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