x86/atomic: fix cmpxchg16b inline assembly to work with clang

clang doesn't understand the "=A" register constrain when used with 64bits
assembly and spits out an internal error:

fatal error: error in backend: Cannot select: 0x7f9fb89c9390: i64 = build_pair 0x7f9fb89c92b0,
      0x7f9fb89c9320
  0x7f9fb89c92b0: i32,ch,glue = CopyFromReg 0x7f9fb89c9240, Register:i32 %EAX, 0x7f9fb89c9240:1
    0x7f9fb89c8c20: i32 = Register %EAX
    0x7f9fb89c9240: ch,glue = inlineasm 0x7f9fb89c90f0,
TargetExternalSymbol:i64'lock; cmpxchg16b $1', MDNode:ch<0x7f9fb8476c38>,
TargetConstant:i64<25>, TargetConstant:i32<18>, Register:i32 %EAX, Register:i32
%EDX, TargetConstant:i32<196622>, 0x7f9fb89c87c0, TargetConstant:i32<9>,
Register:i64 %RCX, TargetConstant:i32<9>, Register:i64 %RBX,
TargetConstant:i32<9>, Register:i64 %RDX, TargetConstant:i32<9>, Register:i64
%RAX, TargetConstant:i32<196622>, 0x7f9fb89c87c0, TargetConstant:i32<12>,
Register:i32 %EFLAGS, 0x7f9fb89c90f0:1
      0x7f9fb89c8a60: i64 = TargetExternalSymbol'lock; cmpxchg16b $1'
      0x7f9fb89c8b40: i64 = TargetConstant<25>
      0x7f9fb89c8bb0: i32 = TargetConstant<18>
      0x7f9fb89c8c20: i32 = Register %EAX
      0x7f9fb89c8c90: i32 = Register %EDX
      0x7f9fb89c8d00: i32 = TargetConstant<196622>
      0x7f9fb89c87c0: i64,ch = load<LD8[%4]> 0x7f9fb9053da0, FrameIndex:i64<1>, undef:i64
        0x7f9fb9053a90: i64 = FrameIndex<1>
        0x7f9fb9053e80: i64 = undef
      0x7f9fb89c8e50: i32 = TargetConstant<9>
      0x7f9fb89c8d70: i64 = Register %RCX
      0x7f9fb89c8e50: i32 = TargetConstant<9>
      0x7f9fb89c8ec0: i64 = Register %RBX
      0x7f9fb89c8e50: i32 = TargetConstant<9>
      0x7f9fb89c8fa0: i64 = Register %RDX
      0x7f9fb89c8e50: i32 = TargetConstant<9>
      0x7f9fb89c9080: i64 = Register %RAX
[...]

Fix this by specifying "rdx:rax" manually using the "d" and "a" constraints.

Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>

diff --git a/xen/include/asm-x86/x86_64/system.h b/xen/include/asm-x86/x86_64/system.h
index 7026c05a258c2da1c4d99a3b41d93b04d377f092..88beae130bfdbb7fd63d3ecacc1d3faa80ecfb57 100644 (file)
--- a/xen/include/asm-x86/x86_64/system.h
+++ b/xen/include/asm-x86/x86_64/system.h
@@ -14,20 +14,21 @@
  */
 
 static always_inline __uint128_t __cmpxchg16b(
-    volatile void *ptr, const __uint128_t *old, const __uint128_t *new)
+    volatile void *ptr, const __uint128_t *oldp, const __uint128_t *newp)
 {
-    __uint128_t prev;
-    uint64_t new_high = *new >> 64;
-    uint64_t new_low = *new;
+    union {
+        struct { uint64_t lo, hi; };
+        __uint128_t raw;
+    } new = { .raw = *newp }, old = { .raw = *oldp }, prev;
 
     ASSERT(cpu_has_cx16);
 
-    asm volatile ( "lock; cmpxchg16b %1"
-                   : "=A" (prev), "+m" (*__xg(ptr))
-                   : "c" (new_high), "b" (new_low),
-                     "0" (*old) );
+    /* Don't use "=A" here - clang can't deal with that. */
+    asm volatile ( "lock; cmpxchg16b %2"
+                   : "=d" (prev.hi), "=a" (prev.lo), "+m" (*__xg(ptr))
+                   : "c" (new.hi), "b" (new.lo), "0" (old.hi), "1" (old.lo) );
 
-    return prev;
+    return prev.raw;
 }
 
 #define cmpxchg16b(ptr, o, n) ({                           \