[PATCH 1/5] x86/xstate: Elide redundant writes in set_xcr0()

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

XSETBV is an expensive instruction as, amongst other things, it involves
reconfiguring the instruction decode at the frontend of the pipeline.

We have several paths which reconfigure %xcr0 in quick succession (the context
switch path has 5, including the fpu save/restore helpers), and only a single
caller takes any care to try to skip redundant writes.

Update set_xcr0() to perform amortisation automatically, and simplify the
__context_switch() path as a consequence.

Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
---
CC: Jan Beulich <JBeulich@xxxxxxxx>
CC: Roger Pau Monné <roger.pau@xxxxxxxxxx>
CC: Wei Liu <wl@xxxxxxx>
---
 xen/arch/x86/domain.c |  4 +---
 xen/arch/x86/xstate.c | 15 +++++++++++----
 2 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/xen/arch/x86/domain.c b/xen/arch/x86/domain.c
index 4dc27f798e..50a27197b5 100644
--- a/xen/arch/x86/domain.c
+++ b/xen/arch/x86/domain.c
@@ -1977,9 +1977,7 @@ static void __context_switch(void)
         memcpy(stack_regs, &n->arch.user_regs, CTXT_SWITCH_STACK_BYTES);
         if ( cpu_has_xsave )
         {
-            u64 xcr0 = n->arch.xcr0 ?: XSTATE_FP_SSE;
-
-            if ( xcr0 != get_xcr0() && !set_xcr0(xcr0) )
+            if ( !set_xcr0(n->arch.xcr0 ?: XSTATE_FP_SSE) )
                 BUG();
 
             if ( cpu_has_xsaves && is_hvm_vcpu(n) )
diff --git a/xen/arch/x86/xstate.c b/xen/arch/x86/xstate.c
index 3794d9a5a5..f82dae8053 100644
--- a/xen/arch/x86/xstate.c
+++ b/xen/arch/x86/xstate.c
@@ -55,11 +55,18 @@ static inline bool xsetbv(u32 index, u64 xfeatures)
     return lo != 0;
 }
 
-bool set_xcr0(u64 xfeatures)
+bool set_xcr0(u64 val)
 {
-    if ( !xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures) )
-        return false;
-    this_cpu(xcr0) = xfeatures;
+    uint64_t *this_xcr0 = &this_cpu(xcr0);
+
+    if ( *this_xcr0 != val )
+    {
+        if ( !xsetbv(XCR_XFEATURE_ENABLED_MASK, val) )
+            return false;
+
+        *this_xcr0 = val;
+    }
+
     return true;
 }
 
-- 
2.11.0