Skip to contentSkip to conversation

x86emul: avoid using _PRE_EFLAGS() in a few cases

PATCH In review Jan Beulich 2 participants 5 messages Jun 2, 2021 – Jun 29, 2021

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86emul: avoid using _PRE_EFLAGS() in a few cases

The macro expanding to quite a few insns, replace its use by simply clearing the status flags when the to be executed insn doesn't depend on their initial state, in cases where this is easily possible. (There are more cases where the uses are hidden inside macros, and where some of the users of the macros want guest flags put in place before running the insn, i.e. the macros can't be updated as easily.)

xen/arch/x86/x86_emulate/x86_emulate.c +6 -3
@@ -6863,7 +6863,8 @@ x86_emulate(
         }
         opc[2] = 0xc3;
 
-        invoke_stub(_PRE_EFLAGS("[eflags]", "[mask]", "[tmp]"),
+        _regs.eflags &= ~EFLAGS_MASK;
+        invoke_stub("",
                     _POST_EFLAGS("[eflags]", "[mask]", "[tmp]"),
                     [eflags] "+g" (_regs.eflags),
                     [tmp] "=&r" (dummy), "+m" (*mmvalp)
@@ -8111,7 +8112,8 @@ x86_emulate(
         opc[2] = 0xc3;
 
         copy_VEX(opc, vex);
-        invoke_stub(_PRE_EFLAGS("[eflags]", "[mask]", "[tmp]"),
+        _regs.eflags &= ~EFLAGS_MASK;
+        invoke_stub("",
                     _POST_EFLAGS("[eflags]", "[mask]", "[tmp]"),
                     [eflags] "+g" (_regs.eflags),
                     "=a" (dst.val), [tmp] "=&r" (dummy)
@@ -11698,13 +11700,14 @@ int x86_emul_rmw(
         break;
 
     case rmw_xadd:
+        *eflags &= ~EFLAGS_MASK;
         switch ( state->op_bytes )
         {
             unsigned long dummy;
 
 #define XADD(sz, cst, mod) \
         case sz: \
-            asm ( _PRE_EFLAGS("[efl]", "[msk]", "[tmp]") \
+            asm ( "" \
                   COND_LOCK(xadd) " %"#mod"[reg], %[mem]; " \
                   _POST_EFLAGS("[efl]", "[msk]", "[tmp]") \
                   : [reg] "+" #cst (state->ea.val), \
#1 lore raw

Ping: [PATCH] x86emul: avoid using _PRE_EFLAGS() in a few cases

Quoted 8 lines · Jan Beulich

The macro expanding to quite a few insns, replace its use by simply clearing the status flags when the to be executed insn doesn't depend on their initial state, in cases where this is easily possible. (There are more cases where the uses are hidden inside macros, and where some of the users of the macros want guest flags put in place before running the insn, i.e. the macros can't be updated as easily.)

Anyone?

Thanks, Jan

Quoted 40 lines
xen/arch/x86/x86_emulate/x86_emulate.c +6 -3
@@ -6863,7 +6863,8 @@ x86_emulate(
         }
         opc[2] = 0xc3;
 
-        invoke_stub(_PRE_EFLAGS("[eflags]", "[mask]", "[tmp]"),
+        _regs.eflags &= ~EFLAGS_MASK;
+        invoke_stub("",
                     _POST_EFLAGS("[eflags]", "[mask]", "[tmp]"),
                     [eflags] "+g" (_regs.eflags),
                     [tmp] "=&r" (dummy), "+m" (*mmvalp)
@@ -8111,7 +8112,8 @@ x86_emulate(
         opc[2] = 0xc3;
 
         copy_VEX(opc, vex);
-        invoke_stub(_PRE_EFLAGS("[eflags]", "[mask]", "[tmp]"),
+        _regs.eflags &= ~EFLAGS_MASK;
+        invoke_stub("",
                     _POST_EFLAGS("[eflags]", "[mask]", "[tmp]"),
                     [eflags] "+g" (_regs.eflags),
                     "=a" (dst.val), [tmp] "=&r" (dummy)
@@ -11698,13 +11700,14 @@ int x86_emul_rmw(
         break;
 
     case rmw_xadd:
+        *eflags &= ~EFLAGS_MASK;
         switch ( state->op_bytes )
         {
             unsigned long dummy;
 
 #define XADD(sz, cst, mod) \
         case sz: \
-            asm ( _PRE_EFLAGS("[efl]", "[msk]", "[tmp]") \
+            asm ( "" \
                   COND_LOCK(xadd) " %"#mod"[reg], %[mem]; " \
                   _POST_EFLAGS("[efl]", "[msk]", "[tmp]") \
                   : [reg] "+" #cst (state->ea.val), \
#2 lore raw
In reply to Jan Beulich
Quoted 8 lines · Jan Beulich

The macro expanding to quite a few insns, replace its use by simply clearing the status flags when the to be executed insn doesn't depend on their initial state, in cases where this is easily possible. (There are more cases where the uses are hidden inside macros, and where some of the users of the macros want guest flags put in place before running the insn, i.e. the macros can't be updated as easily.)

Honestly, this is the first time I've looked into _PRE/_POST_EFLAGS() in detail.  Why is most of this not in C to begin with?

The only bits which need to be in asm are the popf to establish the stub's flags context, and the pushf to save the resulting state. Everything else is better off done by the compiler especially as it would remove a load of work on temporaries from the stack.

Nevertheless, ...

Quoted 13 lines
xen/arch/x86/x86_emulate/x86_emulate.c +2 -1
@@ -6863,7 +6863,8 @@ x86_emulate(
         }
         opc[2] = 0xc3;
 
-        invoke_stub(_PRE_EFLAGS("[eflags]", "[mask]", "[tmp]"),
+        _regs.eflags &= ~EFLAGS_MASK;
+        invoke_stub("",
                     _POST_EFLAGS("[eflags]", "[mask]", "[tmp]"),
                     [eflags] "+g" (_regs.eflags),
                     [tmp] "=&r" (dummy), "+m" (*mmvalp)

xen/arch/x86/x86_emulate/x86_emulate.c x86_emulate( · line 8119

Quoted 9 lines
xen/arch/x86/x86_emulate/x86_emulate.c +2 -1
@@ -8111,7 +8112,8 @@ x86_emulate(
         opc[2] = 0xc3;
 
         copy_VEX(opc, vex);
-        invoke_stub(_PRE_EFLAGS("[eflags]", "[mask]", "[tmp]"),
+        _regs.eflags &= ~EFLAGS_MASK;
+        invoke_stub("",
                     _POST_EFLAGS("[eflags]", "[mask]", "[tmp]"),
                     [eflags] "+g" (_regs.eflags),
                     "=a" (dst.val), [tmp] "=&r" (dummy)

... this hunk is k{,or}test, which only modified ZF and CF according to the SDM.

The other flags are not listed as modified, which means they're preserved, unless you're planning to have Intel issue a correction to the SDM.

The flags logic for both instructions is custom, so it wouldn't be a surprise to me if it really did deviate from the normal behaviour of a test operation.

~Andrew

Quoted 17 lines
xen/arch/x86/x86_emulate/x86_emulate.c +2 -1
@@ -11698,13 +11700,14 @@ int x86_emul_rmw(
         break;
 
     case rmw_xadd:
+        *eflags &= ~EFLAGS_MASK;
         switch ( state->op_bytes )
         {
             unsigned long dummy;
 
 #define XADD(sz, cst, mod) \
         case sz: \
-            asm ( _PRE_EFLAGS("[efl]", "[msk]", "[tmp]") \
+            asm ( "" \
                   COND_LOCK(xadd) " %"#mod"[reg], %[mem]; " \
                   _POST_EFLAGS("[efl]", "[msk]", "[tmp]") \
                   : [reg] "+" #cst (state->ea.val), \
#3 lore raw
Quoted 8 lines · Andrew Cooper
Quoted 7 lines · Jan Beulich

The macro expanding to quite a few insns, replace its use by simply clearing the status flags when the to be executed insn doesn't depend on their initial state, in cases where this is easily possible. (There are more cases where the uses are hidden inside macros, and where some of the users of the macros want guest flags put in place before running the insn, i.e. the macros can't be updated as easily.)

Quoted 4 lines
Quoted 1 line

Honestly, this is the first time I've looked into _PRE/_POST_EFLAGS() in detail.  Why is most of this not in C to begin with?

Ask Keir?

Quoted 4 lines

The only bits which need to be in asm are the popf to establish the stub's flags context, and the pushf to save the resulting state. Everything else is better off done by the compiler especially as it would remove a load of work on temporaries from the stack.

I'll try to find time to do something along these lines.

Quoted 28 lines

Nevertheless, ...

Quoted 13 lines
xen/arch/x86/x86_emulate/x86_emulate.c +2 -1
@@ -6863,7 +6863,8 @@ x86_emulate(
         }
         opc[2] = 0xc3;
 
-        invoke_stub(_PRE_EFLAGS("[eflags]", "[mask]", "[tmp]"),
+        _regs.eflags &= ~EFLAGS_MASK;
+        invoke_stub("",
                     _POST_EFLAGS("[eflags]", "[mask]", "[tmp]"),
                     [eflags] "+g" (_regs.eflags),
                     [tmp] "=&r" (dummy), "+m" (*mmvalp)

xen/arch/x86/x86_emulate/x86_emulate.c x86_emulate( · line 8119

Quoted 9 lines
xen/arch/x86/x86_emulate/x86_emulate.c +2 -1
@@ -8111,7 +8112,8 @@ x86_emulate(
         opc[2] = 0xc3;
 
         copy_VEX(opc, vex);
-        invoke_stub(_PRE_EFLAGS("[eflags]", "[mask]", "[tmp]"),
+        _regs.eflags &= ~EFLAGS_MASK;
+        invoke_stub("",
                     _POST_EFLAGS("[eflags]", "[mask]", "[tmp]"),
                     [eflags] "+g" (_regs.eflags),
                     "=a" (dst.val), [tmp] "=&r" (dummy)

... this hunk is k{,or}test, which only modified ZF and CF according to the SDM.

Quoted 3 lines

The other flags are not listed as modified, which means they're preserved, unless you're planning to have Intel issue a correction to the SDM.

kortest has

"The OF, SF, AF, and PF flags are set to 0."

in its "Flags Affected" section. ktest has

"AF := OF := PF := SF := 0;"

in its "Operation" section.

Jan

Reviewed-by
#4 lore raw
Quoted 52 lines · Jan Beulich
Quoted 11 lines · Andrew Cooper
Quoted 8 lines · Jan Beulich

The macro expanding to quite a few insns, replace its use by simply clearing the status flags when the to be executed insn doesn't depend on their initial state, in cases where this is easily possible. (There are more cases where the uses are hidden inside macros, and where some of the users of the macros want guest flags put in place before running the insn, i.e. the macros can't be updated as easily.)

Honestly, this is the first time I've looked into _PRE/_POST_EFLAGS() in detail.  Why is most of this not in C to begin with?

Ask Keir?

Quoted 4 lines

The only bits which need to be in asm are the popf to establish the stub's flags context, and the pushf to save the resulting state. Everything else is better off done by the compiler especially as it would remove a load of work on temporaries from the stack.

I'll try to find time to do something along these lines.

Quoted 27 lines

Nevertheless, ...

Quoted 13 lines
xen/arch/x86/x86_emulate/x86_emulate.c +2 -1
@@ -6863,7 +6863,8 @@ x86_emulate(
         }
         opc[2] = 0xc3;
 
-        invoke_stub(_PRE_EFLAGS("[eflags]", "[mask]", "[tmp]"),
+        _regs.eflags &= ~EFLAGS_MASK;
+        invoke_stub("",
                     _POST_EFLAGS("[eflags]", "[mask]", "[tmp]"),
                     [eflags] "+g" (_regs.eflags),
                     [tmp] "=&r" (dummy), "+m" (*mmvalp)

xen/arch/x86/x86_emulate/x86_emulate.c x86_emulate( · line 8119

Quoted 9 lines
xen/arch/x86/x86_emulate/x86_emulate.c +2 -1
@@ -8111,7 +8112,8 @@ x86_emulate(
         opc[2] = 0xc3;
 
         copy_VEX(opc, vex);
-        invoke_stub(_PRE_EFLAGS("[eflags]", "[mask]", "[tmp]"),
+        _regs.eflags &= ~EFLAGS_MASK;
+        invoke_stub("",
                     _POST_EFLAGS("[eflags]", "[mask]", "[tmp]"),
                     [eflags] "+g" (_regs.eflags),
                     "=a" (dst.val), [tmp] "=&r" (dummy)

... this hunk is k{,or}test, which only modified ZF and CF according to the SDM.

Quoted 3 lines

The other flags are not listed as modified, which means they're preserved, unless you're planning to have Intel issue a correction to the SDM.

kortest has

Quoted 7 lines

"The OF, SF, AF, and PF flags are set to 0."

in its "Flags Affected" section. ktest has

"AF := OF := PF := SF := 0;"

in its "Operation" section.

Oh - the pseudocode and the text don't match.  How helpful.