On Fri, Nov 9, 2012 at 2:28 PM, Uros Bizjak <[email protected]> wrote:
> Finally, having a post-reload mode-switching pass, we can double-check
> that there are no live SSE registers at vzeroupper insertion point. As
> vzeroupper is only an optimization, we want to play safe and cancel
> vzeroupper insertion in this case
>
> There is no degradation for x86_64 gABI targets, since all SSE
> registers are call-clobbered. Vzeroupper is conditionally inserted
> just before call insn, where all registers are saved to stack and
> already dead. The vzeroupper at function exit is not problematic.
Patch was committed to mainline SVN with the following ChangeLog:
2012-11-16 Uros Bizjak <[email protected]>
* config/i386/i386-protos.h (ix86_emit_mode_set): Add third argument.
* config/i386/i386.h (EMIT_MODE_SET): Update.
* config/i386/i386.c (ix86_avx_emit_vzeroupper): New function.
(ix86_emit_mode_set) <AVX_U128>: Call ix86_avx_emit_vzeroupper.
Bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32},
configured with --with-arch=corei7-avx --with-tune=corei7-avx.
Uros.
Index: i386-protos.h
===================================================================
--- i386-protos.h (revision 193549)
+++ i386-protos.h (working copy)
@@ -172,8 +172,11 @@
extern int ix86_mode_after (int, int, rtx);
extern int ix86_mode_entry (int);
extern int ix86_mode_exit (int);
-extern void ix86_emit_mode_set (int, int);
+#ifdef HARD_CONST
+extern void ix86_emit_mode_set (int, int, HARD_REG_SET);
+#endif
+
extern void x86_order_regs_for_local_alloc (void);
extern void x86_function_profiler (FILE *, int);
extern void x86_emit_floatuns (rtx [2]);
Index: i386.c
===================================================================
--- i386.c (revision 193549)
+++ i386.c (working copy)
@@ -15477,16 +15477,38 @@
emit_move_insn (new_mode, reg);
}
+/* Emit vzeroupper. */
+
+void
+ix86_avx_emit_vzeroupper (HARD_REG_SET regs_live)
+{
+ int i;
+
+ /* Cancel automatic vzeroupper insertion if there are
+ live call-saved SSE registers at the insertion point. */
+
+ for (i = FIRST_SSE_REG; i <= LAST_SSE_REG; i++)
+ if (TEST_HARD_REG_BIT (regs_live, i) && !call_used_regs[i])
+ return;
+
+ if (TARGET_64BIT)
+ for (i = FIRST_REX_SSE_REG; i <= LAST_REX_SSE_REG; i++)
+ if (TEST_HARD_REG_BIT (regs_live, i) && !call_used_regs[i])
+ return;
+
+ emit_insn (gen_avx_vzeroupper ());
+}
+
/* Generate one or more insns to set ENTITY to MODE. */
void
-ix86_emit_mode_set (int entity, int mode)
+ix86_emit_mode_set (int entity, int mode, HARD_REG_SET regs_live)
{
switch (entity)
{
case AVX_U128:
if (mode == AVX_U128_CLEAN)
- emit_insn (gen_avx_vzeroupper ());
+ ix86_avx_emit_vzeroupper (regs_live);
break;
case I387_TRUNC:
case I387_FLOOR:
Index: i386.h
===================================================================
--- i386.h (revision 193549)
+++ i386.h (working copy)
@@ -2226,7 +2226,7 @@
are to be inserted. */
#define EMIT_MODE_SET(ENTITY, MODE, HARD_REGS_LIVE) \
- ix86_emit_mode_set ((ENTITY), (MODE))
+ ix86_emit_mode_set ((ENTITY), (MODE), (HARD_REGS_LIVE))
/* Avoid renaming of stack registers, as doing so in combination with
scheduling just increases amount of live registers at time and in
