Bug 493970 - s390x: Drop saving/restoring FPC upon helper call

Saving the FPC before each helper call and restoring it afterwards creates
unnecessary overhead, and it may also not be desirable.

Drop it.  Also remove the functions in host_s390_defs.c responsible for
emitting LFPC and STFPC instructions.  And since this frees up the FPC
save slot on the stack, adjust the stack layout accordingly.

diff --git a/NEWS b/NEWS
index 59b987d6330709973d52ce597ccd3a0c8729bd99..319caf68f89ce35b5bff6e909778d09f12dc2376 100644 (file)
--- a/NEWS
+++ b/NEWS
@@ -70,6 +70,7 @@ are not entered into bugzilla tend to get forgotten about or ignored.
 493454  Missing FUSE_COMPATIBLE_MAY_BLOCK markers
 493507  direct readlink syscall from PRE handler is incompatible with
         FUSE_COMPATIBLE_MAY_BLOCK
+493970  s390x: Store/restore FPC upon helper call causes slowdown
 n-i-bz  Improve messages for sigaltstack errors, use specific
         stack_t member names
 

diff --git a/VEX/priv/host_s390_defs.c b/VEX/priv/host_s390_defs.c
index ee240347d63b9441c314f2cd5f7e9800a8c00b70..583f116c3c82139b11da0573fb7abcdcc3dbc49f 100644 (file)
--- a/VEX/priv/host_s390_defs.c
+++ b/VEX/priv/host_s390_defs.c
@@ -1706,18 +1706,6 @@ emit_RXY(UChar *p, ULong op, UChar r1, UChar x2, UChar b2, UShort dl2, UChar dh2
 }
 
 
-static UChar *
-emit_S(UChar *p, UInt op, UChar b2, UShort d2)
-{
-   ULong the_insn = op;
-
-   the_insn |= ((ULong)b2) << 12;
-   the_insn |= ((ULong)d2) << 0;
-
-   return emit_4bytes(p, the_insn);
-}
-
-
 static UChar *
 emit_SI(UChar *p, UInt op, UChar i2, UChar b1, UShort d1)
 {
@@ -3791,16 +3779,6 @@ s390_emit_LDY(UChar *p, UChar r1, UChar x2, UChar b2, UShort dl2, UChar dh2)
 }
 
 
-static UChar *
-s390_emit_LFPC(UChar *p, UChar b2, UShort d2)
-{
-   if (UNLIKELY(vex_traceflags & VEX_TRACE_ASM))
-      s390_disasm(ENC2(MNM, UDXB), "lfpc", d2, 0, b2);
-
-   return emit_S(p, 0xb29d0000, b2, d2);
-}
-
-
 static UChar *
 s390_emit_LDGR(UChar *p, UChar r1, UChar r2)
 {
@@ -3895,16 +3873,6 @@ s390_emit_STDY(UChar *p, UChar r1, UChar x2, UChar b2, UShort dl2, UChar dh2)
 }
 
 
-static UChar *
-s390_emit_STFPC(UChar *p, UChar b2, UShort d2)
-{
-   if (UNLIKELY(vex_traceflags & VEX_TRACE_ASM))
-      s390_disasm(ENC2(MNM, UDXB), "stfpc", d2, 0, b2);
-
-   return emit_S(p, 0xb29c0000, b2, d2);
-}
-
-
 static UChar *
 s390_emit_AEBR(UChar *p, UChar r1, UChar r2)
 {
@@ -10354,14 +10322,7 @@ s390_insn_helper_call_emit(UChar *buf, const s390_insn *insn)
       Also, need to arrange for the return address be put into the
       link-register */
    buf = s390_emit_load_64imm(buf, 1, target);
-
-   /* Stash away the client's FPC register because the helper might change it. */
-   buf = s390_emit_STFPC(buf, S390_REGNO_STACK_POINTER, S390_OFFSET_SAVED_FPC_C);
-
-   buf = s390_emit_BASR(buf, S390_REGNO_LINK_REGISTER, 1);      // call helper
-
-   buf = s390_emit_LFPC(buf, S390_REGNO_STACK_POINTER,          // restore FPC
-                        S390_OFFSET_SAVED_FPC_C);
+   buf = s390_emit_BASR(buf, S390_REGNO_LINK_REGISTER, 1);
 
    // preElse:
    UChar* pPreElse = buf;

diff --git a/VEX/pub/libvex_s390x_common.h b/VEX/pub/libvex_s390x_common.h
index 944347242dbd6bbb69f0cf6891939226390dca57..dda5ffd3121c87f153c11cd0ae904033b7256800 100644 (file)
--- a/VEX/pub/libvex_s390x_common.h
+++ b/VEX/pub/libvex_s390x_common.h
@@ -52,10 +52,7 @@
 /* Dispatcher will save 8 FPRs at offsets 160 + 0 ... 160 + 56 */
 
 /* Where the dispatcher saves the r2 contents. */
-#define S390_OFFSET_SAVED_R2 160+80
-
-/* Where client's FPC register is saved. */
-#define S390_OFFSET_SAVED_FPC_C 160+72
+#define S390_OFFSET_SAVED_R2 160+72
 
 /* Where valgrind's FPC register is saved. */
 #define S390_OFFSET_SAVED_FPC_V 160+64
@@ -64,11 +61,11 @@
    Need size for
        8 FPRs
      + 1 GPR  (SAVED_R2)
-     + 2 FPCs (SAVED_FPC_C and SAVED_FPC_V).
+     + 1 FPC  (SAVED_FPC_V)
 
    Additionally, we need a standard frame for helper functions being called
    from client code. (See figure 1-16 in zSeries ABI) */
-#define S390_INNERLOOP_FRAME_SIZE ((8+1+2)*8 + 160)
+#define S390_INNERLOOP_FRAME_SIZE ((8+1+1)*8 + 160)
 
 
 /*--------------------------------------------------------------*/