diff mbox series

[RS6000] PR94145, make PLT loads volatile

Message ID 20200323015103.GS4583@bubble.grove.modra.org
State New
Headers show
Series [RS6000] PR94145, make PLT loads volatile | expand

Commit Message

Jeff Law via Gcc-patches March 23, 2020, 1:51 a.m. UTC
On Wed, Mar 18, 2020 at 04:53:59PM -0500, Segher Boessenkool wrote:
> Could you please send a new patch (could be the same patch even) that
> is easier to review for me?

The PLT is volatile.  On PowerPC it is a bss style section which the
dynamic loader initialises to point at resolver stubs (called glink on
PowerPC64) to support lazy resolution of function addresses.  The
first call to a given function goes via the dynamic loader symbol
resolver, which updates the PLT entry for that function and calls the
function.  The second call, if there is one and we don't have a
multi-threaded race, will use the updated PLT entry and thus avoid
the relatively slow symbol resolver path.

Calls via the PLT are like calls via a function pointer, except that
no initialised function pointer is volatile like the PLT.  All
initialised function pointers are resolved at program startup to point
at the function or are left as NULL.  There is no support for lazy
resolution of any user visible function pointer.

So why does any of this matter to gcc?  Well, normally the PLT call
mechanism happens entirely behind gcc's back, but since we implemented
inline PLT calls (effectively putting the PLT code stub that loads the
PLT entry inline and making that code sequence scheduled), the load of
the PLT entry is visible to gcc.  That load then is subject to gcc
optimization, for example in

/* -S -mcpu=future -mpcrel -mlongcall -O2.  */
int foo (int);
void bar (void)
{
  while (foo(0))
    foo (99);
}

we see the PLT load for foo being hoisted out of the loop and stashed
in a call-saved register.  If that happens to be the first call to
foo, then the stashed value is that for the resolver stub, and every
call to foo in the loop will then go via the slow resolver path.  Not
a good idea.  Also, if foo turns out to be a local function and the
linker replaces the PLT calls with direct calls to foo then gcc has
just wasted a call-saved register.

This patch teaches gcc that the PLT loads are volatile.  The change
doesn't affect other loads of function pointers and thus has no effect
on normal indirect function calls.  Note that because the
"optimization" this patch prevents can only occur over function calls,
the only place gcc can stash PLT loads is in call-saved registers or
in other memory.  I'm reasonably confident that this change will be
neutral or positive for the "ld -z now" case where the PLT is not
volatile, in code where there is any register pressure.  Even if gcc
could be taught to recognise cases where the PLT is resolved, you'd
need to discount use of registers to cache PLT loads by some factor
involving the chance that those calls would be converted to direct
calls..

	PR target/94145
	* config/rs6000/rs6000.c (rs6000_longcall_ref): Use unspec_volatile
	for PLT16_LO and PLT_PCREL.
	* config/rs6000/rs6000.md (UNSPEC_PLT16_LO, UNSPEC_PLT_PCREL): Remove.
	(UNSPECV_PLT16_LO, UNSPECV_PLT_PCREL): Define.
	(pltseq_plt16_lo_, pltseq_plt_pcrel): Use unspec_volatile.

Comments

Segher Boessenkool March 26, 2020, 5:12 p.m. UTC | #1
Hi!

On Mon, Mar 23, 2020 at 12:21:03PM +1030, Alan Modra wrote:
> On Wed, Mar 18, 2020 at 04:53:59PM -0500, Segher Boessenkool wrote:
> > Could you please send a new patch (could be the same patch even) that
> > is easier to review for me?
> 
> The PLT is volatile.  On PowerPC it is a bss style section which the
> dynamic loader initialises to point at resolver stubs (called glink on
> PowerPC64) to support lazy resolution of function addresses.  The
> first call to a given function goes via the dynamic loader symbol
> resolver, which updates the PLT entry for that function and calls the
> function.  The second call, if there is one and we don't have a
> multi-threaded race, will use the updated PLT entry and thus avoid
> the relatively slow symbol resolver path.

Okay, so it isn't volatile, we have the guarantee that it will stay the
same after we have called the function once (on this same execution
thread)?

> Calls via the PLT are like calls via a function pointer, except that
> no initialised function pointer is volatile like the PLT.  All
> initialised function pointers are resolved at program startup to point
> at the function or are left as NULL.  There is no support for lazy
> resolution of any user visible function pointer.
> 
> So why does any of this matter to gcc?  Well, normally the PLT call
> mechanism happens entirely behind gcc's back, but since we implemented
> inline PLT calls (effectively putting the PLT code stub that loads the
> PLT entry inline and making that code sequence scheduled), the load of
> the PLT entry is visible to gcc.  That load then is subject to gcc
> optimization, for example in
> 
> /* -S -mcpu=future -mpcrel -mlongcall -O2.  */
> int foo (int);
> void bar (void)
> {
>   while (foo(0))
>     foo (99);
> }
> 
> we see the PLT load for foo being hoisted out of the loop and stashed
> in a call-saved register.  If that happens to be the first call to
> foo, then the stashed value is that for the resolver stub, and every
> call to foo in the loop will then go via the slow resolver path.  Not
> a good idea.  Also, if foo turns out to be a local function and the
> linker replaces the PLT calls with direct calls to foo then gcc has
> just wasted a call-saved register.

So you are saying that calling the PLT directly is always faster than
calling via a function pointer, even if that is the correct resolved
address?  That is the part I am worried about.

I think that is right, but I don't quite see it, and I don't know what
is done at runtime nearly well enough :-/

> This patch teaches gcc that the PLT loads are volatile.  The change
> doesn't affect other loads of function pointers and thus has no effect
> on normal indirect function calls.  Note that because the
> "optimization" this patch prevents can only occur over function calls,
> the only place gcc can stash PLT loads is in call-saved registers or
> in other memory.  I'm reasonably confident that this change will be
> neutral or positive for the "ld -z now" case where the PLT is not
> volatile, in code where there is any register pressure.

That is good enough for me :-)

> Even if gcc
> could be taught to recognise cases where the PLT is resolved, you'd
> need to discount use of registers to cache PLT loads by some factor
> involving the chance that those calls would be converted to direct
> calls..

> 	PR target/94145
> 	* config/rs6000/rs6000.c (rs6000_longcall_ref): Use unspec_volatile
> 	for PLT16_LO and PLT_PCREL.
> 	* config/rs6000/rs6000.md (UNSPEC_PLT16_LO, UNSPEC_PLT_PCREL): Remove.
> 	(UNSPECV_PLT16_LO, UNSPECV_PLT_PCREL): Define.
> 	(pltseq_plt16_lo_, pltseq_plt_pcrel): Use unspec_volatile.

Okay for trunk.  Thank you!


Segher
diff mbox series

Patch

diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
index 07f7cf516ba..68046fdb5ee 100644
--- a/gcc/config/rs6000/rs6000.c
+++ b/gcc/config/rs6000/rs6000.c
@@ -19274,8 +19274,9 @@  rs6000_longcall_ref (rtx call_ref, rtx arg)
       if (rs6000_pcrel_p (cfun))
 	{
 	  rtx reg = gen_rtx_REG (Pmode, regno);
-	  rtx u = gen_rtx_UNSPEC (Pmode, gen_rtvec (3, base, call_ref, arg),
-				  UNSPEC_PLT_PCREL);
+	  rtx u = gen_rtx_UNSPEC_VOLATILE (Pmode,
+					   gen_rtvec (3, base, call_ref, arg),
+					   UNSPECV_PLT_PCREL);
 	  emit_insn (gen_rtx_SET (reg, u));
 	  return reg;
 	}
@@ -19294,8 +19295,9 @@  rs6000_longcall_ref (rtx call_ref, rtx arg)
       rtx reg = gen_rtx_REG (Pmode, regno);
       rtx hi = gen_rtx_UNSPEC (Pmode, gen_rtvec (3, base, call_ref, arg),
 			       UNSPEC_PLT16_HA);
-      rtx lo = gen_rtx_UNSPEC (Pmode, gen_rtvec (3, reg, call_ref, arg),
-			       UNSPEC_PLT16_LO);
+      rtx lo = gen_rtx_UNSPEC_VOLATILE (Pmode,
+					gen_rtvec (3, reg, call_ref, arg),
+					UNSPECV_PLT16_LO);
       emit_insn (gen_rtx_SET (reg, hi));
       emit_insn (gen_rtx_SET (reg, lo));
       return reg;
diff --git a/gcc/config/rs6000/rs6000.md b/gcc/config/rs6000/rs6000.md
index ad88b6783af..5a8e9de670b 100644
--- a/gcc/config/rs6000/rs6000.md
+++ b/gcc/config/rs6000/rs6000.md
@@ -148,8 +148,6 @@ 
    UNSPEC_SI_FROM_SF
    UNSPEC_PLTSEQ
    UNSPEC_PLT16_HA
-   UNSPEC_PLT16_LO
-   UNSPEC_PLT_PCREL
   ])
 
 ;;
@@ -178,6 +176,8 @@ 
    UNSPECV_MTFSB1		; Set FPSCR Field bit to 1
    UNSPECV_SPLIT_STACK_RETURN   ; A camouflaged return
    UNSPECV_SPEC_BARRIER         ; Speculation barrier
+   UNSPECV_PLT16_LO
+   UNSPECV_PLT_PCREL
   ])
 
 ; The three different kinds of epilogue.
@@ -10359,10 +10359,10 @@ 
 
 (define_insn "*pltseq_plt16_lo_<mode>"
   [(set (match_operand:P 0 "gpc_reg_operand" "=r")
-	(unspec:P [(match_operand:P 1 "gpc_reg_operand" "b")
-		   (match_operand:P 2 "symbol_ref_operand" "s")
-		   (match_operand:P 3 "" "")]
-		  UNSPEC_PLT16_LO))]
+	(unspec_volatile:P [(match_operand:P 1 "gpc_reg_operand" "b")
+			    (match_operand:P 2 "symbol_ref_operand" "s")
+			    (match_operand:P 3 "" "")]
+			   UNSPECV_PLT16_LO))]
   "TARGET_PLTSEQ"
 {
   return rs6000_pltseq_template (operands, RS6000_PLTSEQ_PLT16_LO);
@@ -10382,10 +10382,10 @@ 
 
 (define_insn "*pltseq_plt_pcrel<mode>"
   [(set (match_operand:P 0 "gpc_reg_operand" "=r")
-	(unspec:P [(match_operand:P 1 "" "")
-		   (match_operand:P 2 "symbol_ref_operand" "s")
-		   (match_operand:P 3 "" "")]
-		  UNSPEC_PLT_PCREL))]
+	(unspec_volatile:P [(match_operand:P 1 "" "")
+			    (match_operand:P 2 "symbol_ref_operand" "s")
+			    (match_operand:P 3 "" "")]
+			   UNSPECV_PLT_PCREL))]
   "HAVE_AS_PLTSEQ && TARGET_ELF
    && rs6000_pcrel_p (cfun)"
 {