[committed] pa: Revise memory barriers to use strongly ordered ldcw instruction

Message ID a92ee005-17e4-530f-8742-499966437aa5@bell.net
State New
Headers show
  • [committed] pa: Revise memory barriers to use strongly ordered ldcw instruction
Related show

Commit Message

John David Anglin Nov. 8, 2019, 12:51 a.m.
This change revises the memory barrier patterns to use the ldcw instruction instead of
the sync instruction.  The sync instruction performs better and I have more confidence
in it than sync.

We use a location just above the top of the stack for these operations.  The stack address
is aligned to a 16-byte boundary if the system is not coherent.

I have added two new options.  The first is the -mcoherent-ldcw option.  The majority of
PA 2.0 system have coherent caches and as a result the coherent ldcw completer can be used.
In that case, the ldcw address doesn't require 16-byte alignment.  We set the default to

The second option is the -mordered option.  Although all PA 1.x systems have ordered memory
accesses, PA 2.0 systems are weakly ordered.  Since PA 2.0 are now prevalent, we set the
default to -mno-ordered.  For ordered systems, we fall back to just a compiler memory barrier.

I believe acquire and release fences can be defined in a similar way using an ordered load
and an ordered store, respectively.

Tested on hppa-unknown-linux-gnu, hppa2.0w-hp-hpux11.11 and hppa64-hp-hpux11.11.

Committed to trunk.


2019-11-07  John David Anglin  <danglin@gcc.gnu.org>

	* config/pa/pa.md (memory_barrier): Revise to use ldcw barriers.
	Enhance comment.
	(memory_barrier_coherent, memory_barrier_64, memory_barrier_32): New
	insn patterns using ldcw instruction.
	(memory_barrier): Remove insn pattern using sync instruction.
	* config/pa/pa.opt (coherent-ldcw): New option.
	(ordered): New option.


Index: config/pa/pa.md
--- config/pa/pa.md	(revision 277870)
+++ config/pa/pa.md	(working copy)
@@ -10086,23 +10086,55 @@ 
    (set_attr "length" "4,16")])

 ;; PA 2.0 hardware supports out-of-order execution of loads and stores, so
-;; we need a memory barrier to enforce program order for memory references.
-;; Since we want PA 1.x code to be PA 2.0 compatible, we also need the
-;; barrier when generating PA 1.x code.
+;; we need memory barriers to enforce program order for memory references
+;; when the TLB and PSW O bits are not set.  We assume all PA 2.0 systems
+;; are weakly ordered since neither HP-UX or Linux set the PSW O bit.  Since
+;; we want PA 1.x code to be PA 2.0 compatible, we also need barriers when
+;; generating PA 1.x code even though all PA 1.x systems are strongly ordered.

+;; When barriers are needed, we use a strongly ordered ldcw instruction as
+;; the barrier.  Most PA 2.0 targets are cache coherent.  In that case, we
+;; can use the coherent cache control hint and avoid aligning the ldcw
+;; address.  In spite of its description, it is not clear that the sync
+;; instruction works as a barrier.
 (define_expand "memory_barrier"
-  [(set (match_dup 0)
-        (unspec:BLK [(match_dup 0)] UNSPEC_MEMORY_BARRIER))]
+  [(parallel
+     [(set (match_dup 0) (unspec:BLK [(match_dup 0)] UNSPEC_MEMORY_BARRIER))
+      (clobber (match_dup 1))])]
-  operands[0] = gen_rtx_MEM (BLKmode, gen_rtx_SCRATCH (Pmode));
+  /* We don't need a barrier if the target uses ordered memory references.  */
+    FAIL;
+  operands[1] = gen_reg_rtx (Pmode);
+  operands[0] = gen_rtx_MEM (BLKmode, operands[1]);
   MEM_VOLATILE_P (operands[0]) = 1;

-(define_insn "*memory_barrier"
+(define_insn "*memory_barrier_coherent"
   [(set (match_operand:BLK 0 "" "")
-        (unspec:BLK [(match_dup 0)] UNSPEC_MEMORY_BARRIER))]
+        (unspec:BLK [(match_dup 0)] UNSPEC_MEMORY_BARRIER))
+   (clobber (match_operand 1 "pmode_register_operand" "=r"))]
+  "ldcw,co 0(%%sp),%1"
+  [(set_attr "type" "binary")
+   (set_attr "length" "4")])
+(define_insn "*memory_barrier_64"
+  [(set (match_operand:BLK 0 "" "")
+        (unspec:BLK [(match_dup 0)] UNSPEC_MEMORY_BARRIER))
+    (clobber (match_operand 1 "pmode_register_operand" "=&r"))]
+  "ldo 15(%%sp),%1\n\tdepd %%r0,63,3,%1\n\tldcw 0(%1),%1"
+  [(set_attr "type" "binary")
+   (set_attr "length" "12")])
+(define_insn "*memory_barrier_32"
+  [(set (match_operand:BLK 0 "" "")
+        (unspec:BLK [(match_dup 0)] UNSPEC_MEMORY_BARRIER))
+    (clobber (match_operand 1 "pmode_register_operand" "=&r"))]
-  "sync"
+  "ldo 15(%%sp),%1\n\t{dep|depw} %%r0,31,3,%1\n\tldcw 0(%1),%1"
   [(set_attr "type" "binary")
-   (set_attr "length" "4")])
+   (set_attr "length" "12")])
Index: config/pa/pa.opt
--- config/pa/pa.opt	(revision 277870)
+++ config/pa/pa.opt	(working copy)
@@ -45,6 +45,10 @@ 
 Target Report Mask(CALLER_COPIES)
 Caller copies function arguments passed by hidden reference.

+Target Report Var(TARGET_COHERENT_LDCW) Init(1)
+Use ldcw/ldcd coherent cache-control hint.
 Target Report Mask(DISABLE_FPREGS)
 Disable FP regs.
@@ -90,6 +94,10 @@ 
 Target RejectNegative Report Mask(NO_SPACE_REGS)
 Disable space regs.

+Target Report Var(TARGET_ORDERED) Init(0)
+Assume memory references are ordered and barriers are not needed.
 Target RejectNegative
 Generate PA1.0 code.