[cxx-mem-model 2/3] i386: Implement atomic_load/store.
Richard Henderson
rth@redhat.com
Thu Nov 3 08:23:00 GMT 2011
* config/i386/i386.md (UNSPEC_MOVA): New.
* config/i386/sync.md (ATOMIC): New mode iterator.
(atomic_load<ATOMIC>, atomic_store<ATOMIC>): New.
(atomic_loaddi_fpu, atomic_storedi_fpu, movdi_via_fpu): New.
---
gcc/ChangeLog.mm | 7 ++
gcc/config/i386/i386.md | 3 +
gcc/config/i386/sync.md | 163 +++++++++++++++++++++++++++++++++++++++++++++++
3 files changed, 173 insertions(+), 0 deletions(-)
diff --git a/gcc/ChangeLog.mm b/gcc/ChangeLog.mm
index 4454e93..24137ab 100644
--- a/gcc/ChangeLog.mm
+++ b/gcc/ChangeLog.mm
@@ -1,5 +1,12 @@
2011-11-02 Richard Henderson <rth@redhat.com>
+ * config/i386/i386.md (UNSPEC_MOVA): New.
+ * config/i386/sync.md (ATOMIC): New mode iterator.
+ (atomic_load<ATOMIC>, atomic_store<ATOMIC>): New.
+ (atomic_loaddi_fpu, atomic_storedi_fpu, movdi_via_fpu): New.
+
+2011-11-02 Richard Henderson <rth@redhat.com>
+
* optabs.c (expand_atomic_load): Don't try compare-and-swap.
2011-11-02 Aldy Hernandez <aldyh@redhat.com>
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index f11998c..d073cb9 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -248,6 +248,9 @@
;; For BMI2 support
UNSPEC_PDEP
UNSPEC_PEXT
+
+ ;; For __atomic support
+ UNSPEC_MOVA
])
(define_c_enum "unspecv" [
diff --git a/gcc/config/i386/sync.md b/gcc/config/i386/sync.md
index e5579b1..0ff1712 100644
--- a/gcc/config/i386/sync.md
+++ b/gcc/config/i386/sync.md
@@ -46,6 +46,169 @@
"lock{%;} or{l}\t{$0, (%%esp)|DWORD PTR [esp], 0}"
[(set_attr "memory" "unknown")])
+;; ??? From volume 3 section 7.1.1 Guaranteed Atomic Operations,
+;; Only beginning at Pentium family processors do we get any guarantee of
+;; atomicity in aligned 64-bit quantities. Beginning at P6, we get a
+;; guarantee for 64-bit accesses that do not cross a cacheline boundary.
+;;
+;; Note that the TARGET_CMPXCHG8B test below is a stand-in for "Pentium".
+;;
+;; Importantly, *no* processor makes atomicity guarantees for larger
+;; accesses. In particular, there's no way to perform an atomic TImode
+;; move, despite the apparent applicability of MOVDQA et al.
+
+(define_mode_iterator ATOMIC
+ [QI HI SI
+ (DI "TARGET_64BIT || (TARGET_CMPXCHG8B && (TARGET_80387 || TARGET_SSE))")
+ ])
+
+(define_expand "atomic_load<mode>"
+ [(set (match_operand:ATOMIC 0 "register_operand" "")
+ (unspec:ATOMIC [(match_operand:ATOMIC 1 "memory_operand" "")
+ (match_operand:SI 2 "const_int_operand" "")]
+ UNSPEC_MOVA))]
+ ""
+{
+ /* For DImode on 32-bit, we can use the FPU to perform the load. */
+ if (<MODE>mode == DImode && !TARGET_64BIT)
+ emit_insn (gen_atomic_loaddi_fpu
+ (operands[0], operands[1],
+ assign_386_stack_local (DImode,
+ (virtuals_instantiated
+ ? SLOT_TEMP : SLOT_VIRTUAL))));
+ else
+ emit_move_insn (operands[0], operands[1]);
+ DONE;
+})
+
+(define_insn_and_split "atomic_loaddi_fpu"
+ [(set (match_operand:DI 0 "nonimmediate_operand" "=x,m,?r")
+ (unspec:DI [(match_operand:DI 1 "memory_operand" "m,m,m")]
+ UNSPEC_MOVA))
+ (clobber (match_operand:DI 2 "memory_operand" "=X,X,m"))
+ (clobber (match_scratch:DF 3 "=X,xf,xf"))]
+ "!TARGET_64BIT && (TARGET_80387 || TARGET_SSE)"
+ "#"
+ "&& reload_completed"
+ [(const_int 0)]
+{
+ rtx dst = operands[0], src = operands[1];
+ rtx mem = operands[2], tmp = operands[3];
+
+ if (SSE_REG_P (dst))
+ emit_move_insn (dst, src);
+ else
+ {
+ if (MEM_P (dst))
+ mem = dst;
+
+ if (FP_REG_P (tmp))
+ emit_insn (gen_movdi_via_fpu (mem, src, tmp));
+ else
+ {
+ adjust_reg_mode (tmp, DImode);
+ emit_move_insn (tmp, src);
+ emit_move_insn (mem, tmp);
+ }
+
+ if (mem != dst)
+ emit_move_insn (dst, mem);
+ }
+ DONE;
+})
+
+(define_expand "atomic_store<mode>"
+ [(set (match_operand:ATOMIC 0 "memory_operand" "")
+ (unspec:ATOMIC [(match_operand:ATOMIC 1 "register_operand" "")
+ (match_operand:SI 2 "const_int_operand" "")]
+ UNSPEC_MOVA))]
+ ""
+{
+ enum memmodel model = (enum memmodel) INTVAL (operands[2]);
+
+ if (<MODE>mode == DImode && !TARGET_64BIT)
+ {
+ /* For DImode on 32-bit, we can use the FPU to perform the store. */
+ /* Note that while we could perform a cmpxchg8b loop, that turns
+ out to be significantly larger than this plus a barrier. */
+ emit_insn (gen_atomic_storedi_fpu
+ (operands[0], operands[1],
+ assign_386_stack_local (DImode,
+ (virtuals_instantiated
+ ? SLOT_TEMP : SLOT_VIRTUAL))));
+ }
+ else
+ {
+ /* For seq-cst stores, when we lack MFENCE, use XCHG. */
+ if (model == MEMMODEL_SEQ_CST && !(TARGET_64BIT || TARGET_SSE2))
+ {
+ emit_insn (gen_atomic_exchange<mode> (gen_reg_rtx (<MODE>mode),
+ operands[0], operands[1],
+ operands[2]));
+ DONE;
+ }
+
+ /* Otherwise use a normal store. */
+ emit_move_insn (operands[0], operands[1]);
+ }
+ /* ... followed by an MFENCE, if required. */
+ if (model == MEMMODEL_SEQ_CST)
+ emit_insn (gen_mem_thread_fence (operands[2]));
+ DONE;
+})
+
+(define_insn_and_split "atomic_storedi_fpu"
+ [(set (match_operand:DI 0 "memory_operand" "=m,m,m")
+ (unspec:DI [(match_operand:DI 1 "register_operand" "x,m,?r")]
+ UNSPEC_MOVA))
+ (clobber (match_operand:DI 2 "memory_operand" "=X,X,m"))
+ (clobber (match_scratch:DF 3 "=X,xf,xf"))]
+ "!TARGET_64BIT && (TARGET_80387 || TARGET_SSE)"
+ "#"
+ "&& reload_completed"
+ [(const_int 0)]
+{
+ rtx dst = operands[0], src = operands[1];
+ rtx mem = operands[2], tmp = operands[3];
+
+ if (!SSE_REG_P (src))
+ {
+ if (REG_P (src))
+ {
+ emit_move_insn (mem, src);
+ src = mem;
+ }
+
+ if (FP_REG_P (tmp))
+ {
+ emit_insn (gen_movdi_via_fpu (dst, src, tmp));
+ DONE;
+ }
+ else
+ {
+ adjust_reg_mode (tmp, DImode);
+ emit_move_insn (tmp, mem);
+ src = tmp;
+ }
+ }
+ emit_move_insn (dst, src);
+ DONE;
+})
+
+;; ??? You'd think that we'd be able to perform this via FLOAT + FIX_TRUNC
+;; operations. But the fix_trunc patterns want way more setup than we want
+;; to provide. Note that the scratch is DFmode instead of XFmode in order
+;; to make it easy to allocate a scratch in either SSE or FP_REGs above.
+(define_insn "movdi_via_fpu"
+ [(set (match_operand:DI 0 "memory_operand" "=m")
+ (unspec:DI [(match_operand:DI 1 "memory_operand" "m")] UNSPEC_MOVA))
+ (clobber (match_operand:DF 2 "register_operand" "=f"))]
+ "TARGET_80387"
+ "fild\t%1\;fistp\t%0"
+ [(set_attr "type" "multi")
+ ;; Worst case based on full sib+offset32 addressing modes
+ (set_attr "length" "14")])
+
(define_expand "atomic_compare_and_swap<mode>"
[(match_operand:QI 0 "register_operand" "") ;; bool success output
(match_operand:SWI124 1 "register_operand" "") ;; oldval output
--
1.7.6.4
More information about the Gcc-patches
mailing list