When forking to call a partitioned routine there is no need to propagate local
state from the active thread to the forked threads.
I've committed this patch to implement that optimization.
nathan
2015-08-25 Nathan Sidwell <nat...@codesourcery.com>
* config/nvptx/nvptx.c (nvptx_emit_forking): Add is_call argument,
propagate it into mask.
(nvptx_emit_joining): Likewise.
(nvptx_expand_call): Move emit_forking call to later. Add is_call
argument.
(nvptx_expand_oacc_fork, nvptx_expand_oacc_join): Asjust calls.
(nvptx_discover_pars): Don't look for predecessor insn in call
forks and joins.
(nvptx_process_pars): Don't emit propagation code for a call.
Index: gcc/config/nvptx/nvptx.c
===================================================================
--- gcc/config/nvptx/nvptx.c (revision 227159)
+++ gcc/config/nvptx/nvptx.c (working copy)
@@ -1047,16 +1047,16 @@ nvptx_expand_compare (rtx compare)
/* Emit forking instructions for MASK. */
static void
-nvptx_emit_forking (unsigned mask)
+nvptx_emit_forking (unsigned mask, bool is_call)
{
mask &= (GOMP_DIM_MASK (GOMP_DIM_WORKER)
| GOMP_DIM_MASK (GOMP_DIM_VECTOR));
if (mask)
{
- rtx op = GEN_INT (mask);
+ rtx op = GEN_INT (mask | (is_call << GOMP_DIM_MAX));
/* Emit fork for worker level. */
- if (mask & GOMP_DIM_MASK (GOMP_DIM_WORKER))
+ if (!is_call && mask & GOMP_DIM_MASK (GOMP_DIM_WORKER))
emit_insn (gen_nvptx_fork (op));
emit_insn (gen_nvptx_forked (op));
}
@@ -1065,16 +1065,19 @@ nvptx_emit_forking (unsigned mask)
/* Emit joining instructions for MASK. */
static void
-nvptx_emit_joining (unsigned mask)
+nvptx_emit_joining (unsigned mask, bool is_call)
{
mask &= (GOMP_DIM_MASK (GOMP_DIM_WORKER)
| GOMP_DIM_MASK (GOMP_DIM_VECTOR));
if (mask)
{
- rtx op = GEN_INT (mask);
+ rtx op = GEN_INT (mask | (is_call << GOMP_DIM_MAX));
- /* Emit joining for all pars. */
- emit_insn (gen_nvptx_joining (op));
+ /* Emit joining for all non-call pars to ensure there's a single
+ predecessor for the block the join insn ends up in. This is
+ needed for skipping entire loops. */
+ if (!is_call)
+ emit_insn (gen_nvptx_joining (op));
emit_insn (gen_nvptx_join (op));
}
}
@@ -1135,8 +1138,6 @@ nvptx_expand_call (rtx retval, rtx addre
}
}
- nvptx_emit_forking (parallel);
-
if (cfun->machine->funtype
/* It's possible to construct testcases where we call a variable.
See compile/20020129-1.c. stdarg_p will crash so avoid calling it
@@ -1195,11 +1196,12 @@ nvptx_expand_call (rtx retval, rtx addre
write_func_decl_from_insn (func_decls, retval, pat, callee);
}
}
+ nvptx_emit_forking (parallel, true);
emit_call_insn (pat);
if (tmp_retval != retval)
emit_move_insn (retval, tmp_retval);
- nvptx_emit_joining (parallel);
+ nvptx_emit_joining (parallel, true);
}
/* Expand the oacc fork & join primitive into ptx-required unspecs. */
@@ -1207,13 +1209,13 @@ nvptx_expand_call (rtx retval, rtx addre
void
nvptx_expand_oacc_fork (unsigned mode)
{
- nvptx_emit_forking (GOMP_DIM_MASK (mode));
+ nvptx_emit_forking (GOMP_DIM_MASK (mode), false);
}
void
nvptx_expand_oacc_join (unsigned mode)
{
- nvptx_emit_joining (GOMP_DIM_MASK (mode));
+ nvptx_emit_joining (GOMP_DIM_MASK (mode), false);
}
/* Expander for reduction locking and unlocking. We expect SRC to be
@@ -2611,7 +2613,8 @@ nvptx_discover_pars (bb_insn_map_t *map)
l = new parallel (l, mask);
l->forked_block = block;
l->forked_insn = end;
- if (mask & GOMP_DIM_MASK (GOMP_DIM_WORKER))
+ if (!(mask & GOMP_DIM_MASK (GOMP_DIM_MAX))
+ && (mask & GOMP_DIM_MASK (GOMP_DIM_WORKER)))
l->fork_insn
= nvptx_discover_pre (block, CODE_FOR_nvptx_fork);
}
@@ -2626,7 +2629,8 @@ nvptx_discover_pars (bb_insn_map_t *map)
gcc_assert (l->mask == mask);
l->join_block = block;
l->join_insn = end;
- if (mask & GOMP_DIM_MASK (GOMP_DIM_WORKER))
+ if (!(mask & GOMP_DIM_MASK (GOMP_DIM_MAX))
+ && (mask & GOMP_DIM_MASK (GOMP_DIM_WORKER)))
l->joining_insn
= nvptx_discover_pre (block, CODE_FOR_nvptx_joining);
l = l->parent;
@@ -3013,7 +3017,9 @@ nvptx_process_pars (parallel *par)
inner_mask |= par->inner_mask;
}
- if (par->mask & GOMP_DIM_MASK (GOMP_DIM_WORKER))
+ if (par->mask & GOMP_DIM_MASK (GOMP_DIM_MAX))
+ { /* No propagation needed for a call. */ }
+ else if (par->mask & GOMP_DIM_MASK (GOMP_DIM_WORKER))
{
nvptx_wpropagate (false, par->forked_block, par->forked_insn);
nvptx_wpropagate (true, par->forked_block, par->fork_insn);
