Skip to content
GitLab
Commit b152c943 authored by Matthias Braun's avatar Matthias Braun
Browse files

lower_dw: Shift lowering without intrinsics 

We can still use ia32 ShlD, ShrD functions by matching Or(Shl,Shr)
patterns in the ia32 transformer.
parent 34c916bd
Expand all Hide whitespace changes
Inline Side-by-side
ir/be/ia32/ia32_intrinsics.c
View file @ b152c943
......@@ -242,275 +242,6 @@ static int map_Sub(ir_node *call, void *ctx)
return 1;
}
/**
* Map a Shl (a_l, a_h, count)
*/
static int map_Shl(ir_node *call, void *ctx)
{
ir_graph *irg = current_ir_graph;
dbg_info *dbg = get_irn_dbg_info(call);
ir_node *block = get_nodes_block(call);
ir_node **params = get_Call_param_arr(call);
ir_type *method = get_Call_type(call);
ir_node *a_l = params[BINOP_Left_Low];
ir_node *a_h = params[BINOP_Left_High];
ir_node *cnt = params[BINOP_Right_Low];
ir_mode *l_mode = get_type_mode(get_method_res_type(method, 0));
ir_mode *h_mode = get_type_mode(get_method_res_type(method, 1));
ir_mode *c_mode;
ir_node *l_res, *h_res, *irn, *cond, *upper, *n_block, *l1, *l2, *h1, *h2, *in[2];
(void) ctx;
if (is_Const(cnt)) {
/* the shift count is a const, create better code */
ir_tarval *tv = get_Const_tarval(cnt);
if (tarval_cmp(tv, new_tarval_from_long(32, l_mode))
& (ir_relation_greater_equal)) {
/* simplest case: shift only the lower bits. Note that there is no
need to reduce the constant here, this is done by the hardware. */
ir_node *conv = new_rd_Conv(dbg, block, a_l, h_mode);
h_res = new_rd_Shl(dbg, block, conv, cnt, h_mode);
l_res = new_rd_Const(dbg, irg, get_mode_null(l_mode));
} else {
/* h_res = SHLD a_h, a_l, cnt */
h_res = new_bd_ia32_l_ShlD(dbg, block, a_h, a_l, cnt, h_mode);
/* l_res = SHL a_l, cnt */
l_res = new_bd_ia32_l_ShlDep(dbg, block, a_l, cnt, h_res, l_mode);
}
resolve_call(call, l_res, h_res, irg, block);
return 1;
}
part_block(call);
upper = get_nodes_block(call);
/* h_res = SHLD a_h, a_l, cnt */
h1 = new_bd_ia32_l_ShlD(dbg, upper, a_h, a_l, cnt, h_mode);
/* l_res = SHL a_l, cnt */
l1 = new_bd_ia32_l_ShlDep(dbg, upper, a_l, cnt, h1, l_mode);
c_mode = get_irn_mode(cnt);
irn = new_r_Const_long(irg, c_mode, 32);
irn = new_rd_And(dbg, upper, cnt, irn, c_mode);
irn = new_rd_Cmp(dbg, upper, irn, new_r_Const(irg, get_mode_null(c_mode)), ir_relation_equal);
cond = new_rd_Cond(dbg, upper, irn);
in[0] = new_r_Proj(cond, mode_X, pn_Cond_true);
in[1] = new_r_Proj(cond, mode_X, pn_Cond_false);
/* the block for cnt >= 32 */
n_block = new_rd_Block(dbg, irg, 1, &in[1]);
h2 = new_rd_Conv(dbg, n_block, l1, h_mode);
l2 = new_r_Const(irg, get_mode_null(l_mode));
in[1] = new_r_Jmp(n_block);
set_irn_in(block, 2, in);
in[0] = l1;
in[1] = l2;
l_res = new_r_Phi(block, 2, in, l_mode);
set_Block_phis(block, l_res);
in[0] = h1;
in[1] = h2;
h_res = new_r_Phi(block, 2, in, h_mode);
set_Phi_next(l_res, h_res);
set_Phi_next(h_res, NULL);
/* move it down */
set_nodes_block(call, block);
for (irn = (ir_node*)get_irn_link(call); irn != NULL;
irn = (ir_node*)get_irn_link(irn)) {
set_nodes_block(irn, block);
}
resolve_call(call, l_res, h_res, irg, block);
return 1;
}
/**
* Map a Shr (a_l, a_h, count)
*/
static int map_Shr(ir_node *call, void *ctx)
{
ir_graph *irg = current_ir_graph;
dbg_info *dbg = get_irn_dbg_info(call);
ir_node *block = get_nodes_block(call);
ir_node **params = get_Call_param_arr(call);
ir_type *method = get_Call_type(call);
ir_node *a_l = params[BINOP_Left_Low];
ir_node *a_h = params[BINOP_Left_High];
ir_node *cnt = params[BINOP_Right_Low];
ir_mode *l_mode = get_type_mode(get_method_res_type(method, 0));
ir_mode *h_mode = get_type_mode(get_method_res_type(method, 1));
ir_mode *c_mode;
ir_node *l_res, *h_res, *irn, *cond, *upper, *n_block, *l1, *l2, *h1, *h2, *in[2];
(void) ctx;
if (is_Const(cnt)) {
/* the shift count is a const, create better code */
ir_tarval *tv = get_Const_tarval(cnt);
if (tarval_cmp(tv, new_tarval_from_long(32, l_mode)) & (ir_relation_greater_equal)) {
/* simplest case: shift only the higher bits. Note that there is no
need to reduce the constant here, this is done by the hardware. */
ir_node *conv = new_rd_Conv(dbg, block, a_h, l_mode);
h_res = new_rd_Const(dbg, irg, get_mode_null(h_mode));
l_res = new_rd_Shr(dbg, block, conv, cnt, l_mode);
} else {
/* l_res = SHRD a_h:a_l, cnt */
l_res = new_bd_ia32_l_ShrD(dbg, block, a_l, a_h, cnt, l_mode);
/* h_res = SHR a_h, cnt */
h_res = new_bd_ia32_l_ShrDep(dbg, block, a_h, cnt, l_res, h_mode);
}
resolve_call(call, l_res, h_res, irg, block);
return 1;
}
part_block(call);
upper = get_nodes_block(call);
/* l_res = SHRD a_h:a_l, cnt */
l1 = new_bd_ia32_l_ShrD(dbg, upper, a_l, a_h, cnt, l_mode);
/* h_res = SHR a_h, cnt */
h1 = new_bd_ia32_l_ShrDep(dbg, upper, a_h, cnt, l1, h_mode);
c_mode = get_irn_mode(cnt);
irn = new_r_Const_long(irg, c_mode, 32);
irn = new_rd_And(dbg, upper, cnt, irn, c_mode);
irn = new_rd_Cmp(dbg, upper, irn, new_r_Const(irg, get_mode_null(c_mode)), ir_relation_equal);
cond = new_rd_Cond(dbg, upper, irn);
in[0] = new_r_Proj(cond, mode_X, pn_Cond_true);
in[1] = new_r_Proj(cond, mode_X, pn_Cond_false);
/* the block for cnt >= 32 */
n_block = new_rd_Block(dbg, irg, 1, &in[1]);
l2 = new_rd_Conv(dbg, n_block, h1, l_mode);
h2 = new_r_Const(irg, get_mode_null(h_mode));
in[1] = new_r_Jmp(n_block);
set_irn_in(block, 2, in);
in[0] = l1;
in[1] = l2;
l_res = new_r_Phi(block, 2, in, l_mode);
set_Block_phis(block, l_res);
in[0] = h1;
in[1] = h2;
h_res = new_r_Phi(block, 2, in, h_mode);
set_Phi_next(l_res, h_res);
set_Phi_next(h_res, NULL);
/* move it down */
set_nodes_block(call, block);
for (irn = (ir_node*)get_irn_link(call); irn != NULL;
irn = (ir_node*)get_irn_link(irn)) {
set_nodes_block(irn, block);
}
resolve_call(call, l_res, h_res, irg, block);
return 1;
}
/**
* Map a Shrs (a_l, a_h, count)
*/
static int map_Shrs(ir_node *call, void *ctx)
{
ir_graph *irg = current_ir_graph;
dbg_info *dbg = get_irn_dbg_info(call);
ir_node *block = get_nodes_block(call);
ir_node **params = get_Call_param_arr(call);
ir_type *method = get_Call_type(call);
ir_node *a_l = params[BINOP_Left_Low];
ir_node *a_h = params[BINOP_Left_High];
ir_node *cnt = params[BINOP_Right_Low];
ir_mode *l_mode = get_type_mode(get_method_res_type(method, 0));
ir_mode *h_mode = get_type_mode(get_method_res_type(method, 1));
ir_mode *c_mode;
ir_node *l_res, *h_res, *irn, *cond, *upper, *n_block, *l1, *l2, *h1, *h2, *in[2];
(void) ctx;
if (is_Const(cnt)) {
/* the shift count is a const, create better code */
ir_tarval *tv = get_Const_tarval(cnt);
if (tarval_cmp(tv, new_tarval_from_long(32, l_mode)) & (ir_relation_greater_equal)) {
/* simplest case: shift only the higher bits. Note that there is no
need to reduce the constant here, this is done by the hardware. */
ir_node *conv = new_rd_Conv(dbg, block, a_h, l_mode);
ir_mode *c_mode = get_irn_mode(cnt);
h_res = new_rd_Shrs(dbg, block, a_h, new_r_Const_long(irg, c_mode, 31), h_mode);
l_res = new_rd_Shrs(dbg, block, conv, cnt, l_mode);
} else {
/* l_res = SHRD a_h:a_l, cnt */
l_res = new_bd_ia32_l_ShrD(dbg, block, a_l, a_h, cnt, l_mode);
/* h_res = SAR a_h, cnt */
h_res = new_bd_ia32_l_SarDep(dbg, block, a_h, cnt, l_res, h_mode);
}
resolve_call(call, l_res, h_res, irg, block);
return 1;
}
part_block(call);
upper = get_nodes_block(call);
/* l_res = SHRD a_h:a_l, cnt */
l1 = new_bd_ia32_l_ShrD(dbg, upper, a_l, a_h, cnt, l_mode);
/* h_res = SAR a_h, cnt */
h1 = new_bd_ia32_l_SarDep(dbg, upper, a_h, cnt, l1, h_mode);
c_mode = get_irn_mode(cnt);
irn = new_r_Const_long(irg, c_mode, 32);
irn = new_rd_And(dbg, upper, cnt, irn, c_mode);
irn = new_rd_Cmp(dbg, upper, irn, new_r_Const(irg, get_mode_null(c_mode)), ir_relation_equal);
cond = new_rd_Cond(dbg, upper, irn);
in[0] = new_r_Proj(cond, mode_X, pn_Cond_true);
in[1] = new_r_Proj(cond, mode_X, pn_Cond_false);
/* the block for cnt >= 32 */
n_block = new_rd_Block(dbg, irg, 1, &in[1]);
l2 = new_rd_Conv(dbg, n_block, h1, l_mode);
h2 = new_rd_Shrs(dbg, n_block, a_h, new_r_Const_long(irg, c_mode, 31), h_mode);
in[1] = new_r_Jmp(n_block);
set_irn_in(block, 2, in);
in[0] = l1;
in[1] = l2;
l_res = new_r_Phi(block, 2, in, l_mode);
set_Block_phis(block, l_res);
in[0] = h1;
in[1] = h2;
h_res = new_r_Phi(block, 2, in, h_mode);
set_Phi_next(l_res, h_res);
set_Phi_next(h_res, NULL);
/* move it down */
set_nodes_block(call, block);
for (irn = (ir_node*)get_irn_link(call); irn != NULL;
irn = (ir_node*)get_irn_link(irn)) {
set_nodes_block(irn, block);
}
resolve_call(call, l_res, h_res, irg, block);
return 1;
}
/**
* Checks where node high is a sign extension of low.
*/
......@@ -904,18 +635,6 @@ ir_entity *ia32_create_intrinsic_fkt(ir_type *method, const ir_op *op,
ent = &i_ents[iro_Sub];
mapper = map_Sub;
break;
case iro_Shl:
ent = &i_ents[iro_Shl];
mapper = map_Shl;
break;
case iro_Shr:
ent = &i_ents[iro_Shr];
mapper = map_Shr;
break;
case iro_Shrs:
ent = &i_ents[iro_Shrs];
mapper = map_Shrs;
break;
case iro_Mul:
ent = &i_ents[iro_Mul];
mapper = map_Mul;
......
ir/be/ia32/ia32_spec.pl
View file @ b152c943
......@@ -643,12 +643,6 @@ ShlMem => {
modified_flags => $status_flags
},
l_ShlDep => {
cmp_attr => "return 1;",
reg_req => { in => [ "none", "none", "none" ], out => [ "none" ] },
ins => [ "val", "count", "dep" ],
},
ShlD => {
irn_flags => [ "rematerializable" ],
reg_req => { in => [ "gp", "gp", "ecx" ],
......@@ -662,12 +656,6 @@ ShlD => {
modified_flags => $status_flags
},
l_ShlD => {
cmp_attr => "return 1;",
reg_req => { in => [ "none", "none", "none" ], out => [ "none" ] },
ins => [ "val_high", "val_low", "count" ],
},
Shr => {
irn_flags => [ "rematerializable" ],
reg_req => { in => [ "gp", "ecx" ],
......@@ -693,12 +681,6 @@ ShrMem => {
modified_flags => $status_flags
},
l_ShrDep => {
cmp_attr => "return 1;",
reg_req => { in => [ "none", "none", "none" ], out => [ "none" ] },
ins => [ "val", "count", "dep" ],
},
ShrD => {
irn_flags => [ "rematerializable" ],
reg_req => { in => [ "gp", "gp", "ecx" ],
......@@ -712,12 +694,6 @@ ShrD => {
modified_flags => $status_flags
},
l_ShrD => {
cmp_attr => "return 1;",
reg_req => { in => [ "none", "none", "none" ], out => [ "none" ] },
ins => [ "val_high", "val_low", "count" ],
},
Sar => {
irn_flags => [ "rematerializable" ],
reg_req => { in => [ "gp", "ecx" ],
......@@ -743,12 +719,6 @@ SarMem => {
modified_flags => $status_flags
},
l_SarDep => {
cmp_attr => "return 1;",
ins => [ "val", "count", "dep" ],
reg_req => { in => [ "none", "none", "none" ], out => [ "none" ] },
},
Ror => {
irn_flags => [ "rematerializable" ],
reg_req => { in => [ "gp", "ecx" ],
......
ir/be/ia32/ia32_transform.c
View file @ b152c943
......@@ -43,6 +43,7 @@
#include "irprintf.h"
#include "debug.h"
#include "irdom.h"
#include "iropt.h"
#include "error.h"
#include "array_t.h"
#include "heights.h"
......@@ -1418,7 +1419,116 @@ static ir_node *gen_And(ir_node *node)
match_commutative | match_mode_neutral | match_am | match_immediate);
}
/**
* test wether 2 values result in 'x' and '32-x' when interpreted as a shift
* value.
*/
static bool is_complementary_shifts(ir_node *value1, ir_node *value2)
{
if (is_Const(value1) && is_Const(value2)) {
ir_tarval *tv1 = get_Const_tarval(value1);
ir_tarval *tv2 = get_Const_tarval(value2);
if (tarval_is_long(tv1) && tarval_is_long(tv2)) {
long v1 = get_tarval_long(tv1);
long v2 = get_tarval_long(tv2);
return v1 < v2 && v2 == 32-v1;
}
}
return false;
}
typedef ir_node* (*new_shiftd_func)(dbg_info *dbgi, ir_node *block,
ir_node *high, ir_node *low,
ir_node *count);
/**
* Transforms a l_ShlD/l_ShrD into a ShlD/ShrD. Those nodes have 3 data inputs:
* op1 - target to be shifted
* op2 - contains bits to be shifted into target
* op3 - shift count
* Only op3 can be an immediate.
*/
static ir_node *gen_64bit_shifts(dbg_info *dbgi, ir_node *block,
ir_node *high, ir_node *low, ir_node *count,
new_shiftd_func func)
{
ir_node *new_block = be_transform_node(block);
ir_node *new_high = be_transform_node(high);
ir_node *new_low = be_transform_node(low);
ir_node *new_count;
ir_node *new_node;
/* the shift amount can be any mode that is bigger than 5 bits, since all
* other bits are ignored anyway */
while (is_Conv(count) &&
get_irn_n_edges(count) == 1 &&
mode_is_int(get_irn_mode(count))) {
assert(get_mode_size_bits(get_irn_mode(count)) >= 5);
count = get_Conv_op(count);
}
new_count = create_immediate_or_transform(count, 0);
new_node = func(dbgi, new_block, new_high, new_low, new_count);
return new_node;
}
static ir_node *match_64bit_shift(ir_node *node)
{
ir_node *op1 = get_Or_left(node);
ir_node *op2 = get_Or_right(node);
if (is_Shr(op1)) {
ir_node *tmp = op1;
op1 = op2;
op2 = tmp;
}
/* match ShlD operation */
if (is_Shl(op1) && is_Shr(op2)) {
ir_node *shl_right = get_Shl_right(op1);
ir_node *shl_left = get_Shl_left(op1);
ir_node *shr_right = get_Shr_right(op2);
ir_node *shr_left = get_Shr_left(op2);
/* constant ShlD operation */
if (is_complementary_shifts(shl_right, shr_right)) {
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *block = get_nodes_block(node);
return gen_64bit_shifts(dbgi, block, shl_left, shr_left, shl_right,
new_bd_ia32_ShlD);
}
/* constant ShrD operation */
if (is_complementary_shifts(shr_right, shl_right)) {
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *block = get_nodes_block(node);
return gen_64bit_shifts(dbgi, block, shr_left, shl_left, shr_right,
new_bd_ia32_ShrD);
}
/* lower_dw produces the following for ShlD:
* Or(Shr(Shr(high,1),Not(c)),Shl(low,c)) */
if (is_Shr(shr_left) && is_Not(shr_right)
&& is_Const_1(get_Shr_right(shr_left))
&& get_Not_op(shr_right) == shl_right) {
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *block = get_nodes_block(node);
ir_node *val_h = get_Shr_left(shr_left);
return gen_64bit_shifts(dbgi, block, shl_left, val_h, shl_right,
new_bd_ia32_ShlD);
}
/* lower_dw produces the following for ShrD:
* Or(Shl(Shl(high,1),Not(c)), Shr(low,c)) */
if (is_Shl(shl_left) && is_Not(shl_right)
&& is_Const_1(get_Shl_right(shl_left))
&& get_Not_op(shl_right) == shr_right) {
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *block = get_nodes_block(node);
ir_node *val_h = get_Shl_left(shl_left);
return gen_64bit_shifts(dbgi, block, shr_left, val_h, shr_right,
new_bd_ia32_ShrD);
}
}
return NULL;
}
/**
* Creates an ia32 Or.
......@@ -1429,6 +1539,11 @@ static ir_node *gen_Or(ir_node *node)
{
ir_node *op1 = get_Or_left(node);
ir_node *op2 = get_Or_right(node);
ir_node *res;
res = match_64bit_shift(node);
if (res != NULL)
return res;
assert (! mode_is_float(get_irn_mode(node)));
return gen_binop(node, op1, op2, new_bd_ia32_Or, match_commutative
......@@ -4139,31 +4254,6 @@ static ir_node *gen_IJmp(ir_node *node)
return new_node;
}
static ir_node *gen_ia32_l_ShlDep(ir_node *node)
{
ir_node *left = get_irn_n(node, n_ia32_l_ShlDep_val);
ir_node *right = get_irn_n(node, n_ia32_l_ShlDep_count);
return gen_shift_binop(node, left, right, new_bd_ia32_Shl,
match_immediate | match_mode_neutral);
}
static ir_node *gen_ia32_l_ShrDep(ir_node *node)
{
ir_node *left = get_irn_n(node, n_ia32_l_ShrDep_val);
ir_node *right = get_irn_n(node, n_ia32_l_ShrDep_count);
return gen_shift_binop(node, left, right, new_bd_ia32_Shr,
match_immediate);
}
static ir_node *gen_ia32_l_SarDep(ir_node *node)
{
ir_node *left = get_irn_n(node, n_ia32_l_SarDep_val);
ir_node *right = get_irn_n(node, n_ia32_l_SarDep_count);
return gen_shift_binop(node, left, right, new_bd_ia32_Sar,
match_immediate);
}
static ir_node *gen_ia32_l_Add(ir_node *node)
{
ir_node *left = get_irn_n(node, n_ia32_l_Add_left);
......@@ -4240,62 +4330,6 @@ static ir_node *gen_ia32_l_Sbb(ir_node *node)
match_am | match_immediate | match_mode_neutral);
}
/**
* Transforms a l_ShlD/l_ShrD into a ShlD/ShrD. Those nodes have 3 data inputs:
* op1 - target to be shifted
* op2 - contains bits to be shifted into target
* op3 - shift count
* Only op3 can be an immediate.
*/
static ir_node *gen_lowered_64bit_shifts(ir_node *node, ir_node *high,
ir_node *low, ir_node *count)
{
ir_node *block = get_nodes_block(node);
ir_node *new_block = be_transform_node(block);
dbg_info *dbgi = get_irn_dbg_info(node);
ir_node *new_high = be_transform_node(high);
ir_node *new_low = be_transform_node(low);
ir_node *new_count;
ir_node *new_node;
/* the shift amount can be any mode that is bigger than 5 bits, since all
* other bits are ignored anyway */
while (is_Conv(count) &&
get_irn_n_edges(count) == 1 &&
mode_is_int(get_irn_mode(count))) {
assert(get_mode_size_bits(get_irn_mode(count)) >= 5);
count = get_Conv_op(count);
}
new_count = create_immediate_or_transform(count, 0);
if (is_ia32_l_ShlD(node)) {
new_node = new_bd_ia32_ShlD(dbgi, new_block, new_high, new_low,
new_count);
} else {
new_node = new_bd_ia32_ShrD(dbgi, new_block, new_high, new_low,
new_count);
}
SET_IA32_ORIG_NODE(new_node, node);
return new_node;
}
static ir_node *gen_ia32_l_ShlD(ir_node *node)
{
ir_node *high = get_irn_n(node, n_ia32_l_ShlD_val_high);
ir_node *low = get_irn_n(node, n_ia32_l_ShlD_val_low);
ir_node *count = get_irn_n(node, n_ia32_l_ShlD_count);
return gen_lowered_64bit_shifts(node, high, low, count);
}
static ir_node *gen_ia32_l_ShrD(ir_node *node)
{
ir_node *high = get_irn_n(node, n_ia32_l_ShrD_val_high);
ir_node *low = get_irn_n(node, n_ia32_l_ShrD_val_low);
ir_node *count = get_irn_n(node, n_ia32_l_ShrD_count);
return gen_lowered_64bit_shifts(node, high, low, count);
}
static ir_node *gen_ia32_l_LLtoFloat(ir_node *node)
{
ir_node *src_block = get_nodes_block(node);
......@@ -5665,12 +5699,7 @@ static void register_transformers(void)
be_set_transform_function(op_ia32_l_IMul, gen_ia32_l_IMul);
be_set_transform_function(op_ia32_l_LLtoFloat, gen_ia32_l_LLtoFloat);
be_set_transform_function(op_ia32_l_Mul, gen_ia32_l_Mul);
be_set_transform_function(op_ia32_l_SarDep, gen_ia32_l_SarDep);
be_set_transform_function(op_ia32_l_Sbb, gen_ia32_l_Sbb);
be_set_transform_function(op_ia32_l_ShlDep, gen_ia32_l_ShlDep);
be_set_transform_function(op_ia32_l_ShlD, gen_ia32_l_ShlD);
be_set_transform_function(op_ia32_l_ShrDep, gen_ia32_l_ShrDep);
be_set_transform_function(op_ia32_l_ShrD, gen_ia32_l_ShrD);
be_set_transform_function(op_ia32_l_Sub, gen_ia32_l_Sub);
be_set_transform_function(op_ia32_GetEIP, be_duplicate_node);