* [PATCH] PR middle-end/95126: Expand small const structs as immediate constants.
@ 2022-02-26 21:35 Roger Sayle
2022-06-01 21:16 ` Jeff Law
0 siblings, 1 reply; 2+ messages in thread
From: Roger Sayle @ 2022-02-26 21:35 UTC (permalink / raw)
To: gcc-patches
[-- Attachment #1: Type: text/plain, Size: 2291 bytes --]
This patch resolves PR middle-end/95126 which is a code quality regression,
by teaching the RTL expander to emit small const structs/unions as integer
immediate constants.
The motivating example from the bugzilla PR is:
struct small{ short a,b; signed char c; };
extern int func(struct small X);
void call_func(void)
{
static struct small const s = { 1, 2, 0 };
func(s);
}
which on x86_64 is currently compiled to:
call_func:
movzwl s.0+2(%rip), %eax
movzwl s.0(%rip), %edx
movzwl s.0+4(%rip), %edi
salq $16, %rax
orq %rdx, %rax
salq $32, %rdi
orq %rax, %rdi
jmp func
but with this patch is now optimized to:
call_func:
movl $131073, %edi
jmp func
This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check with no new failures. Ok for mainline?
2022-02-26 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR middle-end/95126
* calls.cc (load_register_parameters): When loading a suitable
immediate_const_ctor_p VAR_DECL into a single word_mode register,
construct it directly in a pseudo rather than read it (by parts)
from memory.
* expr.cc (int_expr_size): Make tree argument a const_tree.
(immediate_const_ctor_p): Helper predicate. Return true for
simple constructors that may be materialized in a register.
(expand_expr_real_1) [VAR_DECL]: When expanding a constant
VAR_DECL with a suitable immediate_const_ctor_p constructor
use store_constructor to materialize it directly in a pseudo.
* expr.h (immediate_const_ctor_p): Prototype here.
* varasm.cc (initializer_constant_valid_for_bitfield_p): Change
VALUE argument from tree to const_tree.
* varasm.h (initializer_constant_valid_for_bitfield_p): Update
prototype.
gcc/testsuite/ChangeLog
PR middle-end/95126
* gcc.target/i386/pr95126-m32-1.c: New test case.
* gcc.target/i386/pr95126-m32-2.c: New test case.
* gcc.target/i386/pr95126-m32-3.c: New test case.
* gcc.target/i386/pr95126-m32-4.c: New test case.
* gcc.target/i386/pr95126-m64-1.c: New test case.
* gcc.target/i386/pr95126-m64-2.c: New test case.
* gcc.target/i386/pr95126-m64-3.c: New test case.
* gcc.target/i386/pr95126-m64-4.c: New test case.
Thanks in advance,
Roger
--
[-- Attachment #2: patchvd4b.txt --]
[-- Type: text/plain, Size: 12168 bytes --]
diff --git a/gcc/calls.cc b/gcc/calls.cc
index e64a937..3417005 100644
--- a/gcc/calls.cc
+++ b/gcc/calls.cc
@@ -2093,7 +2093,8 @@ load_register_parameters (struct arg_data *args, int num_actuals,
poly_int64 size = 0;
HOST_WIDE_INT const_size = 0;
rtx_insn *before_arg = get_last_insn ();
- tree type = TREE_TYPE (args[i].tree_value);
+ tree tree_value = args[i].tree_value;
+ tree type = TREE_TYPE (tree_value);
if (RECORD_OR_UNION_TYPE_P (type) && TYPE_TRANSPARENT_AGGR (type))
type = TREE_TYPE (first_field (type));
/* Set non-negative if we must move a word at a time, even if
@@ -2170,6 +2171,24 @@ load_register_parameters (struct arg_data *args, int num_actuals,
emit_move_insn (gen_rtx_REG (word_mode, REGNO (reg) + j),
args[i].aligned_regs[j]);
+ /* If we need a single register and the source is a constant
+ VAR_DECL with a simple constructor, expand that constructor
+ via a pseudo rather than read from (possibly misaligned)
+ memory. PR middle-end/95126. */
+ else if (nregs == 1
+ && partial == 0
+ && !args[i].pass_on_stack
+ && VAR_P (tree_value)
+ && TREE_READONLY (tree_value)
+ && !TREE_SIDE_EFFECTS (tree_value)
+ && immediate_const_ctor_p (DECL_INITIAL (tree_value)))
+ {
+ rtx target = gen_reg_rtx (word_mode);
+ rtx x = expand_expr (DECL_INITIAL (tree_value),
+ target, word_mode, EXPAND_NORMAL);
+ reg = gen_rtx_REG (word_mode, REGNO (reg));
+ emit_move_insn (reg, x);
+ }
else if (partial == 0 || args[i].pass_on_stack)
{
/* SIZE and CONST_SIZE are 0 for partial arguments and
diff --git a/gcc/expr.cc b/gcc/expr.cc
index 35e4029..b6d8d8c 100644
--- a/gcc/expr.cc
+++ b/gcc/expr.cc
@@ -99,7 +99,7 @@ static void do_tablejump (rtx, machine_mode, rtx, rtx, rtx,
profile_probability);
static rtx const_vector_from_tree (tree);
static tree tree_expr_size (const_tree);
-static HOST_WIDE_INT int_expr_size (tree);
+static HOST_WIDE_INT int_expr_size (const_tree);
static void convert_mode_scalar (rtx, rtx, int);
\f
@@ -4866,7 +4866,22 @@ emit_push_insn (rtx x, machine_mode mode, tree type, rtx size,
return false;
}
}
- emit_block_move (target, xinner, size, BLOCK_OP_CALL_PARM);
+
+ /* If source is a constant VAR_DECL with a simple constructor,
+ store the constructor to the stack instead of moving it. */
+ const_tree decl;
+ if (partial == 0
+ && MEM_P (xinner)
+ && SYMBOL_REF_P (XEXP (xinner, 0))
+ && (decl = SYMBOL_REF_DECL (XEXP (xinner, 0))) != NULL_TREE
+ && VAR_P (decl)
+ && TREE_READONLY (decl)
+ && !TREE_SIDE_EFFECTS (decl)
+ && immediate_const_ctor_p (DECL_INITIAL (decl), 2))
+ store_constructor (DECL_INITIAL (decl), target, 0,
+ int_expr_size (DECL_INITIAL (decl)), false);
+ else
+ emit_block_move (target, xinner, size, BLOCK_OP_CALL_PARM);
}
}
else if (partial > 0)
@@ -6575,6 +6590,25 @@ categorize_ctor_elements (const_tree ctor, HOST_WIDE_INT *p_nz_elts,
p_init_elts, p_complete);
}
+/* Return true if constructor CTOR is simple enough to be materialized
+ in an integer mode register. Limit the size to WORDS words, which
+ is 1 by default. */
+
+bool
+immediate_const_ctor_p (const_tree ctor, unsigned int words)
+{
+ /* Allow function to be called with a VAR_DECL's DECL_INITIAL. */
+ if (!ctor || TREE_CODE (ctor) != CONSTRUCTOR)
+ return false;
+
+ return TREE_CONSTANT (ctor)
+ && !TREE_ADDRESSABLE (ctor)
+ && CONSTRUCTOR_NELTS (ctor)
+ && TREE_CODE (TREE_TYPE (ctor)) != ARRAY_TYPE
+ && int_expr_size (ctor) <= words * UNITS_PER_WORD
+ && initializer_constant_valid_for_bitfield_p (ctor);
+}
+
/* TYPE is initialized by a constructor with NUM_ELTS elements, the last
of which had type LAST_TYPE. Each element was itself a complete
initializer, in the sense that every meaningful byte was explicitly
@@ -10534,6 +10568,21 @@ expand_expr_real_1 (tree exp, rtx target, machine_mode tmode,
if (temp)
return temp;
}
+ /* Expand const VAR_DECLs with CONSTRUCTOR initializers that
+ have scalar integer modes to a reg via store_constructor. */
+ if (TREE_READONLY (exp)
+ && !TREE_SIDE_EFFECTS (exp)
+ && (modifier == EXPAND_NORMAL || modifier == EXPAND_STACK_PARM)
+ && immediate_const_ctor_p (DECL_INITIAL (exp))
+ && SCALAR_INT_MODE_P (TYPE_MODE (TREE_TYPE (exp)))
+ && crtl->emit.regno_pointer_align_length
+ && !target)
+ {
+ target = gen_reg_rtx (TYPE_MODE (TREE_TYPE (exp)));
+ store_constructor (DECL_INITIAL (exp), target, 0,
+ int_expr_size (DECL_INITIAL (exp)), false);
+ return target;
+ }
/* ... fall through ... */
case PARM_DECL:
@@ -13128,7 +13177,7 @@ expr_size (tree exp)
if the size can vary or is larger than an integer. */
static HOST_WIDE_INT
-int_expr_size (tree exp)
+int_expr_size (const_tree exp)
{
tree size;
diff --git a/gcc/expr.h b/gcc/expr.h
index 7e5cf49..d777c28 100644
--- a/gcc/expr.h
+++ b/gcc/expr.h
@@ -338,6 +338,7 @@ extern unsigned HOST_WIDE_INT highest_pow2_factor (const_tree);
extern bool categorize_ctor_elements (const_tree, HOST_WIDE_INT *,
HOST_WIDE_INT *, HOST_WIDE_INT *,
bool *);
+extern bool immediate_const_ctor_p (const_tree, unsigned int words = 1);
extern void expand_operands (tree, tree, rtx, rtx*, rtx*,
enum expand_modifier);
diff --git a/gcc/varasm.cc b/gcc/varasm.cc
index d3d9daf..a00f11d 100644
--- a/gcc/varasm.cc
+++ b/gcc/varasm.cc
@@ -5069,7 +5069,7 @@ initializer_constant_valid_p (tree value, tree endtype, bool reverse)
an element of a "constant" initializer. */
bool
-initializer_constant_valid_for_bitfield_p (tree value)
+initializer_constant_valid_for_bitfield_p (const_tree value)
{
/* For bitfields we support integer constants or possibly nested aggregates
of such. */
@@ -5078,7 +5078,7 @@ initializer_constant_valid_for_bitfield_p (tree value)
case CONSTRUCTOR:
{
unsigned HOST_WIDE_INT idx;
- tree elt;
+ const_tree elt;
FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (value), idx, elt)
if (!initializer_constant_valid_for_bitfield_p (elt))
diff --git a/gcc/varasm.h b/gcc/varasm.h
index d5d8c4e..fc6c753 100644
--- a/gcc/varasm.h
+++ b/gcc/varasm.h
@@ -65,7 +65,7 @@ extern tree initializer_constant_valid_p (tree, tree, bool = false);
/* Return true if VALUE is a valid constant-valued expression
for use in initializing a static bit-field; one that can be
an element of a "constant" initializer. */
-extern bool initializer_constant_valid_for_bitfield_p (tree);
+extern bool initializer_constant_valid_for_bitfield_p (const_tree);
/* Whether a constructor CTOR is a valid static constant initializer if all
its elements are. This used to be internal to initializer_constant_valid_p
diff --git a/gcc/testsuite/gcc.target/i386/pr95126-m32-1.c b/gcc/testsuite/gcc.target/i386/pr95126-m32-1.c
new file mode 100644
index 0000000..1d6acd6
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr95126-m32-1.c
@@ -0,0 +1,16 @@
+/* { dg-do compile { target { ia32 } } } */
+/* { dg-options "-O2" } */
+
+struct small{ short a,b; signed char c; };
+
+void call_func(void)
+{
+ extern int func(struct small X);
+ static struct small const s = { 1,2,0 };
+ func(s);
+}
+
+/* { dg-final { scan-assembler "movl\[ \\t]*\\\$" } } */
+/* { dg-final { scan-assembler "movb\[ \\t]*\\\$0, " } } */
+/* { dg-final { scan-assembler-not "movzwl" } } */
+
diff --git a/gcc/testsuite/gcc.target/i386/pr95126-m32-2.c b/gcc/testsuite/gcc.target/i386/pr95126-m32-2.c
new file mode 100644
index 0000000..b46be9d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr95126-m32-2.c
@@ -0,0 +1,16 @@
+/* { dg-do compile { target { ia32 } } } */
+/* { dg-options "-O2" } */
+
+struct small{ short a,b; signed char c; };
+static const struct small s = { 1,2,0 };
+extern int func(struct small X);
+
+void call_func(void)
+{
+ func(s);
+}
+
+/* { dg-final { scan-assembler "movl\[ \\t]*\\\$" } } */
+/* { dg-final { scan-assembler "movb\[ \\t]*\\\$0, " } } */
+/* { dg-final { scan-assembler-not "movzwl" } } */
+
diff --git a/gcc/testsuite/gcc.target/i386/pr95126-m32-3.c b/gcc/testsuite/gcc.target/i386/pr95126-m32-3.c
new file mode 100644
index 0000000..cc2fe94
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr95126-m32-3.c
@@ -0,0 +1,15 @@
+/* { dg-do compile { target { ia32 } } } */
+/* { dg-options "-O2" } */
+
+struct small{ short a; };
+
+void call_func(void)
+{
+ extern int func(struct small X);
+ static struct small const s = { 2 };
+ func(s);
+}
+
+/* { dg-final { scan-assembler "pushl\[ \\t]*\\\$2" } } */
+/* { dg-final { scan-assembler-not "movzwl" } } */
+
diff --git a/gcc/testsuite/gcc.target/i386/pr95126-m32-4.c b/gcc/testsuite/gcc.target/i386/pr95126-m32-4.c
new file mode 100644
index 0000000..e829335
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr95126-m32-4.c
@@ -0,0 +1,13 @@
+/* { dg-do compile { target { ia32 } } } */
+/* { dg-options "-O2" } */
+
+struct small{ short a,b; };
+
+void call_func(void)
+{
+ extern int func(struct small X);
+ static struct small const s = { 1,2 };
+ func(s);
+}
+
+/* { dg-final { scan-assembler "pushl\[ \\t]*\\\$131073" } } */
diff --git a/gcc/testsuite/gcc.target/i386/pr95126-m64-1.c b/gcc/testsuite/gcc.target/i386/pr95126-m64-1.c
new file mode 100644
index 0000000..d5c6dded
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr95126-m64-1.c
@@ -0,0 +1,17 @@
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2" } */
+
+struct small{ short a,b; signed char c; };
+
+void call_func(void)
+{
+ extern int func(struct small X);
+ static struct small const s = { 1,2,0 };
+ func(s);
+}
+
+/* { dg-final { scan-assembler "movl\[ \\t]*\\\$131073, " } } */
+/* { dg-final { scan-assembler-not "movzwl" } } */
+/* { dg-final { scan-assembler-not "salq" } } */
+/* { dg-final { scan-assembler-not "orq" } } */
+
diff --git a/gcc/testsuite/gcc.target/i386/pr95126-m64-2.c b/gcc/testsuite/gcc.target/i386/pr95126-m64-2.c
new file mode 100644
index 0000000..0230ffc
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr95126-m64-2.c
@@ -0,0 +1,17 @@
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2" } */
+
+struct small{ short a,b; signed char c; };
+static const struct small s = { 1,2,0 };
+extern int func(struct small X);
+
+void call_func(void)
+{
+ func(s);
+}
+
+/* { dg-final { scan-assembler "movl\[ \\t]*\\\$131073, " } } */
+/* { dg-final { scan-assembler-not "movzwl" } } */
+/* { dg-final { scan-assembler-not "salq" } } */
+/* { dg-final { scan-assembler-not "orq" } } */
+
diff --git a/gcc/testsuite/gcc.target/i386/pr95126-m64-3.c b/gcc/testsuite/gcc.target/i386/pr95126-m64-3.c
new file mode 100644
index 0000000..25afe3a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr95126-m64-3.c
@@ -0,0 +1,14 @@
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2" } */
+
+struct small{ short a; };
+
+void call_func(void)
+{
+ extern int func(struct small X);
+ static struct small const s = { 2 };
+ func(s);
+}
+
+/* { dg-final { scan-assembler "movl\[ \\t]*\\\$2, " } } */
+/* { dg-final { scan-assembler-not "movzwl" } } */
diff --git a/gcc/testsuite/gcc.target/i386/pr95126-m64-4.c b/gcc/testsuite/gcc.target/i386/pr95126-m64-4.c
new file mode 100644
index 0000000..71c7908
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr95126-m64-4.c
@@ -0,0 +1,13 @@
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2" } */
+
+struct small{ short a,b; };
+
+void call_func(void)
+{
+ extern int func(struct small X);
+ static struct small const s = { 1,2 };
+ func(s);
+}
+
+/* { dg-final { scan-assembler "movl\[ \\t]*\\\$131073, " } } */
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] PR middle-end/95126: Expand small const structs as immediate constants.
2022-02-26 21:35 [PATCH] PR middle-end/95126: Expand small const structs as immediate constants Roger Sayle
@ 2022-06-01 21:16 ` Jeff Law
0 siblings, 0 replies; 2+ messages in thread
From: Jeff Law @ 2022-06-01 21:16 UTC (permalink / raw)
To: gcc-patches
On 2/26/2022 2:35 PM, Roger Sayle wrote:
> This patch resolves PR middle-end/95126 which is a code quality regression,
> by teaching the RTL expander to emit small const structs/unions as integer
> immediate constants.
>
> The motivating example from the bugzilla PR is:
>
> struct small{ short a,b; signed char c; };
> extern int func(struct small X);
> void call_func(void)
> {
> static struct small const s = { 1, 2, 0 };
> func(s);
> }
>
> which on x86_64 is currently compiled to:
>
> call_func:
> movzwl s.0+2(%rip), %eax
> movzwl s.0(%rip), %edx
> movzwl s.0+4(%rip), %edi
> salq $16, %rax
> orq %rdx, %rax
> salq $32, %rdi
> orq %rax, %rdi
> jmp func
>
> but with this patch is now optimized to:
>
> call_func:
> movl $131073, %edi
> jmp func
>
>
> This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> and make -k check with no new failures. Ok for mainline?
>
>
> 2022-02-26 Roger Sayle <roger@nextmovesoftware.com>
>
> gcc/ChangeLog
> PR middle-end/95126
> * calls.cc (load_register_parameters): When loading a suitable
> immediate_const_ctor_p VAR_DECL into a single word_mode register,
> construct it directly in a pseudo rather than read it (by parts)
> from memory.
> * expr.cc (int_expr_size): Make tree argument a const_tree.
> (immediate_const_ctor_p): Helper predicate. Return true for
> simple constructors that may be materialized in a register.
> (expand_expr_real_1) [VAR_DECL]: When expanding a constant
> VAR_DECL with a suitable immediate_const_ctor_p constructor
> use store_constructor to materialize it directly in a pseudo.
> * expr.h (immediate_const_ctor_p): Prototype here.
> * varasm.cc (initializer_constant_valid_for_bitfield_p): Change
> VALUE argument from tree to const_tree.
> * varasm.h (initializer_constant_valid_for_bitfield_p): Update
> prototype.
>
> gcc/testsuite/ChangeLog
> PR middle-end/95126
> * gcc.target/i386/pr95126-m32-1.c: New test case.
> * gcc.target/i386/pr95126-m32-2.c: New test case.
> * gcc.target/i386/pr95126-m32-3.c: New test case.
> * gcc.target/i386/pr95126-m32-4.c: New test case.
> * gcc.target/i386/pr95126-m64-1.c: New test case.
> * gcc.target/i386/pr95126-m64-2.c: New test case.
> * gcc.target/i386/pr95126-m64-3.c: New test case.
> * gcc.target/i386/pr95126-m64-4.c: New test case.
OK after a fresh bootstrap & regression test. Sorry for the long wait.
jeff
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2022-06-01 21:16 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-02-26 21:35 [PATCH] PR middle-end/95126: Expand small const structs as immediate constants Roger Sayle
2022-06-01 21:16 ` Jeff Law
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).