[PATCH] PR middle-end/105853: Call store_constructor directly from calls.cc.

[PATCH] PR middle-end/105853: Call store_constructor directly from calls.cc.

[Roger Sayle]

[-- Attachment #1: Type: text/plain, Size: 3222 bytes --]


This patch fixes both ICE regressions PR middle-end/105853 and
PR target/105856 caused by my recent patch to expand small const structs
as immediate constants.  That patch updated code generation in three
places: two in expr.cc that call store_constructor directly, and the
third in calls.cc's load_register_parameters that expands its CONSTRUCTOR
via expand_expr, as store_constructor is local/static to expr.cc, and
the "public" API, should usually simply forward the constructor to the
appropriate store_constructor function.

Alas, despite the clean regression testing on multiple targets, the above
ICEs show that expand_expr isn't a suitable proxy for store_constructor,
and things that (I'd assumed) shouldn't affect how/whether a struct is
placed in a register [such as whether the struct is considered packed/
aligned or not] actually interfere with the optimization that is being
attempted.

The (proposed) solution is to export store_constructor (and it's helper
function int_expr_size) from expr.cc, by removing their static qualifier
and prototyping both functions in expr.h, so they can be called directly
from load_register_parameters in calls.cc.  This cures both ICEs, but
almost as important produces much better code generation than GCC 12.

For PR 105853, GCC 12 generates:

compose_nd_na_ipv6_src:
        movzx eax, WORD PTR eth_addr_zero[rip+2]
        movzx edx, WORD PTR eth_addr_zero[rip]
        movzx edi, WORD PTR eth_addr_zero[rip+4]
        sal rax, 16
        or rax, rdx
        sal rdi, 32
        or rdi, rax
        xor eax, eax
        jmp packet_set_nd
eth_addr_zero:  .zero 6

where now (with this fix) GCC 13 generates:
compose_nd_na_ipv6_src:
        xorl    %edi, %edi
        xorl    %eax, %eax
        jmp     packet_set_nd

Likewise, for PR 105856 on ARM, we'd previously generate:
g_329_3:
        movw r3, #:lower16:.LANCHOR0
        movt r3, #:upper16:.LANCHOR0
        ldr r0, [r3]
        b func_19

but with this optimization we now generate:
g_329_3:
        mov     r0, #6
        b       func_19


This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check with no new failures.  I've also confirmed that on a
cross-compiler to arm-linux-gnueabihf --with-arch=armv6 this fixes the
target specific ICE in PR105856.  The make check is currently running
with --target_board=unix{-m32}, OK for mainline if that also passes?

My sincere apologies for the inconvenience.


2022-06-06  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
        PR middle-end/105853
        PR target/105856
        * calls.cc (load_register_parameters): Call store_constructor
        (and int_Expr_size) directly instead of expanding via expand_expr.
        * expr.cc (static void store_constructor): Don't prototype here.
        (static HOST_WIDE_INT int_expr_size): Likewise.
        (store_constructor): No longer static.
        (int_expr_size): Likewise, no longer static.
        * expr.h (store_constructor): Prototype here.
        (int_expr_size): Prototype here.

gcc/testsuite/ChangeLog
        PR middle-end/105853
        PR target/105856
        * gcc.dg/pr105853.c: New test case.
        * gcc.dg/pr105856.c: New test case.


Roger
--


[-- Attachment #2: patchar.txt --]
[-- Type: text/plain, Size: 3656 bytes --]

diff --git a/gcc/calls.cc b/gcc/calls.cc
index a4336c1..f4e1299 100644
--- a/gcc/calls.cc
+++ b/gcc/calls.cc
@@ -2186,10 +2186,11 @@ load_register_parameters (struct arg_data *args, int num_actuals,
 		   && 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);
+	      store_constructor (DECL_INITIAL (tree_value), target, 0,
+				 int_expr_size (DECL_INITIAL (tree_value)),
+				 false);
 	      reg = gen_rtx_REG (word_mode, REGNO (reg));
-	      emit_move_insn (reg, x);
+	      emit_move_insn (reg, target);
 	    }
 	  else if (partial == 0 || args[i].pass_on_stack)
 	    {
diff --git a/gcc/expr.cc b/gcc/expr.cc
index fb062dc..85cb414 100644
--- a/gcc/expr.cc
+++ b/gcc/expr.cc
@@ -84,7 +84,6 @@ static void emit_block_move_via_loop (rtx, rtx, rtx, unsigned);
 static void clear_by_pieces (rtx, unsigned HOST_WIDE_INT, unsigned int);
 static rtx_insn *compress_float_constant (rtx, rtx);
 static rtx get_subtarget (rtx);
-static void store_constructor (tree, rtx, int, poly_int64, bool);
 static rtx store_field (rtx, poly_int64, poly_int64, poly_uint64, poly_uint64,
 			machine_mode, tree, alias_set_type, bool, bool);
 
@@ -100,7 +99,6 @@ 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 (const_tree);
 static void convert_mode_scalar (rtx, rtx, int);
 
 \f
@@ -6757,7 +6755,7 @@ fields_length (const_tree type)
    which has been packed to exclude padding bits.
    If REVERSE is true, the store is to be done in reverse order.  */
 
-static void
+void
 store_constructor (tree exp, rtx target, int cleared, poly_int64 size,
 		   bool reverse)
 {
@@ -13209,7 +13207,7 @@ expr_size (tree exp)
 /* Return a wide integer for the size in bytes of the value of EXP, or -1
    if the size can vary or is larger than an integer.  */
 
-static HOST_WIDE_INT
+HOST_WIDE_INT
 int_expr_size (const_tree exp)
 {
   tree size;
diff --git a/gcc/expr.h b/gcc/expr.h
index d777c28..0351183 100644
--- a/gcc/expr.h
+++ b/gcc/expr.h
@@ -339,6 +339,8 @@ 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 store_constructor (tree, rtx, int, poly_int64, bool);
+extern HOST_WIDE_INT int_expr_size (const_tree exp);
 
 extern void expand_operands (tree, tree, rtx, rtx*, rtx*,
 			     enum expand_modifier);
diff --git a/gcc/testsuite/gcc.dg/pr105853.c b/gcc/testsuite/gcc.dg/pr105853.c
new file mode 100644
index 0000000..4f234ac
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr105853.c
@@ -0,0 +1,11 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+struct {
+  struct {
+    short e16[3];
+  }
+} const eth_addr_zero = {{}};  /* { dg-warning "no semicolon at" } */
+void compose_nd_na_ipv6_src() {
+  packet_set_nd(eth_addr_zero); /* { dg-warning "implicit declaration" } */
+}
diff --git a/gcc/testsuite/gcc.dg/pr105856.c b/gcc/testsuite/gcc.dg/pr105856.c
new file mode 100644
index 0000000..dd3aa2f
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr105856.c
@@ -0,0 +1,10 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+#pragma pack(1)
+struct {
+  unsigned f0;
+} static g_251 = {6};
+void g_329_3() {
+  func_19(g_251);  /* { dg-warning "implicit declaration" } */
+}
+

Re: [PATCH] PR middle-end/105853: Call store_constructor directly from calls.cc.

[Eric Botcazou]

> The (proposed) solution is to export store_constructor (and it's helper
> function int_expr_size) from expr.cc, by removing their static qualifier
> and prototyping both functions in expr.h, so they can be called directly
> from load_register_parameters in calls.cc.  This cures both ICEs, but
> almost as important produces much better code generation than GCC 12.
> 
> For PR 105853, GCC 12 generates:
> 
> compose_nd_na_ipv6_src:
>         movzx eax, WORD PTR eth_addr_zero[rip+2]
>         movzx edx, WORD PTR eth_addr_zero[rip]
>         movzx edi, WORD PTR eth_addr_zero[rip+4]
>         sal rax, 16
>         or rax, rdx
>         sal rdi, 32
>         or rdi, rax
>         xor eax, eax
>         jmp packet_set_nd
> eth_addr_zero:  .zero 6
> 
> where now (with this fix) GCC 13 generates:
> compose_nd_na_ipv6_src:
>         xorl    %edi, %edi
>         xorl    %eax, %eax
>         jmp     packet_set_nd
> 
> Likewise, for PR 105856 on ARM, we'd previously generate:
> g_329_3:
>         movw r3, #:lower16:.LANCHOR0
>         movt r3, #:upper16:.LANCHOR0
>         ldr r0, [r3]
>         b func_19
> 
> but with this optimization we now generate:
> g_329_3:
>         mov     r0, #6
>         b       func_19

Nice work!  Fiddling with RTL expansion like this is never easy...

> This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> and make -k check with no new failures.  I've also confirmed that on a
> cross-compiler to arm-linux-gnueabihf --with-arch=armv6 this fixes the
> target specific ICE in PR105856.  The make check is currently running
> with --target_board=unix{-m32}, OK for mainline if that also passes?

Yes, it looks quite uncontroversial if you ask me and will unblock people.

-- 
Eric Botcazou

Re: [PATCH] PR middle-end/105853: Call store_constructor directly from calls.cc.

[Jeff Law]

On 6/6/2022 7:50 AM, Roger Sayle wrote:
> This patch fixes both ICE regressions PR middle-end/105853 and
> PR target/105856 caused by my recent patch to expand small const structs
> as immediate constants.  That patch updated code generation in three
> places: two in expr.cc that call store_constructor directly, and the
> third in calls.cc's load_register_parameters that expands its CONSTRUCTOR
> via expand_expr, as store_constructor is local/static to expr.cc, and
> the "public" API, should usually simply forward the constructor to the
> appropriate store_constructor function.
>
> Alas, despite the clean regression testing on multiple targets, the above
> ICEs show that expand_expr isn't a suitable proxy for store_constructor,
> and things that (I'd assumed) shouldn't affect how/whether a struct is
> placed in a register [such as whether the struct is considered packed/
> aligned or not] actually interfere with the optimization that is being
> attempted.
>
> The (proposed) solution is to export store_constructor (and it's helper
> function int_expr_size) from expr.cc, by removing their static qualifier
> and prototyping both functions in expr.h, so they can be called directly
> from load_register_parameters in calls.cc.  This cures both ICEs, but
> almost as important produces much better code generation than GCC 12.
>
> For PR 105853, GCC 12 generates:
>
> compose_nd_na_ipv6_src:
>          movzx eax, WORD PTR eth_addr_zero[rip+2]
>          movzx edx, WORD PTR eth_addr_zero[rip]
>          movzx edi, WORD PTR eth_addr_zero[rip+4]
>          sal rax, 16
>          or rax, rdx
>          sal rdi, 32
>          or rdi, rax
>          xor eax, eax
>          jmp packet_set_nd
> eth_addr_zero:  .zero 6
>
> where now (with this fix) GCC 13 generates:
> compose_nd_na_ipv6_src:
>          xorl    %edi, %edi
>          xorl    %eax, %eax
>          jmp     packet_set_nd
>
> Likewise, for PR 105856 on ARM, we'd previously generate:
> g_329_3:
>          movw r3, #:lower16:.LANCHOR0
>          movt r3, #:upper16:.LANCHOR0
>          ldr r0, [r3]
>          b func_19
>
> but with this optimization we now generate:
> g_329_3:
>          mov     r0, #6
>          b       func_19
>
>
> This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> and make -k check with no new failures.  I've also confirmed that on a
> cross-compiler to arm-linux-gnueabihf --with-arch=armv6 this fixes the
> target specific ICE in PR105856.  The make check is currently running
> with --target_board=unix{-m32}, OK for mainline if that also passes?
>
> My sincere apologies for the inconvenience.
>
>
> 2022-06-06  Roger Sayle  <roger@nextmovesoftware.com>
>
> gcc/ChangeLog
>          PR middle-end/105853
>          PR target/105856
>          * calls.cc (load_register_parameters): Call store_constructor
>          (and int_Expr_size) directly instead of expanding via expand_expr.
>          * expr.cc (static void store_constructor): Don't prototype here.
>          (static HOST_WIDE_INT int_expr_size): Likewise.
>          (store_constructor): No longer static.
>          (int_expr_size): Likewise, no longer static.
>          * expr.h (store_constructor): Prototype here.
>          (int_expr_size): Prototype here.
>
> gcc/testsuite/ChangeLog
>          PR middle-end/105853
>          PR target/105856
>          * gcc.dg/pr105853.c: New test case.
>          * gcc.dg/pr105856.c: New test case.
OK
jeff