[Bug regression/43616] New: Extra register move

["carrot at google dot com" <gcc-bugzilla at gcc dot gnu dot org>]

Compile the following code with options -Os -march=armv7-a -mthumb

extern long long foo();
void bar2(long long* p)
{
  long long t = foo();
  *p = t;
}

GCC generates:

bar2:
        push    {r4, lr}
        mov     r4, r0
        bl      foo
        mov     r2, r0     // A
        mov     r3, r1     // B
        strd    r2, [r4]   // C
        pop     {r4, pc}

1. The register moves in instructions AB is a regression relate to gcc4.4.0.
The result of gcc4.4 is:

        push    {r4, lr}
        mov     r4, r0
        bl      foo
        strd    r0, [r4]
        pop     {r4, pc}

The regression may be caused by some changes in ira. Before ira both versions
have similar rtx sequence:

(call_insn 6 3 7 2 src/tb.c:4 (parallel [
            (set (reg:DI 0 r0)
                (call (mem:SI (symbol_ref:SI ("foo") [flags 0x41] 
<function_decl 0x7f623ba6a600 foo>) [0 S4 A32])
                    (const_int 0 [0x0])))
            (use (const_int 0 [0x0]))
            (clobber (reg:SI 14 lr))
        ]) 255 {*call_value_insn} (nil)
    (nil))

(insn 7 6 8 2 src/tb.c:4 (set (reg/v:DI 133 [ t ])
        (reg:DI 0 r0)) 657 {*thumb2_movdi} (nil))

(insn 8 7 0 2 src/tb.c:5 (set (mem:DI (reg/v/f:SI 134 [ p ]) [2 S8 A64])
        (reg/v:DI 133 [ t ])) 657 {*thumb2_movdi} (expr_list:REG_DEAD
(reg/v/f:SI 134 [ p ])
        (expr_list:REG_DEAD (reg/v:DI 133 [ t ])
            (nil))))

After ira, gcc4.5 generates:

(call_insn 6 3 7 2 src/tb.c:4 (parallel [
            (set (reg:DI 0 r0)
                (call (mem:SI (symbol_ref:SI ("foo") [flags 0x41] 
<function_decl 0x7f623ba6a600 foo>) [0 S4 A32])
                    (const_int 0 [0x0])))
            (use (const_int 0 [0x0]))
            (clobber (reg:SI 14 lr))
        ]) 255 {*call_value_insn} (nil)
    (nil))

(insn 7 6 8 2 src/tb.c:4 (set (reg/v:DI 2 r2 [orig:133 t ] [133])
        (reg:DI 0 r0)) 657 {*thumb2_movdi} (expr_list:REG_EQUIV (mem:DI
(reg/v/f:SI 4 r4 [orig:134 p ] [134]) [2 S8 A64])
        (nil)))

(insn 8 7 11 2 src/tb.c:5 (set (mem:DI (reg/v/f:SI 4 r4 [orig:134 p ] [134]) [2
S8 A64])
        (reg/v:DI 2 r2 [orig:133 t ] [133])) 657 {*thumb2_movdi} (nil))

But gcc4.4 generates:

(call_insn 6 3 8 2 src/tb.c:4 (parallel [
            (set (reg:DI 0 r0)
                (call (mem:SI (symbol_ref:SI ("foo") [flags 0x41]
<function_decl 0xf7d1c880 foo>) [0 S4 A32])
                    (const_int 0 [0x0])))
            (use (const_int 0 [0x0]))
            (clobber (reg:SI 14 lr))
        ]) 257 {*call_value_insn} (nil)
    (nil))

(insn 8 6 16 2 src/tb.c:5 (set (mem:DI (reg/v/f:SI 4 r4 [orig:134 p ] [134]) [2
S8 A64])
        (reg/v:DI 0 r0 [orig:133 t ] [133])) 651 {*thumb2_movdi} (nil))

2. Since r4 is never used again after instruction C, it can also be written as
a stm instruction. In thumb2, strd is a 32bit instruction, but stm is 16 bit.


-- 
           Summary: Extra register move
           Product: gcc
           Version: 4.5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: regression
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: carrot at google dot com
 GCC build triplet: i686-linux
  GCC host triplet: i686-linux
GCC target triplet: arm-eabi


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43616