[Bug middle-end/37364] New: [4.4 Regression] IRA generates ineffient code

["hjl dot tools at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

bash-3.2$ cat /tmp/x.c
#include <mmintrin.h>
extern __m64 x, y;
unsigned long long  foo(__m64 m) {
  return _mm_cvtm64_si64(_mm_add_pi32(x, y));
}
bash-3.2$ ./xgcc -B./ -march=core2 -S -O2 /tmp/x.c
bash-3.2$ cat x.s
        .file   "x.c"
        .text
        .p2align 4,,15
.globl foo
        .type   foo, @function
foo:
.LFB129:
        .cfi_startproc
        movq    x(%rip), %mm0
        paddd   y(%rip), %mm0
        movq    %mm0, -8(%rsp)
        movq    -8(%rsp), %rax
        ret
        .cfi_endproc
.LFE129:
        .size   foo, .-foo
        .ident  "GCC: (GNU) 4.4.0 20080903 (experimental) [trunk revision
139952]"
        .section        .note.GNU-stack,"",@progbits
bash-3.2$ ./xgcc -B./ -march=core2 -S -O2 /tmp/x.c -fno-ira
bash-3.2$ cat x.s
        .file   "x.c"
        .text
        .p2align 4,,15
.globl foo
        .type   foo, @function
foo:
.LFB129:
        .cfi_startproc
        movq    x(%rip), %mm0
        paddd   y(%rip), %mm0
        movd    %mm0, %rax
        ret
        .cfi_endproc
.LFE129:
        .size   foo, .-foo
        .ident  "GCC: (GNU) 4.4.0 20080903 (experimental) [trunk revision
139952]"
        .section        .note.GNU-stack,"",@progbits
bash-3.2$ 

The problem is IRA turns

(insn 8 7 14 2 ../../include/mmintrin.h:300 (set (reg:V2SI 61)
        (plus:V2SI (reg:V2SI 63 [ x ])
            (mem/c/i:V2SI (symbol_ref:DI ("y") <var_decl 0x7fd36e03cc80 y>) [2 
y+0 S8 A64]))) 992 {*mmx_addv2si3} (expr_list:REG_DEAD (reg:V2SI 63 [ x ])
        (nil)))

(insn 14 8 20 2 /tmp/x.c:12 (set (reg/i:DI 0 ax)
        (subreg:DI (reg:V2SI 61) 0)) 89 {*movdi_1_rex64} (expr_list:REG_DEAD
(re
g:V2SI 61)
        (nil)))

into

(insn 26 8 14 2 ../../include/mmintrin.h:300 (set (mem/c:V2SI (plus:DI
(reg/f:DI
 7 sp)
                (const_int -8 [0xfffffffffffffff8])) [3 S8 A64])
        (reg:V2SI 29 mm0 [orig:63 x ] [63])) 946 {*movv2si_internal_rex64}
(nil)
)

(insn:HI 14 26 20 2 /tmp/x.c:12 (set (reg/i:DI 0 ax)
        (mem/c:DI (plus:DI (reg/f:DI 7 sp)
                (const_int -8 [0xfffffffffffffff8])) [3 S8 A64])) 89
{*movdi_1_r
ex64} (nil))

while the old RA generates

(insn:HI 14 8 20 2 /tmp/x.c:12 (set (reg/i:DI 0 ax)
        (reg:DI 29 mm0 [orig:63 x ] [63])) 89 {*movdi_1_rex64} (nil))

The outputs from regmove pass are different. With IRA, we got

(insn:HI 8 7 14 2 ../../include/mmintrin.h:300 (set (reg:V2SI 61)
        (plus:V2SI (reg:V2SI 63 [ x ])
            (mem/c/i:V2SI (symbol_ref:DI ("y") <var_decl 0x7f66abfb5c80 y>) [2
y
+0 S8 A64]))) 992 {*mmx_addv2si3} (expr_list:REG_DEAD (reg:V2SI 63 [ x ])
        (nil)))

(insn:HI 14 8 20 2 /tmp/x.c:12 (set (reg/i:DI 0 ax)
        (subreg:DI (reg:V2SI 61) 0)) 89 {*movdi_1_rex64} (expr_list:REG_DEAD
(re
g:V2SI 61)
        (nil)))

Without IRA, we got
(insn:HI 8 7 14 2 ../../include/mmintrin.h:300 (set (reg:V2SI 63 [ x ])
        (plus:V2SI (reg:V2SI 63 [ x ])
            (mem/c/i:V2SI (symbol_ref:DI ("y") <var_decl 0x7fd36e03cc80 y>) [2 
y+0 S8 A64]))) 992 {*mmx_addv2si3} (nil))

(insn:HI 14 8 20 2 /tmp/x.c:12 (set (reg/i:DI 0 ax)
        (subreg:DI (reg:V2SI 63 [ x ]) 0)) 89 {*movdi_1_rex64} (expr_list:REG_D
EAD (reg:V2SI 63 [ x ])
        (nil)))

Does it have any impact on code generation?


-- 
           Summary: [4.4 Regression] IRA generates ineffient code
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: hjl dot tools at gmail dot com
GCC target triplet: x86_64-unknown-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37364