https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67124
--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Richard Biener from comment #2)
> Hmm, looks like a target issue. We expand from
Not this time. It looks to me that generic code is at fault, performing DImode
move from %rdx to %xmm0, instead of inserting %rdx into %xmm0.
This can be seen with -O2 -msse2 -mtune=corei7 (please note that arrow points
to an instruction *before* it is executed):
0x0000000000400440 <+0>: movzbl 0x200c4b(%rip),%eax
0x0000000000400447 <+7>: movdqa 0x200c41(%rip),%xmm0
0x000000000040044f <+15>: movq %xmm0,%rdx
=> 0x0000000000400454 <+20>: mov %rax,%rcx
0x0000000000400457 <+23>: and $0xffffffffff00ffff,%rdx
0x000000000040045e <+30>: shl $0x10,%rax
0x0000000000400462 <+34>: or %rax,%rdx
0x0000000000400465 <+37>: xor %eax,%eax
0x0000000000400467 <+39>: movq %rdx,%xmm0
0x000000000040046c <+44>: pinsrw $0x2,%eax,%xmm0
0x0000000000400471 <+49>: mov 0x200be8(%rip),%rax
0x0000000000400478 <+56>: movups %xmm0,(%rax)
(gdb) i r rax rdx xmm0
rax 0x1
rdx 0x10000
xmm0 {uint128 = 0x00000000000000010000000000010000}
0x0000000000400465 <+37>: xor %eax,%eax
=> 0x0000000000400467 <+39>: movq %rdx,%xmm0
0x000000000040046c <+44>: pinsrw $0x2,%eax,%xmm0
(gdb) i r rdx xmm0
rdx 0x10000
xmm0 {uint128 = 0x00000000000000010000000000010000}
0x0000000000400465 <+37>: xor %eax,%eax
0x0000000000400467 <+39>: movq %rdx,%xmm0
=> 0x000000000040046c <+44>: pinsrw $0x2,%eax,%xmm0
(gdb) i r rdx xmm0
rdx 0x10000
xmm0 {uint128 = 0x00000000000000000000000000010000}
The same code can be compiled with -msse4 instead of -msse2, so we will have
pinsrb insn available:
0x0000000000400440 <+0>: movzbl 0x200c4b(%rip),%eax
0x0000000000400447 <+7>: xor %edx,%edx
0x0000000000400449 <+9>: movdqa 0x200c3f(%rip),%xmm0
=> 0x0000000000400451 <+17>: pinsrb $0x2,%eax,%xmm0
0x0000000000400457 <+23>: pinsrw $0x2,%edx,%xmm0
0x000000000040045c <+28>: mov 0x200bfd(%rip),%rdx # 0x601060
<b>
0x0000000000400463 <+35>: movups %xmm0,(%rdx)
(gdb) i r rax xmm0
rax 0x1
xmm0 {uint128 = 0x00000000000000010000000000010000}
0x0000000000400449 <+9>: movdqa 0x200c3f(%rip),%xmm0
0x0000000000400451 <+17>: pinsrb $0x2,%eax,%xmm0
=> 0x0000000000400457 <+23>: pinsrw $0x2,%edx,%xmm0
i r xmm0
xmm0 {uint128 = 0x00000000000000010000000000010000}
As shown in the first case, the value in %xmm0 is already wrong before pinsrw
is executed.
