https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80799
--- Comment #2 from Uroš Bizjak <ubizjak at gmail dot com> ---
There is something going on in the middle end, the problematic alternatives of
the relevant patterns:
movq (%ebx), %mm0 # 6 *movv2si_internal/8 [length = 3]
movq2dq %mm0, %xmm1 # 13 vec_concatv2di/7 [length = 4]
pcmpeqd %xmm1, %xmm0 # 17 *sse2_eqv4si3/1 [length = 4]
pmovmskb %xmm0, %eax # 18 sse2_pmovmskb [length = 4]
cmpl $65535, %eax # 20 *cmpsi_1/1 [length = 5]
je .L2 # 21 *jcc_1 [length = 2]
call _Z10dummy_callv # 23 *call [length = 5]
movq (%ebx), %mm0 # 24 *movv2si_internal/8 [length = 3]
.L2:
movq %mm0, TRXPOS # 29 *movv2si_internal/9 [length = 8]
are all decorated with "!".
(define_insn "*mov<mode>_internal"
[(set (match_operand:MMXMODE 0 "nonimmediate_operand"
"=r ,o ,r,r ,m ,?!y,!y,?!y,m ,r ,?!Ym,v,v,v,m,*x,*x,*x,m ,r
,Yi,!Ym,*Yi")
(match_operand:MMXMODE 1 "vector_move_operand"
"rCo,rC,C,rm,rC,C ,!y,m ,?!y,?!Yn,r ,C,v,m,v,C ,*x,m ,*x,Yj,r
,*Yj,!Yn"))]
(define_insn "vec_concatv2di"
[(set (match_operand:V2DI 0 "register_operand"
"=Yr,*x,x ,v ,Yi,v ,!x,x,v ,x,x,v")
(vec_concat:V2DI
(match_operand:DI 1 "nonimmediate_operand"
" 0, 0,x ,Yv,r ,vm,*y,0,Yv,0,0,v")
(match_operand:DI 2 "vector_move_operand"
"*rm,rm,rm,rm,C ,C ,C ,x,Yv,x,m,m")))]
I'm afraid this is all target can do to avoid (but still allow) certain
alternative. The testcase works for gcc-6, and those patterns were not touched
for a long time.
Can someone please do a reghunt here?