https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84201
Martin Liška <marxin at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|ASSIGNED |NEW
Assignee|marxin at gcc dot gnu.org |unassigned at gcc dot
gnu.org
--- Comment #10 from Martin Liška <marxin at gcc dot gnu.org> ---
Using -fdbg-cnt option:
-fdbg-cnt=vect_loop:0 -fdbg-cnt=vect_loop:4:5:power.fppized
I was able to track that to a single vectorization that happens here:
308 !! Set up frequency vector(s)
309 tmppower => first_power
310 ipower = 1
311 do while(associated(tmppower))
312 if ( tmppower%nofreq>1 ) then
313 freqstep = (tmppower%freqlast-tmppower%freqfirst)
&
314 / real(tmppower%nofreq - 1, kind=rfp)
315 else
316 freqstep = 0.0_rfp
317 end if
318 freq = tmppower%freqfirst
319 do ifreq = 1, tmppower%nofreq <------ HERE
320 frequency(ifreq,ipower) = freq
321 freq = freq + freqstep
322 end do
323 tmppower => tmppower%next
324 ipower = ipower + 1
325 end do
vect dump:
...
power.fppized.f90:319:0: note: Runtime profitability threshold = 4
power.fppized.f90:319:0: note: Static estimate profitability threshold = 9
power.fppized.f90:319:0: note: epilog loop required
power.fppized.f90:319:0: note: vect_can_advance_ivs_p:
power.fppized.f90:319:0: note: Analyze phi: freq_20 = PHI <pretmp_1948(136),
freq_728(180)>
power.fppized.f90:319:0: note: Analyze phi: ifreq_1653 = PHI <1(136),
ifreq_729(180)>
power.fppized.f90:319:0: note: Analyze phi: .MEM_137 = PHI <.MEM_134(136),
.MEM_727(180)>
power.fppized.f90:319:0: note: reduc or virtual phi. skip.
***dbgcnt: upper limit 5 reached for vect_loop.***
power.fppized.f90:319:0: optimized: loop vectorized using 32 byte vectors
power.fppized.f90:319:0: note: === vec_transform_loop ===
power.fppized.f90:319:0: note: Profitability threshold is 4 loop iterations.
...
which leads to following assembly:
.L248:
movl 88(%rdi), %esi
vmovsd 96(%rdi), %xmm2
cmpl $1, %esi
jle .L243
vmovsd 104(%rdi), %xmm1
leal -1(%rsi), %eax
vcvtsi2sdl %eax, %xmm5, %xmm3
vsubsd %xmm2, %xmm1, %xmm1
testl %esi, %esi
movl %r10d, %r9d
vdivsd %xmm3, %xmm1, %xmm3
cmovg %esi, %r9d
cmpl $3, %esi
jle .L371
vaddsd %xmm2, %xmm3, %xmm0
movl %r9d, %ecx
shrl $2, %ecx
vaddsd %xmm3, %xmm0, %xmm1
salq $5, %rcx
vunpcklpd %xmm0, %xmm2, %xmm0
vaddsd %xmm3, %xmm1, %xmm4
addq %r8, %rcx
movq %r8, %rax
vunpcklpd %xmm4, %xmm1, %xmm1
vmulsd %xmm6, %xmm3, %xmm4
vinsertf128 $0x1, %xmm1, %ymm0, %ymm0
vbroadcastsd %xmm4, %ymm4
.p2align 4,,10
.p2align 3
.L250:
vmovapd %ymm0, %ymm1
vmovupd %ymm1, (%rax)
addq $32, %rax
vaddpd %ymm4, %ymm0, %ymm0
cmpq %rax, %rcx
jne .L250
movl %r9d, %eax
andl $-4, %eax
vcvtsi2sdl %eax, %xmm5, %xmm0
leal 1(%rax), %ecx
vfmadd231sd %xmm3, %xmm0, %xmm2
cmpl %r9d, %eax
je .L251
movslq %ecx, %rcx
addq %rdx, %rcx
addl $2, %eax
vmovsd %xmm2, (%r11,%rcx,8)
vaddsd %xmm3, %xmm2, %xmm2
cmpl %esi, %eax
jg .L251
.L270:
movslq %eax, %rcx
addq %rdx, %rcx
incl %eax
vmovsd %xmm2, (%r11,%rcx,8)
vaddsd %xmm3, %xmm2, %xmm2
cmpl %eax, %esi
jl .L251
cltq
addq %rdx, %rax
vmovsd %xmm2, (%r11,%rax,8)
.L251:
movq 136(%rdi), %rdi
addq %r15, %rdx
addq %r12, %r8
testq %rdi, %rdi
jne .L248
So there's probably nothing we can do right now? I'm removing wrong-code as
it's about floating point precision
when using -Ofast and vectors, which is a known limitation if I'm correct.
