https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69873
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |ASSIGNED
Last reconfirmed| |2016-02-19
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 37740
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=37740&action=edit
patch
Patch fixing this. Also fixing some incosistencies in cost estimation which
unfortunately "improves" the threshold to > 6 iterations.
Issues with the cost model include the odd separation between taken vs.
not-taken
branch cost (as opposed to fallthru, not fallthru or well predicted vs. not
predicted). This artificially raises the cost of the vectorized path (on
x86_64
generic model, vectorized path has a "taken" branch of cost 3 while
unvectorized
has an "not taken" branch of cost 1).
Another issue is the prologue cost for emitting the vector double constant
{ 1., 1. }. The scalar version also has a constant pool entry but we assume
none exist for the scalar path - reasonable only if insns with immediate forms
are available which IMHO is not reasonable to assume for FP modes. This skews
the cost by one.
If you disable cunroll then generated code is
cmpl $6, %edi
jbe .L3
subl $2, %edi
movapd .LC0(%rip), %xmm0
shrl %edi
xorl %eax, %eax
leal 1(%rdi), %ecx
.L4:
movq %rax, %rdx
addq $1, %rax
salq $4, %rdx
cmpl %eax, %ecx
movaps %xmm0, a(%rdx)
ja .L4
rep ret
.L3:
movsd .LC1(%rip), %xmm0
xorl %eax, %eax
.L6:
movsd %xmm0, a(,%rax,8)
addq $1, %rax
cmpl %eax, %edi
ja .L6
.L1:
rep ret
so the stupid IV choice makes the runtime check reasonable.
