https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89114
Bug ID: 89114
Summary: rtx_cost of VEC_SELECT, VEC_CONCAT and VEC_DUPLICATE
with memory operands is wrong
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
Split out from PR89049. On its testcase combine is willing to elide an
unnecessary %ymm build-up but the targets RTX cost makes that not profitable.
See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89049#c5
So with (the bogus)
Index: gcc/config/i386/i386.c
===================================================================
--- gcc/config/i386/i386.c (revision 268383)
+++ gcc/config/i386/i386.c (working copy)
@@ -40848,7 +40848,7 @@ ix86_rtx_costs (rtx x, machine_mode mode
recognizable. In which case they all pretty much have the
same cost. */
*total = cost->sse_op;
- return true;
+ return false;
case VEC_MERGE:
mask = XEXP (x, 2);
/* This is masked instruction, assume the same cost,
we get combine to do
Trying 11 -> 25:
11: r105:V8SF=vec_concat(r106:V4SF,[r85:DI+0x10])
25: r111:V4SF=vec_select(r105:V8SF,parallel)
REG_DEAD r105:V8SF
Successfully matched this instruction:
(set (reg:V4SF 111)
(mem:V4SF (plus:DI (reg:DI 85 [ ivtmp.11 ])
(const_int 16 [0x10])) [1 MEM[base: _2, offset: 0B]+16 S16 A32]))
allowing combination of insns 11 and 25
original costs 16 + 12 = 28
replacement cost 12
and we elide the %ymm build:
.L2:
vmovups (%rdi), %xmm1
addq $32, %rdi
vaddss %xmm1, %xmm0, %xmm0
vshufps $85, %xmm1, %xmm1, %xmm2
vaddss %xmm2, %xmm0, %xmm0
vunpckhps %xmm1, %xmm1, %xmm2
vshufps $255, %xmm1, %xmm1, %xmm1
vaddss %xmm2, %xmm0, %xmm0
vaddss %xmm1, %xmm0, %xmm0
vmovups -16(%rdi), %xmm1
vshufps $85, %xmm1, %xmm1, %xmm2
vaddss %xmm1, %xmm0, %xmm0
vaddss %xmm2, %xmm0, %xmm0
vunpckhps %xmm1, %xmm1, %xmm2
vshufps $255, %xmm1, %xmm1, %xmm1
vaddss %xmm2, %xmm0, %xmm0
vaddss %xmm1, %xmm0, %xmm0
cmpq %rdi, %rax
jne .L2
the patch is bogus because the intention of not scanning sub-rtxen was
to match the various shuffle patterns which do sth like
(vec_select (vec_concat ..) ...).
Not sure if there's a helper in i386.c to extract/cost a single MEM
sub-rtx, but the course of action would be to properly do this
somehow.
