https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87545
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|ASSIGNED |NEW
Assignee|rguenth at gcc dot gnu.org |unassigned at gcc dot
gnu.org
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
note: === vect_compute_single_scalar_iteration_cost ===
0x341e830 v1[i_11] 1 times scalar_load costs 13 in body
0x341e830 v2[i_11] 1 times scalar_load costs 13 in body
0x341e830 MAX_EXPR <_1, _2> 1 times scalar_stmt costs 6 in body
0x341e830 _3 1 times scalar_store costs 20 in body
0x3409090 v1[i_11] 1 times vector_load costs 20 in body
0x3409090 v2[i_11] 1 times vector_load costs 20 in body
0x3409090 MAX_EXPR <_1, _2> 1 times vector_stmt costs 54 in body
0x3409090 _3 1 times unaligned_store costs 20 in body
ix86_add_stmt_cost costs MAX_EXPR as
if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
stmt_cost = ix86_cost->sse_op;
else if (VECTOR_MODE_P (mode))
stmt_cost = ix86_vec_cost (mode, ix86_cost->sse_op, true);
else
stmt_cost = ix86_cost->add;
where
COSTS_N_INSNS (1), /* cost of an add instruction */
COSTS_N_INSNS (8), /* cost of cheap SSE instruction. */
I see that most cost models have COSTS_N_INSNS (1) for a cheap SSE instruction
which generally is used for most integer vector ops.
core_cost has COSTS_N_INSNS (1) for cheap SSE instruction.
nocona_cost has 2, so do pentium4_cost, bdver[1234]_cost, amdfam10_cost,
k8_cost, athlon_cost. All other costs (besides size_cost) have 1.
So why the heck does intel_cost have 8!?
I didn't look at other costs besides cost of cheap SSE instruction but I wonder
how intel_cost was derived? I expected it to be a blend of other existing
intel CPU cost tables, not a completely unrelated outlier?
Not my bug but a pre-existing target one I happened to uncover.
Suggested "fix" (but the other vector costs appear high compared to other
intel costs as well).
diff --git a/gcc/config/i386/x86-tune-costs.h
b/gcc/config/i386/x86-tune-costs.h
index 71a5854c09a..c7f3945d72c 100644
--- a/gcc/config/i386/x86-tune-costs.h
+++ b/gcc/config/i386/x86-tune-costs.h
@@ -2294,7 +2294,7 @@ struct processor_costs intel_cost = {
COSTS_N_INSNS (8), /* cost of FCHS instruction. */
COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
- COSTS_N_INSNS (8), /* cost of cheap SSE instruction. */
+ COSTS_N_INSNS (1), /* cost of cheap SSE instruction. */
COSTS_N_INSNS (8), /* cost of ADDSS/SD SUBSS/SD insns. */
COSTS_N_INSNS (8), /* cost of MULSS instruction. */
COSTS_N_INSNS (8), /* cost of MULSD instruction. */
HJ, can you make Intel people double-check intel_cost and fix it up?
