Re: [PATCH v2 10/14] perf arm-spe: Refactor event type handling

From: Leo Yan
Date: Wed Oct 21 2020 - 00:55:04 EST


On Tue, Oct 20, 2020 at 10:54:16PM +0100, André Przywara wrote:
> On 29/09/2020 14:39, Leo Yan wrote:
>
> Hi,
>
> > Use macros instead of the enum values for event types, this is more
> > directive and without bit shifting when parse packet.
> >
> > Signed-off-by: Leo Yan <leo.yan@xxxxxxxxxx>
> > ---
> > .../util/arm-spe-decoder/arm-spe-decoder.c | 16 +++++++-------
> > .../util/arm-spe-decoder/arm-spe-decoder.h | 17 --------------
> > .../arm-spe-decoder/arm-spe-pkt-decoder.c | 22 +++++++++----------
> > .../arm-spe-decoder/arm-spe-pkt-decoder.h | 16 ++++++++++++++
> > 4 files changed, 35 insertions(+), 36 deletions(-)
> >
> > diff --git a/tools/perf/util/arm-spe-decoder/arm-spe-decoder.c b/tools/perf/util/arm-spe-decoder/arm-spe-decoder.c
> > index 9d3de163d47c..ac66e7f42a58 100644
> > --- a/tools/perf/util/arm-spe-decoder/arm-spe-decoder.c
> > +++ b/tools/perf/util/arm-spe-decoder/arm-spe-decoder.c
> > @@ -168,31 +168,31 @@ static int arm_spe_read_record(struct arm_spe_decoder *decoder)
> > case ARM_SPE_OP_TYPE:
> > break;
> > case ARM_SPE_EVENTS:
> > - if (payload & BIT(EV_L1D_REFILL))
> > + if (payload & SPE_EVT_PKT_L1D_REFILL)
>
> Not sure this (and the others below) are an improvement? I liked the
> enum below, and reading BIT() here tells me that it's a bitmask.

Agreed.

> > decoder->record.type |= ARM_SPE_L1D_MISS;
> >
> > - if (payload & BIT(EV_L1D_ACCESS))
> > + if (payload & SPE_EVT_PKT_L1D_ACCESS)
> > decoder->record.type |= ARM_SPE_L1D_ACCESS;
> >
> > - if (payload & BIT(EV_TLB_WALK))
> > + if (payload & SPE_EVT_PKT_TLB_WALK)
> > decoder->record.type |= ARM_SPE_TLB_MISS;
> >
> > - if (payload & BIT(EV_TLB_ACCESS))
> > + if (payload & SPE_EVT_PKT_TLB_ACCESS)
> > decoder->record.type |= ARM_SPE_TLB_ACCESS;
> >
> > if ((idx == 2 || idx == 4 || idx == 8) &&
> > - (payload & BIT(EV_LLC_MISS)))
> > + (payload & SPE_EVT_PKT_LLC_MISS))
> > decoder->record.type |= ARM_SPE_LLC_MISS;
> >
> > if ((idx == 2 || idx == 4 || idx == 8) &&
> > - (payload & BIT(EV_LLC_ACCESS)))
> > + (payload & SPE_EVT_PKT_LLC_ACCESS))
> > decoder->record.type |= ARM_SPE_LLC_ACCESS;
> >
> > if ((idx == 2 || idx == 4 || idx == 8) &&
> > - (payload & BIT(EV_REMOTE_ACCESS)))
> > + (payload & SPE_EVT_PKT_REMOTE_ACCESS))
> > decoder->record.type |= ARM_SPE_REMOTE_ACCESS;
> >
> > - if (payload & BIT(EV_MISPRED))
> > + if (payload & SPE_EVT_PKT_MISPREDICTED)
> > decoder->record.type |= ARM_SPE_BRANCH_MISS;
> >
> > break;
> > diff --git a/tools/perf/util/arm-spe-decoder/arm-spe-decoder.h b/tools/perf/util/arm-spe-decoder/arm-spe-decoder.h
> > index a5111a8d4360..24727b8ca7ff 100644
> > --- a/tools/perf/util/arm-spe-decoder/arm-spe-decoder.h
> > +++ b/tools/perf/util/arm-spe-decoder/arm-spe-decoder.h
> > @@ -13,23 +13,6 @@
> >
> > #include "arm-spe-pkt-decoder.h"
> >
> > -enum arm_spe_events {
> > - EV_EXCEPTION_GEN = 0,
> > - EV_RETIRED = 1,
> > - EV_L1D_ACCESS = 2,
> > - EV_L1D_REFILL = 3,
> > - EV_TLB_ACCESS = 4,
> > - EV_TLB_WALK = 5,
> > - EV_NOT_TAKEN = 6,
> > - EV_MISPRED = 7,
> > - EV_LLC_ACCESS = 8,
> > - EV_LLC_MISS = 9,
> > - EV_REMOTE_ACCESS = 10,
> > - EV_ALIGNMENT = 11,
> > - EV_PARTIAL_PREDICATE = 17,
> > - EV_EMPTY_PREDICATE = 18,
> > -};
>
> So what about keeping this, but moving it into the other header file?

Will do. This is more directive, especially if we consider every bit
indicates an event type :)

> coding-style.rst says: "Enums are preferred when defining several
> related constants."

Thanks for pasting the coding style, it's useful. I agree that using
BIT() + enum is better form, will refine the patch for this.

> > -
> > enum arm_spe_sample_type {
> > ARM_SPE_L1D_ACCESS = 1 << 0,
> > ARM_SPE_L1D_MISS = 1 << 1,
> > diff --git a/tools/perf/util/arm-spe-decoder/arm-spe-pkt-decoder.c b/tools/perf/util/arm-spe-decoder/arm-spe-pkt-decoder.c
> > index ed0f4c74dfc5..b8f343320abf 100644
> > --- a/tools/perf/util/arm-spe-decoder/arm-spe-pkt-decoder.c
> > +++ b/tools/perf/util/arm-spe-decoder/arm-spe-pkt-decoder.c
> > @@ -284,58 +284,58 @@ int arm_spe_pkt_desc(const struct arm_spe_pkt *packet, char *buf,
> > if (ret < 0)
> > return ret;
> >
> > - if (payload & 0x1) {
> > + if (payload & SPE_EVT_PKT_GEN_EXCEPTION) {
>
> Having the bitmask here directly is indeed not very nice and error
> prone. But I would rather see the above solution:
> if (payload & BIT(EV_EXCEPTION_GEN)) {

Will do.

> > ret = arm_spe_pkt_snprintf(&buf, &blen, " EXCEPTION-GEN");
> > if (ret < 0)
> > return ret;
> > }
> > - if (payload & 0x2) {
> > + if (payload & SPE_EVT_PKT_ARCH_RETIRED) {
> > ret = arm_spe_pkt_snprintf(&buf, &blen, " RETIRED");
> > if (ret < 0)
> > return ret;
> > }
> > - if (payload & 0x4) {
> > + if (payload & SPE_EVT_PKT_L1D_ACCESS) {
> > ret = arm_spe_pkt_snprintf(&buf, &blen, " L1D-ACCESS");
> > if (ret < 0)
> > return ret;
> > }
> > - if (payload & 0x8) {
> > + if (payload & SPE_EVT_PKT_L1D_REFILL) {
> > ret = arm_spe_pkt_snprintf(&buf, &blen, " L1D-REFILL");
> > if (ret < 0)
> > return ret;
> > }
> > - if (payload & 0x10) {
> > + if (payload & SPE_EVT_PKT_TLB_ACCESS) {
> > ret = arm_spe_pkt_snprintf(&buf, &blen, " TLB-ACCESS");
> > if (ret < 0)
> > return ret;
> > }
> > - if (payload & 0x20) {
> > + if (payload & SPE_EVT_PKT_TLB_WALK) {
> > ret = arm_spe_pkt_snprintf(&buf, &blen, " TLB-REFILL");
> > if (ret < 0)
> > return ret;
> > }
> > - if (payload & 0x40) {
> > + if (payload & SPE_EVT_PKT_NOT_TAKEN) {
> > ret = arm_spe_pkt_snprintf(&buf, &blen, " NOT-TAKEN");
> > if (ret < 0)
> > return ret;
> > }
> > - if (payload & 0x80) {
> > + if (payload & SPE_EVT_PKT_MISPREDICTED) {
> > ret = arm_spe_pkt_snprintf(&buf, &blen, " MISPRED");
> > if (ret < 0)
> > return ret;
> > }
> > if (idx > 1) {
>
> Do you know what the purpose of this comparison is? Surely payload would
> not contain more bits than would fit in "idx" bytes? So is this some
> attempt of an optimisation?

Here "idx" is for payload size (in bytes); you could see function
arm_spe_get_events() calculate the payload size:

packet->index = PAYLOAD_LEN(buf[0]);

Please note, the raw payload size (field "sz" in header) value is:

0b00 Byte.
0b01 Halfword.
0b10 Word.
0b11 Doubleword.

After using PAYLOAD_LEN(), the payload size is converted to value in
byte, so:

packet->index = 1 << "sz";

1 Byte
2 Halfword
4 Word
8 Doubleword

In Armv8 ARM, chapter "D10.2.6 Events packet", we can see the events
"Remote access", "Last Level cache miss" and "Last Level cache access"
are only valid when "sz" is equal or longer than Halfword, thus idx is
2/4/8; this is why here checks the condition "if (idx > 1)".

> If so, I doubt it's really useful, the
> compiler might find a smarter solution to the problem. Just continuing
> with the bit mask comparison would make it look nicer, I think.

ARMv8 ARM gives out "Otherwise this bit reads-as-zero.", IIUC this
suggests to firstly check the size, if cannot meet the size requirement,
then the Event bit should be reads-as-zero.

Thanks,
Leo