From: "Dino Papararo" <dino.papararo@gmail.com>
To: 'Mark Cave-Ayland' <mark.cave-ayland@ilande.co.uk>,
qemu-devel@nongnu.org, qemu-ppc@nongnu.org,
richard.henderson@linaro.org, david@gibson.dropbear.id.au
Subject: [Qemu-devel] R: [PATCH v5 8/8] target/ppc: remove various HOST_WORDS_BIGENDIAN hacks in int_helper.c
Date: Sun, 3 Feb 2019 15:58:42 +0100 [thread overview]
Message-ID: <000301d4bbd0$f50d7ac0$df287040$@gmail.com> (raw)
In-Reply-To: <20190130203638.21179-9-mark.cave-ayland@ilande.co.uk>
Hello Mark,
I have a question about improving speed manually unrolling loops like this
Assuming ARRAY_SIZE(r->u8) is always multiple of 4 you can manually improve loop in this way, on modern CPU non sequential instructions can be computed nearly for free:
> {
> int i, j = (sh & 0xf);
>
> - VECTOR_FOR_INORDER_I(i, u8) {
> - r->u8[i] = j++;
> + for (i = 0; i < ARRAY_SIZE(r->u8); i+=4,j+=4) {
> + r->VsrB(i) = j;
> + r->VsrB(i+1) = j+1;
> + r->VsrB(i+2) = j+2;
> + r->VsrB(i+3) = j+3; }
> }
In this patch there are a lot of functions can benefit by unrolling loops, with a huge speed improvement.
Maybe compiler could do it itself but aren't humans still better? 😊
Best Regards,
Dino Papararo
-----Messaggio originale-----
Da: Qemu-devel <qemu-devel-bounces+skizzato73=msn.com@nongnu.org> Per conto di Mark Cave-Ayland
Inviato: mercoledì 30 gennaio 2019 21:37
A: qemu-devel@nongnu.org; qemu-ppc@nongnu.org; richard.henderson@linaro.org; david@gibson.dropbear.id.au
Oggetto: [Qemu-devel] [PATCH v5 8/8] target/ppc: remove various HOST_WORDS_BIGENDIAN hacks in int_helper.c
Following on from the previous work, there are numerous endian-related hacks in int_helper.c that can now be replaced with Vsr* macros.
There are also a few places where the VECTOR_FOR_INORDER_I macro can be replaced with a normal iterator since the processing order is irrelevant.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
---
target/ppc/int_helper.c | 155 ++++++++++++++----------------------------------
1 file changed, 45 insertions(+), 110 deletions(-)
diff --git a/target/ppc/int_helper.c b/target/ppc/int_helper.c index 916d10c25b..8efc283388 100644
--- a/target/ppc/int_helper.c
+++ b/target/ppc/int_helper.c
@@ -443,8 +443,8 @@ void helper_lvsl(ppc_avr_t *r, target_ulong sh) {
int i, j = (sh & 0xf);
- VECTOR_FOR_INORDER_I(i, u8) {
- r->u8[i] = j++;
+ for (i = 0; i < ARRAY_SIZE(r->u8); i++) {
+ r->VsrB(i) = j++;
}
}
@@ -452,18 +452,14 @@ void helper_lvsr(ppc_avr_t *r, target_ulong sh) {
int i, j = 0x10 - (sh & 0xf);
- VECTOR_FOR_INORDER_I(i, u8) {
- r->u8[i] = j++;
+ for (i = 0; i < ARRAY_SIZE(r->u8); i++) {
+ r->VsrB(i) = j++;
}
}
void helper_mtvscr(CPUPPCState *env, ppc_avr_t *r) { -#if defined(HOST_WORDS_BIGENDIAN)
- env->vscr = r->u32[3];
-#else
- env->vscr = r->u32[0];
-#endif
+ env->vscr = r->VsrW(3);
set_flush_to_zero(vscr_nj, &env->vec_status); }
@@ -870,8 +866,8 @@ target_ulong helper_vclzlsbb(ppc_avr_t *r) {
target_ulong count = 0;
int i;
- VECTOR_FOR_INORDER_I(i, u8) {
- if (r->u8[i] & 0x01) {
+ for (i = 0; i < ARRAY_SIZE(r->u8); i++) {
+ if (r->VsrB(i) & 0x01) {
break;
}
count++;
@@ -883,12 +879,8 @@ target_ulong helper_vctzlsbb(ppc_avr_t *r) {
target_ulong count = 0;
int i;
-#if defined(HOST_WORDS_BIGENDIAN)
for (i = ARRAY_SIZE(r->u8) - 1; i >= 0; i--) { -#else
- for (i = 0; i < ARRAY_SIZE(r->u8); i++) {
-#endif
- if (r->u8[i] & 0x01) {
+ if (r->VsrB(i) & 0x01) {
break;
}
count++;
@@ -1137,18 +1129,14 @@ void helper_vperm(CPUPPCState *env, ppc_avr_t *r, ppc_avr_t *a, ppc_avr_t *b,
ppc_avr_t result;
int i;
- VECTOR_FOR_INORDER_I(i, u8) {
- int s = c->u8[i] & 0x1f;
-#if defined(HOST_WORDS_BIGENDIAN)
+ for (i = 0; i < ARRAY_SIZE(r->u8); i++) {
+ int s = c->VsrB(i) & 0x1f;
int index = s & 0xf;
-#else
- int index = 15 - (s & 0xf);
-#endif
if (s & 0x10) {
- result.u8[i] = b->u8[index];
+ result.VsrB(i) = b->VsrB(index);
} else {
- result.u8[i] = a->u8[index];
+ result.VsrB(i) = a->VsrB(index);
}
}
*r = result;
@@ -1160,18 +1148,14 @@ void helper_vpermr(CPUPPCState *env, ppc_avr_t *r, ppc_avr_t *a, ppc_avr_t *b,
ppc_avr_t result;
int i;
- VECTOR_FOR_INORDER_I(i, u8) {
- int s = c->u8[i] & 0x1f;
-#if defined(HOST_WORDS_BIGENDIAN)
+ for (i = 0; i < ARRAY_SIZE(r->u8); i++) {
+ int s = c->VsrB(i) & 0x1f;
int index = 15 - (s & 0xf);
-#else
- int index = s & 0xf;
-#endif
if (s & 0x10) {
- result.u8[i] = a->u8[index];
+ result.VsrB(i) = a->VsrB(index);
} else {
- result.u8[i] = b->u8[index];
+ result.VsrB(i) = b->VsrB(index);
}
}
*r = result;
@@ -1868,25 +1852,14 @@ void helper_vsldoi(ppc_avr_t *r, ppc_avr_t *a, ppc_avr_t *b, uint32_t shift)
int i;
ppc_avr_t result;
-#if defined(HOST_WORDS_BIGENDIAN)
for (i = 0; i < ARRAY_SIZE(r->u8); i++) {
int index = sh + i;
if (index > 0xf) {
- result.u8[i] = b->u8[index - 0x10];
- } else {
- result.u8[i] = a->u8[index];
- }
- }
-#else
- for (i = 0; i < ARRAY_SIZE(r->u8); i++) {
- int index = (16 - sh) + i;
- if (index > 0xf) {
- result.u8[i] = a->u8[index - 0x10];
+ result.VsrB(i) = b->VsrB(index - 0x10);
} else {
- result.u8[i] = b->u8[index];
+ result.VsrB(i) = a->VsrB(index);
}
}
-#endif
*r = result;
}
@@ -1905,25 +1878,20 @@ void helper_vslo(ppc_avr_t *r, ppc_avr_t *a, ppc_avr_t *b)
/* Experimental testing shows that hardware masks the immediate. */ #define _SPLAT_MASKED(element) (splat & (ARRAY_SIZE(r->element) - 1)) -#if defined(HOST_WORDS_BIGENDIAN) #define SPLAT_ELEMENT(element) _SPLAT_MASKED(element) -#else
-#define SPLAT_ELEMENT(element) \
- (ARRAY_SIZE(r->element) - 1 - _SPLAT_MASKED(element))
-#endif
-#define VSPLT(suffix, element) \
+#define VSPLT(suffix, element, access) \
void helper_vsplt##suffix(ppc_avr_t *r, ppc_avr_t *b, uint32_t splat) \
{ \
- uint32_t s = b->element[SPLAT_ELEMENT(element)]; \
+ uint32_t s = b->access(SPLAT_ELEMENT(element)); \
int i; \
\
for (i = 0; i < ARRAY_SIZE(r->element); i++) { \
- r->element[i] = s; \
+ r->access(i) = s; \
} \
}
-VSPLT(b, u8)
-VSPLT(h, u16)
-VSPLT(w, u32)
+VSPLT(b, u8, VsrB)
+VSPLT(h, u16, VsrH)
+VSPLT(w, u32, VsrW)
#undef VSPLT
#undef SPLAT_ELEMENT
#undef _SPLAT_MASKED
@@ -1984,17 +1952,10 @@ void helper_xxextractuw(CPUPPCState *env, target_ulong xtn,
getVSR(xbn, &xb, env);
memset(&xt, 0, sizeof(xt));
-#if defined(HOST_WORDS_BIGENDIAN)
ext_index = index;
for (i = 0; i < es; i++, ext_index++) {
- xt.u8[8 - es + i] = xb.u8[ext_index % 16];
- }
-#else
- ext_index = 15 - index;
- for (i = es - 1; i >= 0; i--, ext_index--) {
- xt.u8[8 + i] = xb.u8[ext_index % 16];
+ xt.VsrB(8 - es + i) = xb.VsrB(ext_index % 16);
}
-#endif
putVSR(xtn, &xt, env);
}
@@ -2009,17 +1970,10 @@ void helper_xxinsertw(CPUPPCState *env, target_ulong xtn,
getVSR(xbn, &xb, env);
getVSR(xtn, &xt, env);
-#if defined(HOST_WORDS_BIGENDIAN)
ins_index = index;
for (i = 0; i < es && ins_index < 16; i++, ins_index++) {
- xt.u8[ins_index] = xb.u8[8 - es + i];
- }
-#else
- ins_index = 15 - index;
- for (i = es - 1; i >= 0 && ins_index >= 0; i--, ins_index--) {
- xt.u8[ins_index] = xb.u8[8 + i];
+ xt.VsrB(ins_index) = xb.VsrB(8 - es + i);
}
-#endif
putVSR(xtn, &xt, env);
}
@@ -2028,7 +1982,7 @@ void helper_xxinsertw(CPUPPCState *env, target_ulong xtn,
void helper_##name(ppc_avr_t *r, ppc_avr_t *b) \
{ \
int i; \
- VECTOR_FOR_INORDER_I(i, element) { \
+ for (i = 0; i < ARRAY_SIZE(r->element); i++) { \
r->element[i] = (cast)b->element[i]; \
} \
}
@@ -2043,7 +1997,7 @@ VEXT_SIGNED(vextsw2d, s64, int32_t)
void helper_##name(ppc_avr_t *r, ppc_avr_t *b) \
{ \
int i; \
- VECTOR_FOR_INORDER_I(i, element) { \
+ for (i = 0; i < ARRAY_SIZE(r->element); i++) { \
r->element[i] = -b->element[i]; \
} \
}
@@ -2115,17 +2069,13 @@ void helper_vsumsws(CPUPPCState *env, ppc_avr_t *r, ppc_avr_t *a, ppc_avr_t *b)
ppc_avr_t result;
int sat = 0;
-#if defined(HOST_WORDS_BIGENDIAN)
- upper = ARRAY_SIZE(r->s32)-1;
-#else
- upper = 0;
-#endif
- t = (int64_t)b->s32[upper];
+ upper = ARRAY_SIZE(r->s32) - 1;
+ t = (int64_t)b->VsrSW(upper);
for (i = 0; i < ARRAY_SIZE(r->s32); i++) {
- t += a->s32[i];
- result.s32[i] = 0;
+ t += a->VsrSW(i);
+ result.VsrSW(i) = 0;
}
- result.s32[upper] = cvtsdsw(t, &sat);
+ result.VsrSW(upper) = cvtsdsw(t, &sat);
*r = result;
if (sat) {
@@ -2139,19 +2089,15 @@ void helper_vsum2sws(CPUPPCState *env, ppc_avr_t *r, ppc_avr_t *a, ppc_avr_t *b)
ppc_avr_t result;
int sat = 0;
-#if defined(HOST_WORDS_BIGENDIAN)
upper = 1;
-#else
- upper = 0;
-#endif
for (i = 0; i < ARRAY_SIZE(r->u64); i++) {
- int64_t t = (int64_t)b->s32[upper + i * 2];
+ int64_t t = (int64_t)b->VsrSW(upper + i * 2);
- result.u64[i] = 0;
+ result.VsrW(i) = 0;
for (j = 0; j < ARRAY_SIZE(r->u64); j++) {
- t += a->s32[2 * i + j];
+ t += a->VsrSW(2 * i + j);
}
- result.s32[upper + i * 2] = cvtsdsw(t, &sat);
+ result.VsrSW(upper + i * 2) = cvtsdsw(t, &sat);
}
*r = result;
@@ -2276,7 +2222,7 @@ VUPK(lsw, s64, s32, UPKLO)
{ \
int i; \
\
- VECTOR_FOR_INORDER_I(i, element) { \
+ for (i = 0; i < ARRAY_SIZE(r->element); i++) { \
r->element[i] = name(b->element[i]); \
} \
}
@@ -2616,20 +2562,12 @@ static int bcd_cmp_zero(ppc_avr_t *bcd)
static uint16_t get_national_digit(ppc_avr_t *reg, int n) { -#if defined(HOST_WORDS_BIGENDIAN)
- return reg->u16[7 - n];
-#else
- return reg->u16[n];
-#endif
+ return reg->VsrH(7 - n);
}
static void set_national_digit(ppc_avr_t *reg, uint8_t val, int n) { -#if defined(HOST_WORDS_BIGENDIAN)
- reg->u16[7 - n] = val;
-#else
- reg->u16[n] = val;
-#endif
+ reg->VsrH(7 - n) = val;
}
static int bcd_cmp_mag(ppc_avr_t *a, ppc_avr_t *b) @@ -3373,14 +3311,11 @@ void helper_vpermxor(ppc_avr_t *r, ppc_avr_t *a, ppc_avr_t *b, ppc_avr_t *c)
ppc_avr_t result;
int i;
- VECTOR_FOR_INORDER_I(i, u8) {
- int indexA = c->u8[i] >> 4;
- int indexB = c->u8[i] & 0xF;
-#if defined(HOST_WORDS_BIGENDIAN)
- result.u8[i] = a->u8[indexA] ^ b->u8[indexB];
-#else
- result.u8[i] = a->u8[15-indexA] ^ b->u8[15-indexB];
-#endif
+ for (i = 0; i < ARRAY_SIZE(r->u8); i++) {
+ int indexA = c->VsrB(i) >> 4;
+ int indexB = c->VsrB(i) & 0xF;
+
+ result.VsrB(i) = a->VsrB(indexA) ^ b->VsrB(indexB);
}
*r = result;
}
--
2.11.0
next prev parent reply other threads:[~2019-02-03 14:59 UTC|newest]
Thread overview: 12+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-01-30 20:36 [Qemu-devel] [PATCH v5 0/8] target/ppc: remove various endian hacks from int_helper.c Mark Cave-Ayland
2019-01-30 20:36 ` [Qemu-devel] [PATCH v5 1/8] target/ppc: implement complete set of Vsr* macros Mark Cave-Ayland
2019-01-30 20:36 ` [Qemu-devel] [PATCH v5 2/8] target/ppc: rework vmrg{l, h}{b, h, w} instructions to use " Mark Cave-Ayland
2019-02-01 5:25 ` Richard Henderson
2019-01-30 20:36 ` [Qemu-devel] [PATCH v5 3/8] target/ppc: rework vmul{e, o}{s, u}{b, " Mark Cave-Ayland
2019-01-30 20:36 ` [Qemu-devel] [PATCH v5 4/8] target/ppc: eliminate use of HI_IDX and LO_IDX macros from int_helper.c Mark Cave-Ayland
2019-01-30 20:36 ` [Qemu-devel] [PATCH v5 5/8] target/ppc: eliminate use of EL_IDX " Mark Cave-Ayland
2019-01-30 20:36 ` [Qemu-devel] [PATCH v5 6/8] target/ppc: simplify VEXT_SIGNED macro in int_helper.c Mark Cave-Ayland
2019-01-30 20:36 ` [Qemu-devel] [PATCH v5 7/8] target/ppc: remove ROTRu32 and ROTRu64 macros from int_helper.c Mark Cave-Ayland
2019-01-30 20:36 ` [Qemu-devel] [PATCH v5 8/8] target/ppc: remove various HOST_WORDS_BIGENDIAN hacks in int_helper.c Mark Cave-Ayland
2019-02-03 14:58 ` Dino Papararo [this message]
2019-01-31 3:09 ` [Qemu-devel] [PATCH v5 0/8] target/ppc: remove various endian hacks from int_helper.c David Gibson
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='000301d4bbd0$f50d7ac0$df287040$@gmail.com' \
--to=dino.papararo@gmail.com \
--cc=david@gibson.dropbear.id.au \
--cc=mark.cave-ayland@ilande.co.uk \
--cc=qemu-devel@nongnu.org \
--cc=qemu-ppc@nongnu.org \
--cc=richard.henderson@linaro.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.