From: Richard Henderson <richard.henderson@linaro.org>
To: qemu-devel@nongnu.org
Subject: [Qemu-devel] [PATCH v2 08/10] tcg/i386: Implement vector minmax arithmetic
Date: Sat, 5 Jan 2019 08:31:14 +1000 [thread overview]
Message-ID: <20190104223116.14037-9-richard.henderson@linaro.org> (raw)
In-Reply-To: <20190104223116.14037-1-richard.henderson@linaro.org>
The avx instruction set does not directly provide MO_64.
We can still implement 64-bit with comparison and vpblendvb.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target.h | 2 +-
tcg/i386/tcg-target.inc.c | 81 +++++++++++++++++++++++++++++++++++++++
2 files changed, 82 insertions(+), 1 deletion(-)
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index efbd5a6fc9..7995fe3eab 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -186,7 +186,7 @@ extern bool have_avx2;
#define TCG_TARGET_HAS_cmp_vec 1
#define TCG_TARGET_HAS_mul_vec 1
#define TCG_TARGET_HAS_sat_vec 1
-#define TCG_TARGET_HAS_minmax_vec 0
+#define TCG_TARGET_HAS_minmax_vec 1
#define TCG_TARGET_deposit_i32_valid(ofs, len) \
(((ofs) == 0 && (len) == 8) || ((ofs) == 8 && (len) == 8) || \
diff --git a/tcg/i386/tcg-target.inc.c b/tcg/i386/tcg-target.inc.c
index feec40a412..94007c7aa5 100644
--- a/tcg/i386/tcg-target.inc.c
+++ b/tcg/i386/tcg-target.inc.c
@@ -392,6 +392,18 @@ static inline int tcg_target_const_match(tcg_target_long val, TCGType type,
#define OPC_PCMPGTW (0x65 | P_EXT | P_DATA16)
#define OPC_PCMPGTD (0x66 | P_EXT | P_DATA16)
#define OPC_PCMPGTQ (0x37 | P_EXT38 | P_DATA16)
+#define OPC_PMAXSB (0x3c | P_EXT38 | P_DATA16)
+#define OPC_PMAXSW (0xee | P_EXT | P_DATA16)
+#define OPC_PMAXSD (0x3d | P_EXT38 | P_DATA16)
+#define OPC_PMAXUB (0xde | P_EXT | P_DATA16)
+#define OPC_PMAXUW (0x3e | P_EXT38 | P_DATA16)
+#define OPC_PMAXUD (0x3f | P_EXT38 | P_DATA16)
+#define OPC_PMINSB (0x38 | P_EXT38 | P_DATA16)
+#define OPC_PMINSW (0xea | P_EXT | P_DATA16)
+#define OPC_PMINSD (0x39 | P_EXT38 | P_DATA16)
+#define OPC_PMINUB (0xda | P_EXT | P_DATA16)
+#define OPC_PMINUW (0x3a | P_EXT38 | P_DATA16)
+#define OPC_PMINUD (0x3b | P_EXT38 | P_DATA16)
#define OPC_PMOVSXBW (0x20 | P_EXT38 | P_DATA16)
#define OPC_PMOVSXWD (0x23 | P_EXT38 | P_DATA16)
#define OPC_PMOVSXDQ (0x25 | P_EXT38 | P_DATA16)
@@ -2638,6 +2650,18 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
static int const packus_insn[4] = {
OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
};
+ static int const smin_insn[4] = {
+ OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_UD2
+ };
+ static int const smax_insn[4] = {
+ OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_UD2
+ };
+ static int const umin_insn[4] = {
+ OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_UD2
+ };
+ static int const umax_insn[4] = {
+ OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_UD2
+ };
TCGType type = vecl + TCG_TYPE_V64;
int insn, sub;
@@ -2678,6 +2702,18 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
case INDEX_op_xor_vec:
insn = OPC_PXOR;
goto gen_simd;
+ case INDEX_op_smin_vec:
+ insn = smin_insn[vece];
+ goto gen_simd;
+ case INDEX_op_umin_vec:
+ insn = umin_insn[vece];
+ goto gen_simd;
+ case INDEX_op_smax_vec:
+ insn = smax_insn[vece];
+ goto gen_simd;
+ case INDEX_op_umax_vec:
+ insn = umax_insn[vece];
+ goto gen_simd;
case INDEX_op_x86_punpckl_vec:
insn = punpckl_insn[vece];
goto gen_simd;
@@ -3043,6 +3079,10 @@ static const TCGTargetOpDef *tcg_target_op_def(TCGOpcode op)
case INDEX_op_usadd_vec:
case INDEX_op_sssub_vec:
case INDEX_op_ussub_vec:
+ case INDEX_op_smin_vec:
+ case INDEX_op_umin_vec:
+ case INDEX_op_smax_vec:
+ case INDEX_op_umax_vec:
case INDEX_op_cmp_vec:
case INDEX_op_x86_shufps_vec:
case INDEX_op_x86_blend_vec:
@@ -3115,6 +3155,11 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
case INDEX_op_sssub_vec:
case INDEX_op_ussub_vec:
return vece <= MO_16;
+ case INDEX_op_smin_vec:
+ case INDEX_op_smax_vec:
+ case INDEX_op_umin_vec:
+ case INDEX_op_umax_vec:
+ return vece <= MO_32 ? 1 : -1;
default:
return 0;
@@ -3343,6 +3388,25 @@ static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
}
}
+static void expand_vec_minmax(TCGType type, unsigned vece,
+ TCGCond cond, bool min,
+ TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
+{
+ TCGv_vec t1 = tcg_temp_new_vec(type);
+
+ tcg_debug_assert(vece == MO_64);
+
+ tcg_gen_cmp_vec(cond, vece, t1, v1, v2);
+ if (min) {
+ TCGv_vec t2;
+ t2 = v1, v1 = v2, v2 = t2;
+ }
+ vec_gen_4(INDEX_op_x86_vpblendvb_vec, type, vece,
+ tcgv_vec_arg(v0), tcgv_vec_arg(v1),
+ tcgv_vec_arg(v2), tcgv_vec_arg(t1));
+ tcg_temp_free_vec(t1);
+}
+
void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
TCGArg a0, ...)
{
@@ -3375,6 +3439,23 @@ void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
break;
+ case INDEX_op_smin_vec:
+ v2 = temp_tcgv_vec(arg_temp(a2));
+ expand_vec_minmax(type, vece, TCG_COND_GT, true, v0, v1, v2);
+ break;
+ case INDEX_op_smax_vec:
+ v2 = temp_tcgv_vec(arg_temp(a2));
+ expand_vec_minmax(type, vece, TCG_COND_GT, false, v0, v1, v2);
+ break;
+ case INDEX_op_umin_vec:
+ v2 = temp_tcgv_vec(arg_temp(a2));
+ expand_vec_minmax(type, vece, TCG_COND_GTU, true, v0, v1, v2);
+ break;
+ case INDEX_op_umax_vec:
+ v2 = temp_tcgv_vec(arg_temp(a2));
+ expand_vec_minmax(type, vece, TCG_COND_GTU, false, v0, v1, v2);
+ break;
+
default:
break;
}
--
2.17.2
next prev parent reply other threads:[~2019-01-04 22:31 UTC|newest]
Thread overview: 18+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-01-04 22:31 [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 01/10] tcg: Add logical simplifications during gvec expand Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 02/10] tcg: Add gvec expanders for nand, nor, eqv Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 03/10] tcg: Add write_aofs to GVecGen4 Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 04/10] tcg: Add opcodes for vector saturated arithmetic Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 05/10] tcg: Add opcodes for vector minmax arithmetic Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 06/10] tcg/i386: Split subroutines out of tcg_expand_vec_op Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 07/10] tcg/i386: Implement vector saturating arithmetic Richard Henderson
2019-01-04 22:31 ` Richard Henderson [this message]
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 09/10] tcg/aarch64: " Richard Henderson
2019-01-04 22:31 ` [Qemu-devel] [PATCH v2 10/10] tcg/aarch64: Implement vector minmax arithmetic Richard Henderson
2019-01-07 13:11 ` [Qemu-devel] [PATCH v2 00/10] tcg vector improvements Mark Cave-Ayland
2019-01-23 5:09 ` Richard Henderson
2019-02-05 21:29 ` Mark Cave-Ayland
2019-02-06 3:37 ` Richard Henderson
2019-02-06 7:15 ` Mark Cave-Ayland
2019-02-06 15:44 ` BALATON Zoltan
2019-02-07 22:31 ` Mark Cave-Ayland
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=20190104223116.14037-9-richard.henderson@linaro.org \
--to=richard.henderson@linaro.org \
--cc=qemu-devel@nongnu.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 a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).