From: Richard Henderson <richard.henderson@linaro.org>
To: qemu-devel@nongnu.org
Cc: zhiwei_liu@linux.alibaba.com, tangtiancheng.ttc@alibaba-inc.com,
philmd@linaro.org
Subject: [PATCH v2 15/18] tcg/ppc: Implement cmpsel_vec
Date: Wed, 11 Sep 2024 09:50:44 -0700 [thread overview]
Message-ID: <20240911165047.1035764-16-richard.henderson@linaro.org> (raw)
In-Reply-To: <20240911165047.1035764-1-richard.henderson@linaro.org>
Do not allow cmpsel_vec to be expanded early, so that we can
make the correct decision wrt the sense of the comparison.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/ppc/tcg-target-con-set.h | 1 +
tcg/ppc/tcg-target.h | 2 +-
tcg/ppc/tcg-target.c.inc | 60 +++++++++++++++++++++++++++++++-----
3 files changed, 54 insertions(+), 9 deletions(-)
diff --git a/tcg/ppc/tcg-target-con-set.h b/tcg/ppc/tcg-target-con-set.h
index 9f99bde505..e7ba00c248 100644
--- a/tcg/ppc/tcg-target-con-set.h
+++ b/tcg/ppc/tcg-target-con-set.h
@@ -33,6 +33,7 @@ C_O1_I2(r, r, rU)
C_O1_I2(r, r, rZW)
C_O1_I2(v, v, v)
C_O1_I3(v, v, v, v)
+C_O1_I4(v, v, v, v, v)
C_O1_I4(r, r, rC, rZ, rZ)
C_O1_I4(r, r, r, ri, ri)
C_O2_I1(r, r, r)
diff --git a/tcg/ppc/tcg-target.h b/tcg/ppc/tcg-target.h
index e154fb14df..0b2171d38c 100644
--- a/tcg/ppc/tcg-target.h
+++ b/tcg/ppc/tcg-target.h
@@ -172,7 +172,7 @@ typedef enum {
#define TCG_TARGET_HAS_sat_vec 1
#define TCG_TARGET_HAS_minmax_vec 1
#define TCG_TARGET_HAS_bitsel_vec have_vsx
-#define TCG_TARGET_HAS_cmpsel_vec 0
+#define TCG_TARGET_HAS_cmpsel_vec 1
#define TCG_TARGET_HAS_tst_vec 0
#define TCG_TARGET_DEFAULT_MO (0)
diff --git a/tcg/ppc/tcg-target.c.inc b/tcg/ppc/tcg-target.c.inc
index 497e130581..9d07b4d8e6 100644
--- a/tcg/ppc/tcg-target.c.inc
+++ b/tcg/ppc/tcg-target.c.inc
@@ -3573,6 +3573,7 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
case INDEX_op_rotli_vec:
return vece <= MO_32 || have_isa_2_07 ? -1 : 0;
case INDEX_op_cmp_vec:
+ case INDEX_op_cmpsel_vec:
return vece <= MO_32 || have_isa_2_07 ? 1 : 0;
case INDEX_op_neg_vec:
return vece >= MO_32 && have_isa_3_00;
@@ -3719,6 +3720,33 @@ static void tcg_out_not_vec(TCGContext *s, TCGReg a0, TCGReg a1)
tcg_out32(s, VNOR | VRT(a0) | VRA(a1) | VRB(a1));
}
+static void tcg_out_or_vec(TCGContext *s, TCGReg a0, TCGReg a1, TCGReg a2)
+{
+ tcg_out32(s, VOR | VRT(a0) | VRA(a1) | VRB(a2));
+}
+
+static void tcg_out_and_vec(TCGContext *s, TCGReg a0, TCGReg a1, TCGReg a2)
+{
+ tcg_out32(s, VAND | VRT(a0) | VRA(a1) | VRB(a2));
+}
+
+static void tcg_out_andc_vec(TCGContext *s, TCGReg a0, TCGReg a1, TCGReg a2)
+{
+ tcg_out32(s, VANDC | VRT(a0) | VRA(a1) | VRB(a2));
+}
+
+static void tcg_out_bitsel_vec(TCGContext *s, TCGReg d,
+ TCGReg c, TCGReg t, TCGReg f)
+{
+ if (TCG_TARGET_HAS_bitsel_vec) {
+ tcg_out32(s, XXSEL | VRT(d) | VRC(c) | VRB(t) | VRA(f));
+ } else {
+ tcg_out_and_vec(s, TCG_VEC_TMP2, t, c);
+ tcg_out_andc_vec(s, d, f, c);
+ tcg_out_or_vec(s, d, d, TCG_VEC_TMP2);
+ }
+}
+
static bool tcg_out_cmp_vec_noinv(TCGContext *s, unsigned vece, TCGReg a0,
TCGReg a1, TCGReg a2, TCGCond cond)
{
@@ -3798,6 +3826,18 @@ static void tcg_out_cmp_vec(TCGContext *s, unsigned vece, TCGReg a0,
}
}
+static void tcg_out_cmpsel_vec(TCGContext *s, unsigned vece, TCGReg a0,
+ TCGReg c1, TCGReg c2, TCGReg v3, TCGReg v4,
+ TCGCond cond)
+{
+ if (tcg_out_cmp_vec_noinv(s, vece, TCG_VEC_TMP1, c1, c2, cond)) {
+ TCGReg swap = v3;
+ v3 = v4;
+ v4 = swap;
+ }
+ tcg_out_bitsel_vec(s, a0, TCG_VEC_TMP1, v3, v4);
+}
+
static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
unsigned vecl, unsigned vece,
const TCGArg args[TCG_MAX_OP_ARGS],
@@ -3889,17 +3929,17 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
insn = sarv_op[vece];
break;
case INDEX_op_and_vec:
- insn = VAND;
- break;
+ tcg_out_and_vec(s, a0, a1, a2);
+ return;
case INDEX_op_or_vec:
- insn = VOR;
- break;
+ tcg_out_or_vec(s, a0, a1, a2);
+ return;
case INDEX_op_xor_vec:
insn = VXOR;
break;
case INDEX_op_andc_vec:
- insn = VANDC;
- break;
+ tcg_out_andc_vec(s, a0, a1, a2);
+ return;
case INDEX_op_not_vec:
tcg_out_not_vec(s, a0, a1);
return;
@@ -3919,9 +3959,11 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
case INDEX_op_cmp_vec:
tcg_out_cmp_vec(s, vece, a0, a1, a2, args[3]);
return;
-
+ case INDEX_op_cmpsel_vec:
+ tcg_out_cmpsel_vec(s, vece, a0, a1, a2, args[3], args[4], args[5]);
+ return;
case INDEX_op_bitsel_vec:
- tcg_out32(s, XXSEL | VRT(a0) | VRC(a1) | VRB(a2) | VRA(args[3]));
+ tcg_out_bitsel_vec(s, a0, a1, a2, args[3]);
return;
case INDEX_op_dup2_vec:
@@ -4287,6 +4329,8 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
case INDEX_op_bitsel_vec:
case INDEX_op_ppc_msum_vec:
return C_O1_I3(v, v, v, v);
+ case INDEX_op_cmpsel_vec:
+ return C_O1_I4(v, v, v, v, v);
default:
g_assert_not_reached();
--
2.43.0
next prev parent reply other threads:[~2024-09-11 16:52 UTC|newest]
Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top
2024-09-11 16:50 [PATCH v2 00/18] tcg: Improve support for cmpsel_vec Richard Henderson
2024-09-11 16:50 ` [PATCH v2 01/18] tcg: Export vec_gen_6 Richard Henderson
2024-09-11 16:50 ` [PATCH v2 02/18] tcg/i386: Split out tcg_out_vex_modrm_type Richard Henderson
2024-09-11 16:50 ` [PATCH v2 03/18] tcg/i386: Do not expand cmp_vec early Richard Henderson
2024-09-11 16:50 ` [PATCH v2 04/18] tcg/i386: Do not expand cmpsel_vec early Richard Henderson
2024-09-11 16:50 ` [PATCH v2 05/18] tcg/ppc: Do not expand cmp_vec early Richard Henderson
2024-09-11 16:50 ` [PATCH v2 06/18] tcg/s390x: " Richard Henderson
2024-09-11 16:50 ` [PATCH v2 07/18] tcg/optimize: Fold movcond with true and false values identical Richard Henderson
2024-09-11 16:50 ` [PATCH v2 08/18] tcg/optimize: Optimize cmp_vec and cmpsel_vec Richard Henderson
2024-09-11 16:50 ` [PATCH v2 09/18] tcg/optimize: Optimize bitsel_vec Richard Henderson
2024-09-11 16:50 ` [PATCH v2 10/18] tcg/i386: Optimize cmpsel with constant 0 operand 3 Richard Henderson
2024-09-11 16:50 ` [PATCH v2 11/18] tcg/i386: Implement cmp_vec with avx512 insns Richard Henderson
2024-09-11 16:50 ` [PATCH v2 12/18] tcg/i386: Add predicate parameters to tcg_out_evex_opc Richard Henderson
2024-09-11 16:50 ` [PATCH v2 13/18] tcg/i386: Implement cmpsel_vec with avx512 insns Richard Henderson
2024-09-11 16:50 ` [PATCH v2 14/18] tcg/i386: Implement vector TST{EQ,NE} for avx512 Richard Henderson
2024-09-12 6:04 ` [PATCH v2 14/18] tcg/i386: Implement vector TST{EQ, NE} " Philippe Mathieu-Daudé
2024-09-11 16:50 ` Richard Henderson [this message]
2024-09-11 16:50 ` [PATCH v2 16/18] tcg/ppc: Optimize cmpsel with constant 0/-1 arguments Richard Henderson
2024-09-11 16:50 ` [PATCH v2 17/18] tcg/s390x: Implement cmpsel_vec Richard Henderson
2024-09-11 16:50 ` [PATCH v2 18/18] tcg/s390x: Optimize cmpsel with constant 0/-1 arguments Richard Henderson
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=20240911165047.1035764-16-richard.henderson@linaro.org \
--to=richard.henderson@linaro.org \
--cc=philmd@linaro.org \
--cc=qemu-devel@nongnu.org \
--cc=tangtiancheng.ttc@alibaba-inc.com \
--cc=zhiwei_liu@linux.alibaba.com \
/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).