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 17/18] tcg/s390x: Implement cmpsel_vec
Date: Wed, 11 Sep 2024 09:50:46 -0700 [thread overview]
Message-ID: <20240911165047.1035764-18-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/s390x/tcg-target-con-set.h | 1 +
tcg/s390x/tcg-target.h | 2 +-
tcg/s390x/tcg-target.c.inc | 40 ++++++++++++++++++----------------
3 files changed, 23 insertions(+), 20 deletions(-)
diff --git a/tcg/s390x/tcg-target-con-set.h b/tcg/s390x/tcg-target-con-set.h
index f75955eaa8..670089086d 100644
--- a/tcg/s390x/tcg-target-con-set.h
+++ b/tcg/s390x/tcg-target-con-set.h
@@ -38,6 +38,7 @@ C_O1_I2(r, rZ, r)
C_O1_I2(v, v, r)
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, ri, rI, r)
C_O1_I4(r, r, rC, rI, r)
C_O2_I1(o, m, r)
diff --git a/tcg/s390x/tcg-target.h b/tcg/s390x/tcg-target.h
index 62ce9d792a..86aeca166f 100644
--- a/tcg/s390x/tcg-target.h
+++ b/tcg/s390x/tcg-target.h
@@ -162,7 +162,7 @@ extern uint64_t s390_facilities[3];
#define TCG_TARGET_HAS_sat_vec 0
#define TCG_TARGET_HAS_minmax_vec 1
#define TCG_TARGET_HAS_bitsel_vec 1
-#define TCG_TARGET_HAS_cmpsel_vec 0
+#define TCG_TARGET_HAS_cmpsel_vec 1
#define TCG_TARGET_HAS_tst_vec 0
/* used for function call generation */
diff --git a/tcg/s390x/tcg-target.c.inc b/tcg/s390x/tcg-target.c.inc
index 23935fd0f0..e044168826 100644
--- a/tcg/s390x/tcg-target.c.inc
+++ b/tcg/s390x/tcg-target.c.inc
@@ -46,6 +46,7 @@
/* A scratch register that may be be used throughout the backend. */
#define TCG_TMP0 TCG_REG_R1
+#define TCG_VEC_TMP0 TCG_REG_V31
#define TCG_GUEST_BASE_REG TCG_REG_R13
@@ -2902,6 +2903,18 @@ 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_TMP0, c1, c2, cond)) {
+ TCGReg swap = v3;
+ v3 = v4;
+ v4 = swap;
+ }
+ tcg_out_insn(s, VRRe, VSEL, a0, v3, v4, TCG_VEC_TMP0);
+}
+
static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
unsigned vecl, unsigned vece,
const TCGArg args[TCG_MAX_OP_ARGS],
@@ -3022,6 +3035,9 @@ 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]);
break;
+ case INDEX_op_cmpsel_vec:
+ tcg_out_cmpsel_vec(s, vece, a0, a1, a2, args[3], args[4], args[5]);
+ break;
case INDEX_op_s390_vuph_vec:
tcg_out_insn(s, VRRa, VUPH, a0, a1, vece);
@@ -3074,8 +3090,8 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
case INDEX_op_umin_vec:
case INDEX_op_xor_vec:
case INDEX_op_cmp_vec:
- return 1;
case INDEX_op_cmpsel_vec:
+ return 1;
case INDEX_op_rotrv_vec:
return -1;
case INDEX_op_mul_vec:
@@ -3088,17 +3104,6 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
}
}
-static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGv_vec v0,
- TCGv_vec c1, TCGv_vec c2,
- TCGv_vec v3, TCGv_vec v4, TCGCond cond)
-{
- TCGv_vec t = tcg_temp_new_vec(type);
-
- tcg_gen_cmp_vec(cond, vece, t, c1, c2);
- tcg_gen_bitsel_vec(vece, v0, t, v3, v4);
- tcg_temp_free_vec(t);
-}
-
static void expand_vec_sat(TCGType type, unsigned vece, TCGv_vec v0,
TCGv_vec v1, TCGv_vec v2, TCGOpcode add_sub_opc)
{
@@ -3140,7 +3145,7 @@ void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
TCGArg a0, ...)
{
va_list va;
- TCGv_vec v0, v1, v2, v3, v4, t0;
+ TCGv_vec v0, v1, v2, t0;
va_start(va, a0);
v0 = temp_tcgv_vec(arg_temp(a0));
@@ -3148,12 +3153,6 @@ void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
v2 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
switch (opc) {
- case INDEX_op_cmpsel_vec:
- v3 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
- v4 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
- expand_vec_cmpsel(type, vece, v0, v1, v2, v3, v4, va_arg(va, TCGArg));
- break;
-
case INDEX_op_rotrv_vec:
t0 = tcg_temp_new_vec(type);
tcg_gen_neg_vec(vece, t0, v2);
@@ -3388,6 +3387,8 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
return C_O1_I2(v, v, r);
case INDEX_op_bitsel_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();
@@ -3512,6 +3513,7 @@ static void tcg_target_init(TCGContext *s)
s->reserved_regs = 0;
tcg_regset_set_reg(s->reserved_regs, TCG_TMP0);
+ tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP0);
/* XXX many insns can't be used with R0, so we better avoid it for now */
tcg_regset_set_reg(s->reserved_regs, TCG_REG_R0);
tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
--
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 ` [PATCH v2 15/18] tcg/ppc: Implement cmpsel_vec Richard Henderson
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 ` Richard Henderson [this message]
2024-09-11 16:50 ` [PATCH v2 18/18] tcg/s390x: " 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=20240911165047.1035764-18-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).