From: Richard Henderson <richard.henderson@linaro.org>
To: qemu-devel@nongnu.org
Cc: peter.maydell@linaro.org
Subject: [PULL 13/31] tcg/i386: Optimize cmpsel with constant 0 operand 3.
Date: Sun, 22 Sep 2024 14:00:54 +0200 [thread overview]
Message-ID: <20240922120112.5067-14-richard.henderson@linaro.org> (raw)
In-Reply-To: <20240922120112.5067-1-richard.henderson@linaro.org>
These can be simplified to and/andc, avoiding the load of
the zero into a register.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/i386/tcg-target-con-set.h | 2 +-
tcg/i386/tcg-target-con-str.h | 1 +
tcg/i386/tcg-target.c.inc | 32 +++++++++++++++++++++++++-------
3 files changed, 27 insertions(+), 8 deletions(-)
diff --git a/tcg/i386/tcg-target-con-set.h b/tcg/i386/tcg-target-con-set.h
index da4411d96b..06e6521001 100644
--- a/tcg/i386/tcg-target-con-set.h
+++ b/tcg/i386/tcg-target-con-set.h
@@ -50,7 +50,7 @@ C_N1_I2(r, r, r)
C_N1_I2(r, r, rW)
C_O1_I3(x, 0, x, x)
C_O1_I3(x, x, x, x)
-C_O1_I4(x, x, x, x, x)
+C_O1_I4(x, x, x, xO, x)
C_O1_I4(r, r, reT, r, 0)
C_O1_I4(r, r, r, ri, ri)
C_O2_I1(r, r, L)
diff --git a/tcg/i386/tcg-target-con-str.h b/tcg/i386/tcg-target-con-str.h
index cc22db227b..52142ab121 100644
--- a/tcg/i386/tcg-target-con-str.h
+++ b/tcg/i386/tcg-target-con-str.h
@@ -28,6 +28,7 @@ REGS('s', ALL_BYTEL_REGS & ~SOFTMMU_RESERVE_REGS) /* qemu_st8_i32 data */
*/
CONST('e', TCG_CT_CONST_S32)
CONST('I', TCG_CT_CONST_I32)
+CONST('O', TCG_CT_CONST_ZERO)
CONST('T', TCG_CT_CONST_TST)
CONST('W', TCG_CT_CONST_WSZ)
CONST('Z', TCG_CT_CONST_U32)
diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index a04dc7d270..210389955d 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -133,6 +133,7 @@ static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
#define TCG_CT_CONST_I32 0x400
#define TCG_CT_CONST_WSZ 0x800
#define TCG_CT_CONST_TST 0x1000
+#define TCG_CT_CONST_ZERO 0x2000
/* Registers used with L constraint, which are the first argument
registers on x86_64, and two random call clobbered registers on
@@ -226,6 +227,9 @@ static bool tcg_target_const_match(int64_t val, int ct,
if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
return 1;
}
+ if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
+ return 1;
+ }
return 0;
}
@@ -3119,13 +3123,27 @@ static void tcg_out_cmpsel_vec(TCGContext *s, TCGType type, unsigned vece,
TCGReg v0, TCGReg c1, TCGReg c2,
TCGReg v3, TCGReg v4, TCGCond cond)
{
- if (tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond)) {
- TCGReg swap = v3;
- v3 = v4;
- v4 = swap;
+ bool inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
+
+ /*
+ * Since XMM0 is 16, the only way we get 0 into V3
+ * is via the constant zero constraint.
+ */
+ if (!v3) {
+ if (inv) {
+ tcg_out_vex_modrm_type(s, OPC_PAND, v0, TCG_TMP_VEC, v4, type);
+ } else {
+ tcg_out_vex_modrm_type(s, OPC_PANDN, v0, TCG_TMP_VEC, v4, type);
+ }
+ } else {
+ if (inv) {
+ TCGReg swap = v3;
+ v3 = v4;
+ v4 = swap;
+ }
+ tcg_out_vex_modrm_type(s, OPC_VPBLENDVB, v0, v4, v3, type);
+ tcg_out8(s, (TCG_TMP_VEC - TCG_REG_XMM0) << 4);
}
- tcg_out_vex_modrm_type(s, OPC_VPBLENDVB, v0, v4, v3, type);
- tcg_out8(s, (TCG_TMP_VEC - TCG_REG_XMM0) << 4);
}
static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
@@ -3716,7 +3734,7 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
case INDEX_op_bitsel_vec:
return C_O1_I3(x, x, x, x);
case INDEX_op_cmpsel_vec:
- return C_O1_I4(x, x, x, x, x);
+ return C_O1_I4(x, x, x, xO, x);
default:
g_assert_not_reached();
--
2.43.0
next prev parent reply other threads:[~2024-09-22 12:04 UTC|newest]
Thread overview: 33+ messages / expand[flat|nested] mbox.gz Atom feed top
2024-09-22 12:00 [PULL 00/31] tcg + linux-user patch queue Richard Henderson
2024-09-22 12:00 ` [PULL 01/31] tcg: Return TCGOp from tcg_gen_op[1-6] Richard Henderson
2024-09-22 12:00 ` [PULL 02/31] tcg: Propagate new TCGOp to add_as_label_use Richard Henderson
2024-09-22 12:00 ` [PULL 03/31] tcg: Fix iteration step in 32-bit gvec operation Richard Henderson
2024-09-22 12:00 ` [PULL 04/31] tcg: Export vec_gen_6 Richard Henderson
2024-09-22 12:00 ` [PULL 05/31] tcg/i386: Split out tcg_out_vex_modrm_type Richard Henderson
2024-09-22 12:00 ` [PULL 06/31] tcg/i386: Do not expand cmp_vec early Richard Henderson
2024-09-22 12:00 ` [PULL 07/31] tcg/i386: Do not expand cmpsel_vec early Richard Henderson
2024-09-22 12:00 ` [PULL 08/31] tcg/ppc: Do not expand cmp_vec early Richard Henderson
2024-09-22 12:00 ` [PULL 09/31] tcg/s390x: " Richard Henderson
2024-09-22 12:00 ` [PULL 10/31] tcg/optimize: Fold movcond with true and false values identical Richard Henderson
2024-09-22 12:00 ` [PULL 11/31] tcg/optimize: Optimize cmp_vec and cmpsel_vec Richard Henderson
2024-09-22 12:00 ` [PULL 12/31] tcg/optimize: Optimize bitsel_vec Richard Henderson
2024-09-22 12:00 ` Richard Henderson [this message]
2024-09-22 12:00 ` [PULL 14/31] tcg/i386: Implement cmp_vec with avx512 insns Richard Henderson
2024-09-22 12:00 ` [PULL 15/31] tcg/i386: Add predicate parameters to tcg_out_evex_opc Richard Henderson
2024-09-22 12:00 ` [PULL 16/31] tcg/i386: Implement cmpsel_vec with avx512 insns Richard Henderson
2024-09-22 12:00 ` [PULL 17/31] tcg/i386: Implement vector TST{EQ,NE} for avx512 Richard Henderson
2024-09-22 12:00 ` [PULL 18/31] tcg/ppc: Implement cmpsel_vec Richard Henderson
2024-09-22 12:01 ` [PULL 19/31] tcg/ppc: Optimize cmpsel with constant 0/-1 arguments Richard Henderson
2024-09-22 12:01 ` [PULL 20/31] tcg/s390x: Implement cmpsel_vec Richard Henderson
2024-09-22 12:01 ` [PULL 21/31] tcg/s390x: Optimize cmpsel with constant 0/-1 arguments Richard Henderson
2024-09-22 12:01 ` [PULL 22/31] target/ppc: Fix lxvx/stxvx facility check Richard Henderson
2024-09-22 12:01 ` [PULL 23/31] linux-user: update syscall_nr.h to Linux v6.10 Richard Henderson
2024-09-22 12:01 ` [PULL 24/31] linux-user, mips: update syscall-args-o32.c.inc " Richard Henderson
2024-09-22 12:01 ` [PULL 25/31] linux-user: update syscall.tbl " Richard Henderson
2024-09-22 12:01 ` [PULL 26/31] linux-user,aarch64: move to syscalltbl file Richard Henderson
2024-09-22 12:01 ` [PULL 27/31] linux-user,openrisc: " Richard Henderson
2024-09-22 12:01 ` [PULL 28/31] linux-user,riscv: " Richard Henderson
2024-09-22 12:01 ` [PULL 29/31] linux-user,hexagon: " Richard Henderson
2024-09-22 12:01 ` [PULL 30/31] linux-user,loongarch: " Richard Henderson
2024-09-22 12:01 ` [PULL 31/31] linux-user: update syscall.tbl to Linux v6.11 Richard Henderson
2024-09-27 15:18 ` [PULL 00/31] tcg + linux-user patch queue Peter Maydell
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=20240922120112.5067-14-richard.henderson@linaro.org \
--to=richard.henderson@linaro.org \
--cc=peter.maydell@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).