From: Richard Henderson <rth@twiddle.net>
To: qemu-devel@nongnu.org
Cc: mark.cave-ayland@ilande.co.uk, aurelien@aurel32.net
Subject: [Qemu-devel] [PATCH v2 12/16] tcg-mips: Improve tcg_out_movi for mips64
Date: Mon, 15 Feb 2016 14:42:30 +1100 [thread overview]
Message-ID: <1455507754-8978-13-git-send-email-rth@twiddle.net> (raw)
In-Reply-To: <1455507754-8978-1-git-send-email-rth@twiddle.net>
For r6, the ADDIUPC, ALUIPC, DAHI and DATI instructions can eliminate 2-4
insns off the pre-r6 full 64-bit immediate load sequence.
For pre-r6, NAL can do the same for the special, but common, case of the
return address for the qemu_ld/st slow paths.
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
tcg/mips/tcg-target.c | 117 ++++++++++++++++++++++++++++++++++++++++++++++----
1 file changed, 109 insertions(+), 8 deletions(-)
diff --git a/tcg/mips/tcg-target.c b/tcg/mips/tcg-target.c
index 8fb2ab0..b4239cb 100644
--- a/tcg/mips/tcg-target.c
+++ b/tcg/mips/tcg-target.c
@@ -282,7 +282,9 @@ typedef enum {
OPC_ORI = 015 << 26,
OPC_XORI = 016 << 26,
OPC_LUI = 017 << 26,
+ OPC_AUI = OPC_LUI,
OPC_DADDIU = 031 << 26,
+ OPC_DAUI = 035 << 26,
OPC_LB = 040 << 26,
OPC_LH = 041 << 26,
OPC_LW = 043 << 26,
@@ -362,6 +364,9 @@ typedef enum {
OPC_REGIMM = 001 << 26,
OPC_BLTZ = OPC_REGIMM | (000 << 16),
OPC_BGEZ = OPC_REGIMM | (001 << 16),
+ OPC_DAHI = OPC_REGIMM | (006 << 16),
+ OPC_NAL = OPC_REGIMM | (020 << 16) | 1, /* bltzal zero, .+8 */
+ OPC_DATI = OPC_REGIMM | (036 << 16),
OPC_SPECIAL2 = 034 << 26,
OPC_MUL_R5 = OPC_SPECIAL2 | 002,
@@ -381,6 +386,10 @@ typedef enum {
OPC_SEB = OPC_SPECIAL3 | 02040,
OPC_SEH = OPC_SPECIAL3 | 03040,
+ OPC_PCREL = 073 << 26,
+ OPC_ADDIUPC = OPC_PCREL | (0 << 19),
+ OPC_ALUIPC = OPC_PCREL | (3 << 19) | (7 << 16),
+
/* MIPS r6 doesn't have JR, JALR should be used instead */
OPC_JR = use_mips32r6_instructions ? OPC_JALR : OPC_JR_R5,
@@ -427,6 +436,17 @@ static inline void tcg_out_opc_imm(TCGContext *s, MIPSInsn opc,
tcg_out32(s, inst);
}
+static inline void tcg_out_opc_pc19(TCGContext *s, MIPSInsn opc,
+ TCGReg rs, TCGArg imm)
+{
+ int32_t inst;
+
+ inst = opc;
+ inst |= (rs & 0x1F) << 21;
+ inst |= (imm & 0x7ffff);
+ tcg_out32(s, inst);
+}
+
/*
* Type bitfield
*/
@@ -555,6 +575,9 @@ static inline void tcg_out_mov(TCGContext *s, TCGType type,
static void tcg_out_movi(TCGContext *s, TCGType type,
TCGReg ret, tcg_target_long arg)
{
+ uintptr_t pc = (uintptr_t)s->code_ptr;
+ intptr_t disp;
+
if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I32) {
arg = (int32_t)arg;
}
@@ -566,18 +589,96 @@ static void tcg_out_movi(TCGContext *s, TCGType type,
tcg_out_opc_imm(s, OPC_ORI, ret, TCG_REG_ZERO, arg);
return;
}
+
+ /* PC-relative address load, part 1. The out-of-line slow paths
+ of qemu_ld/st compute the "return address" of the in-line fast
+ path, so that we can find the guest instruction that triggered
+ the memory fault. Here, we can do that in one instruction. */
+ if (use_mips32r6_instructions && (arg & 3) == 0) {
+ disp = arg - pc;
+ if (disp == sextract32(disp, 0, 21)) {
+ tcg_out_opc_pc19(s, OPC_ADDIUPC, ret, disp >> 2);
+ return;
+ }
+ }
+
if (TCG_TARGET_REG_BITS == 32 || arg == (int32_t)arg) {
tcg_out_opc_imm(s, OPC_LUI, ret, TCG_REG_ZERO, arg >> 16);
- } else {
- tcg_out_movi(s, TCG_TYPE_I32, ret, arg >> 31 >> 1);
- if (arg & 0xffff0000ull) {
- tcg_out_dsll(s, ret, ret, 16);
- tcg_out_opc_imm(s, OPC_ORI, ret, ret, arg >> 16);
- tcg_out_dsll(s, ret, ret, 16);
- } else {
- tcg_out_dsll(s, ret, ret, 32);
+ goto do_lo16;
+ }
+
+ if (use_mips32r6_instructions) {
+ tcg_target_long tmp;
+ TCGReg in;
+
+ /* PC-relative address load, part 2. Here we are able to compute
+ a +- 2GB relative address in two instructions. This not likely
+ to be within code_gen_buffer, but a helper function address or
+ other host memory that just happens to be in range. */
+ disp = sextract64(arg, 16, 48) - sextract64(pc, 16, 48);
+ if (disp == (int16_t)disp) {
+ tcg_out_opc_imm(s, OPC_ALUIPC, 0, ret, disp);
+ goto do_lo16;
+ }
+
+ /* The R6 manual recommends construction of immediates in
+ order of low to high (ADDI, AUI, DAHI, DATI) in order
+ to simplify hardware recognizing these sequences. */
+
+ in = TCG_REG_ZERO;
+ tmp = (int16_t)arg;
+ if (tmp) {
+ tcg_out_opc_imm(s, OPC_ADDIU, ret, in, tmp);
+ in = ret;
}
+ arg = (arg - tmp) >> 16;
+ tmp = (int16_t)arg;
+
+ /* Note that DAHI and DATI only have one register operand,
+ and are thus we must put a zero low part in place. Also
+ note that we already eliminated simple 32-bit constants
+ so we know this must happen. */
+ if (tmp || in != ret) {
+ tcg_out_opc_imm(s, OPC_AUI, ret, in, tmp);
+ }
+ arg = (arg - tmp) >> 16;
+ tmp = (int16_t)arg;
+
+ if (tmp) {
+ tcg_out_opc_imm(s, OPC_DAHI, 0, ret, tmp);
+ }
+ arg = (arg - tmp) >> 16;
+ tcg_debug_assert(arg == (int16_t)arg);
+
+ if (arg) {
+ tcg_out_opc_imm(s, OPC_DATI, 0, ret, arg);
+ }
+ return;
+ }
+
+ /* PC-relative address load, part 3. For mips64 pre-r6, we can use
+ NAL (nop and link, aka BLTZAL with a false condition) to load the
+ return address register with pc+8. This lets us compute the
+ return address for qemu_ld/st only in 2 insns. */
+ disp = arg - (pc + 8);
+ if (disp == (int16_t)disp) {
+ tcg_out32(s, OPC_NAL);
+ tcg_out_opc_imm(s, OPC_DADDIU, ret, TCG_REG_RA, disp);
+ return;
+ }
+
+ /* Last resort: To build a full 64-bit constant, we load the high
+ 32 bits, then shift and or in the low 32 bits. This may take
+ up to 6 instructions. */
+ tcg_out_movi(s, TCG_TYPE_I32, ret, arg >> 31 >> 1);
+ if (arg & 0xffff0000ull) {
+ tcg_out_dsll(s, ret, ret, 16);
+ tcg_out_opc_imm(s, OPC_ORI, ret, ret, arg >> 16);
+ tcg_out_dsll(s, ret, ret, 16);
+ } else {
+ tcg_out_dsll(s, ret, ret, 32);
}
+ do_lo16:
if (arg & 0xffff) {
tcg_out_opc_imm(s, OPC_ORI, ret, ret, arg & 0xffff);
}
--
2.5.0
next prev parent reply other threads:[~2016-02-15 3:43 UTC|newest]
Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top
2016-02-15 3:42 [Qemu-devel] [PATCH v2 00/16] tcg mips64 and mips r6 improvements Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 01/16] tcg-mips: Always use tcg_debug_assert Richard Henderson
2016-02-28 23:51 ` Aurelien Jarno
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 02/16] tcg-mips: Move bswap code to a subroutine Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 03/16] tcg-mips: Add mips64 opcodes Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 04/16] tcg-mips: Support 64-bit opcodes Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 05/16] tcg-mips: Add bswap32u and bswap64 Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 06/16] tcg-mips: Adjust move functions for mips64 Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 07/16] tcg-mips: Adjust load/store " Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 08/16] tcg-mips: Adjust prologue " Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 09/16] tcg-mips: Add tcg unwind info Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 10/16] tcg-mips: Adjust qemu_ld/st for mips64 Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 11/16] tcg-mips: Adjust calling conventions " Richard Henderson
2016-02-15 3:42 ` Richard Henderson [this message]
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 13/16] tcg-mips: Use mips64r6 instructions in tcg_out_ldst Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 14/16] tcg-mips: Use mips64r6 instructions in constant addition Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 15/16] tcg-mips: Use mipsr6 instructions in branches Richard Henderson
2016-02-15 3:42 ` [Qemu-devel] [PATCH v2 16/16] tcg-mips: Use mipsr6 instructions in calls Richard Henderson
2016-02-15 3:47 ` [Qemu-devel] [PATCH v2 00/16] tcg mips64 and mips r6 improvements Richard Henderson
2016-02-28 23:51 ` Aurelien Jarno
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=1455507754-8978-13-git-send-email-rth@twiddle.net \
--to=rth@twiddle.net \
--cc=aurelien@aurel32.net \
--cc=mark.cave-ayland@ilande.co.uk \
--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).