* [Qemu-devel] [PATCH 01/20] target/arm: Hoist address increment for vector memory ops
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 02/20] target/arm: Don't call tcg_clear_temp_count Richard Henderson
` (19 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell, Richard Henderson
From: Richard Henderson <rth@twiddle.net>
This can reduce the number of opcodes required for certain
complex forms of load-multiple (e.g. ld4.16b).
Signed-off-by: Richard Henderson <rth@twiddle.net>
---
target/arm/translate-a64.c | 12 ++++++++----
1 file changed, 8 insertions(+), 4 deletions(-)
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index c403b12eb9..76b5ca3606 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -3012,7 +3012,7 @@ static void disas_ldst_multiple_struct(DisasContext *s, uint32_t insn)
bool is_store = !extract32(insn, 22, 1);
bool is_postidx = extract32(insn, 23, 1);
bool is_q = extract32(insn, 30, 1);
- TCGv_i64 tcg_addr, tcg_rn;
+ TCGv_i64 tcg_addr, tcg_rn, tcg_ebytes;
int ebytes = 1 << size;
int elements = (is_q ? 128 : 64) / (8 << size);
@@ -3077,6 +3077,7 @@ static void disas_ldst_multiple_struct(DisasContext *s, uint32_t insn)
tcg_rn = cpu_reg_sp(s, rn);
tcg_addr = tcg_temp_new_i64();
tcg_gen_mov_i64(tcg_addr, tcg_rn);
+ tcg_ebytes = tcg_const_i64(ebytes);
for (r = 0; r < rpt; r++) {
int e;
@@ -3101,7 +3102,7 @@ static void disas_ldst_multiple_struct(DisasContext *s, uint32_t insn)
clear_vec_high(s, is_q, tt);
}
}
- tcg_gen_addi_i64(tcg_addr, tcg_addr, ebytes);
+ tcg_gen_add_i64(tcg_addr, tcg_addr, tcg_ebytes);
tt = (tt + 1) % 32;
}
}
@@ -3115,6 +3116,7 @@ static void disas_ldst_multiple_struct(DisasContext *s, uint32_t insn)
tcg_gen_add_i64(tcg_rn, tcg_rn, cpu_reg(s, rm));
}
}
+ tcg_temp_free_i64(tcg_ebytes);
tcg_temp_free_i64(tcg_addr);
}
@@ -3157,7 +3159,7 @@ static void disas_ldst_single_struct(DisasContext *s, uint32_t insn)
bool replicate = false;
int index = is_q << 3 | S << 2 | size;
int ebytes, xs;
- TCGv_i64 tcg_addr, tcg_rn;
+ TCGv_i64 tcg_addr, tcg_rn, tcg_ebytes;
switch (scale) {
case 3:
@@ -3210,6 +3212,7 @@ static void disas_ldst_single_struct(DisasContext *s, uint32_t insn)
tcg_rn = cpu_reg_sp(s, rn);
tcg_addr = tcg_temp_new_i64();
tcg_gen_mov_i64(tcg_addr, tcg_rn);
+ tcg_ebytes = tcg_const_i64(ebytes);
for (xs = 0; xs < selem; xs++) {
if (replicate) {
@@ -3252,7 +3255,7 @@ static void disas_ldst_single_struct(DisasContext *s, uint32_t insn)
do_vec_st(s, rt, index, tcg_addr, scale);
}
}
- tcg_gen_addi_i64(tcg_addr, tcg_addr, ebytes);
+ tcg_gen_add_i64(tcg_addr, tcg_addr, tcg_ebytes);
rt = (rt + 1) % 32;
}
@@ -3264,6 +3267,7 @@ static void disas_ldst_single_struct(DisasContext *s, uint32_t insn)
tcg_gen_add_i64(tcg_rn, tcg_rn, cpu_reg(s, rm));
}
}
+ tcg_temp_free_i64(tcg_ebytes);
tcg_temp_free_i64(tcg_addr);
}
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 02/20] target/arm: Don't call tcg_clear_temp_count
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 01/20] target/arm: Hoist address increment for vector memory ops Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 23:35 ` Philippe Mathieu-Daudé
2018-10-11 20:51 ` [Qemu-devel] [PATCH 03/20] target/arm: Use tcg_gen_gvec_dup_i64 for LD[1-4]R Richard Henderson
` (18 subsequent siblings)
20 siblings, 1 reply; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
This is done generically in translator_loop.
Reported-by: Laurent Desnogues <laurent.desnogues@gmail.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate-a64.c | 1 -
target/arm/translate.c | 1 -
2 files changed, 2 deletions(-)
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 76b5ca3606..ac9723c1b9 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -13899,7 +13899,6 @@ static void aarch64_tr_init_disas_context(DisasContextBase *dcbase,
static void aarch64_tr_tb_start(DisasContextBase *db, CPUState *cpu)
{
- tcg_clear_temp_count();
}
static void aarch64_tr_insn_start(DisasContextBase *dcbase, CPUState *cpu)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 426db7828a..736880ee71 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -12701,7 +12701,6 @@ static void arm_tr_tb_start(DisasContextBase *dcbase, CPUState *cpu)
tcg_gen_movi_i32(tmp, 0);
store_cpu_field(tmp, condexec_bits);
}
- tcg_clear_temp_count();
}
static void arm_tr_insn_start(DisasContextBase *dcbase, CPUState *cpu)
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* Re: [Qemu-devel] [PATCH 02/20] target/arm: Don't call tcg_clear_temp_count
2018-10-11 20:51 ` [Qemu-devel] [PATCH 02/20] target/arm: Don't call tcg_clear_temp_count Richard Henderson
@ 2018-10-11 23:35 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 29+ messages in thread
From: Philippe Mathieu-Daudé @ 2018-10-11 23:35 UTC (permalink / raw)
To: Richard Henderson, qemu-devel; +Cc: peter.maydell
On 11/10/2018 22:51, Richard Henderson wrote:
> This is done generically in translator_loop.
>
> Reported-by: Laurent Desnogues <laurent.desnogues@gmail.com>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
> ---
> target/arm/translate-a64.c | 1 -
> target/arm/translate.c | 1 -
> 2 files changed, 2 deletions(-)
>
> diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
> index 76b5ca3606..ac9723c1b9 100644
> --- a/target/arm/translate-a64.c
> +++ b/target/arm/translate-a64.c
> @@ -13899,7 +13899,6 @@ static void aarch64_tr_init_disas_context(DisasContextBase *dcbase,
>
> static void aarch64_tr_tb_start(DisasContextBase *db, CPUState *cpu)
> {
> - tcg_clear_temp_count();
> }
>
> static void aarch64_tr_insn_start(DisasContextBase *dcbase, CPUState *cpu)
> diff --git a/target/arm/translate.c b/target/arm/translate.c
> index 426db7828a..736880ee71 100644
> --- a/target/arm/translate.c
> +++ b/target/arm/translate.c
> @@ -12701,7 +12701,6 @@ static void arm_tr_tb_start(DisasContextBase *dcbase, CPUState *cpu)
> tcg_gen_movi_i32(tmp, 0);
> store_cpu_field(tmp, condexec_bits);
> }
> - tcg_clear_temp_count();
> }
>
> static void arm_tr_insn_start(DisasContextBase *dcbase, CPUState *cpu)
>
^ permalink raw reply [flat|nested] 29+ messages in thread
* [Qemu-devel] [PATCH 03/20] target/arm: Use tcg_gen_gvec_dup_i64 for LD[1-4]R
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 01/20] target/arm: Hoist address increment for vector memory ops Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 02/20] target/arm: Don't call tcg_clear_temp_count Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 04/20] target/arm: Promote consecutive memory ops for aa64 Richard Henderson
` (17 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate-a64.c | 28 +++-------------------------
1 file changed, 3 insertions(+), 25 deletions(-)
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index ac9723c1b9..fff99ca303 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -3217,36 +3217,14 @@ static void disas_ldst_single_struct(DisasContext *s, uint32_t insn)
for (xs = 0; xs < selem; xs++) {
if (replicate) {
/* Load and replicate to all elements */
- uint64_t mulconst;
TCGv_i64 tcg_tmp = tcg_temp_new_i64();
tcg_gen_qemu_ld_i64(tcg_tmp, tcg_addr,
get_mem_index(s), s->be_data + scale);
- switch (scale) {
- case 0:
- mulconst = 0x0101010101010101ULL;
- break;
- case 1:
- mulconst = 0x0001000100010001ULL;
- break;
- case 2:
- mulconst = 0x0000000100000001ULL;
- break;
- case 3:
- mulconst = 0;
- break;
- default:
- g_assert_not_reached();
- }
- if (mulconst) {
- tcg_gen_muli_i64(tcg_tmp, tcg_tmp, mulconst);
- }
- write_vec_element(s, tcg_tmp, rt, 0, MO_64);
- if (is_q) {
- write_vec_element(s, tcg_tmp, rt, 1, MO_64);
- }
+ tcg_gen_gvec_dup_i64(scale, vec_full_reg_offset(s, rt),
+ (is_q + 1) * 8, vec_full_reg_size(s),
+ tcg_tmp);
tcg_temp_free_i64(tcg_tmp);
- clear_vec_high(s, is_q, rt);
} else {
/* Load/store one element per register */
if (is_load) {
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 04/20] target/arm: Promote consecutive memory ops for aa64
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (2 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 03/20] target/arm: Use tcg_gen_gvec_dup_i64 for LD[1-4]R Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 05/20] target/arm: Mark some arrays const Richard Henderson
` (16 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
For a sequence of loads or stores from a single register,
little-endian operations can be promoted to an 8-byte op.
This can reduce the number of operations by a factor of 8.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate-a64.c | 66 +++++++++++++++++++++++---------------
1 file changed, 40 insertions(+), 26 deletions(-)
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index fff99ca303..2f4041462e 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -1200,25 +1200,23 @@ static void write_vec_element_i32(DisasContext *s, TCGv_i32 tcg_src,
/* Store from vector register to memory */
static void do_vec_st(DisasContext *s, int srcidx, int element,
- TCGv_i64 tcg_addr, int size)
+ TCGv_i64 tcg_addr, int size, TCGMemOp endian)
{
- TCGMemOp memop = s->be_data + size;
TCGv_i64 tcg_tmp = tcg_temp_new_i64();
read_vec_element(s, tcg_tmp, srcidx, element, size);
- tcg_gen_qemu_st_i64(tcg_tmp, tcg_addr, get_mem_index(s), memop);
+ tcg_gen_qemu_st_i64(tcg_tmp, tcg_addr, get_mem_index(s), endian | size);
tcg_temp_free_i64(tcg_tmp);
}
/* Load from memory to vector register */
static void do_vec_ld(DisasContext *s, int destidx, int element,
- TCGv_i64 tcg_addr, int size)
+ TCGv_i64 tcg_addr, int size, TCGMemOp endian)
{
- TCGMemOp memop = s->be_data + size;
TCGv_i64 tcg_tmp = tcg_temp_new_i64();
- tcg_gen_qemu_ld_i64(tcg_tmp, tcg_addr, get_mem_index(s), memop);
+ tcg_gen_qemu_ld_i64(tcg_tmp, tcg_addr, get_mem_index(s), endian | size);
write_vec_element(s, tcg_tmp, destidx, element, size);
tcg_temp_free_i64(tcg_tmp);
@@ -3013,9 +3011,10 @@ static void disas_ldst_multiple_struct(DisasContext *s, uint32_t insn)
bool is_postidx = extract32(insn, 23, 1);
bool is_q = extract32(insn, 30, 1);
TCGv_i64 tcg_addr, tcg_rn, tcg_ebytes;
+ TCGMemOp endian = s->be_data;
- int ebytes = 1 << size;
- int elements = (is_q ? 128 : 64) / (8 << size);
+ int ebytes; /* bytes per element */
+ int elements; /* elements per vector */
int rpt; /* num iterations */
int selem; /* structure elements */
int r;
@@ -3074,6 +3073,20 @@ static void disas_ldst_multiple_struct(DisasContext *s, uint32_t insn)
gen_check_sp_alignment(s);
}
+ /* For our purposes, bytes are always little-endian. */
+ if (size == 0) {
+ endian = MO_LE;
+ }
+
+ /* Consecutive little-endian elements from a single register
+ * can be promoted to a larger little-endian operation.
+ */
+ if (selem == 1 && endian == MO_LE) {
+ size = 3;
+ }
+ ebytes = 1 << size;
+ elements = (is_q ? 16 : 8) / ebytes;
+
tcg_rn = cpu_reg_sp(s, rn);
tcg_addr = tcg_temp_new_i64();
tcg_gen_mov_i64(tcg_addr, tcg_rn);
@@ -3082,32 +3095,33 @@ static void disas_ldst_multiple_struct(DisasContext *s, uint32_t insn)
for (r = 0; r < rpt; r++) {
int e;
for (e = 0; e < elements; e++) {
- int tt = (rt + r) % 32;
int xs;
for (xs = 0; xs < selem; xs++) {
+ int tt = (rt + r + xs) % 32;
if (is_store) {
- do_vec_st(s, tt, e, tcg_addr, size);
+ do_vec_st(s, tt, e, tcg_addr, size, endian);
} else {
- do_vec_ld(s, tt, e, tcg_addr, size);
-
- /* For non-quad operations, setting a slice of the low
- * 64 bits of the register clears the high 64 bits (in
- * the ARM ARM pseudocode this is implicit in the fact
- * that 'rval' is a 64 bit wide variable).
- * For quad operations, we might still need to zero the
- * high bits of SVE. We optimize by noticing that we only
- * need to do this the first time we touch a register.
- */
- if (e == 0 && (r == 0 || xs == selem - 1)) {
- clear_vec_high(s, is_q, tt);
- }
+ do_vec_ld(s, tt, e, tcg_addr, size, endian);
}
tcg_gen_add_i64(tcg_addr, tcg_addr, tcg_ebytes);
- tt = (tt + 1) % 32;
}
}
}
+ if (!is_store) {
+ /* For non-quad operations, setting a slice of the low
+ * 64 bits of the register clears the high 64 bits (in
+ * the ARM ARM pseudocode this is implicit in the fact
+ * that 'rval' is a 64 bit wide variable).
+ * For quad operations, we might still need to zero the
+ * high bits of SVE.
+ */
+ for (r = 0; r < rpt * selem; r++) {
+ int tt = (rt + r) % 32;
+ clear_vec_high(s, is_q, tt);
+ }
+ }
+
if (is_postidx) {
int rm = extract32(insn, 16, 5);
if (rm == 31) {
@@ -3228,9 +3242,9 @@ static void disas_ldst_single_struct(DisasContext *s, uint32_t insn)
} else {
/* Load/store one element per register */
if (is_load) {
- do_vec_ld(s, rt, index, tcg_addr, scale);
+ do_vec_ld(s, rt, index, tcg_addr, scale, s->be_data);
} else {
- do_vec_st(s, rt, index, tcg_addr, scale);
+ do_vec_st(s, rt, index, tcg_addr, scale, s->be_data);
}
}
tcg_gen_add_i64(tcg_addr, tcg_addr, tcg_ebytes);
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 05/20] target/arm: Mark some arrays const
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (3 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 04/20] target/arm: Promote consecutive memory ops for aa64 Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 23:34 ` Philippe Mathieu-Daudé
2018-10-19 13:05 ` Peter Maydell
2018-10-11 20:51 ` [Qemu-devel] [PATCH 06/20] target/arm: Use gvec for NEON VDUP Richard Henderson
` (15 subsequent siblings)
20 siblings, 2 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 736880ee71..d59ffa1c67 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -72,7 +72,7 @@ static TCGv_i64 cpu_F0d, cpu_F1d;
#include "exec/gen-icount.h"
-static const char *regnames[] =
+static const char * const regnames[] =
{ "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7",
"r8", "r9", "r10", "r11", "r12", "r13", "r14", "pc" };
@@ -4907,7 +4907,7 @@ static struct {
int nregs;
int interleave;
int spacing;
-} neon_ls_element_type[11] = {
+} const neon_ls_element_type[11] = {
{4, 4, 1},
{4, 4, 2},
{4, 1, 1},
@@ -13089,7 +13089,7 @@ void gen_intermediate_code(CPUState *cpu, TranslationBlock *tb)
translator_loop(ops, &dc.base, cpu, tb);
}
-static const char *cpu_mode_names[16] = {
+static const char * const cpu_mode_names[16] = {
"usr", "fiq", "irq", "svc", "???", "???", "mon", "abt",
"???", "???", "hyp", "und", "???", "???", "???", "sys"
};
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* Re: [Qemu-devel] [PATCH 05/20] target/arm: Mark some arrays const
2018-10-11 20:51 ` [Qemu-devel] [PATCH 05/20] target/arm: Mark some arrays const Richard Henderson
@ 2018-10-11 23:34 ` Philippe Mathieu-Daudé
2018-10-19 13:05 ` Peter Maydell
1 sibling, 0 replies; 29+ messages in thread
From: Philippe Mathieu-Daudé @ 2018-10-11 23:34 UTC (permalink / raw)
To: Richard Henderson, qemu-devel; +Cc: peter.maydell
On 11/10/2018 22:51, Richard Henderson wrote:
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
> ---
> target/arm/translate.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/target/arm/translate.c b/target/arm/translate.c
> index 736880ee71..d59ffa1c67 100644
> --- a/target/arm/translate.c
> +++ b/target/arm/translate.c
> @@ -72,7 +72,7 @@ static TCGv_i64 cpu_F0d, cpu_F1d;
>
> #include "exec/gen-icount.h"
>
> -static const char *regnames[] =
> +static const char * const regnames[] =
> { "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7",
> "r8", "r9", "r10", "r11", "r12", "r13", "r14", "pc" };
>
> @@ -4907,7 +4907,7 @@ static struct {
> int nregs;
> int interleave;
> int spacing;
> -} neon_ls_element_type[11] = {
> +} const neon_ls_element_type[11] = {
> {4, 4, 1},
> {4, 4, 2},
> {4, 1, 1},
> @@ -13089,7 +13089,7 @@ void gen_intermediate_code(CPUState *cpu, TranslationBlock *tb)
> translator_loop(ops, &dc.base, cpu, tb);
> }
>
> -static const char *cpu_mode_names[16] = {
> +static const char * const cpu_mode_names[16] = {
> "usr", "fiq", "irq", "svc", "???", "???", "mon", "abt",
> "???", "???", "hyp", "und", "???", "???", "???", "sys"
> };
>
^ permalink raw reply [flat|nested] 29+ messages in thread* Re: [Qemu-devel] [PATCH 05/20] target/arm: Mark some arrays const
2018-10-11 20:51 ` [Qemu-devel] [PATCH 05/20] target/arm: Mark some arrays const Richard Henderson
2018-10-11 23:34 ` Philippe Mathieu-Daudé
@ 2018-10-19 13:05 ` Peter Maydell
1 sibling, 0 replies; 29+ messages in thread
From: Peter Maydell @ 2018-10-19 13:05 UTC (permalink / raw)
To: Richard Henderson; +Cc: QEMU Developers
On 11 October 2018 at 21:51, Richard Henderson
<richard.henderson@linaro.org> wrote:
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> target/arm/translate.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/target/arm/translate.c b/target/arm/translate.c
> index 736880ee71..d59ffa1c67 100644
> --- a/target/arm/translate.c
> +++ b/target/arm/translate.c
> @@ -72,7 +72,7 @@ static TCGv_i64 cpu_F0d, cpu_F1d;
>
> #include "exec/gen-icount.h"
>
> -static const char *regnames[] =
> +static const char * const regnames[] =
> { "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7",
> "r8", "r9", "r10", "r11", "r12", "r13", "r14", "pc" };
>
> @@ -4907,7 +4907,7 @@ static struct {
> int nregs;
> int interleave;
> int spacing;
> -} neon_ls_element_type[11] = {
> +} const neon_ls_element_type[11] = {
> {4, 4, 1},
> {4, 4, 2},
> {4, 1, 1},
> @@ -13089,7 +13089,7 @@ void gen_intermediate_code(CPUState *cpu, TranslationBlock *tb)
> translator_loop(ops, &dc.base, cpu, tb);
> }
>
> -static const char *cpu_mode_names[16] = {
> +static const char * const cpu_mode_names[16] = {
> "usr", "fiq", "irq", "svc", "???", "???", "mon", "abt",
> "???", "???", "hyp", "und", "???", "???", "???", "sys"
> };
I dropped the last hunk here as one of my other patches
has moved this array (and fixed the constness in passing).
thanks
-- PMM
^ permalink raw reply [flat|nested] 29+ messages in thread
* [Qemu-devel] [PATCH 06/20] target/arm: Use gvec for NEON VDUP
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (4 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 05/20] target/arm: Mark some arrays const Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 07/20] target/arm: Use gvec for NEON VMOV, VMVN, VBIC & VORR (immediate) Richard Henderson
` (14 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Also introduces neon_element_offset to find the env offset
of a specific element within a neon register.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 63 ++++++++++++++++++++++++------------------
1 file changed, 36 insertions(+), 27 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index d59ffa1c67..4ac526e298 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -1585,6 +1585,25 @@ neon_reg_offset (int reg, int n)
return vfp_reg_offset(0, sreg);
}
+/* Return the offset of a 2**SIZE piece of a NEON register, at index ELE,
+ * where 0 is the least significant end of the register.
+ */
+static inline long
+neon_element_offset(int reg, int element, TCGMemOp size)
+{
+ int element_size = 1 << size;
+ int ofs = element * element_size;
+#ifdef HOST_WORDS_BIGENDIAN
+ /* Calculate the offset assuming fully little-endian,
+ * then XOR to account for the order of the 8-byte units.
+ */
+ if (element_size < 8) {
+ ofs ^= 8 - element_size;
+ }
+#endif
+ return neon_reg_offset(reg, 0) + ofs;
+}
+
static TCGv_i32 neon_load_reg(int reg, int pass)
{
TCGv_i32 tmp = tcg_temp_new_i32();
@@ -3432,17 +3451,10 @@ static int disas_vfp_insn(DisasContext *s, uint32_t insn)
tmp = load_reg(s, rd);
if (insn & (1 << 23)) {
/* VDUP */
- if (size == 0) {
- gen_neon_dup_u8(tmp, 0);
- } else if (size == 1) {
- gen_neon_dup_low16(tmp);
- }
- for (n = 0; n <= pass * 2; n++) {
- tmp2 = tcg_temp_new_i32();
- tcg_gen_mov_i32(tmp2, tmp);
- neon_store_reg(rn, n, tmp2);
- }
- neon_store_reg(rn, n, tmp);
+ int vec_size = pass ? 16 : 8;
+ tcg_gen_gvec_dup_i32(size, neon_reg_offset(rn, 0),
+ vec_size, vec_size, tmp);
+ tcg_temp_free_i32(tmp);
} else {
/* VMOV */
switch (size) {
@@ -7755,28 +7767,25 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tcg_temp_free_i32(tmp);
} else if ((insn & 0x380) == 0) {
/* VDUP */
+ int element;
+ TCGMemOp size;
+
if ((insn & (7 << 16)) == 0 || (q && (rd & 1))) {
return 1;
}
- if (insn & (1 << 19)) {
- tmp = neon_load_reg(rm, 1);
- } else {
- tmp = neon_load_reg(rm, 0);
- }
if (insn & (1 << 16)) {
- gen_neon_dup_u8(tmp, ((insn >> 17) & 3) * 8);
+ size = MO_8;
+ element = (insn >> 17) & 7;
} else if (insn & (1 << 17)) {
- if ((insn >> 18) & 1)
- gen_neon_dup_high16(tmp);
- else
- gen_neon_dup_low16(tmp);
+ size = MO_16;
+ element = (insn >> 18) & 3;
+ } else {
+ size = MO_32;
+ element = (insn >> 19) & 1;
}
- for (pass = 0; pass < (q ? 4 : 2); pass++) {
- tmp2 = tcg_temp_new_i32();
- tcg_gen_mov_i32(tmp2, tmp);
- neon_store_reg(rd, pass, tmp2);
- }
- tcg_temp_free_i32(tmp);
+ tcg_gen_gvec_dup_mem(size, neon_reg_offset(rd, 0),
+ neon_element_offset(rm, element, size),
+ q ? 16 : 8, q ? 16 : 8);
} else {
return 1;
}
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 07/20] target/arm: Use gvec for NEON VMOV, VMVN, VBIC & VORR (immediate)
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (5 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 06/20] target/arm: Use gvec for NEON VDUP Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 08/20] target/arm: Use gvec for NEON_3R_LOGIC insns Richard Henderson
` (13 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 67 ++++++++++++++++++++++++------------------
1 file changed, 39 insertions(+), 28 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 4ac526e298..109689a286 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -6641,7 +6641,8 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
return 1;
}
} else { /* (insn & 0x00380080) == 0 */
- int invert;
+ int invert, reg_ofs, vec_size;
+
if (q && (rd & 1)) {
return 1;
}
@@ -6681,8 +6682,9 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
break;
case 14:
imm |= (imm << 8) | (imm << 16) | (imm << 24);
- if (invert)
+ if (invert) {
imm = ~imm;
+ }
break;
case 15:
if (invert) {
@@ -6692,36 +6694,45 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
| ((imm & 0x40) ? (0x1f << 25) : (1 << 30));
break;
}
- if (invert)
+ if (invert) {
imm = ~imm;
+ }
- for (pass = 0; pass < (q ? 4 : 2); pass++) {
- if (op & 1 && op < 12) {
- tmp = neon_load_reg(rd, pass);
- if (invert) {
- /* The immediate value has already been inverted, so
- BIC becomes AND. */
- tcg_gen_andi_i32(tmp, tmp, imm);
- } else {
- tcg_gen_ori_i32(tmp, tmp, imm);
- }
+ reg_ofs = neon_reg_offset(rd, 0);
+ vec_size = q ? 16 : 8;
+
+ if (op & 1 && op < 12) {
+ if (invert) {
+ /* The immediate value has already been inverted,
+ * so BIC becomes AND.
+ */
+ tcg_gen_gvec_andi(MO_32, reg_ofs, reg_ofs, imm,
+ vec_size, vec_size);
} else {
- /* VMOV, VMVN. */
- tmp = tcg_temp_new_i32();
- if (op == 14 && invert) {
- int n;
- uint32_t val;
- val = 0;
- for (n = 0; n < 4; n++) {
- if (imm & (1 << (n + (pass & 1) * 4)))
- val |= 0xff << (n * 8);
- }
- tcg_gen_movi_i32(tmp, val);
- } else {
- tcg_gen_movi_i32(tmp, imm);
- }
+ tcg_gen_gvec_ori(MO_32, reg_ofs, reg_ofs, imm,
+ vec_size, vec_size);
+ }
+ } else {
+ /* VMOV, VMVN. */
+ if (op == 14 && invert) {
+ TCGv_i64 t64 = tcg_temp_new_i64();
+
+ for (pass = 0; pass <= q; ++pass) {
+ uint64_t val = 0;
+ int n;
+
+ for (n = 0; n < 8; n++) {
+ if (imm & (1 << (n + pass * 8))) {
+ val |= 0xffull << (n * 8);
+ }
+ }
+ tcg_gen_movi_i64(t64, val);
+ neon_store_reg64(t64, rd + pass);
+ }
+ tcg_temp_free_i64(t64);
+ } else {
+ tcg_gen_gvec_dup32i(reg_ofs, vec_size, vec_size, imm);
}
- neon_store_reg(rd, pass, tmp);
}
}
} else { /* (insn & 0x00800010 == 0x00800000) */
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 08/20] target/arm: Use gvec for NEON_3R_LOGIC insns
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (6 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 07/20] target/arm: Use gvec for NEON VMOV, VMVN, VBIC & VORR (immediate) Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 09/20] target/arm: Use gvec for NEON_3R_VADD_VSUB insns Richard Henderson
` (12 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Move expanders for VBSL, VBIT, and VBIF from translate-a64.c.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.h | 6 +-
target/arm/translate-a64.c | 61 --------------
target/arm/translate.c | 162 +++++++++++++++++++++++++++----------
3 files changed, 123 insertions(+), 106 deletions(-)
diff --git a/target/arm/translate.h b/target/arm/translate.h
index ad911de98c..7cc2685104 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -191,6 +191,11 @@ static inline TCGv_i32 get_ahp_flag(void)
return ret;
}
+/* Vector operations shared between ARM and AArch64. */
+extern const GVecGen3 bsl_op;
+extern const GVecGen3 bit_op;
+extern const GVecGen3 bif_op;
+
#define FORWARD_FEATURE(NAME) \
static inline bool aa32_dc_feature_##NAME(DisasContext *dc) \
{ return aa32_feature_##NAME(dc->cpu); }
@@ -209,5 +214,4 @@ FORWARD_FEATURE(dp)
FORWARD_FEATURE(fp16_arith)
#undef FORWARD_FEATURE
-
#endif /* TARGET_ARM_TRANSLATE_H */
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 2f4041462e..a115546a6a 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -10392,70 +10392,9 @@ static void disas_simd_three_reg_diff(DisasContext *s, uint32_t insn)
}
}
-static void gen_bsl_i64(TCGv_i64 rd, TCGv_i64 rn, TCGv_i64 rm)
-{
- tcg_gen_xor_i64(rn, rn, rm);
- tcg_gen_and_i64(rn, rn, rd);
- tcg_gen_xor_i64(rd, rm, rn);
-}
-
-static void gen_bit_i64(TCGv_i64 rd, TCGv_i64 rn, TCGv_i64 rm)
-{
- tcg_gen_xor_i64(rn, rn, rd);
- tcg_gen_and_i64(rn, rn, rm);
- tcg_gen_xor_i64(rd, rd, rn);
-}
-
-static void gen_bif_i64(TCGv_i64 rd, TCGv_i64 rn, TCGv_i64 rm)
-{
- tcg_gen_xor_i64(rn, rn, rd);
- tcg_gen_andc_i64(rn, rn, rm);
- tcg_gen_xor_i64(rd, rd, rn);
-}
-
-static void gen_bsl_vec(unsigned vece, TCGv_vec rd, TCGv_vec rn, TCGv_vec rm)
-{
- tcg_gen_xor_vec(vece, rn, rn, rm);
- tcg_gen_and_vec(vece, rn, rn, rd);
- tcg_gen_xor_vec(vece, rd, rm, rn);
-}
-
-static void gen_bit_vec(unsigned vece, TCGv_vec rd, TCGv_vec rn, TCGv_vec rm)
-{
- tcg_gen_xor_vec(vece, rn, rn, rd);
- tcg_gen_and_vec(vece, rn, rn, rm);
- tcg_gen_xor_vec(vece, rd, rd, rn);
-}
-
-static void gen_bif_vec(unsigned vece, TCGv_vec rd, TCGv_vec rn, TCGv_vec rm)
-{
- tcg_gen_xor_vec(vece, rn, rn, rd);
- tcg_gen_andc_vec(vece, rn, rn, rm);
- tcg_gen_xor_vec(vece, rd, rd, rn);
-}
-
/* Logic op (opcode == 3) subgroup of C3.6.16. */
static void disas_simd_3same_logic(DisasContext *s, uint32_t insn)
{
- static const GVecGen3 bsl_op = {
- .fni8 = gen_bsl_i64,
- .fniv = gen_bsl_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true
- };
- static const GVecGen3 bit_op = {
- .fni8 = gen_bit_i64,
- .fniv = gen_bit_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true
- };
- static const GVecGen3 bif_op = {
- .fni8 = gen_bif_i64,
- .fniv = gen_bif_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true
- };
-
int rd = extract32(insn, 0, 5);
int rn = extract32(insn, 5, 5);
int rm = extract32(insn, 16, 5);
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 109689a286..4ab9f69b01 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5262,14 +5262,6 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
return 0;
}
-/* Bitwise select. dest = c ? t : f. Clobbers T and F. */
-static void gen_neon_bsl(TCGv_i32 dest, TCGv_i32 t, TCGv_i32 f, TCGv_i32 c)
-{
- tcg_gen_and_i32(t, t, c);
- tcg_gen_andc_i32(f, f, c);
- tcg_gen_or_i32(dest, t, f);
-}
-
static inline void gen_neon_narrow(int size, TCGv_i32 dest, TCGv_i64 src)
{
switch (size) {
@@ -5712,6 +5704,73 @@ static int do_v81_helper(DisasContext *s, gen_helper_gvec_3_ptr *fn,
return 1;
}
+/*
+ * Expanders for VBitOps_VBIF, VBIT, VBSL.
+ */
+static void gen_bsl_i64(TCGv_i64 rd, TCGv_i64 rn, TCGv_i64 rm)
+{
+ tcg_gen_xor_i64(rn, rn, rm);
+ tcg_gen_and_i64(rn, rn, rd);
+ tcg_gen_xor_i64(rd, rm, rn);
+}
+
+static void gen_bit_i64(TCGv_i64 rd, TCGv_i64 rn, TCGv_i64 rm)
+{
+ tcg_gen_xor_i64(rn, rn, rd);
+ tcg_gen_and_i64(rn, rn, rm);
+ tcg_gen_xor_i64(rd, rd, rn);
+}
+
+static void gen_bif_i64(TCGv_i64 rd, TCGv_i64 rn, TCGv_i64 rm)
+{
+ tcg_gen_xor_i64(rn, rn, rd);
+ tcg_gen_andc_i64(rn, rn, rm);
+ tcg_gen_xor_i64(rd, rd, rn);
+}
+
+static void gen_bsl_vec(unsigned vece, TCGv_vec rd, TCGv_vec rn, TCGv_vec rm)
+{
+ tcg_gen_xor_vec(vece, rn, rn, rm);
+ tcg_gen_and_vec(vece, rn, rn, rd);
+ tcg_gen_xor_vec(vece, rd, rm, rn);
+}
+
+static void gen_bit_vec(unsigned vece, TCGv_vec rd, TCGv_vec rn, TCGv_vec rm)
+{
+ tcg_gen_xor_vec(vece, rn, rn, rd);
+ tcg_gen_and_vec(vece, rn, rn, rm);
+ tcg_gen_xor_vec(vece, rd, rd, rn);
+}
+
+static void gen_bif_vec(unsigned vece, TCGv_vec rd, TCGv_vec rn, TCGv_vec rm)
+{
+ tcg_gen_xor_vec(vece, rn, rn, rd);
+ tcg_gen_andc_vec(vece, rn, rn, rm);
+ tcg_gen_xor_vec(vece, rd, rd, rn);
+}
+
+const GVecGen3 bsl_op = {
+ .fni8 = gen_bsl_i64,
+ .fniv = gen_bsl_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true
+};
+
+const GVecGen3 bit_op = {
+ .fni8 = gen_bit_i64,
+ .fniv = gen_bit_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true
+};
+
+const GVecGen3 bif_op = {
+ .fni8 = gen_bif_i64,
+ .fniv = gen_bif_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true
+};
+
+
/* Translate a NEON data processing instruction. Return nonzero if the
instruction is invalid.
We process data in a mixture of 32-bit and 64-bit chunks.
@@ -5721,13 +5780,14 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
{
int op;
int q;
- int rd, rn, rm;
+ int rd, rn, rm, rd_ofs, rn_ofs, rm_ofs;
int size;
int shift;
int pass;
int count;
int pairwise;
int u;
+ int vec_size;
uint32_t imm, mask;
TCGv_i32 tmp, tmp2, tmp3, tmp4, tmp5;
TCGv_ptr ptr1, ptr2, ptr3;
@@ -5751,6 +5811,11 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
VFP_DREG_N(rn, insn);
VFP_DREG_M(rm, insn);
size = (insn >> 20) & 3;
+ vec_size = q ? 16 : 8;
+ rd_ofs = neon_reg_offset(rd, 0);
+ rn_ofs = neon_reg_offset(rn, 0);
+ rm_ofs = neon_reg_offset(rm, 0);
+
if ((insn & (1 << 23)) == 0) {
/* Three register same length. */
op = ((insn >> 7) & 0x1e) | ((insn >> 4) & 1);
@@ -5841,8 +5906,51 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
q, rd, rn, rm);
}
return 1;
+
+ case NEON_3R_LOGIC: /* Logic ops. */
+ switch ((u << 2) | size) {
+ case 0: /* VAND */
+ tcg_gen_gvec_and(0, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ break;
+ case 1: /* VBIC */
+ tcg_gen_gvec_andc(0, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ break;
+ case 2:
+ if (rn == rm) {
+ /* VMOV */
+ tcg_gen_gvec_mov(0, rd_ofs, rn_ofs, vec_size, vec_size);
+ } else {
+ /* VORR */
+ tcg_gen_gvec_or(0, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ }
+ break;
+ case 3: /* VORN */
+ tcg_gen_gvec_orc(0, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ break;
+ case 4: /* VEOR */
+ tcg_gen_gvec_xor(0, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ break;
+ case 5: /* VBSL */
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size, &bsl_op);
+ break;
+ case 6: /* VBIT */
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size, &bit_op);
+ break;
+ case 7: /* VBIF */
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size, &bif_op);
+ break;
+ }
+ return 0;
}
- if (size == 3 && op != NEON_3R_LOGIC) {
+ if (size == 3) {
/* 64-bit element instructions. */
for (pass = 0; pass < (q ? 2 : 1); pass++) {
neon_load_reg64(cpu_V0, rn + pass);
@@ -6000,40 +6108,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
case NEON_3R_VRHADD:
GEN_NEON_INTEGER_OP(rhadd);
break;
- case NEON_3R_LOGIC: /* Logic ops. */
- switch ((u << 2) | size) {
- case 0: /* VAND */
- tcg_gen_and_i32(tmp, tmp, tmp2);
- break;
- case 1: /* BIC */
- tcg_gen_andc_i32(tmp, tmp, tmp2);
- break;
- case 2: /* VORR */
- tcg_gen_or_i32(tmp, tmp, tmp2);
- break;
- case 3: /* VORN */
- tcg_gen_orc_i32(tmp, tmp, tmp2);
- break;
- case 4: /* VEOR */
- tcg_gen_xor_i32(tmp, tmp, tmp2);
- break;
- case 5: /* VBSL */
- tmp3 = neon_load_reg(rd, pass);
- gen_neon_bsl(tmp, tmp, tmp2, tmp3);
- tcg_temp_free_i32(tmp3);
- break;
- case 6: /* VBIT */
- tmp3 = neon_load_reg(rd, pass);
- gen_neon_bsl(tmp, tmp, tmp3, tmp2);
- tcg_temp_free_i32(tmp3);
- break;
- case 7: /* VBIF */
- tmp3 = neon_load_reg(rd, pass);
- gen_neon_bsl(tmp, tmp3, tmp, tmp2);
- tcg_temp_free_i32(tmp3);
- break;
- }
- break;
case NEON_3R_VHSUB:
GEN_NEON_INTEGER_OP(hsub);
break;
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 09/20] target/arm: Use gvec for NEON_3R_VADD_VSUB insns
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (7 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 08/20] target/arm: Use gvec for NEON_3R_LOGIC insns Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 10/20] target/arm: Use gvec for NEON_2RM_VMN, NEON_2RM_VNEG Richard Henderson
` (11 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 29 ++++++++++-------------------
1 file changed, 10 insertions(+), 19 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 4ab9f69b01..4dcd7123e9 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5949,6 +5949,16 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
break;
}
return 0;
+
+ case NEON_3R_VADD_VSUB:
+ if (u) {
+ tcg_gen_gvec_sub(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ } else {
+ tcg_gen_gvec_add(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ }
+ return 0;
}
if (size == 3) {
/* 64-bit element instructions. */
@@ -6006,13 +6016,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
cpu_V1, cpu_V0);
}
break;
- case NEON_3R_VADD_VSUB:
- if (u) {
- tcg_gen_sub_i64(CPU_V001);
- } else {
- tcg_gen_add_i64(CPU_V001);
- }
- break;
default:
abort();
}
@@ -6147,18 +6150,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tmp2 = neon_load_reg(rd, pass);
gen_neon_add(size, tmp, tmp2);
break;
- case NEON_3R_VADD_VSUB:
- if (!u) { /* VADD */
- gen_neon_add(size, tmp, tmp2);
- } else { /* VSUB */
- switch (size) {
- case 0: gen_helper_neon_sub_u8(tmp, tmp, tmp2); break;
- case 1: gen_helper_neon_sub_u16(tmp, tmp, tmp2); break;
- case 2: tcg_gen_sub_i32(tmp, tmp, tmp2); break;
- default: abort();
- }
- }
- break;
case NEON_3R_VTST_VCEQ:
if (!u) { /* VTST */
switch (size) {
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 10/20] target/arm: Use gvec for NEON_2RM_VMN, NEON_2RM_VNEG
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (8 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 09/20] target/arm: Use gvec for NEON_3R_VADD_VSUB insns Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 11/20] target/arm: Use gvec for NEON_3R_VMUL Richard Henderson
` (10 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 16 ++++++++--------
1 file changed, 8 insertions(+), 8 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 4dcd7123e9..fae132791a 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -7501,6 +7501,14 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tcg_temp_free_ptr(ptr1);
tcg_temp_free_ptr(ptr2);
break;
+
+ case NEON_2RM_VMVN:
+ tcg_gen_gvec_not(0, rd_ofs, rm_ofs, vec_size, vec_size);
+ break;
+ case NEON_2RM_VNEG:
+ tcg_gen_gvec_neg(size, rd_ofs, rm_ofs, vec_size, vec_size);
+ break;
+
default:
elementwise:
for (pass = 0; pass < (q ? 4 : 2); pass++) {
@@ -7541,9 +7549,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
case NEON_2RM_VCNT:
gen_helper_neon_cnt_u8(tmp, tmp);
break;
- case NEON_2RM_VMVN:
- tcg_gen_not_i32(tmp, tmp);
- break;
case NEON_2RM_VQABS:
switch (size) {
case 0:
@@ -7616,11 +7621,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
default: abort();
}
break;
- case NEON_2RM_VNEG:
- tmp2 = tcg_const_i32(0);
- gen_neon_rsb(size, tmp, tmp2);
- tcg_temp_free_i32(tmp2);
- break;
case NEON_2RM_VCGT0_F:
{
TCGv_ptr fpstatus = get_fpstatus_ptr(1);
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 11/20] target/arm: Use gvec for NEON_3R_VMUL
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (9 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 10/20] target/arm: Use gvec for NEON_2RM_VMN, NEON_2RM_VNEG Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 12/20] target/arm: Use gvec for VSHR, VSHL Richard Henderson
` (9 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 31 +++++++++++++++----------------
1 file changed, 15 insertions(+), 16 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index fae132791a..e195efcc55 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5959,6 +5959,19 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
vec_size, vec_size);
}
return 0;
+
+ case NEON_3R_VMUL: /* VMUL */
+ if (u) {
+ /* Polynomial case allows only P8 and is handled below. */
+ if (size != 0) {
+ return 1;
+ }
+ } else {
+ tcg_gen_gvec_mul(size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ return 0;
+ }
+ break;
}
if (size == 3) {
/* 64-bit element instructions. */
@@ -6065,12 +6078,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
return 1;
}
break;
- case NEON_3R_VMUL:
- if (u && (size != 0)) {
- /* UNDEF on invalid size for polynomial subcase */
- return 1;
- }
- break;
case NEON_3R_VFM_VQRDMLSH:
if (!arm_dc_feature(s, ARM_FEATURE_VFP4)) {
return 1;
@@ -6183,16 +6190,8 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
}
break;
case NEON_3R_VMUL:
- if (u) { /* polynomial */
- gen_helper_neon_mul_p8(tmp, tmp, tmp2);
- } else { /* Integer */
- switch (size) {
- case 0: gen_helper_neon_mul_u8(tmp, tmp, tmp2); break;
- case 1: gen_helper_neon_mul_u16(tmp, tmp, tmp2); break;
- case 2: tcg_gen_mul_i32(tmp, tmp, tmp2); break;
- default: abort();
- }
- }
+ /* VMUL.P8; other cases already eliminated. */
+ gen_helper_neon_mul_p8(tmp, tmp, tmp2);
break;
case NEON_3R_VPMAX:
GEN_NEON_INTEGER_OP(pmax);
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 12/20] target/arm: Use gvec for VSHR, VSHL
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (10 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 11/20] target/arm: Use gvec for NEON_3R_VMUL Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:51 ` [Qemu-devel] [PATCH 13/20] target/arm: Use gvec for VSRA Richard Henderson
` (8 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 70 +++++++++++++++++++++++++++++-------------
1 file changed, 48 insertions(+), 22 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index e195efcc55..a16f323d52 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -6373,8 +6373,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
size--;
}
shift = (insn >> 16) & ((1 << (3 + size)) - 1);
- /* To avoid excessive duplication of ops we implement shift
- by immediate using the variable shift operations. */
if (op < 8) {
/* Shift by immediate:
VSHR, VSRA, VRSHR, VRSRA, VSRI, VSHL, VQSHL, VQSHLU. */
@@ -6386,37 +6384,62 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
}
/* Right shifts are encoded as N - shift, where N is the
element size in bits. */
- if (op <= 4)
+ if (op <= 4) {
shift = shift - (1 << (size + 3));
+ }
+
+ switch (op) {
+ case 0: /* VSHR */
+ /* Right shift comes here negative. */
+ shift = -shift;
+ /* Shifts larger than the element size are architecturally
+ * valid. Unsigned results in all zeros; signed results
+ * in all sign bits.
+ */
+ if (!u) {
+ tcg_gen_gvec_sari(size, rd_ofs, rm_ofs,
+ MIN(shift, (8 << size) - 1),
+ vec_size, vec_size);
+ } else if (shift >= 8 << size) {
+ tcg_gen_gvec_dup8i(rd_ofs, vec_size, vec_size, 0);
+ } else {
+ tcg_gen_gvec_shri(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
+ }
+ return 0;
+
+ case 5: /* VSHL, VSLI */
+ if (!u) { /* VSHL */
+ /* Shifts larger than the element size are
+ * architecturally valid and results in zero.
+ */
+ if (shift >= 8 << size) {
+ tcg_gen_gvec_dup8i(rd_ofs, vec_size, vec_size, 0);
+ } else {
+ tcg_gen_gvec_shli(size, rd_ofs, rm_ofs, shift,
+ vec_size, vec_size);
+ }
+ return 0;
+ }
+ break;
+ }
+
if (size == 3) {
count = q + 1;
} else {
count = q ? 4: 2;
}
- switch (size) {
- case 0:
- imm = (uint8_t) shift;
- imm |= imm << 8;
- imm |= imm << 16;
- break;
- case 1:
- imm = (uint16_t) shift;
- imm |= imm << 16;
- break;
- case 2:
- case 3:
- imm = shift;
- break;
- default:
- abort();
- }
+
+ /* To avoid excessive duplication of ops we implement shift
+ * by immediate using the variable shift operations.
+ */
+ imm = dup_const(size, shift);
for (pass = 0; pass < count; pass++) {
if (size == 3) {
neon_load_reg64(cpu_V0, rm + pass);
tcg_gen_movi_i64(cpu_V1, imm);
switch (op) {
- case 0: /* VSHR */
case 1: /* VSRA */
if (u)
gen_helper_neon_shl_u64(cpu_V0, cpu_V0, cpu_V1);
@@ -6447,6 +6470,8 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
cpu_V0, cpu_V1);
}
break;
+ default:
+ g_assert_not_reached();
}
if (op == 1 || op == 3) {
/* Accumulate. */
@@ -6475,7 +6500,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tmp2 = tcg_temp_new_i32();
tcg_gen_movi_i32(tmp2, imm);
switch (op) {
- case 0: /* VSHR */
case 1: /* VSRA */
GEN_NEON_INTEGER_OP(shl);
break;
@@ -6513,6 +6537,8 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
case 7: /* VQSHL */
GEN_NEON_INTEGER_OP_ENV(qshl);
break;
+ default:
+ g_assert_not_reached();
}
tcg_temp_free_i32(tmp2);
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 13/20] target/arm: Use gvec for VSRA
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (11 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 12/20] target/arm: Use gvec for VSHR, VSHL Richard Henderson
@ 2018-10-11 20:51 ` Richard Henderson
2018-10-11 20:52 ` [Qemu-devel] [PATCH 14/20] target/arm: Use gvec for VSRI, VSLI Richard Henderson
` (7 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:51 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Move ssra_op and usra_op expanders from translate-a64.c.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.h | 2 +
target/arm/translate-a64.c | 106 ----------------------------
target/arm/translate.c | 139 ++++++++++++++++++++++++++++++++++---
3 files changed, 130 insertions(+), 117 deletions(-)
diff --git a/target/arm/translate.h b/target/arm/translate.h
index 7cc2685104..8487b349ef 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -195,6 +195,8 @@ static inline TCGv_i32 get_ahp_flag(void)
extern const GVecGen3 bsl_op;
extern const GVecGen3 bit_op;
extern const GVecGen3 bif_op;
+extern const GVecGen2i ssra_op[4];
+extern const GVecGen2i usra_op[4];
#define FORWARD_FEATURE(NAME) \
static inline bool aa32_dc_feature_##NAME(DisasContext *dc) \
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index a115546a6a..1a9ae6d5f4 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -9392,66 +9392,6 @@ static void disas_simd_scalar_two_reg_misc(DisasContext *s, uint32_t insn)
}
}
-static void gen_ssra8_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- tcg_gen_vec_sar8i_i64(a, a, shift);
- tcg_gen_vec_add8_i64(d, d, a);
-}
-
-static void gen_ssra16_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- tcg_gen_vec_sar16i_i64(a, a, shift);
- tcg_gen_vec_add16_i64(d, d, a);
-}
-
-static void gen_ssra32_i32(TCGv_i32 d, TCGv_i32 a, int32_t shift)
-{
- tcg_gen_sari_i32(a, a, shift);
- tcg_gen_add_i32(d, d, a);
-}
-
-static void gen_ssra64_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- tcg_gen_sari_i64(a, a, shift);
- tcg_gen_add_i64(d, d, a);
-}
-
-static void gen_ssra_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
-{
- tcg_gen_sari_vec(vece, a, a, sh);
- tcg_gen_add_vec(vece, d, d, a);
-}
-
-static void gen_usra8_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- tcg_gen_vec_shr8i_i64(a, a, shift);
- tcg_gen_vec_add8_i64(d, d, a);
-}
-
-static void gen_usra16_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- tcg_gen_vec_shr16i_i64(a, a, shift);
- tcg_gen_vec_add16_i64(d, d, a);
-}
-
-static void gen_usra32_i32(TCGv_i32 d, TCGv_i32 a, int32_t shift)
-{
- tcg_gen_shri_i32(a, a, shift);
- tcg_gen_add_i32(d, d, a);
-}
-
-static void gen_usra64_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- tcg_gen_shri_i64(a, a, shift);
- tcg_gen_add_i64(d, d, a);
-}
-
-static void gen_usra_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
-{
- tcg_gen_shri_vec(vece, a, a, sh);
- tcg_gen_add_vec(vece, d, d, a);
-}
-
static void gen_shr8_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
{
uint64_t mask = dup_const(MO_8, 0xff >> shift);
@@ -9507,52 +9447,6 @@ static void gen_shr_ins_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
static void handle_vec_simd_shri(DisasContext *s, bool is_q, bool is_u,
int immh, int immb, int opcode, int rn, int rd)
{
- static const GVecGen2i ssra_op[4] = {
- { .fni8 = gen_ssra8_i64,
- .fniv = gen_ssra_vec,
- .load_dest = true,
- .opc = INDEX_op_sari_vec,
- .vece = MO_8 },
- { .fni8 = gen_ssra16_i64,
- .fniv = gen_ssra_vec,
- .load_dest = true,
- .opc = INDEX_op_sari_vec,
- .vece = MO_16 },
- { .fni4 = gen_ssra32_i32,
- .fniv = gen_ssra_vec,
- .load_dest = true,
- .opc = INDEX_op_sari_vec,
- .vece = MO_32 },
- { .fni8 = gen_ssra64_i64,
- .fniv = gen_ssra_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .opc = INDEX_op_sari_vec,
- .vece = MO_64 },
- };
- static const GVecGen2i usra_op[4] = {
- { .fni8 = gen_usra8_i64,
- .fniv = gen_usra_vec,
- .load_dest = true,
- .opc = INDEX_op_shri_vec,
- .vece = MO_8, },
- { .fni8 = gen_usra16_i64,
- .fniv = gen_usra_vec,
- .load_dest = true,
- .opc = INDEX_op_shri_vec,
- .vece = MO_16, },
- { .fni4 = gen_usra32_i32,
- .fniv = gen_usra_vec,
- .load_dest = true,
- .opc = INDEX_op_shri_vec,
- .vece = MO_32, },
- { .fni8 = gen_usra64_i64,
- .fniv = gen_usra_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .opc = INDEX_op_shri_vec,
- .vece = MO_64, },
- };
static const GVecGen2i sri_op[4] = {
{ .fni8 = gen_shr8_ins_i64,
.fniv = gen_shr_ins_vec,
diff --git a/target/arm/translate.c b/target/arm/translate.c
index a16f323d52..da95851370 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5770,6 +5770,113 @@ const GVecGen3 bif_op = {
.load_dest = true
};
+static void gen_ssra8_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ tcg_gen_vec_sar8i_i64(a, a, shift);
+ tcg_gen_vec_add8_i64(d, d, a);
+}
+
+static void gen_ssra16_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ tcg_gen_vec_sar16i_i64(a, a, shift);
+ tcg_gen_vec_add16_i64(d, d, a);
+}
+
+static void gen_ssra32_i32(TCGv_i32 d, TCGv_i32 a, int32_t shift)
+{
+ tcg_gen_sari_i32(a, a, shift);
+ tcg_gen_add_i32(d, d, a);
+}
+
+static void gen_ssra64_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ tcg_gen_sari_i64(a, a, shift);
+ tcg_gen_add_i64(d, d, a);
+}
+
+static void gen_ssra_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
+{
+ tcg_gen_sari_vec(vece, a, a, sh);
+ tcg_gen_add_vec(vece, d, d, a);
+}
+
+const GVecGen2i ssra_op[4] = {
+ { .fni8 = gen_ssra8_i64,
+ .fniv = gen_ssra_vec,
+ .load_dest = true,
+ .opc = INDEX_op_sari_vec,
+ .vece = MO_8 },
+ { .fni8 = gen_ssra16_i64,
+ .fniv = gen_ssra_vec,
+ .load_dest = true,
+ .opc = INDEX_op_sari_vec,
+ .vece = MO_16 },
+ { .fni4 = gen_ssra32_i32,
+ .fniv = gen_ssra_vec,
+ .load_dest = true,
+ .opc = INDEX_op_sari_vec,
+ .vece = MO_32 },
+ { .fni8 = gen_ssra64_i64,
+ .fniv = gen_ssra_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .opc = INDEX_op_sari_vec,
+ .vece = MO_64 },
+};
+
+static void gen_usra8_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ tcg_gen_vec_shr8i_i64(a, a, shift);
+ tcg_gen_vec_add8_i64(d, d, a);
+}
+
+static void gen_usra16_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ tcg_gen_vec_shr16i_i64(a, a, shift);
+ tcg_gen_vec_add16_i64(d, d, a);
+}
+
+static void gen_usra32_i32(TCGv_i32 d, TCGv_i32 a, int32_t shift)
+{
+ tcg_gen_shri_i32(a, a, shift);
+ tcg_gen_add_i32(d, d, a);
+}
+
+static void gen_usra64_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ tcg_gen_shri_i64(a, a, shift);
+ tcg_gen_add_i64(d, d, a);
+}
+
+static void gen_usra_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
+{
+ tcg_gen_shri_vec(vece, a, a, sh);
+ tcg_gen_add_vec(vece, d, d, a);
+}
+
+const GVecGen2i usra_op[4] = {
+ { .fni8 = gen_usra8_i64,
+ .fniv = gen_usra_vec,
+ .load_dest = true,
+ .opc = INDEX_op_shri_vec,
+ .vece = MO_8, },
+ { .fni8 = gen_usra16_i64,
+ .fniv = gen_usra_vec,
+ .load_dest = true,
+ .opc = INDEX_op_shri_vec,
+ .vece = MO_16, },
+ { .fni4 = gen_usra32_i32,
+ .fniv = gen_usra_vec,
+ .load_dest = true,
+ .opc = INDEX_op_shri_vec,
+ .vece = MO_32, },
+ { .fni8 = gen_usra64_i64,
+ .fniv = gen_usra_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .opc = INDEX_op_shri_vec,
+ .vece = MO_64, },
+};
/* Translate a NEON data processing instruction. Return nonzero if the
instruction is invalid.
@@ -6408,6 +6515,25 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
}
return 0;
+ case 1: /* VSRA */
+ /* Right shift comes here negative. */
+ shift = -shift;
+ /* Shifts larger than the element size are architecturally
+ * valid. Unsigned results in all zeros; signed results
+ * in all sign bits.
+ */
+ if (!u) {
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, vec_size, vec_size,
+ MIN(shift, (8 << size) - 1),
+ &ssra_op[size]);
+ } else if (shift >= 8 << size) {
+ /* rd += 0 */
+ } else {
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, vec_size, vec_size,
+ shift, &usra_op[size]);
+ }
+ return 0;
+
case 5: /* VSHL, VSLI */
if (!u) { /* VSHL */
/* Shifts larger than the element size are
@@ -6440,12 +6566,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
neon_load_reg64(cpu_V0, rm + pass);
tcg_gen_movi_i64(cpu_V1, imm);
switch (op) {
- case 1: /* VSRA */
- if (u)
- gen_helper_neon_shl_u64(cpu_V0, cpu_V0, cpu_V1);
- else
- gen_helper_neon_shl_s64(cpu_V0, cpu_V0, cpu_V1);
- break;
case 2: /* VRSHR */
case 3: /* VRSRA */
if (u)
@@ -6473,7 +6593,7 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
default:
g_assert_not_reached();
}
- if (op == 1 || op == 3) {
+ if (op == 3) {
/* Accumulate. */
neon_load_reg64(cpu_V1, rd + pass);
tcg_gen_add_i64(cpu_V0, cpu_V0, cpu_V1);
@@ -6500,9 +6620,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tmp2 = tcg_temp_new_i32();
tcg_gen_movi_i32(tmp2, imm);
switch (op) {
- case 1: /* VSRA */
- GEN_NEON_INTEGER_OP(shl);
- break;
case 2: /* VRSHR */
case 3: /* VRSRA */
GEN_NEON_INTEGER_OP(rshl);
@@ -6542,7 +6659,7 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
}
tcg_temp_free_i32(tmp2);
- if (op == 1 || op == 3) {
+ if (op == 3) {
/* Accumulate. */
tmp2 = neon_load_reg(rd, pass);
gen_neon_add(size, tmp, tmp2);
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 14/20] target/arm: Use gvec for VSRI, VSLI
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (12 preceding siblings ...)
2018-10-11 20:51 ` [Qemu-devel] [PATCH 13/20] target/arm: Use gvec for VSRA Richard Henderson
@ 2018-10-11 20:52 ` Richard Henderson
2018-10-11 20:52 ` [Qemu-devel] [PATCH 15/20] target/arm: Use gvec for NEON_3R_VML Richard Henderson
` (6 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:52 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Move shi_op and sli_op expanders from translate-a64.c.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.h | 2 +
target/arm/translate-a64.c | 152 +----------------------
target/arm/translate.c | 244 ++++++++++++++++++++++++++-----------
3 files changed, 179 insertions(+), 219 deletions(-)
diff --git a/target/arm/translate.h b/target/arm/translate.h
index 8487b349ef..ef44f0a5e5 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -197,6 +197,8 @@ extern const GVecGen3 bit_op;
extern const GVecGen3 bif_op;
extern const GVecGen2i ssra_op[4];
extern const GVecGen2i usra_op[4];
+extern const GVecGen2i sri_op[4];
+extern const GVecGen2i sli_op[4];
#define FORWARD_FEATURE(NAME) \
static inline bool aa32_dc_feature_##NAME(DisasContext *dc) \
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 1a9ae6d5f4..f47c132e6c 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -9392,85 +9392,10 @@ static void disas_simd_scalar_two_reg_misc(DisasContext *s, uint32_t insn)
}
}
-static void gen_shr8_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- uint64_t mask = dup_const(MO_8, 0xff >> shift);
- TCGv_i64 t = tcg_temp_new_i64();
-
- tcg_gen_shri_i64(t, a, shift);
- tcg_gen_andi_i64(t, t, mask);
- tcg_gen_andi_i64(d, d, ~mask);
- tcg_gen_or_i64(d, d, t);
- tcg_temp_free_i64(t);
-}
-
-static void gen_shr16_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- uint64_t mask = dup_const(MO_16, 0xffff >> shift);
- TCGv_i64 t = tcg_temp_new_i64();
-
- tcg_gen_shri_i64(t, a, shift);
- tcg_gen_andi_i64(t, t, mask);
- tcg_gen_andi_i64(d, d, ~mask);
- tcg_gen_or_i64(d, d, t);
- tcg_temp_free_i64(t);
-}
-
-static void gen_shr32_ins_i32(TCGv_i32 d, TCGv_i32 a, int32_t shift)
-{
- tcg_gen_shri_i32(a, a, shift);
- tcg_gen_deposit_i32(d, d, a, 0, 32 - shift);
-}
-
-static void gen_shr64_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- tcg_gen_shri_i64(a, a, shift);
- tcg_gen_deposit_i64(d, d, a, 0, 64 - shift);
-}
-
-static void gen_shr_ins_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
-{
- uint64_t mask = (2ull << ((8 << vece) - 1)) - 1;
- TCGv_vec t = tcg_temp_new_vec_matching(d);
- TCGv_vec m = tcg_temp_new_vec_matching(d);
-
- tcg_gen_dupi_vec(vece, m, mask ^ (mask >> sh));
- tcg_gen_shri_vec(vece, t, a, sh);
- tcg_gen_and_vec(vece, d, d, m);
- tcg_gen_or_vec(vece, d, d, t);
-
- tcg_temp_free_vec(t);
- tcg_temp_free_vec(m);
-}
-
/* SSHR[RA]/USHR[RA] - Vector shift right (optional rounding/accumulate) */
static void handle_vec_simd_shri(DisasContext *s, bool is_q, bool is_u,
int immh, int immb, int opcode, int rn, int rd)
{
- static const GVecGen2i sri_op[4] = {
- { .fni8 = gen_shr8_ins_i64,
- .fniv = gen_shr_ins_vec,
- .load_dest = true,
- .opc = INDEX_op_shri_vec,
- .vece = MO_8 },
- { .fni8 = gen_shr16_ins_i64,
- .fniv = gen_shr_ins_vec,
- .load_dest = true,
- .opc = INDEX_op_shri_vec,
- .vece = MO_16 },
- { .fni4 = gen_shr32_ins_i32,
- .fniv = gen_shr_ins_vec,
- .load_dest = true,
- .opc = INDEX_op_shri_vec,
- .vece = MO_32 },
- { .fni8 = gen_shr64_ins_i64,
- .fniv = gen_shr_ins_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .opc = INDEX_op_shri_vec,
- .vece = MO_64 },
- };
-
int size = 32 - clz32(immh) - 1;
int immhb = immh << 3 | immb;
int shift = 2 * (8 << size) - immhb;
@@ -9566,85 +9491,10 @@ static void handle_vec_simd_shri(DisasContext *s, bool is_q, bool is_u,
clear_vec_high(s, is_q, rd);
}
-static void gen_shl8_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- uint64_t mask = dup_const(MO_8, 0xff << shift);
- TCGv_i64 t = tcg_temp_new_i64();
-
- tcg_gen_shli_i64(t, a, shift);
- tcg_gen_andi_i64(t, t, mask);
- tcg_gen_andi_i64(d, d, ~mask);
- tcg_gen_or_i64(d, d, t);
- tcg_temp_free_i64(t);
-}
-
-static void gen_shl16_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- uint64_t mask = dup_const(MO_16, 0xffff << shift);
- TCGv_i64 t = tcg_temp_new_i64();
-
- tcg_gen_shli_i64(t, a, shift);
- tcg_gen_andi_i64(t, t, mask);
- tcg_gen_andi_i64(d, d, ~mask);
- tcg_gen_or_i64(d, d, t);
- tcg_temp_free_i64(t);
-}
-
-static void gen_shl32_ins_i32(TCGv_i32 d, TCGv_i32 a, int32_t shift)
-{
- tcg_gen_deposit_i32(d, d, a, shift, 32 - shift);
-}
-
-static void gen_shl64_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
-{
- tcg_gen_deposit_i64(d, d, a, shift, 64 - shift);
-}
-
-static void gen_shl_ins_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
-{
- uint64_t mask = (1ull << sh) - 1;
- TCGv_vec t = tcg_temp_new_vec_matching(d);
- TCGv_vec m = tcg_temp_new_vec_matching(d);
-
- tcg_gen_dupi_vec(vece, m, mask);
- tcg_gen_shli_vec(vece, t, a, sh);
- tcg_gen_and_vec(vece, d, d, m);
- tcg_gen_or_vec(vece, d, d, t);
-
- tcg_temp_free_vec(t);
- tcg_temp_free_vec(m);
-}
-
/* SHL/SLI - Vector shift left */
static void handle_vec_simd_shli(DisasContext *s, bool is_q, bool insert,
int immh, int immb, int opcode, int rn, int rd)
{
- static const GVecGen2i shi_op[4] = {
- { .fni8 = gen_shl8_ins_i64,
- .fniv = gen_shl_ins_vec,
- .opc = INDEX_op_shli_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .vece = MO_8 },
- { .fni8 = gen_shl16_ins_i64,
- .fniv = gen_shl_ins_vec,
- .opc = INDEX_op_shli_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .vece = MO_16 },
- { .fni4 = gen_shl32_ins_i32,
- .fniv = gen_shl_ins_vec,
- .opc = INDEX_op_shli_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .vece = MO_32 },
- { .fni8 = gen_shl64_ins_i64,
- .fniv = gen_shl_ins_vec,
- .opc = INDEX_op_shli_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .vece = MO_64 },
- };
int size = 32 - clz32(immh) - 1;
int immhb = immh << 3 | immb;
int shift = immhb - (8 << size);
@@ -9664,7 +9514,7 @@ static void handle_vec_simd_shli(DisasContext *s, bool is_q, bool insert,
}
if (insert) {
- gen_gvec_op2i(s, is_q, rd, rn, shift, &shi_op[size]);
+ gen_gvec_op2i(s, is_q, rd, rn, shift, &sli_op[size]);
} else {
gen_gvec_fn2i(s, is_q, rd, rn, shift, tcg_gen_gvec_shli, size);
}
diff --git a/target/arm/translate.c b/target/arm/translate.c
index da95851370..104b49543d 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5878,6 +5878,160 @@ const GVecGen2i usra_op[4] = {
.vece = MO_64, },
};
+static void gen_shr8_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ uint64_t mask = dup_const(MO_8, 0xff >> shift);
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_shri_i64(t, a, shift);
+ tcg_gen_andi_i64(t, t, mask);
+ tcg_gen_andi_i64(d, d, ~mask);
+ tcg_gen_or_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_shr16_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ uint64_t mask = dup_const(MO_16, 0xffff >> shift);
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_shri_i64(t, a, shift);
+ tcg_gen_andi_i64(t, t, mask);
+ tcg_gen_andi_i64(d, d, ~mask);
+ tcg_gen_or_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_shr32_ins_i32(TCGv_i32 d, TCGv_i32 a, int32_t shift)
+{
+ tcg_gen_shri_i32(a, a, shift);
+ tcg_gen_deposit_i32(d, d, a, 0, 32 - shift);
+}
+
+static void gen_shr64_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ tcg_gen_shri_i64(a, a, shift);
+ tcg_gen_deposit_i64(d, d, a, 0, 64 - shift);
+}
+
+static void gen_shr_ins_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
+{
+ if (sh == 0) {
+ tcg_gen_mov_vec(d, a);
+ } else {
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+ TCGv_vec m = tcg_temp_new_vec_matching(d);
+
+ tcg_gen_dupi_vec(vece, m, MAKE_64BIT_MASK((8 << vece) - sh, sh));
+ tcg_gen_shri_vec(vece, t, a, sh);
+ tcg_gen_and_vec(vece, d, d, m);
+ tcg_gen_or_vec(vece, d, d, t);
+
+ tcg_temp_free_vec(t);
+ tcg_temp_free_vec(m);
+ }
+}
+
+const GVecGen2i sri_op[4] = {
+ { .fni8 = gen_shr8_ins_i64,
+ .fniv = gen_shr_ins_vec,
+ .load_dest = true,
+ .opc = INDEX_op_shri_vec,
+ .vece = MO_8 },
+ { .fni8 = gen_shr16_ins_i64,
+ .fniv = gen_shr_ins_vec,
+ .load_dest = true,
+ .opc = INDEX_op_shri_vec,
+ .vece = MO_16 },
+ { .fni4 = gen_shr32_ins_i32,
+ .fniv = gen_shr_ins_vec,
+ .load_dest = true,
+ .opc = INDEX_op_shri_vec,
+ .vece = MO_32 },
+ { .fni8 = gen_shr64_ins_i64,
+ .fniv = gen_shr_ins_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .opc = INDEX_op_shri_vec,
+ .vece = MO_64 },
+};
+
+static void gen_shl8_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ uint64_t mask = dup_const(MO_8, 0xff << shift);
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_shli_i64(t, a, shift);
+ tcg_gen_andi_i64(t, t, mask);
+ tcg_gen_andi_i64(d, d, ~mask);
+ tcg_gen_or_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_shl16_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ uint64_t mask = dup_const(MO_16, 0xffff << shift);
+ TCGv_i64 t = tcg_temp_new_i64();
+
+ tcg_gen_shli_i64(t, a, shift);
+ tcg_gen_andi_i64(t, t, mask);
+ tcg_gen_andi_i64(d, d, ~mask);
+ tcg_gen_or_i64(d, d, t);
+ tcg_temp_free_i64(t);
+}
+
+static void gen_shl32_ins_i32(TCGv_i32 d, TCGv_i32 a, int32_t shift)
+{
+ tcg_gen_deposit_i32(d, d, a, shift, 32 - shift);
+}
+
+static void gen_shl64_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
+{
+ tcg_gen_deposit_i64(d, d, a, shift, 64 - shift);
+}
+
+static void gen_shl_ins_vec(unsigned vece, TCGv_vec d, TCGv_vec a, int64_t sh)
+{
+ if (sh == 0) {
+ tcg_gen_mov_vec(d, a);
+ } else {
+ TCGv_vec t = tcg_temp_new_vec_matching(d);
+ TCGv_vec m = tcg_temp_new_vec_matching(d);
+
+ tcg_gen_dupi_vec(vece, m, MAKE_64BIT_MASK(0, sh));
+ tcg_gen_shli_vec(vece, t, a, sh);
+ tcg_gen_and_vec(vece, d, d, m);
+ tcg_gen_or_vec(vece, d, d, t);
+
+ tcg_temp_free_vec(t);
+ tcg_temp_free_vec(m);
+ }
+}
+
+const GVecGen2i sli_op[4] = {
+ { .fni8 = gen_shl8_ins_i64,
+ .fniv = gen_shl_ins_vec,
+ .load_dest = true,
+ .opc = INDEX_op_shli_vec,
+ .vece = MO_8 },
+ { .fni8 = gen_shl16_ins_i64,
+ .fniv = gen_shl_ins_vec,
+ .load_dest = true,
+ .opc = INDEX_op_shli_vec,
+ .vece = MO_16 },
+ { .fni4 = gen_shl32_ins_i32,
+ .fniv = gen_shl_ins_vec,
+ .load_dest = true,
+ .opc = INDEX_op_shli_vec,
+ .vece = MO_32 },
+ { .fni8 = gen_shl64_ins_i64,
+ .fniv = gen_shl_ins_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .opc = INDEX_op_shli_vec,
+ .vece = MO_64 },
+};
+
/* Translate a NEON data processing instruction. Return nonzero if the
instruction is invalid.
We process data in a mixture of 32-bit and 64-bit chunks.
@@ -5895,7 +6049,7 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
int pairwise;
int u;
int vec_size;
- uint32_t imm, mask;
+ uint32_t imm;
TCGv_i32 tmp, tmp2, tmp3, tmp4, tmp5;
TCGv_ptr ptr1, ptr2, ptr3;
TCGv_i64 tmp64;
@@ -6534,8 +6688,27 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
}
return 0;
+ case 4: /* VSRI */
+ if (!u) {
+ return 1;
+ }
+ /* Right shift comes here negative. */
+ shift = -shift;
+ /* Shift out of range leaves destination unchanged. */
+ if (shift < 8 << size) {
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, vec_size, vec_size,
+ shift, &sri_op[size]);
+ }
+ return 0;
+
case 5: /* VSHL, VSLI */
- if (!u) { /* VSHL */
+ if (u) { /* VSLI */
+ /* Shift out of range leaves destination unchanged. */
+ if (shift < 8 << size) {
+ tcg_gen_gvec_2i(rd_ofs, rm_ofs, vec_size,
+ vec_size, shift, &sli_op[size]);
+ }
+ } else { /* VSHL */
/* Shifts larger than the element size are
* architecturally valid and results in zero.
*/
@@ -6545,9 +6718,8 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tcg_gen_gvec_shli(size, rd_ofs, rm_ofs, shift,
vec_size, vec_size);
}
- return 0;
}
- break;
+ return 0;
}
if (size == 3) {
@@ -6573,10 +6745,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
else
gen_helper_neon_rshl_s64(cpu_V0, cpu_V0, cpu_V1);
break;
- case 4: /* VSRI */
- case 5: /* VSHL, VSLI */
- gen_helper_neon_shl_u64(cpu_V0, cpu_V0, cpu_V1);
- break;
case 6: /* VQSHLU */
gen_helper_neon_qshlu_s64(cpu_V0, cpu_env,
cpu_V0, cpu_V1);
@@ -6597,21 +6765,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
/* Accumulate. */
neon_load_reg64(cpu_V1, rd + pass);
tcg_gen_add_i64(cpu_V0, cpu_V0, cpu_V1);
- } else if (op == 4 || (op == 5 && u)) {
- /* Insert */
- neon_load_reg64(cpu_V1, rd + pass);
- uint64_t mask;
- if (shift < -63 || shift > 63) {
- mask = 0;
- } else {
- if (op == 4) {
- mask = 0xffffffffffffffffull >> -shift;
- } else {
- mask = 0xffffffffffffffffull << shift;
- }
- }
- tcg_gen_andi_i64(cpu_V1, cpu_V1, ~mask);
- tcg_gen_or_i64(cpu_V0, cpu_V0, cpu_V1);
}
neon_store_reg64(cpu_V0, rd + pass);
} else { /* size < 3 */
@@ -6624,15 +6777,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
case 3: /* VRSRA */
GEN_NEON_INTEGER_OP(rshl);
break;
- case 4: /* VSRI */
- case 5: /* VSHL, VSLI */
- switch (size) {
- case 0: gen_helper_neon_shl_u8(tmp, tmp, tmp2); break;
- case 1: gen_helper_neon_shl_u16(tmp, tmp, tmp2); break;
- case 2: gen_helper_neon_shl_u32(tmp, tmp, tmp2); break;
- default: abort();
- }
- break;
case 6: /* VQSHLU */
switch (size) {
case 0:
@@ -6664,42 +6808,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tmp2 = neon_load_reg(rd, pass);
gen_neon_add(size, tmp, tmp2);
tcg_temp_free_i32(tmp2);
- } else if (op == 4 || (op == 5 && u)) {
- /* Insert */
- switch (size) {
- case 0:
- if (op == 4)
- mask = 0xff >> -shift;
- else
- mask = (uint8_t)(0xff << shift);
- mask |= mask << 8;
- mask |= mask << 16;
- break;
- case 1:
- if (op == 4)
- mask = 0xffff >> -shift;
- else
- mask = (uint16_t)(0xffff << shift);
- mask |= mask << 16;
- break;
- case 2:
- if (shift < -31 || shift > 31) {
- mask = 0;
- } else {
- if (op == 4)
- mask = 0xffffffffu >> -shift;
- else
- mask = 0xffffffffu << shift;
- }
- break;
- default:
- abort();
- }
- tmp2 = neon_load_reg(rd, pass);
- tcg_gen_andi_i32(tmp, tmp, mask);
- tcg_gen_andi_i32(tmp2, tmp2, ~mask);
- tcg_gen_or_i32(tmp, tmp, tmp2);
- tcg_temp_free_i32(tmp2);
}
neon_store_reg(rd, pass, tmp);
}
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 15/20] target/arm: Use gvec for NEON_3R_VML
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (13 preceding siblings ...)
2018-10-11 20:52 ` [Qemu-devel] [PATCH 14/20] target/arm: Use gvec for VSRI, VSLI Richard Henderson
@ 2018-10-11 20:52 ` Richard Henderson
2018-10-11 20:52 ` [Qemu-devel] [PATCH 16/20] target/arm: Use gvec for NEON_3R_VTST_VCEQ, NEON_3R_VCGT, NEON_3R_VCGE Richard Henderson
` (5 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:52 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Move mla_op and mls_op expanders from translate-a64.c.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.h | 2 +
target/arm/translate-a64.c | 106 -----------------------------
target/arm/translate.c | 134 ++++++++++++++++++++++++++++++++-----
3 files changed, 120 insertions(+), 122 deletions(-)
diff --git a/target/arm/translate.h b/target/arm/translate.h
index ef44f0a5e5..d11d504301 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -195,6 +195,8 @@ static inline TCGv_i32 get_ahp_flag(void)
extern const GVecGen3 bsl_op;
extern const GVecGen3 bit_op;
extern const GVecGen3 bif_op;
+extern const GVecGen3 mla_op[4];
+extern const GVecGen3 mls_op[4];
extern const GVecGen2i ssra_op[4];
extern const GVecGen2i usra_op[4];
extern const GVecGen2i sri_op[4];
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index f47c132e6c..94f7127c0e 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -10410,66 +10410,6 @@ static void disas_simd_3same_float(DisasContext *s, uint32_t insn)
}
}
-static void gen_mla8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
-{
- gen_helper_neon_mul_u8(a, a, b);
- gen_helper_neon_add_u8(d, d, a);
-}
-
-static void gen_mla16_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
-{
- gen_helper_neon_mul_u16(a, a, b);
- gen_helper_neon_add_u16(d, d, a);
-}
-
-static void gen_mla32_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
-{
- tcg_gen_mul_i32(a, a, b);
- tcg_gen_add_i32(d, d, a);
-}
-
-static void gen_mla64_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
-{
- tcg_gen_mul_i64(a, a, b);
- tcg_gen_add_i64(d, d, a);
-}
-
-static void gen_mla_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
-{
- tcg_gen_mul_vec(vece, a, a, b);
- tcg_gen_add_vec(vece, d, d, a);
-}
-
-static void gen_mls8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
-{
- gen_helper_neon_mul_u8(a, a, b);
- gen_helper_neon_sub_u8(d, d, a);
-}
-
-static void gen_mls16_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
-{
- gen_helper_neon_mul_u16(a, a, b);
- gen_helper_neon_sub_u16(d, d, a);
-}
-
-static void gen_mls32_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
-{
- tcg_gen_mul_i32(a, a, b);
- tcg_gen_sub_i32(d, d, a);
-}
-
-static void gen_mls64_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
-{
- tcg_gen_mul_i64(a, a, b);
- tcg_gen_sub_i64(d, d, a);
-}
-
-static void gen_mls_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
-{
- tcg_gen_mul_vec(vece, a, a, b);
- tcg_gen_sub_vec(vece, d, d, a);
-}
-
/* Integer op subgroup of C3.6.16. */
static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
{
@@ -10488,52 +10428,6 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
.prefer_i64 = TCG_TARGET_REG_BITS == 64,
.vece = MO_64 },
};
- static const GVecGen3 mla_op[4] = {
- { .fni4 = gen_mla8_i32,
- .fniv = gen_mla_vec,
- .opc = INDEX_op_mul_vec,
- .load_dest = true,
- .vece = MO_8 },
- { .fni4 = gen_mla16_i32,
- .fniv = gen_mla_vec,
- .opc = INDEX_op_mul_vec,
- .load_dest = true,
- .vece = MO_16 },
- { .fni4 = gen_mla32_i32,
- .fniv = gen_mla_vec,
- .opc = INDEX_op_mul_vec,
- .load_dest = true,
- .vece = MO_32 },
- { .fni8 = gen_mla64_i64,
- .fniv = gen_mla_vec,
- .opc = INDEX_op_mul_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .vece = MO_64 },
- };
- static const GVecGen3 mls_op[4] = {
- { .fni4 = gen_mls8_i32,
- .fniv = gen_mls_vec,
- .opc = INDEX_op_mul_vec,
- .load_dest = true,
- .vece = MO_8 },
- { .fni4 = gen_mls16_i32,
- .fniv = gen_mls_vec,
- .opc = INDEX_op_mul_vec,
- .load_dest = true,
- .vece = MO_16 },
- { .fni4 = gen_mls32_i32,
- .fniv = gen_mls_vec,
- .opc = INDEX_op_mul_vec,
- .load_dest = true,
- .vece = MO_32 },
- { .fni8 = gen_mls64_i64,
- .fniv = gen_mls_vec,
- .opc = INDEX_op_mul_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .load_dest = true,
- .vece = MO_64 },
- };
int is_q = extract32(insn, 30, 1);
int u = extract32(insn, 29, 1);
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 104b49543d..5008a24ffd 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5468,7 +5468,7 @@ static void gen_neon_narrow_op(int op, int u, int size,
#define NEON_3R_VABA 15
#define NEON_3R_VADD_VSUB 16
#define NEON_3R_VTST_VCEQ 17
-#define NEON_3R_VML 18 /* VMLA, VMLAL, VMLS, VMLSL */
+#define NEON_3R_VML 18 /* VMLA, VMLS */
#define NEON_3R_VMUL 19
#define NEON_3R_VPMAX 20
#define NEON_3R_VPMIN 21
@@ -6032,6 +6032,117 @@ const GVecGen2i sli_op[4] = {
.vece = MO_64 },
};
+static void gen_mla8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ gen_helper_neon_mul_u8(a, a, b);
+ gen_helper_neon_add_u8(d, d, a);
+}
+
+static void gen_mls8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ gen_helper_neon_mul_u8(a, a, b);
+ gen_helper_neon_sub_u8(d, d, a);
+}
+
+static void gen_mla16_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ gen_helper_neon_mul_u16(a, a, b);
+ gen_helper_neon_add_u16(d, d, a);
+}
+
+static void gen_mls16_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ gen_helper_neon_mul_u16(a, a, b);
+ gen_helper_neon_sub_u16(d, d, a);
+}
+
+static void gen_mla32_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ tcg_gen_mul_i32(a, a, b);
+ tcg_gen_add_i32(d, d, a);
+}
+
+static void gen_mls32_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ tcg_gen_mul_i32(a, a, b);
+ tcg_gen_sub_i32(d, d, a);
+}
+
+static void gen_mla64_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+ tcg_gen_mul_i64(a, a, b);
+ tcg_gen_add_i64(d, d, a);
+}
+
+static void gen_mls64_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+ tcg_gen_mul_i64(a, a, b);
+ tcg_gen_sub_i64(d, d, a);
+}
+
+static void gen_mla_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
+{
+ tcg_gen_mul_vec(vece, a, a, b);
+ tcg_gen_add_vec(vece, d, d, a);
+}
+
+static void gen_mls_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
+{
+ tcg_gen_mul_vec(vece, a, a, b);
+ tcg_gen_sub_vec(vece, d, d, a);
+}
+
+/* Note that while NEON does not support VMLA and VMLS as 64-bit ops,
+ * these tables are shared with AArch64 which does support them.
+ */
+const GVecGen3 mla_op[4] = {
+ { .fni4 = gen_mla8_i32,
+ .fniv = gen_mla_vec,
+ .opc = INDEX_op_mul_vec,
+ .load_dest = true,
+ .vece = MO_8 },
+ { .fni4 = gen_mla16_i32,
+ .fniv = gen_mla_vec,
+ .opc = INDEX_op_mul_vec,
+ .load_dest = true,
+ .vece = MO_16 },
+ { .fni4 = gen_mla32_i32,
+ .fniv = gen_mla_vec,
+ .opc = INDEX_op_mul_vec,
+ .load_dest = true,
+ .vece = MO_32 },
+ { .fni8 = gen_mla64_i64,
+ .fniv = gen_mla_vec,
+ .opc = INDEX_op_mul_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .vece = MO_64 },
+};
+
+const GVecGen3 mls_op[4] = {
+ { .fni4 = gen_mls8_i32,
+ .fniv = gen_mls_vec,
+ .opc = INDEX_op_mul_vec,
+ .load_dest = true,
+ .vece = MO_8 },
+ { .fni4 = gen_mls16_i32,
+ .fniv = gen_mls_vec,
+ .opc = INDEX_op_mul_vec,
+ .load_dest = true,
+ .vece = MO_16 },
+ { .fni4 = gen_mls32_i32,
+ .fniv = gen_mls_vec,
+ .opc = INDEX_op_mul_vec,
+ .load_dest = true,
+ .vece = MO_32 },
+ { .fni8 = gen_mls64_i64,
+ .fniv = gen_mls_vec,
+ .opc = INDEX_op_mul_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .load_dest = true,
+ .vece = MO_64 },
+};
+
/* Translate a NEON data processing instruction. Return nonzero if the
instruction is invalid.
We process data in a mixture of 32-bit and 64-bit chunks.
@@ -6233,7 +6344,13 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
return 0;
}
break;
+
+ case NEON_3R_VML: /* VMLA, VMLS */
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, vec_size, vec_size,
+ u ? &mls_op[size] : &mla_op[size]);
+ return 0;
}
+
if (size == 3) {
/* 64-bit element instructions. */
for (pass = 0; pass < (q ? 2 : 1); pass++) {
@@ -6435,21 +6552,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
}
}
break;
- case NEON_3R_VML: /* VMLA, VMLAL, VMLS,VMLSL */
- switch (size) {
- case 0: gen_helper_neon_mul_u8(tmp, tmp, tmp2); break;
- case 1: gen_helper_neon_mul_u16(tmp, tmp, tmp2); break;
- case 2: tcg_gen_mul_i32(tmp, tmp, tmp2); break;
- default: abort();
- }
- tcg_temp_free_i32(tmp2);
- tmp2 = neon_load_reg(rd, pass);
- if (u) { /* VMLS */
- gen_neon_rsb(size, tmp, tmp2);
- } else { /* VMLA */
- gen_neon_add(size, tmp, tmp2);
- }
- break;
case NEON_3R_VMUL:
/* VMUL.P8; other cases already eliminated. */
gen_helper_neon_mul_p8(tmp, tmp, tmp2);
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 16/20] target/arm: Use gvec for NEON_3R_VTST_VCEQ, NEON_3R_VCGT, NEON_3R_VCGE
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (14 preceding siblings ...)
2018-10-11 20:52 ` [Qemu-devel] [PATCH 15/20] target/arm: Use gvec for NEON_3R_VML Richard Henderson
@ 2018-10-11 20:52 ` Richard Henderson
2018-10-11 20:52 ` [Qemu-devel] [PATCH 17/20] target/arm: Use gvec for NEON VLD all lanes Richard Henderson
` (4 subsequent siblings)
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:52 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Move cmtst_op expanders from translate-a64.c.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.h | 2 +
target/arm/translate-a64.c | 38 ------------------
target/arm/translate.c | 81 +++++++++++++++++++++++++++-----------
3 files changed, 60 insertions(+), 61 deletions(-)
diff --git a/target/arm/translate.h b/target/arm/translate.h
index d11d504301..eb7ecbebe5 100644
--- a/target/arm/translate.h
+++ b/target/arm/translate.h
@@ -197,10 +197,12 @@ extern const GVecGen3 bit_op;
extern const GVecGen3 bif_op;
extern const GVecGen3 mla_op[4];
extern const GVecGen3 mls_op[4];
+extern const GVecGen3 cmtst_op[4];
extern const GVecGen2i ssra_op[4];
extern const GVecGen2i usra_op[4];
extern const GVecGen2i sri_op[4];
extern const GVecGen2i sli_op[4];
+void gen_cmtst_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b);
#define FORWARD_FEATURE(NAME) \
static inline bool aa32_dc_feature_##NAME(DisasContext *dc) \
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 94f7127c0e..b6f54e66fb 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -8031,28 +8031,6 @@ static void disas_simd_scalar_three_reg_diff(DisasContext *s, uint32_t insn)
}
}
-/* CMTST : test is "if (X & Y != 0)". */
-static void gen_cmtst_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
-{
- tcg_gen_and_i32(d, a, b);
- tcg_gen_setcondi_i32(TCG_COND_NE, d, d, 0);
- tcg_gen_neg_i32(d, d);
-}
-
-static void gen_cmtst_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
-{
- tcg_gen_and_i64(d, a, b);
- tcg_gen_setcondi_i64(TCG_COND_NE, d, d, 0);
- tcg_gen_neg_i64(d, d);
-}
-
-static void gen_cmtst_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
-{
- tcg_gen_and_vec(vece, d, a, b);
- tcg_gen_dupi_vec(vece, a, 0);
- tcg_gen_cmp_vec(TCG_COND_NE, vece, d, d, a);
-}
-
static void handle_3same_64(DisasContext *s, int opcode, bool u,
TCGv_i64 tcg_rd, TCGv_i64 tcg_rn, TCGv_i64 tcg_rm)
{
@@ -10413,22 +10391,6 @@ static void disas_simd_3same_float(DisasContext *s, uint32_t insn)
/* Integer op subgroup of C3.6.16. */
static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
{
- static const GVecGen3 cmtst_op[4] = {
- { .fni4 = gen_helper_neon_tst_u8,
- .fniv = gen_cmtst_vec,
- .vece = MO_8 },
- { .fni4 = gen_helper_neon_tst_u16,
- .fniv = gen_cmtst_vec,
- .vece = MO_16 },
- { .fni4 = gen_cmtst_i32,
- .fniv = gen_cmtst_vec,
- .vece = MO_32 },
- { .fni8 = gen_cmtst_i64,
- .fniv = gen_cmtst_vec,
- .prefer_i64 = TCG_TARGET_REG_BITS == 64,
- .vece = MO_64 },
- };
-
int is_q = extract32(insn, 30, 1);
int u = extract32(insn, 29, 1);
int size = extract32(insn, 22, 2);
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 5008a24ffd..a9bd93bba1 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -6143,6 +6143,44 @@ const GVecGen3 mls_op[4] = {
.vece = MO_64 },
};
+/* CMTST : test is "if (X & Y != 0)". */
+static void gen_cmtst_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
+{
+ tcg_gen_and_i32(d, a, b);
+ tcg_gen_setcondi_i32(TCG_COND_NE, d, d, 0);
+ tcg_gen_neg_i32(d, d);
+}
+
+void gen_cmtst_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
+{
+ tcg_gen_and_i64(d, a, b);
+ tcg_gen_setcondi_i64(TCG_COND_NE, d, d, 0);
+ tcg_gen_neg_i64(d, d);
+}
+
+static void gen_cmtst_vec(unsigned vece, TCGv_vec d, TCGv_vec a, TCGv_vec b)
+{
+ tcg_gen_and_vec(vece, d, a, b);
+ tcg_gen_dupi_vec(vece, a, 0);
+ tcg_gen_cmp_vec(TCG_COND_NE, vece, d, d, a);
+}
+
+const GVecGen3 cmtst_op[4] = {
+ { .fni4 = gen_helper_neon_tst_u8,
+ .fniv = gen_cmtst_vec,
+ .vece = MO_8 },
+ { .fni4 = gen_helper_neon_tst_u16,
+ .fniv = gen_cmtst_vec,
+ .vece = MO_16 },
+ { .fni4 = gen_cmtst_i32,
+ .fniv = gen_cmtst_vec,
+ .vece = MO_32 },
+ { .fni8 = gen_cmtst_i64,
+ .fniv = gen_cmtst_vec,
+ .prefer_i64 = TCG_TARGET_REG_BITS == 64,
+ .vece = MO_64 },
+};
+
/* Translate a NEON data processing instruction. Return nonzero if the
instruction is invalid.
We process data in a mixture of 32-bit and 64-bit chunks.
@@ -6349,6 +6387,26 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs, vec_size, vec_size,
u ? &mls_op[size] : &mla_op[size]);
return 0;
+
+ case NEON_3R_VTST_VCEQ:
+ if (u) { /* VCEQ */
+ tcg_gen_gvec_cmp(TCG_COND_EQ, size, rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size);
+ } else { /* VTST */
+ tcg_gen_gvec_3(rd_ofs, rn_ofs, rm_ofs,
+ vec_size, vec_size, &cmtst_op[size]);
+ }
+ return 0;
+
+ case NEON_3R_VCGT:
+ tcg_gen_gvec_cmp(u ? TCG_COND_GTU : TCG_COND_GT, size,
+ rd_ofs, rn_ofs, rm_ofs, vec_size, vec_size);
+ return 0;
+
+ case NEON_3R_VCGE:
+ tcg_gen_gvec_cmp(u ? TCG_COND_GEU : TCG_COND_GE, size,
+ rd_ofs, rn_ofs, rm_ofs, vec_size, vec_size);
+ return 0;
}
if (size == 3) {
@@ -6502,12 +6560,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
case NEON_3R_VQSUB:
GEN_NEON_INTEGER_OP_ENV(qsub);
break;
- case NEON_3R_VCGT:
- GEN_NEON_INTEGER_OP(cgt);
- break;
- case NEON_3R_VCGE:
- GEN_NEON_INTEGER_OP(cge);
- break;
case NEON_3R_VSHL:
GEN_NEON_INTEGER_OP(shl);
break;
@@ -6535,23 +6587,6 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn)
tmp2 = neon_load_reg(rd, pass);
gen_neon_add(size, tmp, tmp2);
break;
- case NEON_3R_VTST_VCEQ:
- if (!u) { /* VTST */
- switch (size) {
- case 0: gen_helper_neon_tst_u8(tmp, tmp, tmp2); break;
- case 1: gen_helper_neon_tst_u16(tmp, tmp, tmp2); break;
- case 2: gen_helper_neon_tst_u32(tmp, tmp, tmp2); break;
- default: abort();
- }
- } else { /* VCEQ */
- switch (size) {
- case 0: gen_helper_neon_ceq_u8(tmp, tmp, tmp2); break;
- case 1: gen_helper_neon_ceq_u16(tmp, tmp, tmp2); break;
- case 2: gen_helper_neon_ceq_u32(tmp, tmp, tmp2); break;
- default: abort();
- }
- }
- break;
case NEON_3R_VMUL:
/* VMUL.P8; other cases already eliminated. */
gen_helper_neon_mul_p8(tmp, tmp, tmp2);
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* [Qemu-devel] [PATCH 17/20] target/arm: Use gvec for NEON VLD all lanes
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (15 preceding siblings ...)
2018-10-11 20:52 ` [Qemu-devel] [PATCH 16/20] target/arm: Use gvec for NEON_3R_VTST_VCEQ, NEON_3R_VCGT, NEON_3R_VCGE Richard Henderson
@ 2018-10-11 20:52 ` Richard Henderson
2018-10-19 13:05 ` Peter Maydell
2018-10-11 20:52 ` [Qemu-devel] [PATCH 18/20] target/arm: Reorg NEON VLD/VST all elements Richard Henderson
` (3 subsequent siblings)
20 siblings, 1 reply; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:52 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 81 ++++++++++++++----------------------------
1 file changed, 26 insertions(+), 55 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index a9bd93bba1..1e79a1eec0 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -2993,19 +2993,6 @@ static void gen_vfp_msr(TCGv_i32 tmp)
tcg_temp_free_i32(tmp);
}
-static void gen_neon_dup_u8(TCGv_i32 var, int shift)
-{
- TCGv_i32 tmp = tcg_temp_new_i32();
- if (shift)
- tcg_gen_shri_i32(var, var, shift);
- tcg_gen_ext8u_i32(var, var);
- tcg_gen_shli_i32(tmp, var, 8);
- tcg_gen_or_i32(var, var, tmp);
- tcg_gen_shli_i32(tmp, var, 16);
- tcg_gen_or_i32(var, var, tmp);
- tcg_temp_free_i32(tmp);
-}
-
static void gen_neon_dup_low16(TCGv_i32 var)
{
TCGv_i32 tmp = tcg_temp_new_i32();
@@ -3024,28 +3011,6 @@ static void gen_neon_dup_high16(TCGv_i32 var)
tcg_temp_free_i32(tmp);
}
-static TCGv_i32 gen_load_and_replicate(DisasContext *s, TCGv_i32 addr, int size)
-{
- /* Load a single Neon element and replicate into a 32 bit TCG reg */
- TCGv_i32 tmp = tcg_temp_new_i32();
- switch (size) {
- case 0:
- gen_aa32_ld8u(s, tmp, addr, get_mem_index(s));
- gen_neon_dup_u8(tmp, 0);
- break;
- case 1:
- gen_aa32_ld16u(s, tmp, addr, get_mem_index(s));
- gen_neon_dup_low16(tmp);
- break;
- case 2:
- gen_aa32_ld32u(s, tmp, addr, get_mem_index(s));
- break;
- default: /* Avoid compiler warnings. */
- abort();
- }
- return tmp;
-}
-
static int handle_vsel(uint32_t insn, uint32_t rd, uint32_t rn, uint32_t rm,
uint32_t dp)
{
@@ -4949,6 +4914,7 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
int load;
int shift;
int n;
+ int vec_size;
TCGv_i32 addr;
TCGv_i32 tmp;
TCGv_i32 tmp2;
@@ -5118,28 +5084,33 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
}
addr = tcg_temp_new_i32();
load_reg_var(s, addr, rn);
- if (nregs == 1) {
- /* VLD1 to all lanes: bit 5 indicates how many Dregs to write */
- tmp = gen_load_and_replicate(s, addr, size);
- tcg_gen_st_i32(tmp, cpu_env, neon_reg_offset(rd, 0));
- tcg_gen_st_i32(tmp, cpu_env, neon_reg_offset(rd, 1));
- if (insn & (1 << 5)) {
- tcg_gen_st_i32(tmp, cpu_env, neon_reg_offset(rd + 1, 0));
- tcg_gen_st_i32(tmp, cpu_env, neon_reg_offset(rd + 1, 1));
- }
- tcg_temp_free_i32(tmp);
- } else {
- /* VLD2/3/4 to all lanes: bit 5 indicates register stride */
- stride = (insn & (1 << 5)) ? 2 : 1;
- for (reg = 0; reg < nregs; reg++) {
- tmp = gen_load_and_replicate(s, addr, size);
- tcg_gen_st_i32(tmp, cpu_env, neon_reg_offset(rd, 0));
- tcg_gen_st_i32(tmp, cpu_env, neon_reg_offset(rd, 1));
- tcg_temp_free_i32(tmp);
- tcg_gen_addi_i32(addr, addr, 1 << size);
- rd += stride;
+
+ /* VLD1 to all lanes: bit 5 indicates how many Dregs to write.
+ * VLD2/3/4 to all lanes: bit 5 indicates register stride.
+ */
+ stride = insn & (1 << 5) ? 2 : 1;
+ vec_size = nregs == 1 ? stride * 8 : 8;
+
+ tmp = tcg_temp_new_i32();
+ for (reg = 0; reg < nregs; reg++) {
+ gen_aa32_ld_i32(s, tmp, addr, get_mem_index(s),
+ s->be_data | size);
+ if ((rd & 1) && vec_size == 16) {
+ /* We cannot write 16 bytes at once because the
+ * destination is unaligned.
+ */
+ tcg_gen_gvec_dup_i32(size, neon_reg_offset(rd, 0),
+ 8, 8, tmp);
+ tcg_gen_gvec_mov(0, neon_reg_offset(rd + 1, 0),
+ neon_reg_offset(rd, 0), 8, 8);
+ } else {
+ tcg_gen_gvec_dup_i32(size, neon_reg_offset(rd, 0),
+ vec_size, vec_size, tmp);
}
+ tcg_gen_addi_i32(addr, addr, 1 << size);
+ rd += stride;
}
+ tcg_temp_free_i32(tmp);
tcg_temp_free_i32(addr);
stride = (1 << size) * nregs;
} else {
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* Re: [Qemu-devel] [PATCH 17/20] target/arm: Use gvec for NEON VLD all lanes
2018-10-11 20:52 ` [Qemu-devel] [PATCH 17/20] target/arm: Use gvec for NEON VLD all lanes Richard Henderson
@ 2018-10-19 13:05 ` Peter Maydell
0 siblings, 0 replies; 29+ messages in thread
From: Peter Maydell @ 2018-10-19 13:05 UTC (permalink / raw)
To: Richard Henderson; +Cc: QEMU Developers
On 11 October 2018 at 21:52, Richard Henderson
<richard.henderson@linaro.org> wrote:
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> target/arm/translate.c | 81 ++++++++++++++----------------------------
> 1 file changed, 26 insertions(+), 55 deletions(-)
> - /* VLD2/3/4 to all lanes: bit 5 indicates register stride */
> - stride = (insn & (1 << 5)) ? 2 : 1;
> + stride = insn & (1 << 5) ? 2 : 1;
You dropped the clarifying parens in the ?: expression here; I've added
them back...
thanks
-- PMM
^ permalink raw reply [flat|nested] 29+ messages in thread
* [Qemu-devel] [PATCH 18/20] target/arm: Reorg NEON VLD/VST all elements
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (16 preceding siblings ...)
2018-10-11 20:52 ` [Qemu-devel] [PATCH 17/20] target/arm: Use gvec for NEON VLD all lanes Richard Henderson
@ 2018-10-11 20:52 ` Richard Henderson
2018-10-19 13:50 ` Peter Maydell
2018-10-11 20:52 ` [Qemu-devel] [PATCH 19/20] target/arm: Promote consecutive memory ops for aa32 Richard Henderson
` (2 subsequent siblings)
20 siblings, 1 reply; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:52 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Instead of shifts and masks, use direct loads and stores from the neon
register file. Mirror the iteration structure of the ARM pseudocode
more closely. Correct the parameters of the VLD2 A2 insn.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 170 ++++++++++++++++++-----------------------
1 file changed, 74 insertions(+), 96 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 1e79a1eec0..12a744b3c3 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -1611,12 +1611,56 @@ static TCGv_i32 neon_load_reg(int reg, int pass)
return tmp;
}
+static void neon_load_element64(TCGv_i64 var, int reg, int ele, TCGMemOp mop)
+{
+ long offset = neon_element_offset(reg, ele, mop & MO_SIZE);
+
+ switch (mop) {
+ case MO_UB:
+ tcg_gen_ld8u_i64(var, cpu_env, offset);
+ break;
+ case MO_UW:
+ tcg_gen_ld16u_i64(var, cpu_env, offset);
+ break;
+ case MO_UL:
+ tcg_gen_ld32u_i64(var, cpu_env, offset);
+ break;
+ case MO_Q:
+ tcg_gen_ld_i64(var, cpu_env, offset);
+ break;
+ default:
+ g_assert_not_reached();
+ }
+}
+
static void neon_store_reg(int reg, int pass, TCGv_i32 var)
{
tcg_gen_st_i32(var, cpu_env, neon_reg_offset(reg, pass));
tcg_temp_free_i32(var);
}
+static void neon_store_element64(int reg, int ele, TCGMemOp size, TCGv_i64 var)
+{
+ long offset = neon_element_offset(reg, ele, size);
+
+ switch (size) {
+ case MO_8:
+ tcg_gen_st8_i64(var, cpu_env, offset);
+ break;
+ case MO_16:
+ tcg_gen_st16_i64(var, cpu_env, offset);
+ break;
+ case MO_32:
+ tcg_gen_st32_i64(var, cpu_env, offset);
+ break;
+ case MO_64:
+ tcg_gen_st_i64(var, cpu_env, offset);
+ break;
+ default:
+ g_assert_not_reached();
+ }
+}
+
static inline void neon_load_reg64(TCGv_i64 var, int reg)
{
tcg_gen_ld_i64(var, cpu_env, vfp_reg_offset(1, reg));
@@ -4885,16 +4929,16 @@ static struct {
int interleave;
int spacing;
} const neon_ls_element_type[11] = {
- {4, 4, 1},
- {4, 4, 2},
+ {1, 4, 1},
+ {1, 4, 2},
{4, 1, 1},
- {4, 2, 1},
- {3, 3, 1},
- {3, 3, 2},
+ {2, 2, 2},
+ {1, 3, 1},
+ {1, 3, 2},
{3, 1, 1},
{1, 1, 1},
- {2, 2, 1},
- {2, 2, 2},
+ {1, 2, 1},
+ {1, 2, 2},
{2, 1, 1}
};
@@ -4915,6 +4959,8 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
int shift;
int n;
int vec_size;
+ int mmu_idx;
+ TCGMemOp endian;
TCGv_i32 addr;
TCGv_i32 tmp;
TCGv_i32 tmp2;
@@ -4936,6 +4982,8 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
rn = (insn >> 16) & 0xf;
rm = insn & 0xf;
load = (insn & (1 << 21)) != 0;
+ endian = s->be_data;
+ mmu_idx = get_mem_index(s);
if ((insn & (1 << 23)) == 0) {
/* Load store all elements. */
op = (insn >> 8) & 0xf;
@@ -4960,104 +5008,34 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
nregs = neon_ls_element_type[op].nregs;
interleave = neon_ls_element_type[op].interleave;
spacing = neon_ls_element_type[op].spacing;
- if (size == 3 && (interleave | spacing) != 1)
+ if (size == 3 && (interleave | spacing) != 1) {
return 1;
+ }
+ tmp64 = tcg_temp_new_i64();
addr = tcg_temp_new_i32();
+ tmp2 = tcg_const_i32(1 << size);
load_reg_var(s, addr, rn);
- stride = (1 << size) * interleave;
for (reg = 0; reg < nregs; reg++) {
- if (interleave > 2 || (interleave == 2 && nregs == 2)) {
- load_reg_var(s, addr, rn);
- tcg_gen_addi_i32(addr, addr, (1 << size) * reg);
- } else if (interleave == 2 && nregs == 4 && reg == 2) {
- load_reg_var(s, addr, rn);
- tcg_gen_addi_i32(addr, addr, 1 << size);
- }
- if (size == 3) {
- tmp64 = tcg_temp_new_i64();
- if (load) {
- gen_aa32_ld64(s, tmp64, addr, get_mem_index(s));
- neon_store_reg64(tmp64, rd);
- } else {
- neon_load_reg64(tmp64, rd);
- gen_aa32_st64(s, tmp64, addr, get_mem_index(s));
- }
- tcg_temp_free_i64(tmp64);
- tcg_gen_addi_i32(addr, addr, stride);
- } else {
- for (pass = 0; pass < 2; pass++) {
- if (size == 2) {
- if (load) {
- tmp = tcg_temp_new_i32();
- gen_aa32_ld32u(s, tmp, addr, get_mem_index(s));
- neon_store_reg(rd, pass, tmp);
- } else {
- tmp = neon_load_reg(rd, pass);
- gen_aa32_st32(s, tmp, addr, get_mem_index(s));
- tcg_temp_free_i32(tmp);
- }
- tcg_gen_addi_i32(addr, addr, stride);
- } else if (size == 1) {
- if (load) {
- tmp = tcg_temp_new_i32();
- gen_aa32_ld16u(s, tmp, addr, get_mem_index(s));
- tcg_gen_addi_i32(addr, addr, stride);
- tmp2 = tcg_temp_new_i32();
- gen_aa32_ld16u(s, tmp2, addr, get_mem_index(s));
- tcg_gen_addi_i32(addr, addr, stride);
- tcg_gen_shli_i32(tmp2, tmp2, 16);
- tcg_gen_or_i32(tmp, tmp, tmp2);
- tcg_temp_free_i32(tmp2);
- neon_store_reg(rd, pass, tmp);
- } else {
- tmp = neon_load_reg(rd, pass);
- tmp2 = tcg_temp_new_i32();
- tcg_gen_shri_i32(tmp2, tmp, 16);
- gen_aa32_st16(s, tmp, addr, get_mem_index(s));
- tcg_temp_free_i32(tmp);
- tcg_gen_addi_i32(addr, addr, stride);
- gen_aa32_st16(s, tmp2, addr, get_mem_index(s));
- tcg_temp_free_i32(tmp2);
- tcg_gen_addi_i32(addr, addr, stride);
- }
- } else /* size == 0 */ {
- if (load) {
- tmp2 = NULL;
- for (n = 0; n < 4; n++) {
- tmp = tcg_temp_new_i32();
- gen_aa32_ld8u(s, tmp, addr, get_mem_index(s));
- tcg_gen_addi_i32(addr, addr, stride);
- if (n == 0) {
- tmp2 = tmp;
- } else {
- tcg_gen_shli_i32(tmp, tmp, n * 8);
- tcg_gen_or_i32(tmp2, tmp2, tmp);
- tcg_temp_free_i32(tmp);
- }
- }
- neon_store_reg(rd, pass, tmp2);
- } else {
- tmp2 = neon_load_reg(rd, pass);
- for (n = 0; n < 4; n++) {
- tmp = tcg_temp_new_i32();
- if (n == 0) {
- tcg_gen_mov_i32(tmp, tmp2);
- } else {
- tcg_gen_shri_i32(tmp, tmp2, n * 8);
- }
- gen_aa32_st8(s, tmp, addr, get_mem_index(s));
- tcg_temp_free_i32(tmp);
- tcg_gen_addi_i32(addr, addr, stride);
- }
- tcg_temp_free_i32(tmp2);
- }
+ for (n = 0; n < 8 >> size; n++) {
+ int xs;
+ for (xs = 0; xs < interleave; xs++) {
+ int tt = rd + reg + spacing * xs;
+
+ if (load) {
+ gen_aa32_ld_i64(s, tmp64, addr, mmu_idx, endian | size);
+ neon_store_element64(tt, n, size, tmp64);
+ } else {
+ neon_load_element64(tmp64, tt, n, size);
+ gen_aa32_st_i64(s, tmp64, addr, mmu_idx, endian | size);
}
+ tcg_gen_add_i32(addr, addr, tmp2);
}
}
- rd += spacing;
}
tcg_temp_free_i32(addr);
- stride = nregs * 8;
+ tcg_temp_free_i32(tmp2);
+ tcg_temp_free_i64(tmp64);
+ stride = nregs * interleave * 8;
} else {
size = (insn >> 10) & 3;
if (size == 3) {
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* Re: [Qemu-devel] [PATCH 18/20] target/arm: Reorg NEON VLD/VST all elements
2018-10-11 20:52 ` [Qemu-devel] [PATCH 18/20] target/arm: Reorg NEON VLD/VST all elements Richard Henderson
@ 2018-10-19 13:50 ` Peter Maydell
2018-10-19 15:15 ` Richard Henderson
0 siblings, 1 reply; 29+ messages in thread
From: Peter Maydell @ 2018-10-19 13:50 UTC (permalink / raw)
To: Richard Henderson; +Cc: QEMU Developers
On 11 October 2018 at 21:52, Richard Henderson
<richard.henderson@linaro.org> wrote:
> Instead of shifts and masks, use direct loads and stores from the neon
> register file. Mirror the iteration structure of the ARM pseudocode
> more closely. Correct the parameters of the VLD2 A2 insn.
So is this actually fixing a bug here? The new code definitely
looks right, but I can't make heads nor tails of the old code,
so I'm not sure whether it actually behaved differently...
thanks
-- PMM
^ permalink raw reply [flat|nested] 29+ messages in thread
* Re: [Qemu-devel] [PATCH 18/20] target/arm: Reorg NEON VLD/VST all elements
2018-10-19 13:50 ` Peter Maydell
@ 2018-10-19 15:15 ` Richard Henderson
0 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-19 15:15 UTC (permalink / raw)
To: Peter Maydell; +Cc: QEMU Developers
On 10/19/18 6:50 AM, Peter Maydell wrote:
> On 11 October 2018 at 21:52, Richard Henderson
> <richard.henderson@linaro.org> wrote:
>> Instead of shifts and masks, use direct loads and stores from the neon
>> register file. Mirror the iteration structure of the ARM pseudocode
>> more closely. Correct the parameters of the VLD2 A2 insn.
>
> So is this actually fixing a bug here? The new code definitely
> looks right, but I can't make heads nor tails of the old code,
> so I'm not sure whether it actually behaved differently...
Hmm, that comment was written in July. Working backward,
- {4, 2, 1},
+ {2, 2, 2},
which is for "VLD2 (multiple 2-element structures) A2", which does have
regs = 2; inc = 2;
in the pseudocode. So, yes, we had the wrong stride (last number) for that insn.
Aside from that the old code and the new code should have the same end result.
The old code computed addresses as it stepped through register elements, and so
accesses memory in a "random" order. The new code accesses memory sequentially
and computes the register element that corresponds. This is for simplicity of
tcg computation of addresses and not correctness.
r~
^ permalink raw reply [flat|nested] 29+ messages in thread
* [Qemu-devel] [PATCH 19/20] target/arm: Promote consecutive memory ops for aa32
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (17 preceding siblings ...)
2018-10-11 20:52 ` [Qemu-devel] [PATCH 18/20] target/arm: Reorg NEON VLD/VST all elements Richard Henderson
@ 2018-10-11 20:52 ` Richard Henderson
2018-10-19 5:17 ` Philippe Mathieu-Daudé
2018-10-11 20:52 ` [Qemu-devel] [PATCH 20/20] target/arm: Reorg NEON VLD/VST single element to one lane Richard Henderson
2018-10-19 13:51 ` [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Peter Maydell
20 siblings, 1 reply; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:52 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
For a sequence of loads or stores from a single register,
little-endian operations can be promoted to an 8-byte op.
This can reduce the number of operations by a factor of 8.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 12a744b3c3..09f2d648b7 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -5011,6 +5011,16 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
if (size == 3 && (interleave | spacing) != 1) {
return 1;
}
+ /* For our purposes, bytes are always little-endian. */
+ if (size == 0) {
+ endian = MO_LE;
+ }
+ /* Consecutive little-endian elements from a single register
+ * can be promoted to a larger little-endian operation.
+ */
+ if (interleave == 1 && endian == MO_LE) {
+ size = 3;
+ }
tmp64 = tcg_temp_new_i64();
addr = tcg_temp_new_i32();
tmp2 = tcg_const_i32(1 << size);
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* Re: [Qemu-devel] [PATCH 19/20] target/arm: Promote consecutive memory ops for aa32
2018-10-11 20:52 ` [Qemu-devel] [PATCH 19/20] target/arm: Promote consecutive memory ops for aa32 Richard Henderson
@ 2018-10-19 5:17 ` Philippe Mathieu-Daudé
0 siblings, 0 replies; 29+ messages in thread
From: Philippe Mathieu-Daudé @ 2018-10-19 5:17 UTC (permalink / raw)
To: Richard Henderson, qemu-devel; +Cc: peter.maydell
On 11/10/2018 22:52, Richard Henderson wrote:
> For a sequence of loads or stores from a single register,
> little-endian operations can be promoted to an 8-byte op.
> This can reduce the number of operations by a factor of 8.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
> ---
> target/arm/translate.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
>
> diff --git a/target/arm/translate.c b/target/arm/translate.c
> index 12a744b3c3..09f2d648b7 100644
> --- a/target/arm/translate.c
> +++ b/target/arm/translate.c
> @@ -5011,6 +5011,16 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
> if (size == 3 && (interleave | spacing) != 1) {
> return 1;
> }
> + /* For our purposes, bytes are always little-endian. */
> + if (size == 0) {
> + endian = MO_LE;
> + }
> + /* Consecutive little-endian elements from a single register
> + * can be promoted to a larger little-endian operation.
> + */
> + if (interleave == 1 && endian == MO_LE) {
> + size = 3;
> + }
> tmp64 = tcg_temp_new_i64();
> addr = tcg_temp_new_i32();
> tmp2 = tcg_const_i32(1 << size);
>
^ permalink raw reply [flat|nested] 29+ messages in thread
* [Qemu-devel] [PATCH 20/20] target/arm: Reorg NEON VLD/VST single element to one lane
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (18 preceding siblings ...)
2018-10-11 20:52 ` [Qemu-devel] [PATCH 19/20] target/arm: Promote consecutive memory ops for aa32 Richard Henderson
@ 2018-10-11 20:52 ` Richard Henderson
2018-10-19 13:51 ` [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Peter Maydell
20 siblings, 0 replies; 29+ messages in thread
From: Richard Henderson @ 2018-10-11 20:52 UTC (permalink / raw)
To: qemu-devel; +Cc: peter.maydell
Instead of shifts and masks, use direct loads and stores from
the neon register file.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
target/arm/translate.c | 92 +++++++++++++++++++++++-------------------
1 file changed, 50 insertions(+), 42 deletions(-)
diff --git a/target/arm/translate.c b/target/arm/translate.c
index 09f2d648b7..0b21c2d201 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -1611,6 +1611,25 @@ static TCGv_i32 neon_load_reg(int reg, int pass)
return tmp;
}
+static void neon_load_element(TCGv_i32 var, int reg, int ele, TCGMemOp mop)
+{
+ long offset = neon_element_offset(reg, ele, mop & MO_SIZE);
+
+ switch (mop) {
+ case MO_UB:
+ tcg_gen_ld8u_i32(var, cpu_env, offset);
+ break;
+ case MO_UW:
+ tcg_gen_ld16u_i32(var, cpu_env, offset);
+ break;
+ case MO_UL:
+ tcg_gen_ld_i32(var, cpu_env, offset);
+ break;
+ default:
+ g_assert_not_reached();
+ }
+}
+
static void neon_load_element64(TCGv_i64 var, int reg, int ele, TCGMemOp mop)
{
long offset = neon_element_offset(reg, ele, mop & MO_SIZE);
@@ -1639,6 +1658,25 @@ static void neon_store_reg(int reg, int pass, TCGv_i32 var)
tcg_temp_free_i32(var);
}
+static void neon_store_element(int reg, int ele, TCGMemOp size, TCGv_i32 var)
+{
+ long offset = neon_element_offset(reg, ele, size);
+
+ switch (size) {
+ case MO_8:
+ tcg_gen_st8_i32(var, cpu_env, offset);
+ break;
+ case MO_16:
+ tcg_gen_st16_i32(var, cpu_env, offset);
+ break;
+ case MO_32:
+ tcg_gen_st_i32(var, cpu_env, offset);
+ break;
+ default:
+ g_assert_not_reached();
+ }
+}
+
static void neon_store_element64(int reg, int ele, TCGMemOp size, TCGv_i64 var)
{
long offset = neon_element_offset(reg, ele, size);
@@ -4954,9 +4992,7 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
int stride;
int size;
int reg;
- int pass;
int load;
- int shift;
int n;
int vec_size;
int mmu_idx;
@@ -5104,18 +5140,18 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
} else {
/* Single element. */
int idx = (insn >> 4) & 0xf;
- pass = (insn >> 7) & 1;
+ int reg_idx;
switch (size) {
case 0:
- shift = ((insn >> 5) & 3) * 8;
+ reg_idx = (insn >> 5) & 7;
stride = 1;
break;
case 1:
- shift = ((insn >> 6) & 1) * 16;
+ reg_idx = (insn >> 6) & 3;
stride = (insn & (1 << 5)) ? 2 : 1;
break;
case 2:
- shift = 0;
+ reg_idx = (insn >> 7) & 1;
stride = (insn & (1 << 6)) ? 2 : 1;
break;
default:
@@ -5155,52 +5191,24 @@ static int disas_neon_ls_insn(DisasContext *s, uint32_t insn)
*/
return 1;
}
+ tmp = tcg_temp_new_i32();
addr = tcg_temp_new_i32();
load_reg_var(s, addr, rn);
for (reg = 0; reg < nregs; reg++) {
if (load) {
- tmp = tcg_temp_new_i32();
- switch (size) {
- case 0:
- gen_aa32_ld8u(s, tmp, addr, get_mem_index(s));
- break;
- case 1:
- gen_aa32_ld16u(s, tmp, addr, get_mem_index(s));
- break;
- case 2:
- gen_aa32_ld32u(s, tmp, addr, get_mem_index(s));
- break;
- default: /* Avoid compiler warnings. */
- abort();
- }
- if (size != 2) {
- tmp2 = neon_load_reg(rd, pass);
- tcg_gen_deposit_i32(tmp, tmp2, tmp,
- shift, size ? 16 : 8);
- tcg_temp_free_i32(tmp2);
- }
- neon_store_reg(rd, pass, tmp);
+ gen_aa32_ld_i32(s, tmp, addr, get_mem_index(s),
+ s->be_data | size);
+ neon_store_element(rd, reg_idx, size, tmp);
} else { /* Store */
- tmp = neon_load_reg(rd, pass);
- if (shift)
- tcg_gen_shri_i32(tmp, tmp, shift);
- switch (size) {
- case 0:
- gen_aa32_st8(s, tmp, addr, get_mem_index(s));
- break;
- case 1:
- gen_aa32_st16(s, tmp, addr, get_mem_index(s));
- break;
- case 2:
- gen_aa32_st32(s, tmp, addr, get_mem_index(s));
- break;
- }
- tcg_temp_free_i32(tmp);
+ neon_load_element(tmp, rd, reg_idx, size);
+ gen_aa32_st_i32(s, tmp, addr, get_mem_index(s),
+ s->be_data | size);
}
rd += stride;
tcg_gen_addi_i32(addr, addr, 1 << size);
}
tcg_temp_free_i32(addr);
+ tcg_temp_free_i32(tmp);
stride = nregs * (1 << size);
}
}
--
2.17.1
^ permalink raw reply related [flat|nested] 29+ messages in thread* Re: [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec
2018-10-11 20:51 [Qemu-devel] [PATCH 00/20] target/arm: Convert some neon insns to gvec Richard Henderson
` (19 preceding siblings ...)
2018-10-11 20:52 ` [Qemu-devel] [PATCH 20/20] target/arm: Reorg NEON VLD/VST single element to one lane Richard Henderson
@ 2018-10-19 13:51 ` Peter Maydell
20 siblings, 0 replies; 29+ messages in thread
From: Peter Maydell @ 2018-10-19 13:51 UTC (permalink / raw)
To: Richard Henderson; +Cc: QEMU Developers
On 11 October 2018 at 21:51, Richard Henderson
<richard.henderson@linaro.org> wrote:
> The conversion to gvec is by no means complete, but it brings aa32
> to parity with the work already done for aa64. Plus some misc changes,
> at least one of which has been sitting on a branch for 2 years. ;-P
>
> I've run the result through Alex's aarch32-all.tar.xz RISU tests.
> There are no regressions, though there are pre-existing failures.
> Our aa32 translator could use some additional work, no surprise.
Applied to target-arm.next, thanks. I made a couple of minor tweaks
which I've noted in replies to the relevant patches.
If you could clarify whether patch 18 is fixing a bug I can tweak
the commit message on that one.
thanks
-- PMM
^ permalink raw reply [flat|nested] 29+ messages in thread