[PATCH 0/3] tcg misc patches

[PATCH 0/3] tcg misc patches

[Richard Henderson]

A couple of changes I'd like to queue for tcg-next,
which are as yet unreviewed.  The final one has been
on list before, as part of the sve2 patch set.


r~


Richard Henderson (3):
  softmmu/cpus: Only set parallel_cpus for SMP
  tcg: Eliminate one store for in-place 128-bit dup_mem
  tcg: Implement 256-bit dup for tcg_gen_gvec_dup_mem

 softmmu/cpus.c    | 11 +++++++++-
 tcg/tcg-op-gvec.c | 56 ++++++++++++++++++++++++++++++++++++++++++-----
 2 files changed, 61 insertions(+), 6 deletions(-)

-- 
2.25.1

[PATCH 1/3] softmmu/cpus: Only set parallel_cpus for SMP

[Richard Henderson]

Do not set parallel_cpus if there is only one cpu instantiated.
This will allow tcg to use serial code to implement atomics.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 softmmu/cpus.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/softmmu/cpus.c b/softmmu/cpus.c
index a802e899ab..e3b98065c9 100644
--- a/softmmu/cpus.c
+++ b/softmmu/cpus.c
@@ -1895,6 +1895,16 @@ static void qemu_tcg_init_vcpu(CPUState *cpu)
     if (!tcg_region_inited) {
         tcg_region_inited = 1;
         tcg_region_init();
+        /*
+         * If MTTCG, and we will create multiple cpus,
+         * then we will have cpus running in parallel.
+         */
+        if (qemu_tcg_mttcg_enabled()) {
+            MachineState *ms = MACHINE(qdev_get_machine());
+            if (ms->smp.max_cpus > 1) {
+                parallel_cpus = true;
+            }
+        }
     }
 
     if (qemu_tcg_mttcg_enabled() || !single_tcg_cpu_thread) {
@@ -1904,7 +1914,6 @@ static void qemu_tcg_init_vcpu(CPUState *cpu)
 
         if (qemu_tcg_mttcg_enabled()) {
             /* create a thread per vCPU with TCG (MTTCG) */
-            parallel_cpus = true;
             snprintf(thread_name, VCPU_THREAD_NAME_SIZE, "CPU %d/TCG",
                  cpu->cpu_index);
 
-- 
2.25.1

Re: [PATCH 1/3] softmmu/cpus: Only set parallel_cpus for SMP

[Philippe Mathieu-Daudé]

[-- Attachment #1: Type: text/plain, Size: 1575 bytes --]

Le ven. 28 août 2020 20:04, Richard Henderson <richard.henderson@linaro.org>
a écrit :

> Do not set parallel_cpus if there is only one cpu instantiated.
> This will allow tcg to use serial code to implement atomics.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>

Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>

---
>  softmmu/cpus.c | 11 ++++++++++-
>  1 file changed, 10 insertions(+), 1 deletion(-)
>
> diff --git a/softmmu/cpus.c b/softmmu/cpus.c
> index a802e899ab..e3b98065c9 100644
> --- a/softmmu/cpus.c
> +++ b/softmmu/cpus.c
> @@ -1895,6 +1895,16 @@ static void qemu_tcg_init_vcpu(CPUState *cpu)
>      if (!tcg_region_inited) {
>          tcg_region_inited = 1;
>          tcg_region_init();
> +        /*
> +         * If MTTCG, and we will create multiple cpus,
> +         * then we will have cpus running in parallel.
> +         */
> +        if (qemu_tcg_mttcg_enabled()) {
> +            MachineState *ms = MACHINE(qdev_get_machine());
> +            if (ms->smp.max_cpus > 1) {
> +                parallel_cpus = true;
> +            }
> +        }
>      }
>
>      if (qemu_tcg_mttcg_enabled() || !single_tcg_cpu_thread) {
> @@ -1904,7 +1914,6 @@ static void qemu_tcg_init_vcpu(CPUState *cpu)
>
>          if (qemu_tcg_mttcg_enabled()) {
>              /* create a thread per vCPU with TCG (MTTCG) */
> -            parallel_cpus = true;
>              snprintf(thread_name, VCPU_THREAD_NAME_SIZE, "CPU %d/TCG",
>                   cpu->cpu_index);
>
> --
> 2.25.1
>
>
>

[-- Attachment #2: Type: text/html, Size: 2723 bytes --]

[PATCH 2/3] tcg: Eliminate one store for in-place 128-bit dup_mem

[Richard Henderson]

Do not store back to the exact memory from which we just loaded.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/tcg-op-gvec.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
index 793d4ba64c..fcc25b04e6 100644
--- a/tcg/tcg-op-gvec.c
+++ b/tcg/tcg-op-gvec.c
@@ -1581,7 +1581,7 @@ void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
 
             tcg_gen_ld_vec(in, cpu_env, aofs);
-            for (i = 0; i < oprsz; i += 16) {
+            for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
                 tcg_gen_st_vec(in, cpu_env, dofs + i);
             }
             tcg_temp_free_vec(in);
@@ -1591,7 +1591,7 @@ void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
 
             tcg_gen_ld_i64(in0, cpu_env, aofs);
             tcg_gen_ld_i64(in1, cpu_env, aofs + 8);
-            for (i = 0; i < oprsz; i += 16) {
+            for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
                 tcg_gen_st_i64(in0, cpu_env, dofs + i);
                 tcg_gen_st_i64(in1, cpu_env, dofs + i + 8);
             }
-- 
2.25.1

Re: [PATCH 2/3] tcg: Eliminate one store for in-place 128-bit dup_mem

[Philippe Mathieu-Daudé]

[-- Attachment #1: Type: text/plain, Size: 1477 bytes --]

Le ven. 28 août 2020 20:04, Richard Henderson <richard.henderson@linaro.org>
a écrit :

> Do not store back to the exact memory from which we just loaded.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>

Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>

---
>  tcg/tcg-op-gvec.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
> index 793d4ba64c..fcc25b04e6 100644
> --- a/tcg/tcg-op-gvec.c
> +++ b/tcg/tcg-op-gvec.c
> @@ -1581,7 +1581,7 @@ void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t
> dofs, uint32_t aofs,
>              TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
>
>              tcg_gen_ld_vec(in, cpu_env, aofs);
> -            for (i = 0; i < oprsz; i += 16) {
> +            for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
>                  tcg_gen_st_vec(in, cpu_env, dofs + i);
>              }
>              tcg_temp_free_vec(in);
> @@ -1591,7 +1591,7 @@ void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t
> dofs, uint32_t aofs,
>
>              tcg_gen_ld_i64(in0, cpu_env, aofs);
>              tcg_gen_ld_i64(in1, cpu_env, aofs + 8);
> -            for (i = 0; i < oprsz; i += 16) {
> +            for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
>                  tcg_gen_st_i64(in0, cpu_env, dofs + i);
>                  tcg_gen_st_i64(in1, cpu_env, dofs + i + 8);
>              }
> --
> 2.25.1
>
>
>

[-- Attachment #2: Type: text/html, Size: 2526 bytes --]

[PATCH 3/3] tcg: Implement 256-bit dup for tcg_gen_gvec_dup_mem

[Richard Henderson]

We already support duplication of 128-bit blocks.  This extends
that support to 256-bit blocks.  This will be needed by SVE2.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/tcg-op-gvec.c | 52 ++++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 49 insertions(+), 3 deletions(-)

diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
index fcc25b04e6..7ebd9e8298 100644
--- a/tcg/tcg-op-gvec.c
+++ b/tcg/tcg-op-gvec.c
@@ -1570,12 +1570,10 @@ void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
             do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
             tcg_temp_free_i64(in);
         }
-    } else {
+    } else if (vece == 4) {
         /* 128-bit duplicate.  */
-        /* ??? Dup to 256-bit vector.  */
         int i;
 
-        tcg_debug_assert(vece == 4);
         tcg_debug_assert(oprsz >= 16);
         if (TCG_TARGET_HAS_v128) {
             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
@@ -1601,6 +1599,54 @@ void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
         if (oprsz < maxsz) {
             expand_clr(dofs + oprsz, maxsz - oprsz);
         }
+    } else if (vece == 5) {
+        /* 256-bit duplicate.  */
+        int i;
+
+        tcg_debug_assert(oprsz >= 32);
+        tcg_debug_assert(oprsz % 32 == 0);
+        if (TCG_TARGET_HAS_v256) {
+            TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V256);
+
+            tcg_gen_ld_vec(in, cpu_env, aofs);
+            for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
+                tcg_gen_st_vec(in, cpu_env, dofs + i);
+            }
+            tcg_temp_free_vec(in);
+        } else if (TCG_TARGET_HAS_v128) {
+            TCGv_vec in0 = tcg_temp_new_vec(TCG_TYPE_V128);
+            TCGv_vec in1 = tcg_temp_new_vec(TCG_TYPE_V128);
+
+            tcg_gen_ld_vec(in0, cpu_env, aofs);
+            tcg_gen_ld_vec(in1, cpu_env, aofs + 16);
+            for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
+                tcg_gen_st_vec(in0, cpu_env, dofs + i);
+                tcg_gen_st_vec(in1, cpu_env, dofs + i + 16);
+            }
+            tcg_temp_free_vec(in0);
+            tcg_temp_free_vec(in1);
+        } else {
+            TCGv_i64 in[4];
+            int j;
+
+            for (j = 0; j < 4; ++j) {
+                in[j] = tcg_temp_new_i64();
+                tcg_gen_ld_i64(in[j], cpu_env, aofs + j * 8);
+            }
+            for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
+                for (j = 0; j < 4; ++j) {
+                    tcg_gen_st_i64(in[j], cpu_env, dofs + i + j * 8);
+                }
+            }
+            for (j = 0; j < 4; ++j) {
+                tcg_temp_free_i64(in[j]);
+            }
+        }
+        if (oprsz < maxsz) {
+            expand_clr(dofs + oprsz, maxsz - oprsz);
+        }
+    } else {
+        g_assert_not_reached();
     }
 }
 
-- 
2.25.1

Re: [PATCH 3/3] tcg: Implement 256-bit dup for tcg_gen_gvec_dup_mem

[Philippe Mathieu-Daudé]

[-- Attachment #1: Type: text/plain, Size: 3312 bytes --]

Le ven. 28 août 2020 20:04, Richard Henderson <richard.henderson@linaro.org>
a écrit :

> We already support duplication of 128-bit blocks.  This extends
> that support to 256-bit blocks.  This will be needed by SVE2.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>

Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>

---
>  tcg/tcg-op-gvec.c | 52 ++++++++++++++++++++++++++++++++++++++++++++---
>  1 file changed, 49 insertions(+), 3 deletions(-)
>
> diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
> index fcc25b04e6..7ebd9e8298 100644
> --- a/tcg/tcg-op-gvec.c
> +++ b/tcg/tcg-op-gvec.c
> @@ -1570,12 +1570,10 @@ void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t
> dofs, uint32_t aofs,
>              do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
>              tcg_temp_free_i64(in);
>          }
> -    } else {
> +    } else if (vece == 4) {
>          /* 128-bit duplicate.  */
> -        /* ??? Dup to 256-bit vector.  */
>          int i;
>
> -        tcg_debug_assert(vece == 4);
>          tcg_debug_assert(oprsz >= 16);
>          if (TCG_TARGET_HAS_v128) {
>              TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
> @@ -1601,6 +1599,54 @@ void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t
> dofs, uint32_t aofs,
>          if (oprsz < maxsz) {
>              expand_clr(dofs + oprsz, maxsz - oprsz);
>          }
> +    } else if (vece == 5) {
> +        /* 256-bit duplicate.  */
> +        int i;
> +
> +        tcg_debug_assert(oprsz >= 32);
> +        tcg_debug_assert(oprsz % 32 == 0);
> +        if (TCG_TARGET_HAS_v256) {
> +            TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V256);
> +
> +            tcg_gen_ld_vec(in, cpu_env, aofs);
> +            for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
> +                tcg_gen_st_vec(in, cpu_env, dofs + i);
> +            }
> +            tcg_temp_free_vec(in);
> +        } else if (TCG_TARGET_HAS_v128) {
> +            TCGv_vec in0 = tcg_temp_new_vec(TCG_TYPE_V128);
> +            TCGv_vec in1 = tcg_temp_new_vec(TCG_TYPE_V128);
> +
> +            tcg_gen_ld_vec(in0, cpu_env, aofs);
> +            tcg_gen_ld_vec(in1, cpu_env, aofs + 16);
> +            for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
> +                tcg_gen_st_vec(in0, cpu_env, dofs + i);
> +                tcg_gen_st_vec(in1, cpu_env, dofs + i + 16);
> +            }
> +            tcg_temp_free_vec(in0);
> +            tcg_temp_free_vec(in1);
> +        } else {
> +            TCGv_i64 in[4];
> +            int j;
> +
> +            for (j = 0; j < 4; ++j) {
> +                in[j] = tcg_temp_new_i64();
> +                tcg_gen_ld_i64(in[j], cpu_env, aofs + j * 8);
> +            }
> +            for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
> +                for (j = 0; j < 4; ++j) {
> +                    tcg_gen_st_i64(in[j], cpu_env, dofs + i + j * 8);
> +                }
> +            }
> +            for (j = 0; j < 4; ++j) {
> +                tcg_temp_free_i64(in[j]);
> +            }
> +        }
> +        if (oprsz < maxsz) {
> +            expand_clr(dofs + oprsz, maxsz - oprsz);
> +        }
> +    } else {
> +        g_assert_not_reached();
>      }
>  }
>
> --
> 2.25.1
>
>
>

[-- Attachment #2: Type: text/html, Size: 4721 bytes --]