diff mbox series

[7/9] target/arm: Expand vector registers for SVE

Message ID 20171218173022.18418-8-richard.henderson@linaro.org
State Superseded
Headers show
Series target/arm: Prepatory work for SVE | expand

Commit Message

Richard Henderson Dec. 18, 2017, 5:30 p.m. UTC
Change vfp.regs as a uint64_t to vfp.zregs as an ARMVectorReg.
The previous patches have made the change in representation
relatively painless.

Add vfp.pregs as an ARMPredicateReg.  Let FFR be P16 to make
it easier to treat it as for any other predicate.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

---
 target/arm/cpu.h           | 64 ++++++++++++++++++++++++++++++++--------------
 target/arm/machine.c       | 37 ++++++++++++++++++++++++++-
 target/arm/translate-a64.c |  8 +++---
 target/arm/translate.c     | 12 ++++-----
 4 files changed, 90 insertions(+), 31 deletions(-)

-- 
2.14.3

Comments

Peter Maydell Jan. 11, 2018, 6:44 p.m. UTC | #1
On 18 December 2017 at 17:30, Richard Henderson
<richard.henderson@linaro.org> wrote:
> Change vfp.regs as a uint64_t to vfp.zregs as an ARMVectorReg.

> The previous patches have made the change in representation

> relatively painless.

>

> Add vfp.pregs as an ARMPredicateReg.  Let FFR be P16 to make

> it easier to treat it as for any other predicate.


> diff --git a/target/arm/machine.c b/target/arm/machine.c

> index a85c2430d3..39f3123b10 100644

> --- a/target/arm/machine.c

> +++ b/target/arm/machine.c

> @@ -50,7 +50,42 @@ static const VMStateDescription vmstate_vfp = {

>      .minimum_version_id = 3,

>      .needed = vfp_needed,

>      .fields = (VMStateField[]) {

> -        VMSTATE_UINT64_ARRAY(env.vfp.regs, ARMCPU, 64),

> +        /* For compatibility, store Qn out of Zn here.  */

> +        /* TODO: store the other quadwords in another subsection,

> +           along with predicate registers and ZCR_EL[1-3].  */


I think we should get the migration state stuff right at the point
where we change the data structure in cpu.h.


> +        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[0].d, ARMCPU, 0, 2),

> +        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[1].d, ARMCPU, 0, 2),

> +        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[2].d, ARMCPU, 0, 2),


David: does the migration code guarantee that we can split an
old VMSTATE_UINT64_ARRAY into being a sequence of SUB_ARRAYs
like this and retain the same on-the-wire format (ie and
keep migration compat) ?

I suspect we've used this clever trick before elsewhere, but thought
it better to check...

thanks
-- PMM
Dr. David Alan Gilbert Jan. 11, 2018, 6:58 p.m. UTC | #2
* Peter Maydell (peter.maydell@linaro.org) wrote:
> On 18 December 2017 at 17:30, Richard Henderson

> <richard.henderson@linaro.org> wrote:

> > Change vfp.regs as a uint64_t to vfp.zregs as an ARMVectorReg.

> > The previous patches have made the change in representation

> > relatively painless.

> >

> > Add vfp.pregs as an ARMPredicateReg.  Let FFR be P16 to make

> > it easier to treat it as for any other predicate.

> 

> > diff --git a/target/arm/machine.c b/target/arm/machine.c

> > index a85c2430d3..39f3123b10 100644

> > --- a/target/arm/machine.c

> > +++ b/target/arm/machine.c

> > @@ -50,7 +50,42 @@ static const VMStateDescription vmstate_vfp = {

> >      .minimum_version_id = 3,

> >      .needed = vfp_needed,

> >      .fields = (VMStateField[]) {

> > -        VMSTATE_UINT64_ARRAY(env.vfp.regs, ARMCPU, 64),

> > +        /* For compatibility, store Qn out of Zn here.  */

> > +        /* TODO: store the other quadwords in another subsection,

> > +           along with predicate registers and ZCR_EL[1-3].  */

> 

> I think we should get the migration state stuff right at the point

> where we change the data structure in cpu.h.

> 

> 

> > +        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[0].d, ARMCPU, 0, 2),

> > +        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[1].d, ARMCPU, 0, 2),

> > +        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[2].d, ARMCPU, 0, 2),

> 

> David: does the migration code guarantee that we can split an

> old VMSTATE_UINT64_ARRAY into being a sequence of SUB_ARRAYs

> like this and retain the same on-the-wire format (ie and

> keep migration compat) ?


There's no on-the-wire protocol for this - it's just the array elements,
so as long as the total adds up the same and it's still uint64's
it should work.

Dave

> I suspect we've used this clever trick before elsewhere, but thought

> it better to check...

> 

> thanks

> -- PMM

--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
Peter Maydell Jan. 12, 2018, 6:38 p.m. UTC | #3
On 18 December 2017 at 17:30, Richard Henderson
<richard.henderson@linaro.org> wrote:
> Change vfp.regs as a uint64_t to vfp.zregs as an ARMVectorReg.

> The previous patches have made the change in representation

> relatively painless.

>

> Add vfp.pregs as an ARMPredicateReg.  Let FFR be P16 to make

> it easier to treat it as for any other predicate.

>

> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

> ---

>  target/arm/cpu.h           | 64 ++++++++++++++++++++++++++++++++--------------

>  target/arm/machine.c       | 37 ++++++++++++++++++++++++++-

>  target/arm/translate-a64.c |  8 +++---

>  target/arm/translate.c     | 12 ++++-----

>  4 files changed, 90 insertions(+), 31 deletions(-)

>

> diff --git a/target/arm/cpu.h b/target/arm/cpu.h

> index e1a8e2880d..150b0d9d84 100644

> --- a/target/arm/cpu.h

> +++ b/target/arm/cpu.h

> @@ -153,6 +153,47 @@ typedef struct {

>      uint32_t base_mask;

>  } TCR;

>

> +/* Define a maximum sized vector register.

> + * For 32-bit, this is a 128-bit NEON/AdvSIMD register.

> + * For 64-bit, this is a 2048-bit SVE register.

> + *

> + * Note that the mapping between S, D, and Q views of the register bank

> + * differs between AArch64 and AArch32.

> + * In AArch32:

> + *  Qn = regs[n].d[1]:regs[n].d[0]

> + *  Dn = regs[n / 2].d[n & 1]

> + *  Sn = regs[n / 4].d[n % 4 / 2],

> + *       bits 31..0 for even n, and bits 63..32 for odd n

> + *       (and regs[16] to regs[31] are inaccessible)

> + * In AArch64:

> + *  Zn = regs[n].d[*]

> + *  Qn = regs[n].d[1]:regs[n].d[0]

> + *  Dn = regs[n].d[0]

> + *  Sn = regs[n].d[0] bits 31..0

> + *

> + * This corresponds to the architecturally defined mapping between

> + * the two execution states, and means we do not need to explicitly

> + * map these registers when changing states.

> + */


The transformations on the data structures (ie the meat of this patch
looks good).

> +

> +#ifdef TARGET_AARCH64

> +# define ARM_MAX_VQ    16

> +#else

> +# define ARM_MAX_VQ    1

> +#endif

> +

> +typedef struct ARMVectorReg {

> +    uint64_t d[2 * ARM_MAX_VQ] QEMU_ALIGNED(16);

> +} ARMVectorReg;

> +

> +typedef struct ARMPredicateReg {

> +#ifdef TARGET_AARCH64

> +    uint64_t p[2 * ARM_MAX_VQ / 8] QEMU_ALIGNED(16);

> +#else

> +    uint64_t p[0];

> +#endif

> +} ARMPredicateReg;


I think introducing the predicate registers should go in a separate
patch.

> --- a/target/arm/translate.c

> +++ b/target/arm/translate.c

> @@ -1513,19 +1513,17 @@ static inline void gen_vfp_st(DisasContext *s, int dp, TCGv_i32 addr)

>      }

>  }

>

> -static inline long

> -vfp_reg_offset (int dp, int reg)

> +static inline long vfp_reg_offset(bool dp, unsigned reg)

>  {

>      if (dp) {

> -        return offsetof(CPUARMState, vfp.regs[reg]);

> +        return offsetof(CPUARMState, vfp.zregs[reg >> 1].d[reg & 1]);

>      } else {

> -        long ofs = offsetof(CPUARMState, vfp.regs[reg >> 1]);

> +        long r = offsetof(CPUARMState, vfp.zregs[reg >> 2].d[(reg >> 1) & 1]);

>          if (reg & 1) {

> -            ofs += offsetof(CPU_DoubleU, l.upper);

> +            return r + offsetof(CPU_DoubleU, l.upper);

>          } else {

> -            ofs += offsetof(CPU_DoubleU, l.lower);

> +            return r + offsetof(CPU_DoubleU, l.lower);

>          }

> -        return ofs;


...I see we're tweaking the logic on this code again. I was
expecting that the changes in the previous patch would have turned
out to be in support of just having to do a one-line change in this
one, but apparently not ?

>      }

>  }


thanks
-- PMM
Richard Henderson Jan. 12, 2018, 6:50 p.m. UTC | #4
On 01/12/2018 10:38 AM, Peter Maydell wrote:
>> -        long ofs = offsetof(CPUARMState, vfp.regs[reg >> 1]);

>> +        long r = offsetof(CPUARMState, vfp.zregs[reg >> 2].d[(reg >> 1) & 1]);

>>          if (reg & 1) {

>> -            ofs += offsetof(CPU_DoubleU, l.upper);

>> +            return r + offsetof(CPU_DoubleU, l.upper);

>>          } else {

>> -            ofs += offsetof(CPU_DoubleU, l.lower);

>> +            return r + offsetof(CPU_DoubleU, l.lower);

>>          }

>> -        return ofs;

> ...I see we're tweaking the logic on this code again. I was

> expecting that the changes in the previous patch would have turned

> out to be in support of just having to do a one-line change in this

> one, but apparently not ?

> 


I don't know why I changed my mind about where the return would go.  It should
have just been the one line change...

r~
diff mbox series

Patch

diff --git a/target/arm/cpu.h b/target/arm/cpu.h
index e1a8e2880d..150b0d9d84 100644
--- a/target/arm/cpu.h
+++ b/target/arm/cpu.h
@@ -153,6 +153,47 @@  typedef struct {
     uint32_t base_mask;
 } TCR;
 
+/* Define a maximum sized vector register.
+ * For 32-bit, this is a 128-bit NEON/AdvSIMD register.
+ * For 64-bit, this is a 2048-bit SVE register.
+ *
+ * Note that the mapping between S, D, and Q views of the register bank
+ * differs between AArch64 and AArch32.
+ * In AArch32:
+ *  Qn = regs[n].d[1]:regs[n].d[0]
+ *  Dn = regs[n / 2].d[n & 1]
+ *  Sn = regs[n / 4].d[n % 4 / 2],
+ *       bits 31..0 for even n, and bits 63..32 for odd n
+ *       (and regs[16] to regs[31] are inaccessible)
+ * In AArch64:
+ *  Zn = regs[n].d[*]
+ *  Qn = regs[n].d[1]:regs[n].d[0]
+ *  Dn = regs[n].d[0]
+ *  Sn = regs[n].d[0] bits 31..0
+ *
+ * This corresponds to the architecturally defined mapping between
+ * the two execution states, and means we do not need to explicitly
+ * map these registers when changing states.
+ */
+
+#ifdef TARGET_AARCH64
+# define ARM_MAX_VQ    16
+#else
+# define ARM_MAX_VQ    1
+#endif
+
+typedef struct ARMVectorReg {
+    uint64_t d[2 * ARM_MAX_VQ] QEMU_ALIGNED(16);
+} ARMVectorReg;
+
+typedef struct ARMPredicateReg {
+#ifdef TARGET_AARCH64
+    uint64_t p[2 * ARM_MAX_VQ / 8] QEMU_ALIGNED(16);
+#else
+    uint64_t p[0];
+#endif
+} ARMPredicateReg;
+
 typedef struct CPUARMState {
     /* Regs for current mode.  */
     uint32_t regs[16];
@@ -477,23 +518,8 @@  typedef struct CPUARMState {
 
     /* VFP coprocessor state.  */
     struct {
-        /* VFP/Neon register state. Note that the mapping between S, D and Q
-         * views of the register bank differs between AArch64 and AArch32:
-         * In AArch32:
-         *  Qn = regs[2n+1]:regs[2n]
-         *  Dn = regs[n]
-         *  Sn = regs[n/2] bits 31..0 for even n, and bits 63..32 for odd n
-         * (and regs[32] to regs[63] are inaccessible)
-         * In AArch64:
-         *  Qn = regs[2n+1]:regs[2n]
-         *  Dn = regs[2n]
-         *  Sn = regs[2n] bits 31..0
-         *  Hn = regs[2n] bits 15..0 for even n, and bits 31..16 for odd n
-         * This corresponds to the architecturally defined mapping between
-         * the two execution states, and means we do not need to explicitly
-         * map these registers when changing states.
-         */
-        uint64_t regs[64] QEMU_ALIGNED(16);
+        ARMVectorReg zregs[32];
+        ARMPredicateReg pregs[17];  /* ffr is pregs[16] */
 
         uint32_t xregs[16];
         /* We store these fpcsr fields separately for convenience.  */
@@ -2905,7 +2931,7 @@  static inline void *arm_get_el_change_hook_opaque(ARMCPU *cpu)
  */
 static inline uint64_t *aa32_vfp_dreg(CPUARMState *env, unsigned regno)
 {
-    return &env->vfp.regs[regno];
+    return &env->vfp.zregs[regno >> 1].d[regno & 1];
 }
 
 /**
@@ -2914,7 +2940,7 @@  static inline uint64_t *aa32_vfp_dreg(CPUARMState *env, unsigned regno)
  */
 static inline uint64_t *aa64_vfp_qreg(CPUARMState *env, unsigned regno)
 {
-    return &env->vfp.regs[2 * regno];
+    return &env->vfp.zregs[regno].d[0];
 }
 
 #endif
diff --git a/target/arm/machine.c b/target/arm/machine.c
index a85c2430d3..39f3123b10 100644
--- a/target/arm/machine.c
+++ b/target/arm/machine.c
@@ -50,7 +50,42 @@  static const VMStateDescription vmstate_vfp = {
     .minimum_version_id = 3,
     .needed = vfp_needed,
     .fields = (VMStateField[]) {
-        VMSTATE_UINT64_ARRAY(env.vfp.regs, ARMCPU, 64),
+        /* For compatibility, store Qn out of Zn here.  */
+        /* TODO: store the other quadwords in another subsection,
+           along with predicate registers and ZCR_EL[1-3].  */
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[0].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[1].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[2].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[3].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[4].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[5].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[6].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[7].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[8].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[9].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[10].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[11].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[12].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[13].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[14].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[15].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[16].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[17].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[18].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[19].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[20].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[21].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[22].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[23].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[24].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[25].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[26].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[27].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[28].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[29].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[30].d, ARMCPU, 0, 2),
+        VMSTATE_UINT64_SUB_ARRAY(env.vfp.zregs[31].d, ARMCPU, 0, 2),
+
         /* The xregs array is a little awkward because element 1 (FPSCR)
          * requires a specific accessor, so we have to split it up in
          * the vmstate:
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index e17c7269d4..487408a7a7 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -523,8 +523,8 @@  static inline int vec_reg_offset(DisasContext *s, int regno,
 {
     int offs = 0;
 #ifdef HOST_WORDS_BIGENDIAN
-    /* This is complicated slightly because vfp.regs[2n] is
-     * still the low half and  vfp.regs[2n+1] the high half
+    /* This is complicated slightly because vfp.zregs[n].d[0] is
+     * still the low half and vfp.zregs[n].d[1] the high half
      * of the 128 bit vector, even on big endian systems.
      * Calculate the offset assuming a fully bigendian 128 bits,
      * then XOR to account for the order of the two 64 bit halves.
@@ -534,7 +534,7 @@  static inline int vec_reg_offset(DisasContext *s, int regno,
 #else
     offs += element * (1 << size);
 #endif
-    offs += offsetof(CPUARMState, vfp.regs[regno * 2]);
+    offs += offsetof(CPUARMState, vfp.zregs[regno]);
     assert_fp_access_checked(s);
     return offs;
 }
@@ -543,7 +543,7 @@  static inline int vec_reg_offset(DisasContext *s, int regno,
 static inline int vec_full_reg_offset(DisasContext *s, int regno)
 {
     assert_fp_access_checked(s);
-    return offsetof(CPUARMState, vfp.regs[regno * 2]);
+    return offsetof(CPUARMState, vfp.zregs[regno]);
 }
 
 /* Return the byte size of the "whole" vector register, VL / 8.  */
diff --git a/target/arm/translate.c b/target/arm/translate.c
index a93e26498e..7a5e493333 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -1513,19 +1513,17 @@  static inline void gen_vfp_st(DisasContext *s, int dp, TCGv_i32 addr)
     }
 }
 
-static inline long
-vfp_reg_offset (int dp, int reg)
+static inline long vfp_reg_offset(bool dp, unsigned reg)
 {
     if (dp) {
-        return offsetof(CPUARMState, vfp.regs[reg]);
+        return offsetof(CPUARMState, vfp.zregs[reg >> 1].d[reg & 1]);
     } else {
-        long ofs = offsetof(CPUARMState, vfp.regs[reg >> 1]);
+        long r = offsetof(CPUARMState, vfp.zregs[reg >> 2].d[(reg >> 1) & 1]);
         if (reg & 1) {
-            ofs += offsetof(CPU_DoubleU, l.upper);
+            return r + offsetof(CPU_DoubleU, l.upper);
         } else {
-            ofs += offsetof(CPU_DoubleU, l.lower);
+            return r + offsetof(CPU_DoubleU, l.lower);
         }
-        return ofs;
     }
 }