diff mbox series

[v1,06/15] tcg/riscv: Implement vector load/store

Message ID 20240813113436.831-7-zhiwei_liu@linux.alibaba.com
State New
Headers show
Series tcg/riscv: Add support for vector | expand

Commit Message

LIU Zhiwei Aug. 13, 2024, 11:34 a.m. UTC 
From: TANG Tiancheng <tangtiancheng.ttc@alibaba-inc.com>

Signed-off-by: TANG Tiancheng <tangtiancheng.ttc@alibaba-inc.com>
Reviewed-by: Liu Zhiwei <zhiwei_liu@linux.alibaba.com>
---
 tcg/riscv/tcg-target-con-set.h |  2 +
 tcg/riscv/tcg-target.c.inc     | 92 ++++++++++++++++++++++++++++++++--
 2 files changed, 91 insertions(+), 3 deletions(-)

Comments

Richard Henderson Aug. 14, 2024, 9:01 a.m. UTC | #1 
On 8/13/24 21:34, LIU Zhiwei wrote:
> @@ -827,14 +850,59 @@ static void tcg_out_ldst(TCGContext *s, RISCVInsn opc, TCGReg data,
>   static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
>                          TCGReg arg1, intptr_t arg2)
>   {
> -    RISCVInsn insn = type == TCG_TYPE_I32 ? OPC_LW : OPC_LD;
> +    RISCVInsn insn;
> +
> +    if (type < TCG_TYPE_V64) {
> +        insn = type == TCG_TYPE_I32 ? OPC_LW : OPC_LD;
> +    } else {
> +        tcg_debug_assert(arg >= TCG_REG_V1);
> +        switch (prev_vece) {
> +        case MO_8:
> +            insn = OPC_VLE8_V;
> +            break;
> +        case MO_16:
> +            insn = OPC_VLE16_V;
> +            break;
> +        case MO_32:
> +            insn = OPC_VLE32_V;
> +            break;
> +        case MO_64:
> +            insn = OPC_VLE64_V;
> +            break;
> +        default:
> +            g_assert_not_reached();
> +        }
> +    }
>       tcg_out_ldst(s, insn, arg, arg1, arg2);

tcg_out_ld/st are called directly from register allocation spill/fill.
You'll need to set vtype here, and cannot rely on this having been done in tcg_out_vec_op.

That said, with a little-endian host, the selected element size doesn't matter *too* much. 
  A write of 8 uint16_t or a write of 2 uint64_t produces the same bits in memory.

Therefore you can examine prev_vtype and adjust only if the vector length changes.  But we 
do that -- e.g. load V256, store V256, store V128 to perform a 384-bit store for AArch64 
SVE when VQ=3.

Is there an advantage to using the vector load/store whole register insns, if the 
requested length is not too small?  IIRC the NF field can be used to store multiples, but 
we can't store half of a register with these.


r~
LIU Zhiwei Aug. 19, 2024, 1:41 a.m. UTC | #2 
On 2024/8/14 17:01, Richard Henderson wrote:
> On 8/13/24 21:34, LIU Zhiwei wrote:
>> @@ -827,14 +850,59 @@ static void tcg_out_ldst(TCGContext *s, 
>> RISCVInsn opc, TCGReg data,
>>   static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
>>                          TCGReg arg1, intptr_t arg2)
>>   {
>> -    RISCVInsn insn = type == TCG_TYPE_I32 ? OPC_LW : OPC_LD;
>> +    RISCVInsn insn;
>> +
>> +    if (type < TCG_TYPE_V64) {
>> +        insn = type == TCG_TYPE_I32 ? OPC_LW : OPC_LD;
>> +    } else {
>> +        tcg_debug_assert(arg >= TCG_REG_V1);
>> +        switch (prev_vece) {
>> +        case MO_8:
>> +            insn = OPC_VLE8_V;
>> +            break;
>> +        case MO_16:
>> +            insn = OPC_VLE16_V;
>> +            break;
>> +        case MO_32:
>> +            insn = OPC_VLE32_V;
>> +            break;
>> +        case MO_64:
>> +            insn = OPC_VLE64_V;
>> +            break;
>> +        default:
>> +            g_assert_not_reached();
>> +        }
>> +    }
>>       tcg_out_ldst(s, insn, arg, arg1, arg2);
>
> tcg_out_ld/st are called directly from register allocation spill/fill.
> You'll need to set vtype here, and cannot rely on this having been 
> done in tcg_out_vec_op.
OK.
>
> That said, with a little-endian host, the selected element size 
> doesn't matter *too* much.  A write of 8 uint16_t or a write of 2 
> uint64_t produces the same bits in memory.
>
> Therefore you can examine prev_vtype and adjust only if the vector 
> length changes.
OK.
>   But we do that -- e.g. load V256, store V256, store V128 to perform 
> a 384-bit store for AArch64 SVE when VQ=3.
>
> Is there an advantage to using the vector load/store whole register 
> insns, if the requested length is not too small?
For vector type equal or bigger than vlen in host, we will use the whole 
register instructions.
>   IIRC the NF field can be used to store multiples, but we can't store 
> half of a register with these.

I think we can still use the unit-stride instructions for them.

Thanks,
Zhiwei

>
>
> r~
diff mbox series

Patch

diff --git a/tcg/riscv/tcg-target-con-set.h b/tcg/riscv/tcg-target-con-set.h
index aac5ceee2b..d73a62b0f2 100644
--- a/tcg/riscv/tcg-target-con-set.h
+++ b/tcg/riscv/tcg-target-con-set.h
@@ -21,3 +21,5 @@  C_O1_I2(r, rZ, rZ)
 C_N1_I2(r, r, rM)
 C_O1_I4(r, r, rI, rM, rM)
 C_O2_I4(r, r, rZ, rZ, rM, rM)
+C_O0_I2(v, r)
+C_O1_I1(v, r)
diff --git a/tcg/riscv/tcg-target.c.inc b/tcg/riscv/tcg-target.c.inc
index d17f523187..f17d679d71 100644
--- a/tcg/riscv/tcg-target.c.inc
+++ b/tcg/riscv/tcg-target.c.inc
@@ -279,6 +279,15 @@  typedef enum {
     OPC_VSETVLI  = 0x57 | V_OPCFG,
     OPC_VSETIVLI = 0xc0000057 | V_OPCFG,
     OPC_VSETVL   = 0x80000057 | V_OPCFG,
+
+    OPC_VLE8_V  = 0x7 | V_LUMOP,
+    OPC_VLE16_V = 0x5007 | V_LUMOP,
+    OPC_VLE32_V = 0x6007 | V_LUMOP,
+    OPC_VLE64_V = 0x7007 | V_LUMOP,
+    OPC_VSE8_V  = 0x27 | V_SUMOP,
+    OPC_VSE16_V = 0x5027 | V_SUMOP,
+    OPC_VSE32_V = 0x6027 | V_SUMOP,
+    OPC_VSE64_V = 0x7027 | V_SUMOP,
 } RISCVInsn;
 
 /*
@@ -810,6 +819,13 @@  static void tcg_out_ldst(TCGContext *s, RISCVInsn opc, TCGReg data,
     case OPC_SD:
         tcg_out_opc_store(s, opc, addr, data, imm12);
         break;
+    case OPC_VSE8_V:
+    case OPC_VSE16_V:
+    case OPC_VSE32_V:
+    case OPC_VSE64_V:
+        tcg_out_opc_imm(s, OPC_ADDI, TCG_REG_TMP0, addr, imm12);
+        tcg_out_opc_ldst_vec(s, opc, data, TCG_REG_TMP0, true);
+        break;
     case OPC_LB:
     case OPC_LBU:
     case OPC_LH:
@@ -819,6 +835,13 @@  static void tcg_out_ldst(TCGContext *s, RISCVInsn opc, TCGReg data,
     case OPC_LD:
         tcg_out_opc_imm(s, opc, data, addr, imm12);
         break;
+    case OPC_VLE8_V:
+    case OPC_VLE16_V:
+    case OPC_VLE32_V:
+    case OPC_VLE64_V:
+        tcg_out_opc_imm(s, OPC_ADDI, TCG_REG_TMP0, addr, imm12);
+        tcg_out_opc_ldst_vec(s, opc, data, TCG_REG_TMP0, true);
+        break;
     default:
         g_assert_not_reached();
     }
@@ -827,14 +850,59 @@  static void tcg_out_ldst(TCGContext *s, RISCVInsn opc, TCGReg data,
 static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
                        TCGReg arg1, intptr_t arg2)
 {
-    RISCVInsn insn = type == TCG_TYPE_I32 ? OPC_LW : OPC_LD;
+    RISCVInsn insn;
+
+    if (type < TCG_TYPE_V64) {
+        insn = type == TCG_TYPE_I32 ? OPC_LW : OPC_LD;
+    } else {
+        tcg_debug_assert(arg >= TCG_REG_V1);
+        switch (prev_vece) {
+        case MO_8:
+            insn = OPC_VLE8_V;
+            break;
+        case MO_16:
+            insn = OPC_VLE16_V;
+            break;
+        case MO_32:
+            insn = OPC_VLE32_V;
+            break;
+        case MO_64:
+            insn = OPC_VLE64_V;
+            break;
+        default:
+            g_assert_not_reached();
+        }
+    }
     tcg_out_ldst(s, insn, arg, arg1, arg2);
 }
 
 static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
                        TCGReg arg1, intptr_t arg2)
 {
-    RISCVInsn insn = type == TCG_TYPE_I32 ? OPC_SW : OPC_SD;
+    RISCVInsn insn;
+
+    if (type < TCG_TYPE_V64) {
+        insn = type == TCG_TYPE_I32 ? OPC_SW : OPC_SD;
+        tcg_out_ldst(s, insn, arg, arg1, arg2);
+    } else {
+        tcg_debug_assert(arg >= TCG_REG_V1);
+        switch (prev_vece) {
+        case MO_8:
+            insn = OPC_VSE8_V;
+            break;
+        case MO_16:
+            insn = OPC_VSE16_V;
+            break;
+        case MO_32:
+            insn = OPC_VSE32_V;
+            break;
+        case MO_64:
+            insn = OPC_VSE64_V;
+            break;
+        default:
+            g_assert_not_reached();
+        }
+    }
     tcg_out_ldst(s, insn, arg, arg1, arg2);
 }
 
@@ -2030,11 +2098,25 @@  static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
                            const int const_args[TCG_MAX_OP_ARGS])
 {
     TCGType type = vecl + TCG_TYPE_V64;
+    TCGArg a0, a1, a2;
+
+    a0 = args[0];
+    a1 = args[1];
+    a2 = args[2];
 
-    if (vec_vtpye_init) {
+    if (!vec_vtpye_init &&
+        (opc == INDEX_op_ld_vec || opc == INDEX_op_st_vec)) {
+        tcg_target_set_vec_config(s, type, prev_vece);
+    } else {
         tcg_target_set_vec_config(s, type, vece);
     }
     switch (opc) {
+    case INDEX_op_ld_vec:
+        tcg_out_ld(s, type, a0, a1, a2);
+        break;
+    case INDEX_op_st_vec:
+        tcg_out_st(s, type, a0, a1, a2);
+        break;
     case INDEX_op_mov_vec: /* Always emitted via tcg_out_mov.  */
     case INDEX_op_dup_vec: /* Always emitted via tcg_out_dup_vec.  */
     default:
@@ -2198,6 +2280,10 @@  static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
     case INDEX_op_qemu_st_a64_i64:
         return C_O0_I2(rZ, r);
 
+    case INDEX_op_st_vec:
+        return C_O0_I2(v, r);
+    case INDEX_op_ld_vec:
+        return C_O1_I1(v, r);
     default:
         g_assert_not_reached();
     }