[v1] Lower TCG vector ops to LSX

[PATCH 03/11] tcg/loongarch64: Lower cmp_vec to vseq/vsle/vslt

Posted by Jiajie Chen 1 year, 3 months ago

Signed-off-by: Jiajie Chen <c@jia.je>
---
 tcg/loongarch64/tcg-target-con-set.h |  1 +
 tcg/loongarch64/tcg-target.c.inc     | 25 +++++++++++++++++++++++++
 2 files changed, 26 insertions(+)

diff --git a/tcg/loongarch64/tcg-target-con-set.h b/tcg/loongarch64/tcg-target-con-set.h
index 37b3f80bf9..e80fc7f3f7 100644
--- a/tcg/loongarch64/tcg-target-con-set.h
+++ b/tcg/loongarch64/tcg-target-con-set.h
@@ -31,4 +31,5 @@ C_O1_I2(r, 0, rZ)
 C_O1_I2(r, rZ, ri)
 C_O1_I2(r, rZ, rJ)
 C_O1_I2(r, rZ, rZ)
+C_O1_I2(w, w, w)
 C_O1_I4(r, rZ, rJ, rZ, rZ)
diff --git a/tcg/loongarch64/tcg-target.c.inc b/tcg/loongarch64/tcg-target.c.inc
index 0f9427572c..cc80e5fa20 100644
--- a/tcg/loongarch64/tcg-target.c.inc
+++ b/tcg/loongarch64/tcg-target.c.inc
@@ -1624,6 +1624,15 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
     TCGReg temp = TCG_REG_TMP0;
     int32_t offset;
 
+    static const LoongArchInsn cmp_vec_insn[16][4] = {
+        [TCG_COND_EQ] = {OPC_VSEQ_B, OPC_VSEQ_H, OPC_VSEQ_W, OPC_VSEQ_D},
+        [TCG_COND_LE] = {OPC_VSLE_B, OPC_VSLE_H, OPC_VSLE_W, OPC_VSLE_D},
+        [TCG_COND_LEU] = {OPC_VSLE_BU, OPC_VSLE_HU, OPC_VSLE_WU, OPC_VSLE_DU},
+        [TCG_COND_LT] = {OPC_VSLT_B, OPC_VSLT_H, OPC_VSLT_W, OPC_VSLT_D},
+        [TCG_COND_LTU] = {OPC_VSLT_BU, OPC_VSLT_HU, OPC_VSLT_WU, OPC_VSLT_DU},
+    };
+    LoongArchInsn insn;
+
     a0 = args[0];
     a1 = args[1];
     a2 = args[2];
@@ -1656,6 +1665,18 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
         }
         tcg_out_opc_vld(s, a0, base, offset);
         break;
+    case INDEX_op_cmp_vec:
+        TCGCond cond = args[3];
+        insn = cmp_vec_insn[cond][vece];
+        if (insn == 0) {
+            TCGArg t;
+            t = a1, a1 = a2, a2 = t;
+            cond = tcg_swap_cond(cond);
+            insn = cmp_vec_insn[cond][vece];
+            tcg_debug_assert(insn != 0);
+        }
+        tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
+        break;
     case INDEX_op_dupm_vec:
         tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
         break;
@@ -1671,6 +1692,7 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
     case INDEX_op_st_vec:
     case INDEX_op_dup_vec:
     case INDEX_op_dupm_vec:
+    case INDEX_op_cmp_vec:
         return 1;
     default:
         return 0;
@@ -1832,6 +1854,9 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
     case INDEX_op_st_vec:
         return C_O0_I2(w, r);
 
+    case INDEX_op_cmp_vec:
+        return C_O1_I2(w, w, w);
+
     default:
         g_assert_not_reached();
     }
-- 
2.42.0

Re: [PATCH 03/11] tcg/loongarch64: Lower cmp_vec to vseq/vsle/vslt

Posted by Richard Henderson 1 year, 3 months ago

On 8/28/23 08:19, Jiajie Chen wrote:
> Signed-off-by: Jiajie Chen <c@jia.je>
> ---
>   tcg/loongarch64/tcg-target-con-set.h |  1 +
>   tcg/loongarch64/tcg-target.c.inc     | 25 +++++++++++++++++++++++++
>   2 files changed, 26 insertions(+)

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

> @@ -1624,6 +1624,15 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
>       TCGReg temp = TCG_REG_TMP0;
>       int32_t offset;
>   
> +    static const LoongArchInsn cmp_vec_insn[16][4] = {
> +        [TCG_COND_EQ] = {OPC_VSEQ_B, OPC_VSEQ_H, OPC_VSEQ_W, OPC_VSEQ_D},
> +        [TCG_COND_LE] = {OPC_VSLE_B, OPC_VSLE_H, OPC_VSLE_W, OPC_VSLE_D},
> +        [TCG_COND_LEU] = {OPC_VSLE_BU, OPC_VSLE_HU, OPC_VSLE_WU, OPC_VSLE_DU},
> +        [TCG_COND_LT] = {OPC_VSLT_B, OPC_VSLT_H, OPC_VSLT_W, OPC_VSLT_D},
> +        [TCG_COND_LTU] = {OPC_VSLT_BU, OPC_VSLT_HU, OPC_VSLT_WU, OPC_VSLT_DU},
> +    };
> +    LoongArchInsn insn;
> +
>       a0 = args[0];
>       a1 = args[1];
>       a2 = args[2];
> @@ -1656,6 +1665,18 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
>           }
>           tcg_out_opc_vld(s, a0, base, offset);
>           break;
> +    case INDEX_op_cmp_vec:
> +        TCGCond cond = args[3];
> +        insn = cmp_vec_insn[cond][vece];
> +        if (insn == 0) {
> +            TCGArg t;
> +            t = a1, a1 = a2, a2 = t;
> +            cond = tcg_swap_cond(cond);
> +            insn = cmp_vec_insn[cond][vece];
> +            tcg_debug_assert(insn != 0);
> +        }
> +        tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
> +        break;
>       case INDEX_op_dupm_vec:
>           tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
>           break;
> @@ -1671,6 +1692,7 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
>       case INDEX_op_st_vec:
>       case INDEX_op_dup_vec:
>       case INDEX_op_dupm_vec:
> +    case INDEX_op_cmp_vec:
>           return 1;
>       default:
>           return 0;
> @@ -1832,6 +1854,9 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
>       case INDEX_op_st_vec:
>           return C_O0_I2(w, r);
>   
> +    case INDEX_op_cmp_vec:
> +        return C_O1_I2(w, w, w);

Further improvement can be done with VS*I, and its 5-bit signed immediate.
While you can't swap operands for GT/GE in that case, you can invert, e.g.
GT -> LE, then invert the result using VNOR.

Comparing vs zero is especially common...


r~