qemu-devel.nongnu.org archive mirror
 help / color / mirror / Atom feed
From: Richard Henderson <richard.henderson@linaro.org>
To: qemu-devel@nongnu.org
Cc: peter.maydell@linaro.org
Subject: [PATCH 11/11] target/arm: Improve do_prewiden_3d
Date: Tue, 27 Oct 2020 20:27:03 -0700	[thread overview]
Message-ID: <20201028032703.201526-12-richard.henderson@linaro.org> (raw)
In-Reply-To: <20201028032703.201526-1-richard.henderson@linaro.org>

We can use proper widening loads to extend 32-bit inputs,
and skip the "widenfn" step.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/arm/translate.c          |  6 +++
 target/arm/translate-neon.c.inc | 66 ++++++++++++++++++---------------
 2 files changed, 43 insertions(+), 29 deletions(-)

diff --git a/target/arm/translate.c b/target/arm/translate.c
index 7611c1f0f1..29ea1eb781 100644
--- a/target/arm/translate.c
+++ b/target/arm/translate.c
@@ -1183,6 +1183,12 @@ static void read_neon_element64(TCGv_i64 dest, int reg, int ele, MemOp memop)
     long off = neon_element_offset(reg, ele, memop);
 
     switch (memop) {
+    case MO_SL:
+        tcg_gen_ld32s_i64(dest, cpu_env, off);
+        break;
+    case MO_UL:
+        tcg_gen_ld32u_i64(dest, cpu_env, off);
+        break;
     case MO_Q:
         tcg_gen_ld_i64(dest, cpu_env, off);
         break;
diff --git a/target/arm/translate-neon.c.inc b/target/arm/translate-neon.c.inc
index 7cd41c79ec..8f33b54067 100644
--- a/target/arm/translate-neon.c.inc
+++ b/target/arm/translate-neon.c.inc
@@ -1788,11 +1788,10 @@ static bool trans_Vimm_1r(DisasContext *s, arg_1reg_imm *a)
 static bool do_prewiden_3d(DisasContext *s, arg_3diff *a,
                            NeonGenWidenFn *widenfn,
                            NeonGenTwo64OpFn *opfn,
-                           bool src1_wide)
+                           int src1_mop, int src2_mop)
 {
     /* 3-regs different lengths, prewidening case (VADDL/VSUBL/VAADW/VSUBW) */
     TCGv_i64 rn0_64, rn1_64, rm_64;
-    TCGv_i32 rm;
 
     if (!arm_dc_feature(s, ARM_FEATURE_NEON)) {
         return false;
@@ -1804,12 +1803,12 @@ static bool do_prewiden_3d(DisasContext *s, arg_3diff *a,
         return false;
     }
 
-    if (!widenfn || !opfn) {
+    if (!opfn) {
         /* size == 3 case, which is an entirely different insn group */
         return false;
     }
 
-    if ((a->vd & 1) || (src1_wide && (a->vn & 1))) {
+    if ((a->vd & 1) || (src1_mop == MO_Q && (a->vn & 1))) {
         return false;
     }
 
@@ -1821,40 +1820,48 @@ static bool do_prewiden_3d(DisasContext *s, arg_3diff *a,
     rn1_64 = tcg_temp_new_i64();
     rm_64 = tcg_temp_new_i64();
 
-    if (src1_wide) {
-        read_neon_element64(rn0_64, a->vn, 0, MO_64);
+    if (src1_mop >= 0) {
+        read_neon_element64(rn0_64, a->vn, 0, src1_mop);
     } else {
         TCGv_i32 tmp = tcg_temp_new_i32();
         read_neon_element32(tmp, a->vn, 0, MO_32);
         widenfn(rn0_64, tmp);
         tcg_temp_free_i32(tmp);
     }
-    rm = tcg_temp_new_i32();
-    read_neon_element32(rm, a->vm, 0, MO_32);
+    if (src2_mop >= 0) {
+        read_neon_element64(rm_64, a->vm, 0, src2_mop);
+    } else {
+        TCGv_i32 tmp = tcg_temp_new_i32();
+        read_neon_element32(tmp, a->vm, 0, MO_32);
+        widenfn(rm_64, tmp);
+        tcg_temp_free_i32(tmp);
+    }
 
-    widenfn(rm_64, rm);
-    tcg_temp_free_i32(rm);
     opfn(rn0_64, rn0_64, rm_64);
 
     /*
      * Load second pass inputs before storing the first pass result, to
      * avoid incorrect results if a narrow input overlaps with the result.
      */
-    if (src1_wide) {
-        read_neon_element64(rn1_64, a->vn, 1, MO_64);
+    if (src1_mop >= 0) {
+        read_neon_element64(rn1_64, a->vn, 1, src1_mop);
     } else {
         TCGv_i32 tmp = tcg_temp_new_i32();
         read_neon_element32(tmp, a->vn, 1, MO_32);
         widenfn(rn1_64, tmp);
         tcg_temp_free_i32(tmp);
     }
-    rm = tcg_temp_new_i32();
-    read_neon_element32(rm, a->vm, 1, MO_32);
+    if (src2_mop >= 0) {
+        read_neon_element64(rm_64, a->vm, 1, src2_mop);
+    } else {
+        TCGv_i32 tmp = tcg_temp_new_i32();
+        read_neon_element32(tmp, a->vm, 1, MO_32);
+        widenfn(rm_64, tmp);
+        tcg_temp_free_i32(tmp);
+    }
 
     write_neon_element64(rn0_64, a->vd, 0, MO_64);
 
-    widenfn(rm_64, rm);
-    tcg_temp_free_i32(rm);
     opfn(rn1_64, rn1_64, rm_64);
     write_neon_element64(rn1_64, a->vd, 1, MO_64);
 
@@ -1865,14 +1872,13 @@ static bool do_prewiden_3d(DisasContext *s, arg_3diff *a,
     return true;
 }
 
-#define DO_PREWIDEN(INSN, S, EXT, OP, SRC1WIDE)                         \
+#define DO_PREWIDEN(INSN, S, OP, SRC1WIDE, SIGN)                        \
     static bool trans_##INSN##_3d(DisasContext *s, arg_3diff *a)        \
     {                                                                   \
         static NeonGenWidenFn * const widenfn[] = {                     \
             gen_helper_neon_widen_##S##8,                               \
             gen_helper_neon_widen_##S##16,                              \
-            tcg_gen_##EXT##_i32_i64,                                    \
-            NULL,                                                       \
+            NULL, NULL,                                                 \
         };                                                              \
         static NeonGenTwo64OpFn * const addfn[] = {                     \
             gen_helper_neon_##OP##l_u16,                                \
@@ -1880,18 +1886,20 @@ static bool do_prewiden_3d(DisasContext *s, arg_3diff *a,
             tcg_gen_##OP##_i64,                                         \
             NULL,                                                       \
         };                                                              \
-        return do_prewiden_3d(s, a, widenfn[a->size],                   \
-                              addfn[a->size], SRC1WIDE);                \
+        int narrow_mop = a->size == MO_32 ? MO_32 | SIGN : -1;          \
+        return do_prewiden_3d(s, a, widenfn[a->size], addfn[a->size],   \
+                              SRC1WIDE ? MO_Q : narrow_mop,             \
+                              narrow_mop);                              \
     }
 
-DO_PREWIDEN(VADDL_S, s, ext, add, false)
-DO_PREWIDEN(VADDL_U, u, extu, add, false)
-DO_PREWIDEN(VSUBL_S, s, ext, sub, false)
-DO_PREWIDEN(VSUBL_U, u, extu, sub, false)
-DO_PREWIDEN(VADDW_S, s, ext, add, true)
-DO_PREWIDEN(VADDW_U, u, extu, add, true)
-DO_PREWIDEN(VSUBW_S, s, ext, sub, true)
-DO_PREWIDEN(VSUBW_U, u, extu, sub, true)
+DO_PREWIDEN(VADDL_S, s, add, false, MO_SIGN)
+DO_PREWIDEN(VADDL_U, u, add, false, 0)
+DO_PREWIDEN(VSUBL_S, s, sub, false, MO_SIGN)
+DO_PREWIDEN(VSUBL_U, u, sub, false, 0)
+DO_PREWIDEN(VADDW_S, s, add, true, MO_SIGN)
+DO_PREWIDEN(VADDW_U, u, add, true, 0)
+DO_PREWIDEN(VSUBW_S, s, sub, true, MO_SIGN)
+DO_PREWIDEN(VSUBW_U, u, sub, true, 0)
 
 static bool do_narrow_3d(DisasContext *s, arg_3diff *a,
                          NeonGenTwo64OpFn *opfn, NeonGenNarrowFn *narrowfn)
-- 
2.25.1



  parent reply	other threads:[~2020-10-28  3:31 UTC|newest]

Thread overview: 19+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-10-28  3:26 [PATCH 00/11] target/arm: Fix neon reg offsets Richard Henderson
2020-10-28  3:26 ` [PATCH 01/11] target/arm: Introduce neon_full_reg_offset Richard Henderson
2020-10-28  3:26 ` [PATCH 02/11] target/arm: Move neon_element_offset to translate.c Richard Henderson
2020-10-28  3:26 ` [PATCH 03/11] target/arm: Use neon_element_offset in neon_load/store_reg Richard Henderson
2020-10-28  3:26 ` [PATCH 04/11] target/arm: Use neon_element_offset in vfp_reg_offset Richard Henderson
2020-10-28  3:26 ` [PATCH 05/11] target/arm: Add read/write_neon_element32 Richard Henderson
2020-10-28 20:22   ` Richard Henderson
2020-10-29 10:15     ` Peter Maydell
2020-10-28  3:26 ` [PATCH 06/11] target/arm: Expand read/write_neon_element32 to all MemOp Richard Henderson
2020-10-28  3:26 ` [PATCH 07/11] target/arm: Rename neon_load_reg32 to vfp_load_reg32 Richard Henderson
2020-10-28  3:27 ` [PATCH 08/11] target/arm: Add read/write_neon_element64 Richard Henderson
2020-10-28  3:27 ` [PATCH 09/11] target/arm: Rename neon_load_reg64 to vfp_load_reg64 Richard Henderson
2020-10-28  3:27 ` [PATCH 10/11] target/arm: Simplify do_long_3d and do_2scalar_long Richard Henderson
2020-10-28  3:27 ` Richard Henderson [this message]
2020-10-28 16:48 ` [PATCH 00/11] target/arm: Fix neon reg offsets Peter Maydell
2020-10-28 18:31   ` Peter Maydell
2020-10-28 19:32   ` Richard Henderson
2020-10-28 20:03 ` Peter Maydell
2020-10-29 11:04   ` Peter Maydell

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20201028032703.201526-12-richard.henderson@linaro.org \
    --to=richard.henderson@linaro.org \
    --cc=peter.maydell@linaro.org \
    --cc=qemu-devel@nongnu.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).