QEMU-Devel Archive on lore.kernel.org
 help / color / Atom feed
From: Richard Henderson <richard.henderson@linaro.org>
To: qemu-devel@nongnu.org
Cc: peter.maydell@linaro.org, "Alex Bennée" <alex.bennee@linaro.org>
Subject: [PULL 08/10] tcg: Improve vector tail clearing
Date: Wed,  6 May 2020 11:30:00 -0700
Message-ID: <20200506183002.3192-9-richard.henderson@linaro.org> (raw)
In-Reply-To: <20200506183002.3192-1-richard.henderson@linaro.org>

Better handling of non-power-of-2 tails as seen with Arm 8-byte
vector operations.

Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/tcg-op-gvec.c | 82 ++++++++++++++++++++++++++++++++++++-----------
 1 file changed, 63 insertions(+), 19 deletions(-)

diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c
index 5a6cc19812..43cac1a0bf 100644
--- a/tcg/tcg-op-gvec.c
+++ b/tcg/tcg-op-gvec.c
@@ -326,11 +326,34 @@ void tcg_gen_gvec_5_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
    in units of LNSZ.  This limits the expansion of inline code.  */
 static inline bool check_size_impl(uint32_t oprsz, uint32_t lnsz)
 {
-    if (oprsz % lnsz == 0) {
-        uint32_t lnct = oprsz / lnsz;
-        return lnct >= 1 && lnct <= MAX_UNROLL;
+    uint32_t q, r;
+
+    if (oprsz < lnsz) {
+        return false;
     }
-    return false;
+
+    q = oprsz / lnsz;
+    r = oprsz % lnsz;
+    tcg_debug_assert((r & 7) == 0);
+
+    if (lnsz < 16) {
+        /* For sizes below 16, accept no remainder. */
+        if (r != 0) {
+            return false;
+        }
+    } else {
+        /*
+         * Recall that ARM SVE allows vector sizes that are not a
+         * power of 2, but always a multiple of 16.  The intent is
+         * that e.g. size == 80 would be expanded with 2x32 + 1x16.
+         * In addition, expand_clr needs to handle a multiple of 8.
+         * Thus we can handle the tail with one more operation per
+         * diminishing power of 2.
+         */
+        q += ctpop32(r);
+    }
+
+    return q <= MAX_UNROLL;
 }
 
 static void expand_clr(uint32_t dofs, uint32_t maxsz);
@@ -402,22 +425,31 @@ static void gen_dup_i64(unsigned vece, TCGv_i64 out, TCGv_i64 in)
 static TCGType choose_vector_type(const TCGOpcode *list, unsigned vece,
                                   uint32_t size, bool prefer_i64)
 {
-    if (TCG_TARGET_HAS_v256 && check_size_impl(size, 32)) {
-        /*
-         * Recall that ARM SVE allows vector sizes that are not a
-         * power of 2, but always a multiple of 16.  The intent is
-         * that e.g. size == 80 would be expanded with 2x32 + 1x16.
-         * It is hard to imagine a case in which v256 is supported
-         * but v128 is not, but check anyway.
-         */
-        if (tcg_can_emit_vecop_list(list, TCG_TYPE_V256, vece)
-            && (size % 32 == 0
-                || tcg_can_emit_vecop_list(list, TCG_TYPE_V128, vece))) {
-            return TCG_TYPE_V256;
-        }
+    /*
+     * Recall that ARM SVE allows vector sizes that are not a
+     * power of 2, but always a multiple of 16.  The intent is
+     * that e.g. size == 80 would be expanded with 2x32 + 1x16.
+     * It is hard to imagine a case in which v256 is supported
+     * but v128 is not, but check anyway.
+     * In addition, expand_clr needs to handle a multiple of 8.
+     */
+    if (TCG_TARGET_HAS_v256 &&
+        check_size_impl(size, 32) &&
+        tcg_can_emit_vecop_list(list, TCG_TYPE_V256, vece) &&
+        (!(size & 16) ||
+         (TCG_TARGET_HAS_v128 &&
+          tcg_can_emit_vecop_list(list, TCG_TYPE_V128, vece))) &&
+        (!(size & 8) ||
+         (TCG_TARGET_HAS_v64 &&
+          tcg_can_emit_vecop_list(list, TCG_TYPE_V64, vece)))) {
+        return TCG_TYPE_V256;
     }
-    if (TCG_TARGET_HAS_v128 && check_size_impl(size, 16)
-        && tcg_can_emit_vecop_list(list, TCG_TYPE_V128, vece)) {
+    if (TCG_TARGET_HAS_v128 &&
+        check_size_impl(size, 16) &&
+        tcg_can_emit_vecop_list(list, TCG_TYPE_V128, vece) &&
+        (!(size & 8) ||
+         (TCG_TARGET_HAS_v64 &&
+          tcg_can_emit_vecop_list(list, TCG_TYPE_V64, vece)))) {
         return TCG_TYPE_V128;
     }
     if (TCG_TARGET_HAS_v64 && !prefer_i64 && check_size_impl(size, 8)
@@ -432,6 +464,18 @@ static void do_dup_store(TCGType type, uint32_t dofs, uint32_t oprsz,
 {
     uint32_t i = 0;
 
+    tcg_debug_assert(oprsz >= 8);
+
+    /*
+     * This may be expand_clr for the tail of an operation, e.g.
+     * oprsz == 8 && maxsz == 64.  The first 8 bytes of this store
+     * are misaligned wrt the maximum vector size, so do that first.
+     */
+    if (dofs & 8) {
+        tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V64);
+        i += 8;
+    }
+
     switch (type) {
     case TCG_TYPE_V256:
         /*
-- 
2.20.1



  parent reply index

Thread overview: 14+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-05-06 18:29 [PULL 00/10] tcg patch queue Richard Henderson
2020-05-06 18:29 ` [PULL 01/10] tcg: Add tcg_gen_gvec_dup_imm Richard Henderson
2020-05-06 18:29 ` [PULL 02/10] target/s390x: Use tcg_gen_gvec_dup_imm Richard Henderson
2020-05-06 18:29 ` [PULL 03/10] target/ppc: " Richard Henderson
2020-05-06 18:29 ` [PULL 04/10] target/arm: " Richard Henderson
2020-05-07 14:39   ` Laurent Desnogues
2020-05-07 17:16     ` Richard Henderson
2020-05-06 18:29 ` [PULL 05/10] tcg: Use tcg_gen_gvec_dup_imm in logical simplifications Richard Henderson
2020-05-06 18:29 ` [PULL 06/10] tcg: Remove tcg_gen_gvec_dup{8,16,32,64}i Richard Henderson
2020-05-06 18:29 ` [PULL 07/10] tcg: Add tcg_gen_gvec_dup_tl Richard Henderson
2020-05-06 18:30 ` Richard Henderson [this message]
2020-05-06 18:30 ` [PULL 09/10] tcg: Add load_dest parameter to GVecGen2 Richard Henderson
2020-05-06 18:30 ` [PULL 10/10] tcg: Fix integral argument type to tcg_gen_rot[rl]i_i{32, 64} Richard Henderson
2020-05-07  9:54 ` [PULL 00/10] tcg patch queue 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=20200506183002.3192-9-richard.henderson@linaro.org \
    --to=richard.henderson@linaro.org \
    --cc=alex.bennee@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

QEMU-Devel Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/qemu-devel/0 qemu-devel/git/0.git
	git clone --mirror https://lore.kernel.org/qemu-devel/1 qemu-devel/git/1.git
	git clone --mirror https://lore.kernel.org/qemu-devel/2 qemu-devel/git/2.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 qemu-devel qemu-devel/ https://lore.kernel.org/qemu-devel \
		qemu-devel@nongnu.org
	public-inbox-index qemu-devel

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.nongnu.qemu-devel


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git