[13/21] Hexagon (target/hexagon) Short-circuit packet HVX writes

Commit Message

Taylor Simpson April 26, 2023, 12:42 a.m. UTC

In certain cases, we can avoid the overhead of writing to future_VRegs
and write directly to VRegs.  We consider HVX reads/writes when computing
ctx->need_commit.  Then, we can early-exit from gen_commit_hvx.

Signed-off-by: Taylor Simpson <[e-mail]>
---
 target/hexagon/genptr.c    |  6 ++++-
 target/hexagon/translate.c | 46 +++++++++++++++++++++++++++++++++++++-
 2 files changed, 50 insertions(+), 2 deletions(-)

Comments

Richard Henderson April 27, 2023, 10:42 a.m. UTC | #1

On 4/26/23 01:42, Taylor Simpson wrote:
> In certain cases, we can avoid the overhead of writing to future_VRegs
> and write directly to VRegs.  We consider HVX reads/writes when computing
> ctx->need_commit.  Then, we can early-exit from gen_commit_hvx.
> 
> Signed-off-by: Taylor Simpson<[e-mail]>
> ---
>   target/hexagon/genptr.c    |  6 ++++-
>   target/hexagon/translate.c | 46 +++++++++++++++++++++++++++++++++++++-
>   2 files changed, 50 insertions(+), 2 deletions(-)

Reviewed-by: Richard Henderson <[e-mail]>

r~

Patch

diff --git a/target/hexagon/genptr.c b/target/hexagon/genptr.c
index da68d19ed3..8e5afab931 100644
--- a/target/hexagon/genptr.c
+++ b/target/hexagon/genptr.c
@@ -1101,7 +1101,11 @@  static void gen_log_vreg_write_pair(DisasContext *ctx, intptr_t srcoff, int num,
 
 static intptr_t get_result_qreg(DisasContext *ctx, int qnum)
 {
-    return  offsetof(CPUHexagonState, future_QRegs[qnum]);
+    if (ctx->need_commit) {
+        return  offsetof(CPUHexagonState, future_QRegs[qnum]);
+    } else {
+        return  offsetof(CPUHexagonState, QRegs[qnum]);
+    }
 }
 
 static void gen_vreg_load(DisasContext *ctx, intptr_t dstoff, TCGv src,
diff --git a/target/hexagon/translate.c b/target/hexagon/translate.c
index 07ed36f6a8..8e024b2cd2 100644
--- a/target/hexagon/translate.c
+++ b/target/hexagon/translate.c
@@ -70,6 +70,10 @@  intptr_t ctx_future_vreg_off(DisasContext *ctx, int regnum,
 {
     intptr_t offset;
 
+    if (!ctx->need_commit) {
+        return offsetof(CPUHexagonState, VRegs[regnum]);
+    }
+
     /* See if it is already allocated */
     for (int i = 0; i < ctx->future_vregs_idx; i++) {
         if (ctx->future_vregs_num[i] == regnum) {
@@ -374,7 +378,7 @@  static bool need_commit(DisasContext *ctx)
         return true;
     }
 
-    if (pkt->num_insns == 1) {
+    if (pkt->num_insns == 1 && !pkt->pkt_has_hvx) {
         return false;
     }
 
@@ -394,6 +398,40 @@  static bool need_commit(DisasContext *ctx)
         }
     }
 
+    /* Check for overlap between HVX reads and writes */
+    for (int i = 0; i < ctx->vreg_log_idx; i++) {
+        int vnum = ctx->vreg_log[i];
+        if (test_bit(vnum, ctx->vregs_read)) {
+            return true;
+        }
+    }
+    if (!bitmap_empty(ctx->vregs_updated_tmp, NUM_VREGS)) {
+        int i = find_first_bit(ctx->vregs_updated_tmp, NUM_VREGS);
+        while (i < NUM_VREGS) {
+            if (test_bit(i, ctx->vregs_read)) {
+                return true;
+            }
+            i = find_next_bit(ctx->vregs_updated_tmp, NUM_VREGS, i + 1);
+        }
+    }
+    if (!bitmap_empty(ctx->vregs_select, NUM_VREGS)) {
+        int i = find_first_bit(ctx->vregs_select, NUM_VREGS);
+        while (i < NUM_VREGS) {
+            if (test_bit(i, ctx->vregs_read)) {
+                return true;
+            }
+            i = find_next_bit(ctx->vregs_select, NUM_VREGS, i + 1);
+        }
+    }
+
+    /* Check for overlap between HVX predicate reads and writes */
+    for (int i = 0; i < ctx->qreg_log_idx; i++) {
+        int qnum = ctx->qreg_log[i];
+        if (test_bit(qnum, ctx->qregs_read)) {
+            return true;
+        }
+    }
+
     return false;
 }
 
@@ -787,6 +825,12 @@  static void gen_commit_hvx(DisasContext *ctx)
 {
     int i;
 
+    /* Early exit if not needed */
+    if (!ctx->need_commit) {
+        g_assert(!pkt_has_hvx_store(ctx->pkt));
+        return;
+    }
+
     /*
      *    for (i = 0; i < ctx->vreg_log_idx; i++) {
      *        int rnum = ctx->vreg_log[i];