[PATCH v2 2/3] Hexagon (target/hexagon) Enable more short-circuit packets (scalar core)

[Taylor Simpson <ltaylorsimpson@gmail.com>]

Look for read-after-write instead of overlap of reads and writes

Here is an example with overalp but no read-after-write:
0x000200fc:  0x38103876	{	R0 = add(R0,R1); R6 = add(R6,R7) }

BEFORE:
 ---- 00000000000200fc
 mov_i32 loc2,$0x0
 mov_i32 loc2,r0
 add_i32 loc3,loc2,r1
 mov_i32 loc2,loc3
 mov_i32 loc4,$0x0
 mov_i32 loc4,r6
 add_i32 loc5,loc4,r7
 mov_i32 loc4,loc5
 mov_i32 r0,loc2
 mov_i32 r6,loc4

AFTER:
 ---- 00000000000200fc
 add_i32 loc2,r0,r1
 mov_i32 r0,loc2
 add_i32 loc3,r6,r7
 mov_i32 r6,loc3

We can also short-circuit packets with .new values by reading from the
real destination instead of the temporary.
0x00020100:  0x78005ff3	{	R19 = #0xff
0x00020104:  0x2002e204		if (cmp.eq(N19.new,R2)) jump:t PC+8 }

BEFORE:
 ---- 0000000000020100
 mov_i32 pc,$0x20108
 mov_i32 loc8,$0x0
 mov_i32 loc8,$0xff
 setcond_i32 loc10,loc8,r2,eq
 mov_i32 loc6,loc10
 mov_i32 r19,loc8
 add_i32 pkt_cnt,pkt_cnt,$0x2
 add_i32 insn_cnt,insn_cnt,$0x4
 brcond_i32 loc6,$0x0,eq,$L1
 goto_tb $0x0
 mov_i32 pc,$0x20108
 exit_tb $0x7fbb54000040
 set_label $L1
 goto_tb $0x1
 exit_tb $0x7fbb54000041
 set_label $L0
 exit_tb $0x7fbb54000043

AFTER:
 ---- 0000000000020100
 mov_i32 pc,$0x20108
 mov_i32 r19,$0xff
 setcond_i32 loc7,r19,r2,eq
 mov_i32 loc4,loc7
 add_i32 pkt_cnt,pkt_cnt,$0x2
 add_i32 insn_cnt,insn_cnt,$0x4
 brcond_i32 loc4,$0x0,eq,$L1
 goto_tb $0x0
 mov_i32 pc,$0x20108
 exit_tb $0x7f9764000040
 set_label $L1
 goto_tb $0x1
 exit_tb $0x7f9764000041
 set_label $L0
 exit_tb $0x7f9764000043

Signed-off-by: Taylor Simpson <ltaylorsimpson@gmail.com>
---
 target/hexagon/translate.h | 13 +++++++------
 target/hexagon/translate.c | 21 ++++-----------------
 2 files changed, 11 insertions(+), 23 deletions(-)

diff --git a/target/hexagon/translate.h b/target/hexagon/translate.h
index f06d71fc53..d5e7f49ad8 100644
--- a/target/hexagon/translate.h
+++ b/target/hexagon/translate.h
@@ -38,12 +38,10 @@ typedef struct DisasContext {
     int reg_log[REG_WRITES_MAX];
     int reg_log_idx;
     DECLARE_BITMAP(regs_written, TOTAL_PER_THREAD_REGS);
-    DECLARE_BITMAP(regs_read, TOTAL_PER_THREAD_REGS);
     DECLARE_BITMAP(predicated_regs, TOTAL_PER_THREAD_REGS);
     int preg_log[PRED_WRITES_MAX];
     int preg_log_idx;
     DECLARE_BITMAP(pregs_written, NUM_PREGS);
-    DECLARE_BITMAP(pregs_read, NUM_PREGS);
     uint8_t store_width[STORES_MAX];
     bool s1_store_processed;
     int future_vregs_idx;
@@ -68,6 +66,7 @@ typedef struct DisasContext {
     bool is_tight_loop;
     bool short_circuit;
     bool has_hvx_helper;
+    bool read_after_write;
     TCGv new_value[TOTAL_PER_THREAD_REGS];
     TCGv new_pred_value[NUM_PREGS];
     TCGv pred_written;
@@ -88,13 +87,14 @@ static inline void ctx_log_pred_write(DisasContext *ctx, int pnum)
 
 static inline void ctx_log_pred_read(DisasContext *ctx, int pnum)
 {
-    set_bit(pnum, ctx->pregs_read);
+    if (test_bit(pnum, ctx->pregs_written)) {
+        ctx->read_after_write = true;
+    }
 }
 
 static inline void ctx_log_pred_read_new(DisasContext *ctx, int pnum)
 {
     g_assert(test_bit(pnum, ctx->pregs_written));
-    set_bit(pnum, ctx->pregs_read);
 }
 
 static inline void ctx_log_reg_write(DisasContext *ctx, int rnum,
@@ -125,13 +125,14 @@ static inline void ctx_log_reg_write_pair(DisasContext *ctx, int rnum,
 
 static inline void ctx_log_reg_read(DisasContext *ctx, int rnum)
 {
-    set_bit(rnum, ctx->regs_read);
+    if (test_bit(rnum, ctx->regs_written)) {
+        ctx->read_after_write = true;
+    }
 }
 
 static inline void ctx_log_reg_read_new(DisasContext *ctx, int rnum)
 {
     g_assert(test_bit(rnum, ctx->regs_written));
-    set_bit(rnum, ctx->regs_read);
 }
 
 static inline void ctx_log_reg_read_pair(DisasContext *ctx, int rnum)
diff --git a/target/hexagon/translate.c b/target/hexagon/translate.c
index 95579ae243..751ca71790 100644
--- a/target/hexagon/translate.c
+++ b/target/hexagon/translate.c
@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2024 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -394,20 +394,8 @@ static bool need_commit(DisasContext *ctx)
         }
     }
 
-    /* Check for overlap between register reads and writes */
-    for (int i = 0; i < ctx->reg_log_idx; i++) {
-        int rnum = ctx->reg_log[i];
-        if (test_bit(rnum, ctx->regs_read)) {
-            return true;
-        }
-    }
-
-    /* Check for overlap between predicate reads and writes */
-    for (int i = 0; i < ctx->preg_log_idx; i++) {
-        int pnum = ctx->preg_log[i];
-        if (test_bit(pnum, ctx->pregs_read)) {
-            return true;
-        }
+    if (ctx->read_after_write) {
+        return true;
     }
 
     /* Check for overlap between HVX reads and writes */
@@ -466,6 +454,7 @@ static void analyze_packet(DisasContext *ctx)
 {
     Packet *pkt = ctx->pkt;
     ctx->has_hvx_helper = false;
+    ctx->read_after_write = false;
     for (int i = 0; i < pkt->num_insns; i++) {
         Insn *insn = &pkt->insn[i];
         ctx->insn = insn;
@@ -490,11 +479,9 @@ static void gen_start_packet(DisasContext *ctx)
     ctx->next_PC = next_PC;
     ctx->reg_log_idx = 0;
     bitmap_zero(ctx->regs_written, TOTAL_PER_THREAD_REGS);
-    bitmap_zero(ctx->regs_read, TOTAL_PER_THREAD_REGS);
     bitmap_zero(ctx->predicated_regs, TOTAL_PER_THREAD_REGS);
     ctx->preg_log_idx = 0;
     bitmap_zero(ctx->pregs_written, NUM_PREGS);
-    bitmap_zero(ctx->pregs_read, NUM_PREGS);
     ctx->future_vregs_idx = 0;
     ctx->tmp_vregs_idx = 0;
     ctx->vreg_log_idx = 0;
-- 
2.34.1