diff mbox series

[09/21] Hexagon (target/hexagon) Don't overlap dest writes with source reads

Message ID 20230426004217.1319317-1-tsimpson@quicinc.com (mailing list archive)
State New, archived
Headers show
Series Hexagon (target/hexagon) short-circuit and move to DisasContext | expand

Commit Message

Taylor Simpson April 26, 2023, 12:42 a.m. UTC 
When generating TCG, make sure we have read all the operand registers
before writing to the destination registers.

This is a prerequesite for short-circuiting where the source and dest
operands could be the same.

Signed-off-by: Taylor Simpson <tsimpson@quicinc.com>
---
 target/hexagon/genptr.c | 45 ++++++++++++++++++++++++++---------------
 1 file changed, 29 insertions(+), 16 deletions(-)

Comments

Richard Henderson April 27, 2023, 8:03 a.m. UTC | #1 
On 4/26/23 01:42, Taylor Simpson wrote:
> When generating TCG, make sure we have read all the operand registers
> before writing to the destination registers.
> 
> This is a prerequesite for short-circuiting where the source and dest
> operands could be the same.
> 
> Signed-off-by: Taylor Simpson<tsimpson@quicinc.com>
> ---
>   target/hexagon/genptr.c | 45 ++++++++++++++++++++++++++---------------
>   1 file changed, 29 insertions(+), 16 deletions(-)

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

r~
diff mbox series

Patch

diff --git a/target/hexagon/genptr.c b/target/hexagon/genptr.c
index b9c4a46e3a..e4c116edc0 100644
--- a/target/hexagon/genptr.c
+++ b/target/hexagon/genptr.c
@@ -968,6 +968,7 @@  static void gen_cmpi_jumpnv(DisasContext *ctx,
 /* Shift left with saturation */
 static void gen_shl_sat(DisasContext *ctx, TCGv dst, TCGv src, TCGv shift_amt)
 {
+    TCGv tmp = tcg_temp_new();    /* In case dst == src */
     TCGv usr = get_result_gpr(ctx, HEX_REG_USR);
     TCGv sh32 = tcg_temp_new();
     TCGv dst_sar = tcg_temp_new();
@@ -992,17 +993,17 @@  static void gen_shl_sat(DisasContext *ctx, TCGv dst, TCGv src, TCGv shift_amt)
      */
 
     tcg_gen_andi_tl(sh32, shift_amt, 31);
-    tcg_gen_movcond_tl(TCG_COND_EQ, dst, sh32, shift_amt,
+    tcg_gen_movcond_tl(TCG_COND_EQ, tmp, sh32, shift_amt,
                        src, tcg_constant_tl(0));
-    tcg_gen_shl_tl(dst, dst, sh32);
-    tcg_gen_sar_tl(dst_sar, dst, sh32);
+    tcg_gen_shl_tl(tmp, tmp, sh32);
+    tcg_gen_sar_tl(dst_sar, tmp, sh32);
     tcg_gen_movcond_tl(TCG_COND_LT, satval, src, tcg_constant_tl(0), min, max);
 
     tcg_gen_setcond_tl(TCG_COND_NE, ovf, dst_sar, src);
     tcg_gen_shli_tl(ovf, ovf, reg_field_info[USR_OVF].offset);
     tcg_gen_or_tl(usr, usr, ovf);
 
-    tcg_gen_movcond_tl(TCG_COND_EQ, dst, dst_sar, src, dst, satval);
+    tcg_gen_movcond_tl(TCG_COND_EQ, dst, dst_sar, src, tmp, satval);
 }
 
 static void gen_sar(TCGv dst, TCGv src, TCGv shift_amt)
@@ -1225,22 +1226,28 @@  void gen_sat_i32(TCGv dest, TCGv source, int width)
 
 void gen_sat_i32_ovfl(TCGv ovfl, TCGv dest, TCGv source, int width)
 {
-    gen_sat_i32(dest, source, width);
-    tcg_gen_setcond_tl(TCG_COND_NE, ovfl, source, dest);
+    TCGv tmp = tcg_temp_new();    /* In case dest == source */
+    gen_sat_i32(tmp, source, width);
+    tcg_gen_setcond_tl(TCG_COND_NE, ovfl, source, tmp);
+    tcg_gen_mov_tl(dest, tmp);
 }
 
 void gen_satu_i32(TCGv dest, TCGv source, int width)
 {
+    TCGv tmp = tcg_temp_new();    /* In case dest == source */
     TCGv max_val = tcg_constant_tl((1 << width) - 1);
     TCGv zero = tcg_constant_tl(0);
-    tcg_gen_movcond_tl(TCG_COND_GTU, dest, source, max_val, max_val, source);
-    tcg_gen_movcond_tl(TCG_COND_LT, dest, source, zero, zero, dest);
+    tcg_gen_movcond_tl(TCG_COND_GTU, tmp, source, max_val, max_val, source);
+    tcg_gen_movcond_tl(TCG_COND_LT, tmp, source, zero, zero, tmp);
+    tcg_gen_mov_tl(dest, tmp);
 }
 
 void gen_satu_i32_ovfl(TCGv ovfl, TCGv dest, TCGv source, int width)
 {
-    gen_satu_i32(dest, source, width);
-    tcg_gen_setcond_tl(TCG_COND_NE, ovfl, source, dest);
+    TCGv tmp = tcg_temp_new();    /* In case dest == source */
+    gen_satu_i32(tmp, source, width);
+    tcg_gen_setcond_tl(TCG_COND_NE, ovfl, source, tmp);
+    tcg_gen_mov_tl(dest, tmp);
 }
 
 void gen_sat_i64(TCGv_i64 dest, TCGv_i64 source, int width)
@@ -1253,27 +1260,33 @@  void gen_sat_i64(TCGv_i64 dest, TCGv_i64 source, int width)
 
 void gen_sat_i64_ovfl(TCGv ovfl, TCGv_i64 dest, TCGv_i64 source, int width)
 {
+    TCGv_i64 tmp = tcg_temp_new_i64(); /* In case dest == source */
     TCGv_i64 ovfl_64;
-    gen_sat_i64(dest, source, width);
+    gen_sat_i64(tmp, source, width);
     ovfl_64 = tcg_temp_new_i64();
-    tcg_gen_setcond_i64(TCG_COND_NE, ovfl_64, dest, source);
+    tcg_gen_setcond_i64(TCG_COND_NE, ovfl_64, tmp, source);
+    tcg_gen_mov_i64(dest, tmp);
     tcg_gen_trunc_i64_tl(ovfl, ovfl_64);
 }
 
 void gen_satu_i64(TCGv_i64 dest, TCGv_i64 source, int width)
 {
+    TCGv_i64 tmp = tcg_temp_new_i64();    /* In case dest == source */
     TCGv_i64 max_val = tcg_constant_i64((1LL << width) - 1LL);
     TCGv_i64 zero = tcg_constant_i64(0);
-    tcg_gen_movcond_i64(TCG_COND_GTU, dest, source, max_val, max_val, source);
-    tcg_gen_movcond_i64(TCG_COND_LT, dest, source, zero, zero, dest);
+    tcg_gen_movcond_i64(TCG_COND_GTU, tmp, source, max_val, max_val, source);
+    tcg_gen_movcond_i64(TCG_COND_LT, tmp, source, zero, zero, tmp);
+    tcg_gen_mov_i64(dest, tmp);
 }
 
 void gen_satu_i64_ovfl(TCGv ovfl, TCGv_i64 dest, TCGv_i64 source, int width)
 {
+    TCGv_i64 tmp = tcg_temp_new_i64();    /* In case dest == source */
     TCGv_i64 ovfl_64;
-    gen_satu_i64(dest, source, width);
+    gen_satu_i64(tmp, source, width);
     ovfl_64 = tcg_temp_new_i64();
-    tcg_gen_setcond_i64(TCG_COND_NE, ovfl_64, dest, source);
+    tcg_gen_setcond_i64(TCG_COND_NE, ovfl_64, tmp, source);
+    tcg_gen_mov_i64(dest, tmp);
     tcg_gen_trunc_i64_tl(ovfl, ovfl_64);
 }