Re: [PATCH][AArch64] Implement ALU_BRANCH fusion

[Hurugalawadi, Naveen]

Hi James,

>> My reason for asking is that the instruction fusion implemented in LLVM
>> ( lib/Target/AArch64/AArch64MacroFusion.cpp::shouldScheduleAdjacent )

Sorry. There seems to be some confusion in the branch instructions.
The branch should be conditional for ALU_BRANCH fusion.

Please find attached the modified patch that fuses ALU instructions and
conditional branches.

Bootstrapped and Regression tested on aarch64-thunder-linux.
Please review the patch and let us know if its okay?

Thanks,
Naveen
    
diff --git a/gcc/config/aarch64/aarch64-fusion-pairs.def b/gcc/config/aarch64/aarch64-fusion-pairs.def
index f0e6dbc..300cd00 100644
--- a/gcc/config/aarch64/aarch64-fusion-pairs.def
+++ b/gcc/config/aarch64/aarch64-fusion-pairs.def
@@ -34,5 +34,6 @@ AARCH64_FUSION_PAIR ("movk+movk", MOVK_MOVK)
 AARCH64_FUSION_PAIR ("adrp+ldr", ADRP_LDR)
 AARCH64_FUSION_PAIR ("cmp+branch", CMP_BRANCH)
 AARCH64_FUSION_PAIR ("aes+aesmc", AES_AESMC)
+AARCH64_FUSION_PAIR ("alu+branch", ALU_BRANCH)
 
 #undef AARCH64_FUSION_PAIR
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index a069427..f76a2ff 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -792,7 +792,8 @@ static const struct tune_params thunderx2t99_tunings =
   &generic_approx_modes,
   4, /* memmov_cost.  */
   4, /* issue_rate.  */
-  (AARCH64_FUSE_CMP_BRANCH | AARCH64_FUSE_AES_AESMC), /* fusible_ops  */
+  (AARCH64_FUSE_CMP_BRANCH | AARCH64_FUSE_AES_AESMC
+   | AARCH64_FUSE_ALU_BRANCH), /* fusible_ops  */
   16,	/* function_align.  */
   8,	/* jump_align.  */
   16,	/* loop_align.  */
@@ -13981,6 +13982,37 @@ aarch_macro_fusion_pair_p (rtx_insn *prev, rtx_insn *curr)
         return true;
     }
 
+  if (aarch64_fusion_enabled_p (AARCH64_FUSE_ALU_BRANCH)
+      && any_condjump_p (curr))
+    {
+      /* These types correspond to the reservation "vulcan_alu_basic" for
+	 Broadcom Vulcan: these are ALU operations that produce a single uop
+	 during instruction decoding.  */
+      switch (get_attr_type (prev))
+	{
+	case TYPE_ALU_IMM:
+	case TYPE_ALU_SREG:
+	case TYPE_ADC_REG:
+	case TYPE_ADC_IMM:
+	case TYPE_ADCS_REG:
+	case TYPE_ADCS_IMM:
+	case TYPE_LOGIC_REG:
+	case TYPE_LOGIC_IMM:
+	case TYPE_CSEL:
+	case TYPE_ADR:
+	case TYPE_MOV_IMM:
+	case TYPE_SHIFT_REG:
+	case TYPE_SHIFT_IMM:
+	case TYPE_BFM:
+	case TYPE_RBIT:
+	case TYPE_REV:
+	case TYPE_EXTEND:
+	  return true;
+
+	default:;
+	}
+    }
+
   return false;
 }