From: Kyrylo Tkachov <kyrylo.tkachov@arm.com>
Date: Mon, 8 Mar 2021 09:35:14 +0000 (+0000)
Subject: aarch64: Add internal tune flag to minimise VL-based scalar ops
X-Git-Tag: releases/gcc-10.3.0~241
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=e1da328ec157ded50659d07f044767abea79cda2;p=thirdparty%2Fgcc.git

aarch64: Add internal tune flag to minimise VL-based scalar ops

This is a backport of the cse_sve_vl_constants tuning param to GCC 10.

Bootstrapped and tested on the branch on aarch64-none-linux-gnu.

gcc/ChangeLog:

	* config/aarch64/aarch64-tuning-flags.def (cse_sve_vl_constants):
	Define.
	* config/aarch64/aarch64.md (add<mode>3): Force CONST_POLY_INT immediates
	into a register when the above is enabled.
	* config/aarch64/aarch64.c (neoversev1_tunings):
	AARCH64_EXTRA_TUNE_CSE_SVE_VL_CONSTANTS.
	(aarch64_rtx_costs): Use AARCH64_EXTRA_TUNE_CSE_SVE_VL_CONSTANTS.

gcc/testsuite/

	* gcc.target/aarch64/sve/cse_sve_vl_constants_1.c: New test.
---

diff --git a/gcc/config/aarch64/aarch64-tuning-flags.def b/gcc/config/aarch64/aarch64-tuning-flags.def
index 52c7f4763f58..7677ec0b0b40 100644
--- a/gcc/config/aarch64/aarch64-tuning-flags.def
+++ b/gcc/config/aarch64/aarch64-tuning-flags.def
@@ -49,4 +49,6 @@ AARCH64_EXTRA_TUNING_OPTION ("rename_load_regs", RENAME_LOAD_REGS)
 /* Prefer Advanced SIMD over SVE for auto-vectorization.  */
 AARCH64_EXTRA_TUNING_OPTION ("prefer_advsimd_autovec", PREFER_ADVSIMD_AUTOVEC)
 
+AARCH64_EXTRA_TUNING_OPTION ("cse_sve_vl_constants", CSE_SVE_VL_CONSTANTS)
+
 #undef AARCH64_EXTRA_TUNING_OPTION
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index 46fe68355068..775f79d88ff8 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -1403,7 +1403,8 @@ static const struct tune_params neoversev1_tunings =
   2,	/* min_div_recip_mul_df.  */
   0,	/* max_case_values.  */
   tune_params::AUTOPREFETCHER_WEAK,	/* autoprefetcher_model.  */
-  (AARCH64_EXTRA_TUNE_PREFER_ADVSIMD_AUTOVEC),	/* tune_flags.  */
+  (AARCH64_EXTRA_TUNE_PREFER_ADVSIMD_AUTOVEC
+   | AARCH64_EXTRA_TUNE_CSE_SVE_VL_CONSTANTS),	/* tune_flags.  */
   &generic_prefetch_tune
 };
 
@@ -12532,8 +12533,18 @@ cost_plus:
 	    *cost += rtx_cost (op0, mode, PLUS, 0, speed);
 
 	    if (speed)
-	      /* ADD (immediate).  */
-	      *cost += extra_cost->alu.arith;
+	      {
+		/* ADD (immediate).  */
+		*cost += extra_cost->alu.arith;
+
+		/* Some tunings prefer to not use the VL-based scalar ops.
+		   Increase the cost of the poly immediate to prevent their
+		   formation.  */
+		if (GET_CODE (op1) == CONST_POLY_INT
+		    && (aarch64_tune_params.extra_tuning_flags
+			& AARCH64_EXTRA_TUNE_CSE_SVE_VL_CONSTANTS))
+		  *cost += COSTS_N_INSNS (1);
+	      }
 	    return true;
 	  }
 
diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md
index 7f262e067ce9..8f6bbcde9042 100644
--- a/gcc/config/aarch64/aarch64.md
+++ b/gcc/config/aarch64/aarch64.md
@@ -1898,6 +1898,14 @@
       && (!REG_P (op1)
 	 || !REGNO_PTR_FRAME_P (REGNO (op1))))
     operands[2] = force_reg (<MODE>mode, operands[2]);
+  /* Some tunings prefer to avoid VL-based operations.
+     Split off the poly immediate here.  The rtx costs hook will reject attempts
+     to combine them back.  */
+  else if (GET_CODE (operands[2]) == CONST_POLY_INT
+	   && can_create_pseudo_p ()
+	   && (aarch64_tune_params.extra_tuning_flags
+	       & AARCH64_EXTRA_TUNE_CSE_SVE_VL_CONSTANTS))
+    operands[2] = force_reg (<MODE>mode, operands[2]);
   /* Expand polynomial additions now if the destination is the stack
      pointer, since we don't want to use that as a temporary.  */
   else if (operands[0] == stack_pointer_rtx
diff --git a/gcc/testsuite/gcc.target/aarch64/sve/cse_sve_vl_constants_1.c b/gcc/testsuite/gcc.target/aarch64/sve/cse_sve_vl_constants_1.c
new file mode 100644
index 000000000000..dd04b667ee70
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sve/cse_sve_vl_constants_1.c
@@ -0,0 +1,12 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -moverride=tune=cse_sve_vl_constants" } */
+
+void __attribute__((noinline, noclone))
+vadd (int *dst, int *op1, int *op2, int count)
+{
+  for (int i = 0; i < count; ++i)
+    dst[i] = op1[i] + op2[i];
+}
+
+/* { dg-final { scan-assembler-not {\tincw\tx[0-9]+} } } */
+