public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH] i386: Roundeven expansion for SSE4.1+
@ 2019-07-24 10:45 Tejas Joshi
  2019-07-27  9:22 ` Uros Bizjak
  0 siblings, 1 reply; 6+ messages in thread
From: Tejas Joshi @ 2019-07-24 10:45 UTC (permalink / raw)
  To: gcc-patches; +Cc: Martin Jambor, hubicka, ubizjak

[-- Attachment #1: Type: text/plain, Size: 1632 bytes --]

Hi.
This is a patch that Uros suggested for roundeven expansion, here.
Thanks for the heads up.
<https://gcc.gnu.org/ml/gcc-patches/2019-07/msg01038.html>
I have rerun the testsuite on the patch, it survives the regression
tests and bootstraps on x86_64-linux-gnu. Note, patch to be applied on
top of
<https://gcc.gnu.org/ml/gcc-patches/2019-06/msg01828.html>

Thanks,
Tejas

gcc/ChangeLog:

2019-07-24  Tejas Joshi  <tejasjoshi9673@gmail.com>

    * builtins.c (mathfn_built_in_2): Change CASE_MATHFN to
    CASE_MATHFN_FLOATN for roundeven.
    * config/i386/i386.c (ix86_i387_mode_needed): Add case I387_ROUNDEVEN.
    (ix86_mode_needed): Likewise.
    (ix86_mode_after): Likewise.
    (ix86_mode_entry): Likewise.
    (ix86_mode_exit): Likewise.
    (ix86_emit_mode_set): Likewise.
    (emit_i387_cw_initialization): Add case I387_CW_ROUNDEVEN.
    * config/i386/i386.h (ix86_stack_slot) : Add SLOT_CW_ROUNDEVEN.
    (ix86_entry): Add I387_ROUNDEVEN.
    (avx_u128_state): Add I387_CW_ANY.
    * config/i386/i386.md: Define UNSPEC_FRNDINT_ROUNDEVEN.
    (define_int_iterator): Likewise.
    (define_int_attr): Likewise for rounding_insn, rounding and ROUNDING.
    (define_constant): Define ROUND_ROUNDEVEN mode.
    (define_attr): Add roundeven mode for i387_cw.
    (<rouding_insn><mode>2): Add condition for ROUND_ROUNDEVEN.
    * internal-fn.def (ROUNDEVEN): New builtin function.
    * optabs.def (roundeven_optab): New optab.

gcc/testsuite/ChangeLog:

2019-07-24  Tejas Joshi  <tejasjoshi9673@gmail.com>

    * gcc.target/i386/avx-vround-roundeven-1.c: New test.
    * gcc.target/i386/avx-vround-roundeven-2.c: New test.

[-- Attachment #2: roundeven-md1.diff --]
[-- Type: text/x-patch, Size: 8553 bytes --]

diff --git a/gcc/builtins.c b/gcc/builtins.c
index 8ceb077b0bf..f61f10422fd 100644
--- a/gcc/builtins.c
+++ b/gcc/builtins.c
@@ -2056,7 +2056,7 @@ mathfn_built_in_2 (tree type, combined_fn fn)
     CASE_MATHFN (REMQUO)
     CASE_MATHFN_FLOATN (RINT)
     CASE_MATHFN_FLOATN (ROUND)
-    CASE_MATHFN (ROUNDEVEN)
+    CASE_MATHFN_FLOATN (ROUNDEVEN)
     CASE_MATHFN (SCALB)
     CASE_MATHFN (SCALBLN)
     CASE_MATHFN (SCALBN)
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 49f49c5f8d0..aa32a61f416 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -13593,6 +13593,11 @@ ix86_i387_mode_needed (int entity, rtx_insn *insn)
 
   switch (entity)
     {
+    case I387_ROUNDEVEN:
+      if (mode == I387_CW_ROUNDEVEN)
+	return mode;
+      break;
+
     case I387_TRUNC:
       if (mode == I387_CW_TRUNC)
 	return mode;
@@ -13627,6 +13632,7 @@ ix86_mode_needed (int entity, rtx_insn *insn)
       return ix86_dirflag_mode_needed (insn);
     case AVX_U128:
       return ix86_avx_u128_mode_needed (insn);
+    case I387_ROUNDEVEN:
     case I387_TRUNC:
     case I387_FLOOR:
     case I387_CEIL:
@@ -13687,6 +13693,7 @@ ix86_mode_after (int entity, int mode, rtx_insn *insn)
       return mode;
     case AVX_U128:
       return ix86_avx_u128_mode_after (mode, insn);
+    case I387_ROUNDEVEN:
     case I387_TRUNC:
     case I387_FLOOR:
     case I387_CEIL:
@@ -13739,6 +13746,7 @@ ix86_mode_entry (int entity)
       return ix86_dirflag_mode_entry ();
     case AVX_U128:
       return ix86_avx_u128_mode_entry ();
+    case I387_ROUNDEVEN:
     case I387_TRUNC:
     case I387_FLOOR:
     case I387_CEIL:
@@ -13776,6 +13784,7 @@ ix86_mode_exit (int entity)
       return X86_DIRFLAG_ANY;
     case AVX_U128:
       return ix86_avx_u128_mode_exit ();
+    case I387_ROUNDEVEN:
     case I387_TRUNC:
     case I387_FLOOR:
     case I387_CEIL:
@@ -13810,6 +13819,12 @@ emit_i387_cw_initialization (int mode)
 
   switch (mode)
     {
+    case I387_CW_ROUNDEVEN:
+      /* round to nearest */
+      emit_insn (gen_andhi3 (reg, reg, GEN_INT (0x0c00)));
+      slot = SLOT_CW_ROUNDEVEN;
+      break;
+
     case I387_CW_TRUNC:
       /* round toward zero (truncate) */
       emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0c00)));
@@ -13856,6 +13871,7 @@ ix86_emit_mode_set (int entity, int mode, int prev_mode ATTRIBUTE_UNUSED,
       if (mode == AVX_U128_CLEAN)
 	emit_insn (gen_avx_vzeroupper ());
       break;
+    case I387_ROUNDEVEN:
     case I387_TRUNC:
     case I387_FLOOR:
     case I387_CEIL:
diff --git a/gcc/config/i386/i386.h b/gcc/config/i386/i386.h
index 1f70844fc45..a8e1cd048a8 100644
--- a/gcc/config/i386/i386.h
+++ b/gcc/config/i386/i386.h
@@ -2470,6 +2470,7 @@ enum ix86_stack_slot
 {
   SLOT_TEMP = 0,
   SLOT_CW_STORED,
+  SLOT_CW_ROUNDEVEN,
   SLOT_CW_TRUNC,
   SLOT_CW_FLOOR,
   SLOT_CW_CEIL,
@@ -2481,6 +2482,7 @@ enum ix86_entity
 {
   X86_DIRFLAG = 0,
   AVX_U128,
+  I387_ROUNDEVEN,
   I387_TRUNC,
   I387_FLOOR,
   I387_CEIL,
@@ -2516,7 +2518,7 @@ enum avx_u128_state
 
 #define NUM_MODES_FOR_MODE_SWITCHING			\
   { X86_DIRFLAG_ANY, AVX_U128_ANY,			\
-    I387_CW_ANY, I387_CW_ANY, I387_CW_ANY }
+    I387_CW_ANY, I387_CW_ANY, I387_CW_ANY, I387_CW_ANY  }
 
 \f
 /* Avoid renaming of stack registers, as doing so in combination with
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index db5fa9ae3ca..21532494f6d 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -141,6 +141,7 @@
   UNSPEC_FXAM
 
   ;; x87 Rounding
+  UNSPEC_FRNDINT_ROUNDEVEN
   UNSPEC_FRNDINT_FLOOR
   UNSPEC_FRNDINT_CEIL
   UNSPEC_FRNDINT_TRUNC
@@ -303,7 +304,8 @@
 
 ;; Constants to represent rounding modes in the ROUND instruction
 (define_constants
-  [(ROUND_FLOOR			0x1)
+  [(ROUND_ROUNDEVEN		0x0)
+   (ROUND_FLOOR			0x1)
    (ROUND_CEIL			0x2)
    (ROUND_TRUNC			0x3)
    (ROUND_MXCSR			0x4)
@@ -779,7 +781,7 @@
 
 ;; Defines rounding mode of an FP operation.
 
-(define_attr "i387_cw" "trunc,floor,ceil,uninitialized,any"
+(define_attr "i387_cw" "roundeven,floor,ceil,trunc,uninitialized,any"
   (const_string "any"))
 
 ;; Define attribute to indicate AVX insns with partial XMM register update.
@@ -16251,7 +16253,8 @@
 })
 
 (define_int_iterator FRNDINT_ROUNDING
-	[UNSPEC_FRNDINT_FLOOR
+	[UNSPEC_FRNDINT_ROUNDEVEN
+	 UNSPEC_FRNDINT_FLOOR
 	 UNSPEC_FRNDINT_CEIL
 	 UNSPEC_FRNDINT_TRUNC])
 
@@ -16261,21 +16264,24 @@
 
 ;; Base name for define_insn
 (define_int_attr rounding_insn
-	[(UNSPEC_FRNDINT_FLOOR "floor")
+	[(UNSPEC_FRNDINT_ROUNDEVEN "roundeven")
+	 (UNSPEC_FRNDINT_FLOOR "floor")
 	 (UNSPEC_FRNDINT_CEIL "ceil")
 	 (UNSPEC_FRNDINT_TRUNC "btrunc")
 	 (UNSPEC_FIST_FLOOR "floor")
 	 (UNSPEC_FIST_CEIL "ceil")])
 
 (define_int_attr rounding
-	[(UNSPEC_FRNDINT_FLOOR "floor")
+	[(UNSPEC_FRNDINT_ROUNDEVEN "roundeven")
+	 (UNSPEC_FRNDINT_FLOOR "floor")
 	 (UNSPEC_FRNDINT_CEIL "ceil")
 	 (UNSPEC_FRNDINT_TRUNC "trunc")
 	 (UNSPEC_FIST_FLOOR "floor")
 	 (UNSPEC_FIST_CEIL "ceil")])
 
 (define_int_attr ROUNDING
-	[(UNSPEC_FRNDINT_FLOOR "FLOOR")
+	[(UNSPEC_FRNDINT_ROUNDEVEN "ROUNDEVEN")
+	 (UNSPEC_FRNDINT_FLOOR "FLOOR")
 	 (UNSPEC_FRNDINT_CEIL "CEIL")
 	 (UNSPEC_FRNDINT_TRUNC "TRUNC")
 	 (UNSPEC_FIST_FLOOR "FLOOR")
@@ -16338,8 +16344,9 @@
 	|| TARGET_MIX_SSE_I387)
     && (flag_fp_int_builtin_inexact || !flag_trapping_math))
    || (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
-       && (TARGET_SSE4_1 || flag_fp_int_builtin_inexact
-	   || !flag_trapping_math))"
+       && (TARGET_SSE4_1
+	  || (ROUND_<ROUNDING> != ROUND_ROUNDEVEN
+	      && (flag_fp_int_builtin_inexact || !flag_trapping_math))))"
 {
   if (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
       && (TARGET_SSE4_1 || flag_fp_int_builtin_inexact || !flag_trapping_math))
diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
index 906d74b1d08..15f019b9b49 100644
--- a/gcc/internal-fn.def
+++ b/gcc/internal-fn.def
@@ -234,6 +234,7 @@ DEF_INTERNAL_FLT_FLOATN_FN (FLOOR, ECF_CONST, floor, unary)
 DEF_INTERNAL_FLT_FLOATN_FN (NEARBYINT, ECF_CONST, nearbyint, unary)
 DEF_INTERNAL_FLT_FLOATN_FN (RINT, ECF_CONST, rint, unary)
 DEF_INTERNAL_FLT_FLOATN_FN (ROUND, ECF_CONST, round, unary)
+DEF_INTERNAL_FLT_FLOATN_FN (ROUNDEVEN, ECF_CONST, roundeven, unary)
 DEF_INTERNAL_FLT_FLOATN_FN (TRUNC, ECF_CONST, btrunc, unary)
 
 /* Binary math functions.  */
diff --git a/gcc/optabs.def b/gcc/optabs.def
index 4ffd0f35a40..065e3f64dda 100644
--- a/gcc/optabs.def
+++ b/gcc/optabs.def
@@ -268,6 +268,7 @@ OPTAB_D (fnms_optab, "fnms$a4")
 
 OPTAB_D (rint_optab, "rint$a2")
 OPTAB_D (round_optab, "round$a2")
+OPTAB_D (roundeven_optab, "roundeven$a2")
 OPTAB_D (floor_optab, "floor$a2")
 OPTAB_D (ceil_optab, "ceil$a2")
 OPTAB_D (btrunc_optab, "btrunc$a2")
diff --git a/gcc/reg-stack.c b/gcc/reg-stack.c
index 710f14a9544..0f0089acdea 100644
--- a/gcc/reg-stack.c
+++ b/gcc/reg-stack.c
@@ -1817,6 +1817,7 @@ subst_stack_regs_pat (rtx_insn *insn, stack_ptr regstack, rtx pat)
 	      case UNSPEC_FRNDINT:
 	      case UNSPEC_F2XM1:
 
+	      case UNSPEC_FRNDINT_ROUNDEVEN:
 	      case UNSPEC_FRNDINT_FLOOR:
 	      case UNSPEC_FRNDINT_CEIL:
 	      case UNSPEC_FRNDINT_TRUNC:
diff --git a/gcc/testsuite/gcc.target/i386/avx-vround-roundeven-1.c b/gcc/testsuite/gcc.target/i386/avx-vround-roundeven-1.c
new file mode 100644
index 00000000000..072d0f0e73a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx-vround-roundeven-1.c
@@ -0,0 +1,17 @@
+/* { dg-do compile } */
+/* { dg-options "-mavx" } */
+
+__attribute__((noinline, noclone)) double
+f1 (double x)
+{
+  return __builtin_roundeven (x);
+}
+
+__attribute__((noinline, noclone)) float
+f2 (float x)
+{
+  return __builtin_roundevenf (x);
+}
+
+/* { dg-final { scan-assembler-times "vroundsd\[^\n\r\]*xmm" 1 } } */
+/* { dg-final { scan-assembler-times "vroundss\[^\n\r\]*xmm" 1 } } */
diff --git a/gcc/testsuite/gcc.target/i386/avx-vround-roundeven-2.c b/gcc/testsuite/gcc.target/i386/avx-vround-roundeven-2.c
new file mode 100644
index 00000000000..211758d026a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx-vround-roundeven-2.c
@@ -0,0 +1,21 @@
+/* { dg-do run } */
+/* { dg-require-effective-target avx } */
+/* { dg-options "-mavx" } */
+
+#ifndef CHECK_H
+#define CHECK_H "avx-check.h"
+#define TEST avx_test
+#define SRC "avx-vround-roundeven-1.c"
+#endif
+
+#include CHECK_H
+#include SRC
+
+static void
+TEST (void)
+{
+  if (f1 (0.5) != 0.0 || f1 (1.5) != 2.0 || f1 (-0.5) != 0.0 || f1 (-1.5) != -2.0)
+    abort ();
+  if (f2 (0.5f) != 0.0f || f2 (1.5f) != 2.0f || f2 (-0.5f) != 0.0f || f2 (-1.5f) != -2.0f)
+    abort ();
+}

^ permalink raw reply	[flat|nested] 6+ messages in thread
end of thread, other threads:[~2019-08-22 14:26 UTC | newest]

Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2019-07-24 10:45 [PATCH] i386: Roundeven expansion for SSE4.1+ Tejas Joshi
2019-07-27  9:22 ` Uros Bizjak
     [not found]   ` <CACMrGjATykFB-Noe+eTdunPUfEPFQk15uHc4gABdE7B+mGD5Og@mail.gmail.com>
     [not found]     ` <CACMrGjCwVOMf=pSjnHnusVyQ=dT73g-XYX3UH7iecBSCn8Pr0g@mail.gmail.com>
2019-07-31  7:10       ` Tejas Joshi
2019-07-31  8:51         ` Uros Bizjak
2019-08-22 14:43           ` Martin Jambor
2019-08-08 21:54   ` Jeff Law

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).