rs6000.c (rs6000_emit_swdivsf): New function.

        * config/rs6000/rs6000.c (rs6000_emit_swdivsf): New function.
        (rs6000_emit_swdivdf): New function.
        * config/rs6000/rs6000.md (fres): New pattern.
        (divsf3): Add approximation through rs6000_emit_swdivsf.
        (fred): New pattern.
        (divdf3): Add approximation through rs6000_emit_swdivdf.
        * config/rs6000/rs6000-protos.h (rs6000_emit_swdivsf): Declare.
        (rs6000_emit_swdivdf): Declare.
        * config/rs6000/rs6000.opt (mswdiv): New option.
        * doc/invoke.texi (RS/6000 and PowerPC Options): Document mswdiv.

From-SVN: r100645

6 files changed, 165 insertions, 3 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 50916ee..3d87d0a 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,16 @@
+2005-06-05  David Edelsohn  <edelsohn@gnu.org>
+
+	* config/rs6000/rs6000.c (rs6000_emit_swdivsf): New function.
+	(rs6000_emit_swdivdf): New function.
+	* config/rs6000/rs6000.md (fres): New pattern.
+	(divsf3): Add approximation through rs6000_emit_swdivsf.
+	(fred): New pattern.
+	(divdf3): Add approximation through rs6000_emit_swdivdf.
+	* config/rs6000/rs6000-protos.h (rs6000_emit_swdivsf): Declare.
+	(rs6000_emit_swdivdf): Declare.
+	* config/rs6000/rs6000.opt (mswdiv): New option.
+	* doc/invoke.texi (RS/6000 and PowerPC Options): Document mswdiv.
+
 2005-06-05  Zdenek Dvorak  <dvorakz@suse.cz>
 
 	PR tree-optimization/21846
diff --git a/gcc/config/rs6000/rs6000-protos.h b/gcc/config/rs6000/rs6000-protos.h
index 63ffe4f..041169b 100644
--- a/gcc/config/rs6000/rs6000-protos.h
+++ b/gcc/config/rs6000/rs6000-protos.h
@@ -82,6 +82,8 @@ extern int rs6000_emit_vector_cond_expr (rtx, rtx, rtx, rtx, rtx, rtx);
 extern void rs6000_emit_minmax (rtx, enum rtx_code, rtx, rtx);
 extern void rs6000_emit_sync (enum rtx_code, enum machine_mode,
 			      rtx, rtx, rtx, rtx, bool);
+extern void rs6000_emit_swdivsf (rtx, rtx, rtx);
+extern void rs6000_emit_swdivdf (rtx, rtx, rtx);
 extern void output_toc (FILE *, rtx, int, enum machine_mode);
 extern void rs6000_initialize_trampoline (rtx, rtx, rtx);
 extern rtx rs6000_longcall_ref (rtx);
diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
index 2506445..9626885 100644
--- a/gcc/config/rs6000/rs6000.c
+++ b/gcc/config/rs6000/rs6000.c
@@ -17955,6 +17955,109 @@ rs6000_memory_move_cost (enum machine_mode mode, enum reg_class class,
     return 4 + rs6000_register_move_cost (mode, class, GENERAL_REGS);
 }
 
+/* Newton-Raphson approximation of single-precision floating point divide n/d.
+   Assumes no trapping math and finite arguments.  */
+
+void
+rs6000_emit_swdivsf (rtx res, rtx n, rtx d)
+{
+  rtx x0, e0, e1, y1, u0, v0, one;
+
+  x0 = gen_reg_rtx (SFmode);
+  e0 = gen_reg_rtx (SFmode);
+  e1 = gen_reg_rtx (SFmode);
+  y1 = gen_reg_rtx (SFmode);
+  u0 = gen_reg_rtx (SFmode);
+  v0 = gen_reg_rtx (SFmode);
+  one = force_reg (SFmode, CONST_DOUBLE_FROM_REAL_VALUE (dconst1, SFmode));
+
+  /* x0 = 1./d estimate */
+  emit_insn (gen_rtx_SET (VOIDmode, x0,
+			  gen_rtx_UNSPEC (SFmode, gen_rtvec (1, d),
+					  UNSPEC_FRES)));
+  /* e0 = 1. - d * x0 */
+  emit_insn (gen_rtx_SET (VOIDmode, e0,
+			  gen_rtx_MINUS (SFmode, one,
+					 gen_rtx_MULT (SFmode, d, x0))));
+  /* e1 = e0 + e0 * e0 */
+  emit_insn (gen_rtx_SET (VOIDmode, e1,
+			  gen_rtx_PLUS (SFmode,
+					gen_rtx_MULT (SFmode, e0, e0), e0)));
+  /* y1 = x0 + e1 * x0 */
+  emit_insn (gen_rtx_SET (VOIDmode, y1,
+			  gen_rtx_PLUS (SFmode,
+					gen_rtx_MULT (SFmode, e1, x0), x0)));
+  /* u0 = n * y1 */
+  emit_insn (gen_rtx_SET (VOIDmode, u0,
+			  gen_rtx_MULT (SFmode, n, y1)));
+  /* v0 = n - d * u0 */
+  emit_insn (gen_rtx_SET (VOIDmode, v0,
+			  gen_rtx_MINUS (SFmode, n,
+					 gen_rtx_MULT (SFmode, d, u0))));
+  /* res = u0 + v0 * y1 */
+  emit_insn (gen_rtx_SET (VOIDmode, res,
+			  gen_rtx_PLUS (SFmode,
+					gen_rtx_MULT (SFmode, v0, y1), u0)));
+}
+
+/* Newton-Raphson approximation of double-precision floating point divide n/d.
+   Assumes no trapping math and finite arguments.  */
+
+void
+rs6000_emit_swdivdf (rtx res, rtx n, rtx d)
+{
+  rtx x0, e0, e1, e2, y1, y2, y3, u0, v0, one;
+
+  x0 = gen_reg_rtx (DFmode);
+  e0 = gen_reg_rtx (DFmode);
+  e1 = gen_reg_rtx (DFmode);
+  e2 = gen_reg_rtx (DFmode);
+  y1 = gen_reg_rtx (DFmode);
+  y2 = gen_reg_rtx (DFmode);
+  y3 = gen_reg_rtx (DFmode);
+  u0 = gen_reg_rtx (DFmode);
+  v0 = gen_reg_rtx (DFmode);
+  one = force_reg (DFmode, CONST_DOUBLE_FROM_REAL_VALUE (dconst1, DFmode));
+
+  /* x0 = 1./d estimate */
+  emit_insn (gen_rtx_SET (VOIDmode, x0,
+			  gen_rtx_UNSPEC (DFmode, gen_rtvec (1, d),
+					  UNSPEC_FRES)));
+  /* e0 = 1. - d * x0 */
+  emit_insn (gen_rtx_SET (VOIDmode, e0,
+			  gen_rtx_MINUS (DFmode, one,
+					 gen_rtx_MULT (SFmode, d, x0))));
+  /* y1 = x0 + e0 * x0 */
+  emit_insn (gen_rtx_SET (VOIDmode, y1,
+			  gen_rtx_PLUS (DFmode,
+					gen_rtx_MULT (DFmode, e0, x0), x0)));
+  /* e1 = e0 * e0 */
+  emit_insn (gen_rtx_SET (VOIDmode, e1,
+			  gen_rtx_MULT (DFmode, e0, e0)));
+  /* y2 = y1 + e1 * y1 */
+  emit_insn (gen_rtx_SET (VOIDmode, y2,
+			  gen_rtx_PLUS (DFmode,
+					gen_rtx_MULT (DFmode, e1, y1), y1)));
+  /* e2 = e1 * e1 */
+  emit_insn (gen_rtx_SET (VOIDmode, e2,
+			  gen_rtx_MULT (DFmode, e1, e1)));
+  /* y3 = y2 + e2 * y2 */
+  emit_insn (gen_rtx_SET (VOIDmode, y3,
+			  gen_rtx_PLUS (DFmode,
+					gen_rtx_MULT (DFmode, e2, y2), y2)));
+  /* u0 = n * y3 */
+  emit_insn (gen_rtx_SET (VOIDmode, u0,
+			  gen_rtx_MULT (DFmode, n, y3)));
+  /* v0 = n - d * u0 */
+  emit_insn (gen_rtx_SET (VOIDmode, v0,
+			  gen_rtx_MINUS (DFmode, n,
+					 gen_rtx_MULT (DFmode, d, u0))));
+  /* res = u0 + v0 * y3 */
+  emit_insn (gen_rtx_SET (VOIDmode, res,
+			  gen_rtx_PLUS (DFmode,
+					gen_rtx_MULT (DFmode, v0, y3), u0)));
+}
+
 /* Return an RTX representing where to find the function value of a
    function returning MODE.  */
 static rtx
diff --git a/gcc/config/rs6000/rs6000.md b/gcc/config/rs6000/rs6000.md
index 97ac84b..633dd7c 100644
--- a/gcc/config/rs6000/rs6000.md
+++ b/gcc/config/rs6000/rs6000.md
@@ -58,6 +58,7 @@
    (UNSPEC_LWSYNC		36)
    (UNSPEC_ISYNC		37)
    (UNSPEC_POPCNTB		38)
+   (UNSPEC_FRES			39)
   ])
 
 ;;
@@ -4640,12 +4641,26 @@
   "{fm|fmul} %0,%1,%2"
   [(set_attr "type" "dmul")])
 
+(define_insn "fres"
+  [(set (match_operand:SF 0 "gpc_reg_operand" "=f")
+	(unspec:SF [(match_operand:SF 1 "gpc_reg_operand" "f")] UNSPEC_FRES))]
+  "TARGET_PPC_GFXOPT && flag_finite_math_only"
+  "fres %0,%1"
+  [(set_attr "type" "fp")])
+
 (define_expand "divsf3"
   [(set (match_operand:SF 0 "gpc_reg_operand" "")
 	(div:SF (match_operand:SF 1 "gpc_reg_operand" "")
 		(match_operand:SF 2 "gpc_reg_operand" "")))]
   "TARGET_HARD_FLOAT"
-  "")
+{
+  if (swdiv && !optimize_size && TARGET_PPC_GFXOPT
+  && flag_finite_math_only && !flag_trapping_math)
+    {
+      rs6000_emit_swdivsf (operands[0], operands[1], operands[2]);
+      DONE;
+    }
+})
 
 (define_insn ""
   [(set (match_operand:SF 0 "gpc_reg_operand" "=f")
@@ -5028,12 +5043,26 @@
   "{fm|fmul} %0,%1,%2"
   [(set_attr "type" "dmul")])
 
+(define_insn "fred"
+  [(set (match_operand:DF 0 "gpc_reg_operand" "=f")
+	(unspec:DF [(match_operand:DF 1 "gpc_reg_operand" "f")] UNSPEC_FRES))]
+  "TARGET_POPCNTB && flag_finite_math_only"
+  "fre %0,%1"
+  [(set_attr "type" "fp")])
+
 (define_expand "divdf3"
   [(set (match_operand:DF 0 "gpc_reg_operand" "")
 	(div:DF (match_operand:DF 1 "gpc_reg_operand" "")
 		(match_operand:DF 2 "gpc_reg_operand" "")))]
   "TARGET_HARD_FLOAT && (TARGET_FPRS || TARGET_E500_DOUBLE)"
-  "")
+{
+  if (swdiv && !optimize_size && TARGET_POPCNTB
+  && flag_finite_math_only && !flag_trapping_math)
+    {
+      rs6000_emit_swdivdf (operands[0], operands[1], operands[2]);
+      DONE;
+    }
+})
 
 (define_insn "*divdf3_fpr"
   [(set (match_operand:DF 0 "gpc_reg_operand" "=f")
diff --git a/gcc/config/rs6000/rs6000.opt b/gcc/config/rs6000/rs6000.opt
index fd5d3b0..7a08c67 100644
--- a/gcc/config/rs6000/rs6000.opt
+++ b/gcc/config/rs6000/rs6000.opt
@@ -123,6 +123,10 @@ mxl-compat
 Target Report Var(TARGET_XL_COMPAT)
 Conform more closely to IBM XLC semantics
 
+mswdiv
+Target Report Var(swdiv)
+Generate software floating point divide for better throughput
+
 mno-fp-in-toc
 Target Report RejectNegative Mask(NO_FP_IN_TOC)
 Do not place floating point constants in TOC
diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
index 0bcd3d9..eef9a0f 100644
--- a/gcc/doc/invoke.texi
+++ b/gcc/doc/invoke.texi
@@ -631,7 +631,7 @@ See RS/6000 and PowerPC Options.
 -mstrict-align  -mno-strict-align  -mrelocatable @gol
 -mno-relocatable  -mrelocatable-lib  -mno-relocatable-lib @gol
 -mtoc  -mno-toc  -mlittle  -mlittle-endian  -mbig  -mbig-endian @gol
--mdynamic-no-pic @gol
+-mdynamic-no-pic  -maltivec  -mswdiv @gol
 -mprioritize-restricted-insns=@var{priority} @gol
 -msched-costly-dep=@var{dependence_type} @gol
 -minsert-sched-nops=@var{scheme} @gol
@@ -10720,6 +10720,17 @@ values for @var{cpu_type} are used for @option{-mtune} as for
 architecture, registers, and mnemonics set by @option{-mcpu}, but the
 scheduling parameters set by @option{-mtune}.
 
+@item -mswdiv
+@itemx -mno-swdiv
+@opindex mswdiv
+@opindex mno-swdiv
+Generate code to compute division as reciprocal estimate and iterative
+refinement, creating opportunities for increased throughput.  This
+feature requires: optional PowerPC Graphics instruction set for single
+precision and FRE instruction for double precision, assuming divides
+cannot generate user-visible traps, and the domain values not include
+Infinities, denormals or zero denominator.
+
 @item -maltivec
 @itemx -mno-altivec
 @opindex maltivec