public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH v4] LoongArch:Implement 128-bit floating point functions in gcc.
@ 2023-08-31  2:46 chenxiaolong
  2023-08-31  4:32 ` Xi Ruoyao
  0 siblings, 1 reply; 2+ messages in thread
From: chenxiaolong @ 2023-08-31  2:46 UTC (permalink / raw)
  To: gcc-patches; +Cc: xry111, i, xuchenghua, chenglulu, chenxiaolong

Brief version history of patch set:

v1 -> v2:
   According to the GNU code specification, adjust the format of the
function implementation with "q" as the suffix function.

v2 - >v3:

   1.On the LoongArch architecture, refer to the functionality of 64-bit functions
and modify the underlying implementation of __builtin_{nanq,nansq} functions
in libgcc.

   2.Modify the function's instruction template to use some instructions such as
"bstrins.d" to implement the 128-bit __builtin_{fabsq,copysignq} function
instead of calling libgcc library support, so as to better play the machine's
performance.

v3 -> v4:

   1.The above v1,v2, and v3 all implement 128-bit floating-point functions
with "q" as the suffix, but it is an older implementation. The v4 version
completely abandoned the old implementation by associating the 128-bit
floating-point function with the "q" suffix with the "f128" function that
already existed in GCC.

   2.Modify the code so that both "__float128" and "_Float128" function types
can be supported in compiler gcc.

   3.Associating a function with the suffix "q" to the "f128" function allows
two different forms of the function to produce the same effect, For example,
__builtin_{huge_{valq,valf128},{infq/inff128},{nanq/nanf128},{nansq/nansf128}}.

   4.For the _builtin_{fabsq,copysignq} function, do not call the new "f128"
implementation, but use the "bstrins" and other instructions in the machine
description file to implement the function function, the result is that the
number of assembly instructions can be reduced and the function optimization
to achieve the optimal effect.

During implementation, float128_type_node is bound with the type "__float128"
so that the compiler can correctly identify the type   of the function. The
"q" suffix is associated with the "f128" function, which makes GCC more
flexible to support different user input cases, implementing functions such
as __builtin_{huge_valq infq, fabsq, copysignq, nanq,nansq}.At the same time,
the __builtin_{copysign{q/f128},fabs{q/f128}} functions are optimized by
using "bstrins" and other instructions on LoongArch architecture to better
play the optimization performance of the compiler.

gcc/ChangeLog:

	* config/loongarch/loongarch-builtins.cc (loongarch_init_builtins):
	Associate the __float128 type to float128_type_node so that it can
	be recognized by the compiler.
	* config/loongarch/loongarch-c.cc (loongarch_cpu_cpp_builtins):
	Add the flag "FLOAT128_TYPE" to gcc and associate a function
        with the suffix "q" to "f128".
	* config/loongarch/loongarch.md (abstf2):Modify the instruction
        template to implement the __builtin_{copysignf128/fabsf128} function.
	(abstf_local):Ditto.
	(copysigntf3):Implement the built-in function __builtin_copysignf128().
	* doc/extend.texi:Added support for 128-bit floating-point functions on
	the LoongArch architecture.

gcc/testsuite/ChangeLog:

	* gcc.target/loongarch/math-float-128.c: New test.
---
 gcc/config/loongarch/loongarch-builtins.cc    |   5 +
 gcc/config/loongarch/loongarch-c.cc           |  11 ++
 gcc/config/loongarch/loongarch.md             |  54 ++++++++
 gcc/doc/extend.texi                           |  20 ++-
 .../gcc.target/loongarch/math-float-128.c     | 115 ++++++++++++++++++
 5 files changed, 202 insertions(+), 3 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/loongarch/math-float-128.c

diff --git a/gcc/config/loongarch/loongarch-builtins.cc b/gcc/config/loongarch/loongarch-builtins.cc
index b929f224dfa..58b612bf445 100644
--- a/gcc/config/loongarch/loongarch-builtins.cc
+++ b/gcc/config/loongarch/loongarch-builtins.cc
@@ -256,6 +256,11 @@ loongarch_init_builtins (void)
   unsigned int i;
   tree type;
 
+  /* Register the type float128_type_node as a built-in type and
+     give it an alias "__float128".  */
+  (*lang_hooks.types.register_builtin_type) (float128_type_node,
+					    "__float128");
+
   /* Iterate through all of the bdesc arrays, initializing all of the
      builtin functions.  */
   for (i = 0; i < ARRAY_SIZE (loongarch_builtins); i++)
diff --git a/gcc/config/loongarch/loongarch-c.cc b/gcc/config/loongarch/loongarch-c.cc
index 67911b78f28..6ffbf748316 100644
--- a/gcc/config/loongarch/loongarch-c.cc
+++ b/gcc/config/loongarch/loongarch-c.cc
@@ -99,6 +99,17 @@ loongarch_cpu_cpp_builtins (cpp_reader *pfile)
   else
     builtin_define ("__loongarch_frlen=0");
 
+  /* Add support for FLOAT128_TYPE on the LoongArch architecture.  */
+  builtin_define ("__FLOAT128_TYPE__");
+
+  /* Map the old _Float128 'q' builtins into the new 'f128' builtins.  */
+  builtin_define ("__builtin_fabsq=__builtin_fabsf128");
+  builtin_define ("__builtin_copysignq=__builtin_copysignf128");
+  builtin_define ("__builtin_nanq=__builtin_nanf128");
+  builtin_define ("__builtin_nansq=__builtin_nansf128");
+  builtin_define ("__builtin_infq=__builtin_inff128");
+  builtin_define ("__builtin_huge_valq=__builtin_huge_valf128");
+
   /* Native Data Sizes.  */
   builtin_define_with_int_value ("_LOONGARCH_SZINT", INT_TYPE_SIZE);
   builtin_define_with_int_value ("_LOONGARCH_SZLONG", LONG_TYPE_SIZE);
diff --git a/gcc/config/loongarch/loongarch.md b/gcc/config/loongarch/loongarch.md
index b37e070660f..230d7aac972 100644
--- a/gcc/config/loongarch/loongarch.md
+++ b/gcc/config/loongarch/loongarch.md
@@ -38,6 +38,7 @@ (define_c_enum "unspec" [
   UNSPEC_FMAX
   UNSPEC_FMIN
   UNSPEC_FCOPYSIGN
+  UNSPEC_COPYSIGNF128
   UNSPEC_FTINT
   UNSPEC_FTINTRM
   UNSPEC_FTINTRP
@@ -2008,6 +2009,59 @@ (define_insn "movfcc"
   ""
   "movgr2cf\t%0,$r0")
 
+;; Implement __builtin_fabs128 function.
+
+(define_expand "abstf2"
+  [(match_operand:TF 0 "register_operand")
+   (match_operand:TF 1 "register_operand")]
+  "TARGET_64BIT"
+{
+  loongarch_emit_move (operands[0], operands[1]);
+  emit_insn (gen_abstf_local (operands[0]));
+  DONE;
+})
+
+(define_insn "abstf_local"
+  [(set (match_operand:TF 0 "register_operand" "+r")
+	(abs:TF (match_dup 0)))]
+  "TARGET_64BIT"
+{
+  operands[0] = gen_rtx_REG (DImode, REGNO (operands[0]) + 1);
+  return "bstrins.d\t%0,$r0,0x3f,0x3f";
+})
+
+;; Implement __builtin_copysignf128 function.
+
+(define_insn_and_split "copysigntf3"
+  [(set (match_operand:TF 0 "register_operand" "=&r")
+	(unspec:TF [(match_operand:TF 1 "register_operand" "r")
+		    (match_operand:TF 2 "register_operand" "r")]
+		    UNSPEC_COPYSIGNF128))]
+  "TARGET_64BIT"
+  "#"
+  "reload_completed"
+ [(const_int 0)]
+{
+  rtx op0_lo = gen_rtx_REG (DImode,REGNO (operands[0]) + 0);
+  rtx op0_hi = gen_rtx_REG (DImode,REGNO (operands[0]) + 1);
+  rtx op1_lo = gen_rtx_REG (DImode,REGNO (operands[1]) + 0);
+  rtx op1_hi = gen_rtx_REG (DImode,REGNO (operands[1]) + 1);
+  rtx op2_hi = gen_rtx_REG (DImode,REGNO (operands[2]) + 1);
+
+  if (REGNO (operands[1]) == REGNO (operands[2]))
+    {
+      loongarch_emit_move (operands[0], operands[1]);
+      DONE;
+    }
+  else
+    {
+      loongarch_emit_move (op0_hi, op2_hi);
+      loongarch_emit_move (op0_lo, op1_lo);
+      emit_insn (gen_insvdi (op0_hi, GEN_INT (63), GEN_INT (0), op1_hi));
+      DONE;
+    }
+})
+
 ;; Conditional move instructions.
 
 (define_insn "*sel<code><GPR:mode>_using_<GPR2:mode>"
diff --git a/gcc/doc/extend.texi b/gcc/doc/extend.texi
index 400284b85f5..074f1ee33a0 100644
--- a/gcc/doc/extend.texi
+++ b/gcc/doc/extend.texi
@@ -1093,10 +1093,10 @@ types.
 As an extension, GNU C and GNU C++ support additional floating
 types, which are not supported by all targets.
 @itemize @bullet
-@item @code{__float128} is available on i386, x86_64, IA-64, and
-hppa HP-UX, as well as on PowerPC GNU/Linux targets that enable
+@item @code{__float128} is available on i386, x86_64, IA-64, LoongArch
+and hppa HP-UX, as well as on PowerPC GNU/Linux targets that enable
 the vector scalar (VSX) instruction set.  @code{__float128} supports
-the 128-bit floating type.  On i386, x86_64, PowerPC, and IA-64
+the 128-bit floating type.  On i386, x86_64, PowerPC, LoongArch and IA-64,
 other than HP-UX, @code{__float128} is an alias for @code{_Float128}.
 On hppa and IA-64 HP-UX, @code{__float128} is an alias for @code{long
 double}.
@@ -16657,6 +16657,20 @@ function you need to include @code{larchintrin.h}.
     void __break (imm0_32767)
 @end smallexample
 
+Additional built-in functions are available for LoongArch family
+processors to efficiently use 128-bit floating-point (__float128)
+values.
+
+The following are the basic built-in functions supported.
+@smallexample
+__float128 __builtin_fabsq (__float128);
+__float128 __builtin_copysignq (__float128, __float128);
+__float128 __builtin_infq (void);
+__float128 __builtin_huge_valq (void);
+__float128 __builtin_nanq (void);
+__float128 __builtin_nansq (void);
+@end smallexample
+
 @node MIPS DSP Built-in Functions
 @subsection MIPS DSP Built-in Functions
 
diff --git a/gcc/testsuite/gcc.target/loongarch/math-float-128.c b/gcc/testsuite/gcc.target/loongarch/math-float-128.c
new file mode 100644
index 00000000000..5eab3019278
--- /dev/null
+++ b/gcc/testsuite/gcc.target/loongarch/math-float-128.c
@@ -0,0 +1,115 @@
+/* { dg-do compile } */
+/* { dg-options " -march=loongarch64 " } */
+/* { dg-final { scan-assembler-times {bstrins.d}  6} } */
+/* { dg-final { scan-assembler-not "my_fabsq2:.*\\bl\t%plt\\(__fabsq\\).*my_fabsq2" } } */
+/* { dg-final { scan-assembler-not "my_copysignq2:.*\\bl\t%plt\\(__copysignq\\).*my_copysignq2" } } */
+/* { dg-final { scan-assembler-not "my_nanq2:.*\\bl\t%plt\\(__builtin_nanq\\).*my_nanq2" } } */
+/* { dg-final { scan-assembler-not "my_nansq2:.*\\bl\t%plt\\(__builtin_nansq\\).*my_nansq2" } } */
+
+__float128
+my_fabsq1 (__float128 a)
+{
+  return __builtin_fabsq (a);
+}
+
+_Float128
+my_fabsq2 (_Float128 a)
+{
+  return __builtin_fabsq (a);
+}
+
+_Float128
+my_fabsf128 (_Float128 a)
+{
+  return __builtin_fabsf128 (a);
+}
+
+__float128
+my_copysignq1 (__float128 a, __float128 b)
+{
+  return __builtin_copysignq (a, b);
+}
+
+_Float128
+my_copysignq2 (_Float128 a, _Float128 b)
+{
+  return __builtin_copysignq (a, b);
+}
+
+_Float128
+my_copysignf128 (_Float128 a, _Float128 b)
+{
+  return __builtin_copysignf128 (a, b);
+}
+
+__float128
+my_infq1 (void)
+{
+  return __builtin_infq ();
+}
+
+_Float128
+my_infq2 (void)
+{
+  return __builtin_infq ();
+}
+
+_Float128
+my_inff128 (void)
+{
+  return __builtin_inff128 ();
+}
+
+__float128
+my_huge_valq1 (void)
+{
+  return __builtin_huge_valq ();
+}
+
+_Float128
+my_huge_valq2 (void)
+{
+  return __builtin_huge_valq ();
+}
+
+_Float128
+my_huge_valf128 (void)
+{
+  return __builtin_huge_valf128 ();
+}
+
+__float128
+my_nanq1 (void)
+{
+  return __builtin_nanq ("");
+}
+
+_Float128
+my_nanq2 (void)
+{
+  return __builtin_nanq ("");
+}
+
+_Float128
+my_nanf128 (void)
+{
+  return __builtin_nanf128 ("");
+}
+
+__float128
+my_nansq1 (void)
+{
+  return __builtin_nansq ("");
+}
+
+_Float128
+my_nansq2 (void)
+{
+  return __builtin_nansq ("");
+}
+
+_Float128
+my_nansf128 (void)
+{
+  return __builtin_nansf128 ("");
+}
-- 
2.20.1


^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2023-08-31  4:32 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-08-31  2:46 [PATCH v4] LoongArch:Implement 128-bit floating point functions in gcc chenxiaolong
2023-08-31  4:32 ` Xi Ruoyao

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).