[libc] Template the writing mode for the writer class (#111559)

Summary:
Currently we dispatch the writing mode off of a runtime enum passed in
by the constructor. This causes very unfortunate codegen for the GPU
targets where we get worst-case codegen because of the unused function
pointer for `sprintf`. Instead, this patch moves all of this to a
template so it can be masked out. This results in no dynamic stack and
uses 60 VGPRs instead of 117. It also compiles about 5x as fast.

Author: jhuber6

libc/config/config.json

@@ -41,6 +41,10 @@
     "LIBC_CONF_PRINTF_DISABLE_STRERROR": {
       "value": false,
       "doc": "Disable handling of %m to print strerror in printf and friends."
+    },
+    "LIBC_CONF_PRINTF_RUNTIME_DISPATCH": {
+      "value": true,
+      "doc": "Use dynamic dispatch for the output mechanism to reduce code size."
     }
   },
   "scanf": {

libc/config/gpu/amdgpu/config.json

@@ -19,6 +19,9 @@
     },
     "LIBC_CONF_PRINTF_DISABLE_STRERROR": {
       "value": true
+    },
+    "LIBC_CONF_PRINTF_RUNTIME_DISPATCH": {
+      "value": false
     }
   },
   "scanf": {

libc/config/gpu/nvptx/config.json

@@ -19,6 +19,9 @@
     },
     "LIBC_CONF_PRINTF_DISABLE_STRERROR": {
       "value": true
+    },
+    "LIBC_CONF_PRINTF_RUNTIME_DISPATCH": {
+      "value": false
     }
   },
   "scanf": {

libc/docs/configure.rst

@@ -45,6 +45,7 @@ to learn about the defaults for your platform and target.
     - ``LIBC_CONF_PRINTF_FLOAT_TO_STR_USE_DYADIC_FLOAT``: Use dyadic float for faster and smaller but less accurate printf doubles.
     - ``LIBC_CONF_PRINTF_FLOAT_TO_STR_USE_FLOAT320``: Use an alternative printf float implementation based on 320-bit floats
     - ``LIBC_CONF_PRINTF_FLOAT_TO_STR_USE_MEGA_LONG_DOUBLE_TABLE``: Use large table for better printf long double performance.
+    - ``LIBC_CONF_PRINTF_RUNTIME_DISPATCH``: Use dynamic dispatch for the output mechanism to reduce code size.
 * **"pthread" options**
     - ``LIBC_CONF_RAW_MUTEX_DEFAULT_SPIN_COUNT``: Default number of spins before blocking if a mutex is in contention (default to 100).
     - ``LIBC_CONF_RWLOCK_DEFAULT_SPIN_COUNT``: Default number of spins before blocking if a rwlock is in contention (default to 100).

libc/src/stdio/printf_core/CMakeLists.txt

@@ -25,6 +25,9 @@ endif()
 if(LIBC_CONF_PRINTF_DISABLE_STRERROR)
   list(APPEND printf_config_copts "-DLIBC_COPT_PRINTF_DISABLE_STRERROR")
 endif()
+if(LIBC_CONF_PRINTF_RUNTIME_DISPATCH)
+  list(APPEND printf_config_copts "-DLIBC_COPT_PRINTF_RUNTIME_DISPATCH")
+endif()
 if(printf_config_copts)
   list(PREPEND printf_config_copts "COMPILE_OPTIONS")
 endif()
@@ -62,10 +65,8 @@ add_header_library(
     libc.src.__support.common
 )
 
-add_object_library(
+add_header_library(
   writer
-  SRCS
-    writer.cpp
   HDRS
     writer.h
   DEPENDS
@@ -76,10 +77,8 @@ add_object_library(
     libc.src.string.memory_utils.inline_memset
 )
 
-add_object_library(
+add_header_library(
   converter
-  SRCS
-    converter.cpp
   HDRS
     converter.h
     converter_atlas.h
@@ -113,10 +112,8 @@ add_object_library(
     libc.src.__support.StringUtil.error_to_string
 )
 
-add_object_library(
+add_header_library(
   printf_main
-  SRCS
-    printf_main.cpp
   HDRS
     printf_main.h
   DEPENDS

libc/src/stdio/printf_core/char_converter.h

@@ -17,7 +17,9 @@
 namespace LIBC_NAMESPACE_DECL {
 namespace printf_core {
 
-LIBC_INLINE int convert_char(Writer *writer, const FormatSection &to_conv) {
+template <WriteMode write_mode>
+LIBC_INLINE int convert_char(Writer<write_mode> *writer,
+                             const FormatSection &to_conv) {
   char c = static_cast<char>(to_conv.conv_val_raw);
 
   constexpr int STRING_LEN = 1;

libc/src/stdio/printf_core/converter.cpp

@@ -11,8 +11,18 @@
 
 #include "src/__support/macros/config.h"
 #include "src/stdio/printf_core/core_structs.h"
+#include "src/stdio/printf_core/printf_config.h"
+#include "src/stdio/printf_core/strerror_converter.h"
 #include "src/stdio/printf_core/writer.h"
 
+// This option allows for replacing all of the conversion functions with custom
+// replacements. This allows conversions to be replaced at compile time.
+#ifndef LIBC_COPT_PRINTF_CONV_ATLAS
+#include "src/stdio/printf_core/converter_atlas.h"
+#else
+#include LIBC_COPT_PRINTF_CONV_ATLAS
+#endif
+
 #include <stddef.h>
 
 namespace LIBC_NAMESPACE_DECL {
@@ -21,7 +31,80 @@ namespace printf_core {
 // convert will call a conversion function to convert the FormatSection into
 // its string representation, and then that will write the result to the
 // writer.
-int convert(Writer *writer, const FormatSection &to_conv);
+template <WriteMode write_mode>
+int convert(Writer<write_mode> *writer, const FormatSection &to_conv) {
+  if (!to_conv.has_conv)
+    return writer->write(to_conv.raw_string);
+
+#if !defined(LIBC_COPT_PRINTF_DISABLE_FLOAT) &&                                \
+    defined(LIBC_COPT_PRINTF_HEX_LONG_DOUBLE)
+  if (to_conv.length_modifier == LengthModifier::L) {
+    switch (to_conv.conv_name) {
+    case 'f':
+    case 'F':
+    case 'e':
+    case 'E':
+    case 'g':
+    case 'G':
+      return convert_float_hex_exp(writer, to_conv);
+    default:
+      break;
+    }
+  }
+#endif // LIBC_COPT_PRINTF_DISABLE_FLOAT
+
+  switch (to_conv.conv_name) {
+  case '%':
+    return writer->write("%");
+  case 'c':
+    return convert_char(writer, to_conv);
+  case 's':
+    return convert_string(writer, to_conv);
+  case 'd':
+  case 'i':
+  case 'u':
+  case 'o':
+  case 'x':
+  case 'X':
+  case 'b':
+  case 'B':
+    return convert_int(writer, to_conv);
+#ifndef LIBC_COPT_PRINTF_DISABLE_FLOAT
+  case 'f':
+  case 'F':
+    return convert_float_decimal(writer, to_conv);
+  case 'e':
+  case 'E':
+    return convert_float_dec_exp(writer, to_conv);
+  case 'a':
+  case 'A':
+    return convert_float_hex_exp(writer, to_conv);
+  case 'g':
+  case 'G':
+    return convert_float_dec_auto(writer, to_conv);
+#endif // LIBC_COPT_PRINTF_DISABLE_FLOAT
+#ifdef LIBC_INTERNAL_PRINTF_HAS_FIXED_POINT
+  case 'r':
+  case 'R':
+  case 'k':
+  case 'K':
+    return convert_fixed(writer, to_conv);
+#endif // LIBC_INTERNAL_PRINTF_HAS_FIXED_POINT
+#ifndef LIBC_COPT_PRINTF_DISABLE_STRERROR
+  case 'm':
+    return convert_strerror(writer, to_conv);
+#endif // LIBC_COPT_PRINTF_DISABLE_STRERROR
+#ifndef LIBC_COPT_PRINTF_DISABLE_WRITE_INT
+  case 'n':
+    return convert_write_int(writer, to_conv);
+#endif // LIBC_COPT_PRINTF_DISABLE_WRITE_INT
+  case 'p':
+    return convert_pointer(writer, to_conv);
+  default:
+    return writer->write(to_conv.raw_string);
+  }
+  return -1;
+}
 
 } // namespace printf_core
 } // namespace LIBC_NAMESPACE_DECL

libc/src/stdio/printf_core/converter.h

@@ -63,7 +63,9 @@ LIBC_INLINE constexpr uint32_t const_ten_exp(uint32_t exponent) {
     }                                                                          \
   } while (false)
 
-LIBC_INLINE int convert_fixed(Writer *writer, const FormatSection &to_conv) {
+template <WriteMode write_mode>
+LIBC_INLINE int convert_fixed(Writer<write_mode> *writer,
+                              const FormatSection &to_conv) {
   // Long accum should be the largest type, so we can store all the smaller
   // numbers in things sized for it.
   using LARep = fixed_point::FXRep<unsigned long accum>;