From f70cb7901a43c201242399225d31eed83e7d41ce Mon Sep 17 00:00:00 2001
From: Josh Blum
Date: Mon, 16 Apr 2012 21:56:01 -0700
Subject: volk: use archs.xml to specify compiler flags + support

---
 volk/gen/archs.xml             | 45 +++++++++++++++++++++++++-----------------
 volk/gen/volk_arch_defs.py     | 19 +++++++++++++-----
 volk/gen/volk_compile_utils.py | 17 +++-------------
 3 files changed, 44 insertions(+), 37 deletions(-)

(limited to 'volk/gen')

diff --git a/volk/gen/archs.xml b/volk/gen/archs.xml
index ac4dd9f7f..2c9ab41a5 100644
--- a/volk/gen/archs.xml
+++ b/volk/gen/archs.xml
@@ -5,21 +5,21 @@
 </arch>
 
 <arch name="altivec">
-  <flag>maltivec</flag>
+  <flag compiler="gnu">-maltivec</flag>
   <alignment>16</alignment>
   <check name="has_ppc"></check>
 </arch>
 
 <arch name="neon">
-  <flag>mfpu=neon</flag>
-  <flag>mfloat-abi=softfp</flag>
-  <flag>funsafe-math-optimizations</flag>
+  <flag compiler="gnu">-mfpu=neon</flag>
+  <flag compiler="gnu">-mfloat-abi=softfp</flag>
+  <flag compiler="gnu">-funsafe-math-optimizations</flag>
   <alignment>16</alignment>
   <check name="has_neon"></check>
 </arch>
 
 <arch name="32">
-  <flag>m32</flag>
+  <flag compiler="gnu">-m32</flag>
 </arch>
 
 <arch name="64">
@@ -31,7 +31,7 @@
       <param>0x80000001</param> <!-- cpuid operation -->
       <param>29</param>         <!-- bit shift -->
   </check>
-  <flag>m64</flag>
+  <flag compiler="gnu">-m64</flag>
 </arch>
 
 <arch name="3dnow">
@@ -40,7 +40,7 @@
       <param>0x80000001</param>
       <param>31</param>
   </check>
-  <flag>m3dnow</flag>
+  <flag compiler="gnu">-m3dnow</flag>
   <alignment>8</alignment>
 </arch>
 
@@ -50,7 +50,7 @@
       <param>0x80000001</param>
       <param>5</param>
   </check>
-  <flag>msse4.2</flag>
+  <flag compiler="gnu">-msse4.2</flag>
   <alignment>16</alignment>
 </arch>
 
@@ -60,7 +60,8 @@
       <param>0x00000001</param>
       <param>23</param>
   </check>
-  <flag>mpopcnt</flag>
+  <flag compiler="gnu">-mpopcnt</flag>
+  <flag compiler="msvc">/arch:AVX</flag>
 </arch>
 
 <arch name="mmx">
@@ -69,7 +70,8 @@
       <param>0x00000001</param>
       <param>23</param>
   </check>
-  <flag>mmmx</flag>
+  <flag compiler="gnu">-mmmx</flag>
+  <flag compiler="msvc">/arch:SSE</flag>
   <alignment>8</alignment>
 </arch>
 
@@ -79,7 +81,8 @@
       <param>0x00000001</param>
       <param>25</param>
   </check>
-  <flag>msse</flag>
+  <flag compiler="gnu">-msse</flag>
+  <flag compiler="msvc">/arch:SSE</flag>
   <environment>_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);</environment>
   <include>xmmintrin.h</include>
   <alignment>16</alignment>
@@ -91,7 +94,8 @@
       <param>0x00000001</param>
       <param>26</param>
   </check>
-  <flag>msse2</flag>
+  <flag compiler="gnu">-msse2</flag>
+  <flag compiler="msvc">/arch:SSE2</flag>
   <alignment>16</alignment>
 </arch>
 
@@ -108,7 +112,8 @@
       <param>0x00000001</param>
       <param>0</param>
   </check>
-  <flag>msse3</flag>
+  <flag compiler="gnu">-msse3</flag>
+  <flag compiler="msvc">/arch:AVX</flag>
   <environment>_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);</environment>
   <include>pmmintrin.h</include>
   <alignment>16</alignment>
@@ -120,7 +125,8 @@
       <param>0x00000001</param>
       <param>9</param>
   </check>
-  <flag>mssse3</flag>
+  <flag compiler="gnu">-mssse3</flag>
+  <flag compiler="msvc">/arch:AVX</flag>
   <alignment>16</alignment>
 </arch>
 
@@ -130,7 +136,7 @@
       <param>0x80000001</param>
       <param>6</param>
   </check>
-  <flag>msse4a</flag>
+  <flag compiler="gnu">-msse4a</flag>
   <alignment>16</alignment>
 </arch>
 
@@ -140,7 +146,8 @@
       <param>0x00000001</param>
       <param>19</param>
   </check>
-  <flag>msse4.1</flag>
+  <flag compiler="gnu">-msse4.1</flag>
+  <flag compiler="msvc">/arch:AVX</flag>
   <alignment>16</alignment>
 </arch>
 
@@ -150,7 +157,8 @@
       <param>0x00000001</param>
       <param>20</param>
   </check>
-  <flag>msse4.2</flag>
+  <flag compiler="gnu">-msse4.2</flag>
+  <flag compiler="msvc">/arch:AVX</flag>
   <alignment>16</alignment>
 </arch>
 
@@ -168,7 +176,8 @@
   </check>
   <!-- check to see that the OS has enabled AVX -->
   <check name="get_avx_enabled"></check>
-  <flag>mavx</flag>
+  <flag compiler="gnu">-mavx</flag>
+  <flag compiler="msvc">/arch:AVX</flag>
   <alignment>32</alignment>
 </arch>
 
diff --git a/volk/gen/volk_arch_defs.py b/volk/gen/volk_arch_defs.py
index d64f8def2..8ac99b338 100644
--- a/volk/gen/volk_arch_defs.py
+++ b/volk/gen/volk_arch_defs.py
@@ -33,7 +33,15 @@ class arch_class:
             except: setattr(self, key, failval)
         self.checks = checks
         assert(self.name)
-        self.flags = flags
+        self._flags = flags
+
+    def is_supported(self, compiler):
+        if not self._flags.keys(): return True
+        return compiler in self._flags.keys()
+
+    def get_flags(self, compiler):
+        try: return self._flags[compiler]
+        except KeyError: return list()
 
     def __repr__(self): return self.name
 
@@ -64,17 +72,18 @@ for arch_xml in archs_xml:
             val = arch_xml.getElementsByTagName(name)[0].firstChild.data
             kwargs[name] = val
         except: pass
-    checks = []
+    checks = list()
     for check_xml in arch_xml.getElementsByTagName("check"):
         name = check_xml.attributes["name"].value
         params = list()
         for param_xml in check_xml.getElementsByTagName("param"):
             params.append(param_xml.firstChild.data)
         checks.append([name, params])
-    flags = []
+    flags = dict()
     for flag_xml in arch_xml.getElementsByTagName("flag"):
-        flags.append(flag_xml.firstChild.data)
-    
+        name = flag_xml.attributes["compiler"].value
+        if not flags.has_key(name): flags[name] = list()
+        flags[name].append(flag_xml.firstChild.data)
     register_arch(flags=flags, checks=checks, **kwargs)
 
 if __name__ == '__main__':
diff --git a/volk/gen/volk_compile_utils.py b/volk/gen/volk_compile_utils.py
index a62c5657e..cd9f9f0a9 100644
--- a/volk/gen/volk_compile_utils.py
+++ b/volk/gen/volk_compile_utils.py
@@ -19,21 +19,11 @@ import optparse
 import volk_arch_defs
 import volk_machine_defs
 
-def format_flag(flag, compiler):
-    if compiler == 'msvc' and flag in ('mmmx', 'msse'):
-        return '/arch:SSE'
-    if compiler == 'msvc' and flag in ('msse2',):
-        return '/arch:SSE2'
-    if compiler == 'msvc' and flag in ('msse3', 'mssse3', 'msse4.1', 'msse4.2', 'mpopcnt', 'mavx'):
-        return '/arch:AVX'
-    #otherwise its a gcc compiler
-    return '-' + flag
-
 def do_arch_flags_list(compiler):
     output = list()
     for arch in volk_arch_defs.archs:
-        fields = [arch.name]
-        fields += list(map(lambda f: format_flag(f, compiler), arch.flags))
+        if not arch.is_supported(compiler): continue
+        fields = [arch.name] + arch.get_flags(compiler)
         output.append(','.join(fields))
     print ';'.join(output)
 
@@ -49,8 +39,7 @@ def do_machine_flags_list(compiler, machine_name):
     output = list()
     machine = volk_machine_defs.machine_dict[machine_name]
     for arch in machine.archs:
-        for flag in arch.flags:
-            output.append(format_flag(flag, compiler))
+        output.extend(arch.get_flags(compiler))
     print ' '.join(output)
 
 def main():
-- 
cgit