avx1-i32x4 implementation as sse4-i32x4 with avx target-feature flag

builtins.cpp

@@ -942,11 +942,22 @@ DefineStdlib(SymbolTable *symbolTable, llvm::LLVMContext *ctx, llvm::Module *mod
     case Target::AVX: {
         switch (g->target->getVectorWidth()) {
         case 4:
-            if (runtime32) {
+            if (g->target->getDataTypeWidth() == 32) {
-                EXPORT_MODULE(builtins_bitcode_avx1_i64x4_32bit);
+                if (runtime32) {
-            }
+                    EXPORT_MODULE(builtins_bitcode_sse4_32bit);
-            else {
+                }
-                EXPORT_MODULE(builtins_bitcode_avx1_i64x4_64bit);
+                else {
+                    EXPORT_MODULE(builtins_bitcode_sse4_64bit);
+                }
+            } else if (g->target->getDataTypeWidth() == 64) {
+                if (runtime32) {
+                    EXPORT_MODULE(builtins_bitcode_avx1_i64x4_32bit);
+                }
+                else {
+                    EXPORT_MODULE(builtins_bitcode_avx1_i64x4_64bit);
+                }
+            } else {
+                FATAL("logic error in DefineStdlib");
             }
             break;
         case 8:

ispc.cpp

@@ -169,7 +169,7 @@ static const char *supportedCPUs[] = {
     , "core-avx-i", "core-avx2"
 #endif // LLVM 3.2+
 #if !defined(LLVM_3_1) && !defined(LLVM_3_2) && !defined(LLVM_3_3)
-    , "slm" 
+    , "slm"
 #endif // LLVM 3.4+
 };
 
@@ -191,6 +191,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
     m_tf_attributes(NULL),
 #endif
     m_nativeVectorWidth(-1),
+    m_dataTypeWidth(-1),
     m_vectorWidth(-1),
     m_generatePIC(pic),
     m_maskingIsFree(false),
@@ -308,6 +309,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
         !strcasecmp(isa, "sse2-i32x4")) {
         this->m_isa = Target::SSE2;
         this->m_nativeVectorWidth = 4;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 4;
         this->m_attributes = "+sse,+sse2,-sse3,-sse4a,-ssse3,-popcnt"
 #if defined(LLVM_3_4)
@@ -323,6 +325,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
              !strcasecmp(isa, "sse2-i32x8")) {
         this->m_isa = Target::SSE2;
         this->m_nativeVectorWidth = 4;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 8;
         this->m_attributes = "+sse,+sse2,-sse3,-sse4a,-ssse3,-popcnt"
 #if defined(LLVM_3_4)
@@ -338,11 +341,12 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
              !strcasecmp(isa, "sse4-i32x4")) {
         this->m_isa = Target::SSE4;
         this->m_nativeVectorWidth = 4;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 4;
         // TODO: why not sse42 and popcnt?
         this->m_attributes = "+sse,+sse2,+sse3,-sse4a,+ssse3,-popcnt,+cmov"
 #if defined(LLVM_3_4)
-        ",+sse4.1,-sse4.2"        
+        ",+sse4.1,-sse4.2"
 #else
         ",+sse41,-sse42"
 #endif
@@ -355,10 +359,11 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
              !strcasecmp(isa, "sse4-i32x8")) {
         this->m_isa = Target::SSE4;
         this->m_nativeVectorWidth = 4;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 8;
         this->m_attributes = "+sse,+sse2,+sse3,-sse4a,+ssse3,-popcnt,+cmov"
 #if defined(LLVM_3_4)
-        ",+sse4.1,-sse4.2"        
+        ",+sse4.1,-sse4.2"
 #else
         ",+sse41,-sse42"
 #endif
@@ -369,10 +374,11 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
     else if (!strcasecmp(isa, "sse4-i8x16")) {
         this->m_isa = Target::SSE4;
         this->m_nativeVectorWidth = 16;
+        this->m_dataTypeWidth = 8;
         this->m_vectorWidth = 16;
         this->m_attributes = "+sse,+sse2,+sse3,-sse4a,+ssse3,-popcnt,+cmov"
 #if defined(LLVM_3_4)
-        ",+sse4.1,-sse4.2"        
+        ",+sse4.1,-sse4.2"
 #else
         ",+sse41,-sse42"
 #endif
@@ -383,10 +389,11 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
     else if (!strcasecmp(isa, "sse4-i16x8")) {
         this->m_isa = Target::SSE4;
         this->m_nativeVectorWidth = 8;
+        this->m_dataTypeWidth = 16;
         this->m_vectorWidth = 8;
         this->m_attributes = "+sse,+sse2,+sse3,-sse4a,+ssse3,-popcnt,+cmov"
 #if defined(LLVM_3_4)
-        ",+sse4.1,-sse4.2"        
+        ",+sse4.1,-sse4.2"
 #else
         ",+sse41,-sse42"
 #endif
@@ -457,11 +464,21 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
         this->m_maskingIsFree = false;
         this->m_maskBitCount = 32;
     }
+    else if (!strcasecmp(isa, "avx1-i32x4")) {
+        this->m_isa = Target::AVX;
+        this->m_nativeVectorWidth = 8;
+        this->m_dataTypeWidth = 32;
+        this->m_vectorWidth = 4;
+        this->m_attributes = "+avx,+popcnt,+cmov";
+        this->m_maskingIsFree = false;
+        this->m_maskBitCount = 32;
+    }
     else if (!strcasecmp(isa, "avx") ||
              !strcasecmp(isa, "avx1") ||
              !strcasecmp(isa, "avx1-i32x8")) {
         this->m_isa = Target::AVX;
         this->m_nativeVectorWidth = 8;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 8;
         this->m_attributes = "+avx,+popcnt,+cmov";
         this->m_maskingIsFree = false;
@@ -471,6 +488,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
              !strcasecmp(isa, "avx1-i64x4")) {
         this->m_isa = Target::AVX;
         this->m_nativeVectorWidth = 8;  /* native vector width in terms of floats */
+        this->m_dataTypeWidth = 64;
         this->m_vectorWidth = 4;
         this->m_attributes = "+avx,+popcnt,+cmov";
         this->m_maskingIsFree = false;
@@ -481,6 +499,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
              !strcasecmp(isa, "avx1-i32x16")) {
         this->m_isa = Target::AVX;
         this->m_nativeVectorWidth = 8;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 16;
         this->m_attributes = "+avx,+popcnt,+cmov";
         this->m_maskingIsFree = false;
@@ -490,6 +509,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
              !strcasecmp(isa, "avx1.1-i32x8")) {
         this->m_isa = Target::AVX11;
         this->m_nativeVectorWidth = 8;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 8;
         this->m_attributes = "+avx,+popcnt,+cmov,+f16c"
 #if defined(LLVM_3_4)
@@ -510,6 +530,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
              !strcasecmp(isa, "avx1.1-i32x16")) {
         this->m_isa = Target::AVX11;
         this->m_nativeVectorWidth = 8;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 16;
         this->m_attributes = "+avx,+popcnt,+cmov,+f16c"
 #if defined(LLVM_3_4)
@@ -517,7 +538,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
 #else
         ",+rdrand"
 #endif
-        ;           
+        ;
         this->m_maskingIsFree = false;
         this->m_maskBitCount = 32;
         this->m_hasHalf = true;
@@ -529,6 +550,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
     else if (!strcasecmp(isa, "avx1.1-i64x4")) {
         this->m_isa = Target::AVX11;
         this->m_nativeVectorWidth = 8;  /* native vector width in terms of floats */
+        this->m_dataTypeWidth = 64;
         this->m_vectorWidth = 4;
         this->m_attributes = "+avx,+popcnt,+cmov,+f16c"
 #if defined(LLVM_3_4)
@@ -536,7 +558,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
 #else
         ",+rdrand"
 #endif
-        ;           
+        ;
         this->m_maskingIsFree = false;
         this->m_maskBitCount = 64;
         this->m_hasHalf = true;
@@ -549,6 +571,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
              !strcasecmp(isa, "avx2-i32x8")) {
         this->m_isa = Target::AVX2;
         this->m_nativeVectorWidth = 8;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 8;
         this->m_attributes = "+avx2,+popcnt,+cmov,+f16c"
 #if defined(LLVM_3_4)
@@ -573,6 +596,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
              !strcasecmp(isa, "avx2-i32x16")) {
         this->m_isa = Target::AVX2;
         this->m_nativeVectorWidth = 16;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 16;
         this->m_attributes = "+avx2,+popcnt,+cmov,+f16c"
 #if defined(LLVM_3_4)
@@ -596,6 +620,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
     else if (!strcasecmp(isa, "avx2-i64x4")) {
         this->m_isa = Target::AVX2;
         this->m_nativeVectorWidth = 8;  /* native vector width in terms of floats */
+        this->m_dataTypeWidth = 64;
         this->m_vectorWidth = 4;
         this->m_attributes = "+avx2,+popcnt,+cmov,+f16c"
 #if defined(LLVM_3_4)
@@ -620,6 +645,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
     else if (!strcasecmp(isa, "neon-i8x16")) {
         this->m_isa = Target::NEON8;
         this->m_nativeVectorWidth = 16;
+        this->m_dataTypeWidth = 8;
         this->m_vectorWidth = 16;
         this->m_attributes = "+neon,+fp16";
         this->m_hasHalf = true; // ??
@@ -629,6 +655,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
     else if (!strcasecmp(isa, "neon-i16x8")) {
         this->m_isa = Target::NEON16;
         this->m_nativeVectorWidth = 8;
+        this->m_dataTypeWidth = 16;
         this->m_vectorWidth = 8;
         this->m_attributes = "+neon,+fp16";
         this->m_hasHalf = true; // ??
@@ -639,6 +666,7 @@ Target::Target(const char *arch, const char *cpu, const char *isa, bool pic) :
              !strcasecmp(isa, "neon-i32x4")) {
         this->m_isa = Target::NEON32;
         this->m_nativeVectorWidth = 4;
+        this->m_dataTypeWidth = 32;
         this->m_vectorWidth = 4;
         this->m_attributes = "+neon,+fp16";
         this->m_hasHalf = true; // ??

ispc.h

@@ -253,6 +253,8 @@ public:
 
     int getNativeVectorWidth() const {return m_nativeVectorWidth;}
 
+    int getDataTypeWidth() const {return m_dataTypeWidth;}
+
     int getVectorWidth() const {return m_vectorWidth;}
 
     bool getGeneratePIC() const {return m_generatePIC;}
@@ -319,10 +321,14 @@ private:
 #endif
 
     /** Native vector width of the vector instruction set.  Note that this
-        value is directly derived from the ISA Being used (e.g. it's 4 for
+        value is directly derived from the ISA being used (e.g. it's 4 for
         SSE, 8 for AVX, etc.) */
     int m_nativeVectorWidth;
 
+    /** Data type with in bits. Typically it's 32, but could be 8, 16 or 64.
+        For generic it's -1, which means undefined. */
+    int m_dataTypeWidth;
+
     /** Actual vector width currently being compiled to.  This may be an
         integer multiple of the native vector width, for example if we're
         "doubling up" and compiling 8-wide on a 4-wide SSE system. */