Generate FMA instructions with AVX2 (when possible).

Issue #320.

ispc.cpp

@@ -366,7 +366,11 @@ Target::GetTarget(const char *arch, const char *cpu, const char *isa,
         t->isa = Target::AVX2;
         t->nativeVectorWidth = 8;
         t->vectorWidth = 8;
-        t->attributes = "+avx2,+popcnt,+cmov,+f16c,+rdrand";
+        t->attributes = "+avx2,+popcnt,+cmov,+f16c,+rdrand"
+#ifndef LLVM_3_1
+            ",+fma"
+#endif // !LLVM_3_1
+            ;
         t->maskingIsFree = false;
         t->maskBitCount = 32;
         t->hasHalf = true;
@@ -380,7 +384,11 @@ Target::GetTarget(const char *arch, const char *cpu, const char *isa,
         t->isa = Target::AVX2;
         t->nativeVectorWidth = 16;
         t->vectorWidth = 16;
-        t->attributes = "+avx2,+popcnt,+cmov,+f16c,+rdrand";
+        t->attributes = "+avx2,+popcnt,+cmov,+f16c,+rdrand"
+#ifndef LLVM_3_1
+            ",+fma"
+#endif // !LLVM_3_1
+            ;
         t->maskingIsFree = false;
         t->maskBitCount = 32;
         t->hasHalf = true;
@@ -477,6 +485,10 @@ Target::GetTargetMachine() const {
 #else
     std::string featuresString = attributes;
     llvm::TargetOptions options;
+#if !defined(LLVM_3_1)
+    if (g->opt.disableFMA == false)
+        options.AllowFPOpFusion = llvm::FPOpFusion::Fast;
+#endif // !LLVM_3_1
     llvm::TargetMachine *targetMachine = 
         target->createTargetMachine(triple, cpu, featuresString, options,
                                     relocModel);
@@ -627,6 +639,7 @@ Opt::Opt() {
     force32BitAddressing = true;
     unrollLoops = true;
     disableAsserts = false;
+    disableFMA = false;
     disableMaskAllOnOptimizations = false;
     disableHandlePseudoMemoryOps = false;
     disableBlendedMaskedStores = false;

ispc.h

@@ -304,6 +304,10 @@ struct Opt {
         performance in the generated code). */
     bool disableAsserts;
 
+    /** Indicates whether FMA instructions should be disabled (on targets
+        that support them). */
+    bool disableFMA;
+
     /** If enabled, disables the various optimizations that kick in when
         the execution mask can be determined to be "all on" at compile
         time. */

main.cpp

@@ -115,6 +115,7 @@ usage(int ret) {
     printf("    [-O0/-O1]\t\t\t\tSet optimization level (-O1 is default)\n");
     printf("    [--opt=<option>]\t\t\tSet optimization option\n");
     printf("        disable-assertions\t\tRemove assertion statements from final code.\n");
+    printf("        disable-fma\t\t\tDisable 'fused multiply-add' instructions (on targets that support them)\n");
     printf("        disable-loop-unroll\t\tDisable loop unrolling.\n");
     printf("        fast-masked-vload\t\tFaster masked vector loads on SSE (may go past end of array)\n");
     printf("        fast-math\t\t\tPerform non-IEEE-compliant optimizations of numeric expressions\n");
@@ -333,6 +334,8 @@ int main(int Argc, char *Argv[]) {
                 g->opt.disableAsserts = true;
             else if (!strcmp(opt, "disable-loop-unroll"))
                 g->opt.unrollLoops = false;
+            else if (!strcmp(opt, "disable-fma"))
+                g->opt.disableFMA = true;
 
             // These are only used for performance tests of specific
             // optimizations