Vector Tests (rust-lang#357)

* fix bugs

* add test

* fix test

Author: tgymnich

enzyme/Enzyme/AdjointGenerator.h

@@ -1207,8 +1207,9 @@ class AdjointGenerator
 
       Value *prediff =
           gutils->isConstantValue(orig_vector)
-              ? diffe(orig_vector, Builder2)
-              : ConstantVector::getNullValue(orig_vector->getType());
+              ? ConstantVector::getNullValue(orig_vector->getType())
+              : diffe(orig_vector, Builder2);
+
       auto dindex = Builder2.CreateInsertElement(
           prediff, diff_inserted, gutils->getNewFromOriginal(orig_index));
       setDiffe(&IEI, dindex, Builder2);
@@ -1275,7 +1276,6 @@ class AdjointGenerator
 
       Value *orig_vector1 = SVI.getOperand(0);
       Value *orig_vector2 = SVI.getOperand(1);
-      Value *orig_mask = SVI.getOperand(0);
 
       auto diffe_vector1 =
           gutils->isConstantValue(orig_vector1)
@@ -1286,8 +1286,13 @@ class AdjointGenerator
               ? ConstantVector::getNullValue(orig_vector2->getType())
               : diffe(orig_vector2, Builder2);
 
-      auto diffe = Builder2.CreateShuffleVector(
-          diffe_vector1, diffe_vector2, gutils->getNewFromOriginal(orig_mask));
+#if LLVM_VERSION_MAJOR >= 11
+      auto diffe = Builder2.CreateShuffleVector(diffe_vector1, diffe_vector2,
+                                                SVI.getShuffleMaskForBitcode());
+#else
+      auto diffe = Builder2.CreateShuffleVector(diffe_vector1, diffe_vector2,
+                                                SVI.getOperand(2));
+#endif
 
       setDiffe(&SVI, diffe, Builder2);
       return;
@@ -1465,8 +1470,8 @@ class AdjointGenerator
 
       Value *prediff =
           gutils->isConstantValue(orig_agg)
-              ? diffe(orig_agg, Builder2)
-              : ConstantAggregate::getNullValue(orig_agg->getType());
+              ? ConstantAggregate::getNullValue(orig_agg->getType())
+              : diffe(orig_agg, Builder2);
       auto dindex =
           Builder2.CreateInsertValue(prediff, diff_inserted, IVI.getIndices());
       setDiffe(&IVI, dindex, Builder2);

enzyme/test/Enzyme/ForwardMode/square_array.ll

@@ -27,6 +27,7 @@ declare { double, double } @__enzyme_fwddiff(i8*, ...)
 ; CHECK-NEXT:   %2 = fmul fast double %1, %x
 ; CHECK-NEXT:   %3 = fmul fast double %"x'", %mul
 ; CHECK-NEXT:   %4 = fadd fast double %2, %3
-; CHECK-NEXT:   %5 = insertvalue { double, double } zeroinitializer, double %4, 1
-; CHECK-NEXT:   ret { double, double } %5
+; CHECK-NEXT:   %5 = insertvalue { double, double } zeroinitializer, double %1, 0
+; CHECK-NEXT:   %6 = insertvalue { double, double } %5, double %4, 1
+; CHECK-NEXT:   ret { double, double } %6
 ; CHECK-NEXT: }

enzyme/test/Enzyme/ForwardMode/vecsquare.ll

@@ -0,0 +1,42 @@
+; RUN: %opt < %s %loadEnzyme -enzyme -enzyme-preopt=false -mem2reg -simplifycfg -early-cse -instcombine -S | FileCheck %s
+
+declare {float, float, float} @__enzyme_fwddiff({float, float, float} (<4 x float>)*, <4 x float>, <4 x float>)
+
+define {float, float, float} @square(<4 x float> %x) {
+entry:
+  %vec = insertelement <4 x float> %x, float 1.0, i32 3
+  %sq = fmul <4 x float> %x, %x
+  %cb = fmul <4 x float> %sq, %x          
+  %id = shufflevector <4 x float> %sq, <4 x float> %cb, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
+  %res1 = extractelement <4 x float> %id, i32 1
+  %res2 = extractelement <4 x float> %id, i32 2
+  %res3 = extractelement <4 x float> %id, i32 3
+  %agg1 = insertvalue {float, float, float} undef, float %res1, 0
+  %agg2 = insertvalue {float, float, float} %agg1, float %res2, 1
+  %agg3 = insertvalue {float, float, float} %agg2, float %res3, 2
+  ret {float, float, float} %agg3
+}
+
+define {float, float, float} @dsquare(<4 x float> %x) {
+entry:
+  %call = tail call {float, float, float} @__enzyme_fwddiff({float, float, float} (<4 x float>)* @square, <4 x float> %x, <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>)
+  ret {float, float, float} %call
+}
+
+
+; CHECK: define internal { float, float, float } @fwddiffesquare(<4 x float> %x, <4 x float> %"x'")
+; CHECK-NEXT: entry:
+; CHECK-NEXT:    %sq = fmul <4 x float> %x, %x
+; CHECK-NEXT:    %0 = fmul fast <4 x float> %"x'", %x
+; CHECK-NEXT:    %1 = fadd fast <4 x float> %0, %0
+; CHECK-NEXT:    %2 = fmul fast <4 x float> %1, %x
+; CHECK-NEXT:    %3 = fmul fast <4 x float> %sq, %"x'"
+; CHECK-NEXT:    %4 = fadd fast <4 x float> %2, %3
+; CHECK-NEXT:    %5 = extractelement <4 x float> %1, i32 1
+; CHECK-NEXT:    %6 = extractelement <4 x float> %4, i32 0
+; CHECK-NEXT:    %7 = extractelement <4 x float> %4, i32 1
+; CHECK-NEXT:    %8 = insertvalue { float, float, float } zeroinitializer, float %5, 0
+; CHECK-NEXT:    %9 = insertvalue { float, float, float } %8, float %6, 1
+; CHECK-NEXT:    %10 = insertvalue { float, float, float } %9, float %7, 2
+; CHECK-NEXT:    ret { float, float, float } %10
+; CHECK-NEXT:  }