[LLVM][ConstProp] Enable intrinsic simplifications for vector ConstantInt based operands. (llvm#159358)

paulwalker-arm · web-flow · commit d17d32a6cb6d · 2025-10-17T13:43:00.000+01:00
Simplifcation of vector.reduce intrinsics are prevented by an early
bailout for ConstantInt base operands. This PR removes the bailout and
updates the tests to show matching output when
-use-constant-int-for-*-splat is used.
diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp
@@ -2177,16 +2177,13 @@ Constant *constantFoldVectorReduce(Intrinsic::ID IID, Constant *Op) {
     return PoisonValue::get(VT->getElementType());
 
   // TODO: Handle undef.
-  if (!isa<ConstantVector>(Op) && !isa<ConstantDataVector>(Op))
-    return nullptr;
-
-  auto *EltC = dyn_cast<ConstantInt>(Op->getAggregateElement(0U));
+  auto *EltC = dyn_cast_or_null<ConstantInt>(Op->getAggregateElement(0U));
   if (!EltC)
     return nullptr;
 
   APInt Acc = EltC->getValue();
   for (unsigned I = 1, E = VT->getNumElements(); I != E; I++) {
-    if (!(EltC = dyn_cast<ConstantInt>(Op->getAggregateElement(I))))
+    if (!(EltC = dyn_cast_or_null<ConstantInt>(Op->getAggregateElement(I))))
       return nullptr;
     const APInt &X = EltC->getValue();
     switch (IID) {
@@ -3059,35 +3056,25 @@ static Constant *ConstantFoldScalarCall1(StringRef Name,
       Val = Val | Val << 1;
       return ConstantInt::get(Ty, Val);
     }
-
-    default:
-      return nullptr;
     }
   }
 
-  switch (IntrinsicID) {
-  default: break;
-  case Intrinsic::vector_reduce_add:
-  case Intrinsic::vector_reduce_mul:
-  case Intrinsic::vector_reduce_and:
-  case Intrinsic::vector_reduce_or:
-  case Intrinsic::vector_reduce_xor:
-  case Intrinsic::vector_reduce_smin:
-  case Intrinsic::vector_reduce_smax:
-  case Intrinsic::vector_reduce_umin:
-  case Intrinsic::vector_reduce_umax:
-    if (Constant *C = constantFoldVectorReduce(IntrinsicID, Operands[0]))
-      return C;
-    break;
-  }
-
-  // Support ConstantVector in case we have an Undef in the top.
-  if (isa<ConstantVector>(Operands[0]) ||
-      isa<ConstantDataVector>(Operands[0]) ||
-      isa<ConstantAggregateZero>(Operands[0])) {
+  if (Operands[0]->getType()->isVectorTy()) {
     auto *Op = cast<Constant>(Operands[0]);
     switch (IntrinsicID) {
     default: break;
+    case Intrinsic::vector_reduce_add:
+    case Intrinsic::vector_reduce_mul:
+    case Intrinsic::vector_reduce_and:
+    case Intrinsic::vector_reduce_or:
+    case Intrinsic::vector_reduce_xor:
+    case Intrinsic::vector_reduce_smin:
+    case Intrinsic::vector_reduce_smax:
+    case Intrinsic::vector_reduce_umin:
+    case Intrinsic::vector_reduce_umax:
+      if (Constant *C = constantFoldVectorReduce(IntrinsicID, Operands[0]))
+        return C;
+      break;
     case Intrinsic::x86_sse_cvtss2si:
     case Intrinsic::x86_sse_cvtss2si64:
     case Intrinsic::x86_sse2_cvtsd2si:
@@ -3116,10 +3103,15 @@ static Constant *ConstantFoldScalarCall1(StringRef Name,
     case Intrinsic::wasm_alltrue:
       // Check each element individually
       unsigned E = cast<FixedVectorType>(Op->getType())->getNumElements();
-      for (unsigned I = 0; I != E; ++I)
-        if (Constant *Elt = Op->getAggregateElement(I))
-          if (Elt->isZeroValue())
-            return ConstantInt::get(Ty, 0);
+      for (unsigned I = 0; I != E; ++I) {
+        Constant *Elt = Op->getAggregateElement(I);
+        // Return false as soon as we find a non-true element.
+        if (Elt && Elt->isZeroValue())
+          return ConstantInt::get(Ty, 0);
+        // Bail as soon as we find an element we cannot prove to be true.
+        if (!Elt || !isa<ConstantInt>(Elt))
+          return nullptr;
+      }
 
       return ConstantInt::get(Ty, 1);
     }
diff --git a/llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/any_all_true.ll b/llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/any_all_true.ll
@@ -1,6 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
 
 ; RUN: opt -passes=instsimplify -S < %s | FileCheck %s
+; RUN: opt -passes=instsimplify -use-constant-int-for-fixed-length-splat -S < %s | FileCheck %s
 
 ; Test that intrinsics wasm call are constant folded
 
diff --git a/llvm/test/Transforms/InstSimplify/ConstProp/bitcount.ll b/llvm/test/Transforms/InstSimplify/ConstProp/bitcount.ll
@@ -1,5 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -passes=instsimplify -S | FileCheck %s
+; RUN: opt < %s -passes=instsimplify -use-constant-int-for-fixed-length-splat -use-constant-int-for-scalable-splat -S | FileCheck %s
 
 declare i31 @llvm.ctpop.i31(i31 %val)
 declare i32 @llvm.cttz.i32(i32 %val, i1)
@@ -120,6 +121,22 @@ define <2 x i31> @ctpop_vector() {
   ret <2 x i31> %x
 }
 
+define <2 x i31> @ctpop_vector_splat_v2i31() {
+; CHECK-LABEL: @ctpop_vector_splat_v2i31(
+; CHECK-NEXT:    ret <2 x i31> splat (i31 1)
+;
+  %x = call <2 x i31> @llvm.ctpop.v2i31(<2 x i31> splat(i31 16))
+  ret <2 x i31> %x
+}
+
+define <vscale x 2 x i31> @ctpop_vector_splat_nxv2i31() {
+; CHECK-LABEL: @ctpop_vector_splat_nxv2i31(
+; CHECK-NEXT:    ret <vscale x 2 x i31> splat (i31 1)
+;
+  %x = call <vscale x 2 x i31> @llvm.ctpop.nxv2i31(<vscale x 2 x i31> splat(i31 16))
+  ret <vscale x 2 x i31> %x
+}
+
 define <2 x i31> @ctpop_vector_undef() {
 ; CHECK-LABEL: @ctpop_vector_undef(
 ; CHECK-NEXT:    ret <2 x i31> zeroinitializer
diff --git a/llvm/test/Transforms/InstSimplify/ConstProp/bitreverse.ll b/llvm/test/Transforms/InstSimplify/ConstProp/bitreverse.ll
@@ -0,0 +1,51 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt < %s -passes=instsimplify -S | FileCheck %s
+; RUN: opt < %s -passes=instsimplify -use-constant-int-for-fixed-length-splat -use-constant-int-for-scalable-splat -S | FileCheck %s
+
+define i16 @W() {
+; CHECK-LABEL: define i16 @W() {
+; CHECK-NEXT:    ret i16 -32768
+;
+  %Z = call i16 @llvm.bitreverse.i16(i16 1)
+  ret i16 %Z
+}
+
+define i32 @X() {
+; CHECK-LABEL: define i32 @X() {
+; CHECK-NEXT:    ret i32 -2147483648
+;
+  %Z = call i32 @llvm.bitreverse.i32(i32 1)
+  ret i32 %Z
+}
+
+define i64 @Y() {
+; CHECK-LABEL: define i64 @Y() {
+; CHECK-NEXT:    ret i64 -9223372036854775808
+;
+  %Z = call i64 @llvm.bitreverse.i64(i64 1)
+  ret i64 %Z
+}
+
+define i80 @Z() {
+; CHECK-LABEL: define i80 @Z() {
+; CHECK-NEXT:    ret i80 23777929115895377691656
+;
+  %Z = call i80 @llvm.bitreverse.i80(i80 76151636403560493650080)
+  ret i80 %Z
+}
+
+define <4 x i32> @bitreverse_splat_v4i32() {
+; CHECK-LABEL: define <4 x i32> @bitreverse_splat_v4i32() {
+; CHECK-NEXT:    ret <4 x i32> splat (i32 -2147483648)
+;
+  %Z = call <4 x i32> @llvm.bitreverse.v4i32(<4 x i32> splat(i32 1))
+  ret <4 x i32> %Z
+}
+
+define <vscale x 4 x i32> @bitreverse_splat_nxv4i32() {
+; CHECK-LABEL: define <vscale x 4 x i32> @bitreverse_splat_nxv4i32() {
+; CHECK-NEXT:    ret <vscale x 4 x i32> splat (i32 -2147483648)
+;
+  %Z = call <vscale x 4 x i32> @llvm.bitreverse.v4i32(<vscale x 4 x i32> splat(i32 1))
+  ret <vscale x 4 x i32> %Z
+}
diff --git a/llvm/test/Transforms/InstSimplify/ConstProp/bswap.ll b/llvm/test/Transforms/InstSimplify/ConstProp/bswap.ll
@@ -2,6 +2,7 @@
 ; bswap should be constant folded when it is passed a constant argument
 
 ; RUN: opt < %s -passes=instsimplify -S | FileCheck %s
+; RUN: opt < %s -passes=instsimplify -use-constant-int-for-fixed-length-splat -use-constant-int-for-scalable-splat -S | FileCheck %s
 
 declare i16 @llvm.bswap.i16(i16)
 
@@ -42,3 +43,19 @@ define i80 @Z() {
   %Z = call i80 @llvm.bswap.i80( i80 76151636403560493650080 )
   ret i80 %Z
 }
+
+define <4 x i32> @bswap_splat_v4i32() {
+; CHECK-LABEL: define <4 x i32> @bswap_splat_v4i32() {
+; CHECK-NEXT:    ret <4 x i32> splat (i32 16777216)
+;
+  %Z = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> splat(i32 1))
+  ret <4 x i32> %Z
+}
+
+define <vscale x 4 x i32> @bswap_splat_nxv4i32() {
+; CHECK-LABEL: define <vscale x 4 x i32> @bswap_splat_nxv4i32() {
+; CHECK-NEXT:    ret <vscale x 4 x i32> splat (i32 16777216)
+;
+  %Z = call <vscale x 4 x i32> @llvm.bswap.v4i32(<vscale x 4 x i32> splat(i32 1))
+  ret <vscale x 4 x i32> %Z
+}
diff --git a/llvm/test/Transforms/InstSimplify/ConstProp/vecreduce.ll b/llvm/test/Transforms/InstSimplify/ConstProp/vecreduce.ll
@@ -1,5 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -passes=instsimplify -S | FileCheck %s
+; RUN: opt < %s -passes=instsimplify -use-constant-int-for-fixed-length-splat -S | FileCheck %s
 
 declare i32 @llvm.vector.reduce.add.v1i32(<1 x i32> %a)
 declare i32 @llvm.vector.reduce.add.v8i32(<8 x i32> %a)
