[X86] Add constant folding for AVX512 versions of scalar floating point to integer...

Summary:
We've supported constant folding for sse versions for many years. This patch adds support for the avx512 versions including unsigned with the default rounding mode. We could probably do more with other roundings modes and SAE in the future.

The test cases are largely based on the sse.ll test cases. But I did add some test cases to ensure the unsigned versions don't accept negative values. Also checked the bounds of f64->i32 conversions to make sure unsigned has a larger positive range than signed.

Reviewers: RKSimon, spatel, chandlerc

Reviewed By: RKSimon

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D50553

llvm-svn: 339529

diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp
index c5281c5..f741f66 100644 (file)
--- a/llvm/lib/Analysis/ConstantFolding.cpp
+++ b/llvm/lib/Analysis/ConstantFolding.cpp
@@ -1413,6 +1413,22 @@ bool llvm::canConstantFoldCallTo(ImmutableCallSite CS, const Function *F) {
   case Intrinsic::x86_sse2_cvtsd2si64:
   case Intrinsic::x86_sse2_cvttsd2si:
   case Intrinsic::x86_sse2_cvttsd2si64:
+  case Intrinsic::x86_avx512_vcvtss2si32:
+  case Intrinsic::x86_avx512_vcvtss2si64:
+  case Intrinsic::x86_avx512_cvttss2si:
+  case Intrinsic::x86_avx512_cvttss2si64:
+  case Intrinsic::x86_avx512_vcvtsd2si32:
+  case Intrinsic::x86_avx512_vcvtsd2si64:
+  case Intrinsic::x86_avx512_cvttsd2si:
+  case Intrinsic::x86_avx512_cvttsd2si64:
+  case Intrinsic::x86_avx512_vcvtss2usi32:
+  case Intrinsic::x86_avx512_vcvtss2usi64:
+  case Intrinsic::x86_avx512_cvttss2usi:
+  case Intrinsic::x86_avx512_cvttss2usi64:
+  case Intrinsic::x86_avx512_vcvtsd2usi32:
+  case Intrinsic::x86_avx512_vcvtsd2usi64:
+  case Intrinsic::x86_avx512_cvttsd2usi:
+  case Intrinsic::x86_avx512_cvttsd2usi64:
     return true;
   default:
     return false;
@@ -1553,7 +1569,7 @@ Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), double V,
 /// result. Returns null if the conversion cannot be performed, otherwise
 /// returns the Constant value resulting from the conversion.
 Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,
-                                      Type *Ty) {
+                                      Type *Ty, bool IsSigned) {
   // All of these conversion intrinsics form an integer of at most 64bits.
   unsigned ResultWidth = Ty->getIntegerBitWidth();
   assert(ResultWidth <= 64 &&
@@ -1565,11 +1581,11 @@ Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,
                                               : APFloat::rmNearestTiesToEven;
   APFloat::opStatus status =
       Val.convertToInteger(makeMutableArrayRef(UIntVal), ResultWidth,
-                           /*isSigned=*/true, mode, &isExact);
+                           IsSigned, mode, &isExact);
   if (status != APFloat::opOK &&
       (!roundTowardZero || status != APFloat::opInexact))
     return nullptr;
-  return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
+  return ConstantInt::get(Ty, UIntVal, IsSigned);
 }
 
 double getValueAsDouble(ConstantFP *Op) {
@@ -1849,7 +1865,8 @@ Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID, Type *Ty,
         if (ConstantFP *FPOp =
                 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
           return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
-                                             /*roundTowardZero=*/false, Ty);
+                                             /*roundTowardZero=*/false, Ty,
+                                             /*IsSigned*/true);
         break;
       case Intrinsic::x86_sse_cvttss2si:
       case Intrinsic::x86_sse_cvttss2si64:
@@ -1858,7 +1875,8 @@ Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID, Type *Ty,
         if (ConstantFP *FPOp =
                 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
           return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
-                                             /*roundTowardZero=*/true, Ty);
+                                             /*roundTowardZero=*/true, Ty,
+                                             /*IsSigned*/true);
         break;
       }
     }
@@ -1983,6 +2001,59 @@ Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID, Type *Ty,
 
       return nullptr;
     }
+
+    // Support ConstantVector in case we have an Undef in the top.
+    if ((isa<ConstantVector>(Operands[0]) ||
+         isa<ConstantDataVector>(Operands[0])) &&
+        // Check for default rounding mode.
+        // FIXME: Support other rounding modes?
+        isa<ConstantInt>(Operands[1]) &&
+        cast<ConstantInt>(Operands[1])->getValue() == 4) {
+      auto *Op = cast<Constant>(Operands[0]);
+      switch (IntrinsicID) {
+      default: break;
+      case Intrinsic::x86_avx512_vcvtss2si32:
+      case Intrinsic::x86_avx512_vcvtss2si64:
+      case Intrinsic::x86_avx512_vcvtsd2si32:
+      case Intrinsic::x86_avx512_vcvtsd2si64:
+        if (ConstantFP *FPOp =
+                dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
+          return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
+                                             /*roundTowardZero=*/false, Ty,
+                                             /*IsSigned*/true);
+        break;
+      case Intrinsic::x86_avx512_vcvtss2usi32:
+      case Intrinsic::x86_avx512_vcvtss2usi64:
+      case Intrinsic::x86_avx512_vcvtsd2usi32:
+      case Intrinsic::x86_avx512_vcvtsd2usi64:
+        if (ConstantFP *FPOp =
+                dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
+          return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
+                                             /*roundTowardZero=*/false, Ty,
+                                             /*IsSigned*/false);
+        break;
+      case Intrinsic::x86_avx512_cvttss2si:
+      case Intrinsic::x86_avx512_cvttss2si64:
+      case Intrinsic::x86_avx512_cvttsd2si:
+      case Intrinsic::x86_avx512_cvttsd2si64:
+        if (ConstantFP *FPOp =
+                dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
+          return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
+                                             /*roundTowardZero=*/true, Ty,
+                                             /*IsSigned*/true);
+        break;
+      case Intrinsic::x86_avx512_cvttss2usi:
+      case Intrinsic::x86_avx512_cvttss2usi64:
+      case Intrinsic::x86_avx512_cvttsd2usi:
+      case Intrinsic::x86_avx512_cvttsd2usi64:
+        if (ConstantFP *FPOp =
+                dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
+          return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
+                                             /*roundTowardZero=*/true, Ty,
+                                             /*IsSigned*/false);
+        break;
+      }
+    }
     return nullptr;
   }
 

diff --git a/llvm/test/Transforms/ConstProp/avx512.ll b/llvm/test/Transforms/ConstProp/avx512.ll
new file mode 100644 (file)
index 0000000..7043c23
--- /dev/null
+++ b/llvm/test/Transforms/ConstProp/avx512.ll
@@ -0,0 +1,490 @@
+; RUN: opt < %s -constprop -S | FileCheck %s
+; REQUIRES: x86-registered-target
+
+define i1 @test_avx512_cvts_exact() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvts_exact(
+; CHECK-NOT: call
+; CHECK: ret i1 true
+entry:
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2si32(<4 x float> <float 3.0, float undef, float undef, float undef>, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2si64(<4 x float> <float 3.0, float undef, float undef, float undef>, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2si32(<2 x double> <double 7.0, double undef>, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2si64(<2 x double> <double 7.0, double undef>, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %cmp02 = icmp eq i32 %sum02, 10
+  %cmp13 = icmp eq i64 %sum13, 10
+  %b = and i1 %cmp02, %cmp13
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvts_exact_max() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvts_exact_max(
+; CHECK-NOT: call
+; CHECK: ret i1 true
+entry:
+  %i0 = call i32 @llvm.x86.avx512.vcvtsd2si32(<2 x double> <double 2147483647.0, double undef>, i32 4) nounwind
+  %b = icmp eq i32 %i0, 2147483647
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvts_exact_max_p1() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvts_exact_max_p1(
+; CHECK: call
+entry:
+  %i0 = call i32 @llvm.x86.avx512.vcvtsd2si32(<2 x double> <double 2147483648.0, double undef>, i32 4) nounwind
+  %b = icmp eq i32 %i0, 2147483648
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvts_exact_neg_max() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvts_exact_neg_max(
+; CHECK-NOT: call
+; CHECK: ret i1 true
+entry:
+  %i0 = call i32 @llvm.x86.avx512.vcvtsd2si32(<2 x double> <double -2147483648.0, double undef>, i32 4) nounwind
+  %b = icmp eq i32 %i0, -2147483648
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvts_exact_neg_max_p1() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvts_exact_neg_max_p1(
+; CHECK: call
+entry:
+  %i0 = call i32 @llvm.x86.avx512.vcvtsd2si32(<2 x double> <double -2147483649.0, double undef>, i32 4) nounwind
+  %b = icmp eq i32 %i0, -2147483649
+  ret i1 %b
+}
+
+; Inexact values should not fold as they are dependent on rounding mode
+define i1 @test_avx512_cvts_inexact() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvts_inexact(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2si32(<4 x float> <float 1.75, float undef, float undef, float undef>, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2si64(<4 x float> <float 1.75, float undef, float undef, float undef>, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2si32(<2 x double> <double 1.75, double undef>, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2si64(<2 x double> <double 1.75, double undef>, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %cmp02 = icmp eq i32 %sum02, 4
+  %cmp13 = icmp eq i64 %sum13, 4
+  %b = and i1 %cmp02, %cmp13
+  ret i1 %b
+}
+
+; FLT_MAX/DBL_MAX should not fold
+define i1 @test_avx512_cvts_max() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvts_max(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2139095039, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9218868437227405311, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2si32(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2si64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2si32(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2si64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+; INF should not fold
+define i1 @test_avx512_cvts_inf() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvts_inf(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2139095040, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9218868437227405312, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2si32(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2si64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2si32(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2si64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+; NAN should not fold
+define i1 @test_avx512_cvts_nan() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvts_nan(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2143289344, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9221120237041090560, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2si32(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2si64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2si32(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2si64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvtts_exact() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtts_exact(
+; CHECK-NOT: call
+; CHECK: ret i1 true
+entry:
+  %i0 = tail call i32 @llvm.x86.avx512.cvttss2si(<4 x float> <float 3.0, float undef, float undef, float undef>, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.cvttss2si64(<4 x float> <float 3.0, float undef, float undef, float undef>, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.cvttsd2si(<2 x double> <double 7.0, double undef>, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.cvttsd2si64(<2 x double> <double 7.0, double undef>, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %cmp02 = icmp eq i32 %sum02, 10
+  %cmp13 = icmp eq i64 %sum13, 10
+  %b = and i1 %cmp02, %cmp13
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvtts_inexact() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtts_inexact(
+; CHECK-NOT: call
+; CHECK: ret i1 true
+entry:
+  %i0 = tail call i32 @llvm.x86.avx512.cvttss2si(<4 x float> <float 1.75, float undef, float undef, float undef>, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.cvttss2si64(<4 x float> <float 1.75, float undef, float undef, float undef>, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.cvttsd2si(<2 x double> <double 1.75, double undef>, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.cvttsd2si64(<2 x double> <double 1.75, double undef>, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %cmp02 = icmp eq i32 %sum02, 2
+  %cmp13 = icmp eq i64 %sum13, 2
+  %b = and i1 %cmp02, %cmp13
+  ret i1 %b
+}
+
+; FLT_MAX/DBL_MAX should not fold
+define i1 @test_avx512_cvtts_max() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtts_max(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2139095039, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9218868437227405311, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.cvttss2si(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.cvttss2si64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.cvttsd2si(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.cvttsd2si64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+; INF should not fold
+define i1 @test_avx512_cvtts_inf() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtts_inf(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2139095040, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9218868437227405312, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.cvttss2si(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.cvttss2si64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.cvttsd2si(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.cvttsd2si64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+; NAN should not fold
+define i1 @test_avx512_cvtts_nan() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtts_nan(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2143289344, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9221120237041090560, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.cvttss2si(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.cvttss2si64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.cvttsd2si(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.cvttsd2si64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvtu_exact() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtu_exact(
+; CHECK-NOT: call
+; CHECK: ret i1 true
+entry:
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2usi32(<4 x float> <float 3.0, float undef, float undef, float undef>, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2usi64(<4 x float> <float 3.0, float undef, float undef, float undef>, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2usi32(<2 x double> <double 7.0, double undef>, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2usi64(<2 x double> <double 7.0, double undef>, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %cmp02 = icmp eq i32 %sum02, 10
+  %cmp13 = icmp eq i64 %sum13, 10
+  %b = and i1 %cmp02, %cmp13
+  ret i1 %b
+}
+
+; Negative values should not fold as they can't be represented in an unsigned int.
+define i1 @test_avx512_cvtu_neg() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtu_neg(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2usi32(<4 x float> <float -3.0, float undef, float undef, float undef>, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2usi64(<4 x float> <float -3.0, float undef, float undef, float undef>, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2usi32(<2 x double> <double -7.0, double undef>, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2usi64(<2 x double> <double -7.0, double undef>, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %cmp02 = icmp eq i32 %sum02, -10
+  %cmp13 = icmp eq i64 %sum13, -10
+  %b = and i1 %cmp02, %cmp13
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvtu_exact_max() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtu_exact_max(
+; CHECK-NOT: call
+; CHECK: ret i1 true
+entry:
+  %i0 = call i32 @llvm.x86.avx512.vcvtsd2usi32(<2 x double> <double 4294967295.0, double undef>, i32 4) nounwind
+  %b = icmp eq i32 %i0, 4294967295
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvtu_exact_max_p1() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtu_exact_max_p1(
+; CHECK: call
+entry:
+  %i0 = call i32 @llvm.x86.avx512.vcvtsd2usi32(<2 x double> <double 4294967296.0, double undef>, i32 4) nounwind
+  %b = icmp eq i32 %i0, 4294967296
+  ret i1 %b
+}
+
+; Inexact values should not fold as they are dependent on rounding mode
+define i1 @test_avx512_cvtu_inexact() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtu_inexact(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2usi32(<4 x float> <float 1.75, float undef, float undef, float undef>, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2usi64(<4 x float> <float 1.75, float undef, float undef, float undef>, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2usi32(<2 x double> <double 1.75, double undef>, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2usi64(<2 x double> <double 1.75, double undef>, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %cmp02 = icmp eq i32 %sum02, 4
+  %cmp13 = icmp eq i64 %sum13, 4
+  %b = and i1 %cmp02, %cmp13
+  ret i1 %b
+}
+
+; FLT_MAX/DBL_MAX should not fold
+define i1 @test_avx512_cvtu_max() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtu_max(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2139095039, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9218868437227405311, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2usi32(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2usi64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2usi32(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2usi64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+; INF should not fold
+define i1 @test_avx512_cvtu_inf() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtu_inf(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2139095040, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9218868437227405312, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2usi32(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2usi64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2usi32(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2usi64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+; NAN should not fold
+define i1 @test_avx512_cvtu_nan() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvtu_nan(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2143289344, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9221120237041090560, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.vcvtss2usi32(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.vcvtss2usi64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.vcvtsd2usi32(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.vcvtsd2usi64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvttu_exact() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvttu_exact(
+; CHECK-NOT: call
+; CHECK: ret i1 true
+entry:
+  %i0 = tail call i32 @llvm.x86.avx512.cvttss2usi(<4 x float> <float 3.0, float undef, float undef, float undef>, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.cvttss2usi64(<4 x float> <float 3.0, float undef, float undef, float undef>, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.cvttsd2usi(<2 x double> <double 7.0, double undef>, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.cvttsd2usi64(<2 x double> <double 7.0, double undef>, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %cmp02 = icmp eq i32 %sum02, 10
+  %cmp13 = icmp eq i64 %sum13, 10
+  %b = and i1 %cmp02, %cmp13
+  ret i1 %b
+}
+
+define i1 @test_avx512_cvttu_inexact() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvttu_inexact(
+; CHECK-NOT: call
+; CHECK: ret i1 true
+entry:
+  %i0 = tail call i32 @llvm.x86.avx512.cvttss2usi(<4 x float> <float 1.75, float undef, float undef, float undef>, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.cvttss2usi64(<4 x float> <float 1.75, float undef, float undef, float undef>, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.cvttsd2usi(<2 x double> <double 1.75, double undef>, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.cvttsd2usi64(<2 x double> <double 1.75, double undef>, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %cmp02 = icmp eq i32 %sum02, 2
+  %cmp13 = icmp eq i64 %sum13, 2
+  %b = and i1 %cmp02, %cmp13
+  ret i1 %b
+}
+
+; FLT_MAX/DBL_MAX should not fold
+define i1 @test_avx512_cvttu_max() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvttu_max(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2139095039, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9218868437227405311, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.cvttss2usi(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.cvttss2usi64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.cvttsd2usi(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.cvttsd2usi64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+; INF should not fold
+define i1 @test_avx512_cvttu_inf() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvttu_inf(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2139095040, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9218868437227405312, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.cvttss2usi(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.cvttss2usi64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.cvttsd2usi(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.cvttsd2usi64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+; NAN should not fold
+define i1 @test_avx512_cvttu_nan() nounwind readnone {
+; CHECK-LABEL: @test_avx512_cvttu_nan(
+; CHECK: call
+; CHECK: call
+; CHECK: call
+; CHECK: call
+entry:
+  %fm = bitcast <4 x i32> <i32 2143289344, i32 undef, i32 undef, i32 undef> to <4 x float>
+  %dm = bitcast <2 x i64> <i64 9221120237041090560, i64 undef> to <2 x double>
+  %i0 = tail call i32 @llvm.x86.avx512.cvttss2usi(<4 x float> %fm, i32 4) nounwind
+  %i1 = tail call i64 @llvm.x86.avx512.cvttss2usi64(<4 x float> %fm, i32 4) nounwind
+  %i2 = call i32 @llvm.x86.avx512.cvttsd2usi(<2 x double> %dm, i32 4) nounwind
+  %i3 = call i64 @llvm.x86.avx512.cvttsd2usi64(<2 x double> %dm, i32 4) nounwind
+  %sum02 = add i32 %i0, %i2
+  %sum13 = add i64 %i1, %i3
+  %sum02.sext = sext i32 %sum02 to i64
+  %b = icmp eq i64 %sum02.sext, %sum13
+  ret i1 %b
+}
+
+declare i32 @llvm.x86.avx512.vcvtss2si32(<4 x float>, i32) nounwind readnone
+declare i32 @llvm.x86.avx512.cvttss2si(<4 x float>, i32) nounwind readnone
+declare i64 @llvm.x86.avx512.vcvtss2si64(<4 x float>, i32) nounwind readnone
+declare i64 @llvm.x86.avx512.cvttss2si64(<4 x float>, i32) nounwind readnone
+declare i32 @llvm.x86.avx512.vcvtsd2si32(<2 x double>, i32) nounwind readnone
+declare i32 @llvm.x86.avx512.cvttsd2si(<2 x double>, i32) nounwind readnone
+declare i64 @llvm.x86.avx512.vcvtsd2si64(<2 x double>, i32) nounwind readnone
+declare i64 @llvm.x86.avx512.cvttsd2si64(<2 x double>, i32) nounwind readnone
+declare i32 @llvm.x86.avx512.vcvtss2usi32(<4 x float>, i32) nounwind readnone
+declare i32 @llvm.x86.avx512.cvttss2usi(<4 x float>, i32) nounwind readnone
+declare i64 @llvm.x86.avx512.vcvtss2usi64(<4 x float>, i32) nounwind readnone
+declare i64 @llvm.x86.avx512.cvttss2usi64(<4 x float>, i32) nounwind readnone
+declare i32 @llvm.x86.avx512.vcvtsd2usi32(<2 x double>, i32) nounwind readnone
+declare i32 @llvm.x86.avx512.cvttsd2usi(<2 x double>, i32) nounwind readnone
+declare i64 @llvm.x86.avx512.vcvtsd2usi64(<2 x double>, i32) nounwind readnone
+declare i64 @llvm.x86.avx512.cvttsd2usi64(<2 x double>, i32) nounwind readnone