[Mlir-commits] [mlir] 3a3732c - [mlir][arith] wide integer emulation support for fpto*i ops (#132375)

llvmlistbot at llvm.org llvmlistbot at llvm.org
Thu Mar 27 17:58:59 PDT 2025


Author: egebeysel
Date: 2025-03-27T20:58:56-04:00
New Revision: 3a3732c2527056342051068c3dbaef4e822bd0a6

URL: https://github.com/llvm/llvm-project/commit/3a3732c2527056342051068c3dbaef4e822bd0a6
DIFF: https://github.com/llvm/llvm-project/commit/3a3732c2527056342051068c3dbaef4e822bd0a6.diff

LOG: [mlir][arith]  wide integer emulation support for fpto*i ops (#132375)

Adding wide integer emulation support for `arith.fpto*i` operations. As
the other emulated operations, the upper and lower `N` bits of the `i2N`
integer result are emitted separately.

For the unsigned case we use the following emulation

```c
// example is 64 -> 32 bit emulation, but the implementation is generalized to any 2N -> N case
const double TWO_POW_N = (uint_64_t(1) << N); // 2^N, N is the bitwidth of the widest int supported

// f is a floating-point value representing the input of the fptoui op.
uint32_t hi = (uint32_t)(f / TWO_POW_N);         // Truncates the division result
uint32_t lo = (uint32_t)(f - hi * TWO_POW_N);       // Subtracts to get the lower bits.
```

For the signed case, we defer the emulation of the absolute value to
`fptoui` and handle the sign:

```
fptosi(fp) = sign(fp) * fptoui(abs(fp))
```

The edge cases of `NaNs, +-inf` and overflows/underflows are undefined
behaviour and the resulting numbers are the combination of the lower
bitwidth UB values. These operations also propagate poison values.

Signed-off-by: Ege Beysel <beysel at roofline.ai>

Added: 
    mlir/test/Integration/Dialect/Arith/CPU/test-wide-int-emulation-fptosi-i64.mlir
    mlir/test/Integration/Dialect/Arith/CPU/test-wide-int-emulation-fptoui-i64.mlir

Modified: 
    mlir/lib/Dialect/Arith/Transforms/EmulateWideInt.cpp
    mlir/test/Dialect/Arith/emulate-wide-int.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/lib/Dialect/Arith/Transforms/EmulateWideInt.cpp b/mlir/lib/Dialect/Arith/Transforms/EmulateWideInt.cpp
index 3226b5d99114a..22babe8123617 100644
--- a/mlir/lib/Dialect/Arith/Transforms/EmulateWideInt.cpp
+++ b/mlir/lib/Dialect/Arith/Transforms/EmulateWideInt.cpp
@@ -17,6 +17,7 @@
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/TypeUtilities.h"
 #include "mlir/Transforms/DialectConversion.h"
+#include "llvm/ADT/APFloat.h"
 #include "llvm/ADT/APInt.h"
 #include "llvm/Support/FormatVariadic.h"
 #include "llvm/Support/MathExtras.h"
@@ -1008,6 +1009,128 @@ struct ConvertUIToFP final : OpConversionPattern<arith::UIToFPOp> {
   }
 };
 
+//===----------------------------------------------------------------------===//
+// ConvertFPToSI
+//===----------------------------------------------------------------------===//
+
+struct ConvertFPToSI final : OpConversionPattern<arith::FPToSIOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(arith::FPToSIOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    // Get the input float type.
+    Value inFp = adaptor.getIn();
+    Type fpTy = inFp.getType();
+
+    Type intTy = op.getType();
+
+    auto newTy = getTypeConverter()->convertType<VectorType>(intTy);
+    if (!newTy)
+      return rewriter.notifyMatchFailure(
+          loc, llvm::formatv("unsupported type: {}", intTy));
+
+    // Work on the absolute value and then convert the result to signed integer.
+    // Defer absolute value to fptoui. If minSInt < fp < maxSInt, i.e. if the fp
+    // is representable in signed i2N, emits the correct result. Else, the
+    // result is UB.
+
+    TypedAttr zeroAttr = rewriter.getZeroAttr(fpTy);
+    Value zeroCst = rewriter.create<arith::ConstantOp>(loc, zeroAttr);
+    Value zeroCstInt = createScalarOrSplatConstant(rewriter, loc, intTy, 0);
+
+    // Get the absolute value. One could have used math.absf here, but that
+    // introduces an extra dependency.
+    Value isNeg = rewriter.create<arith::CmpFOp>(loc, arith::CmpFPredicate::OLT,
+                                                 inFp, zeroCst);
+    Value negInFp = rewriter.create<arith::NegFOp>(loc, inFp);
+
+    Value absVal = rewriter.create<arith::SelectOp>(loc, isNeg, negInFp, inFp);
+
+    // Defer the absolute value to fptoui.
+    Value res = rewriter.create<arith::FPToUIOp>(loc, intTy, absVal);
+
+    // Negate the value if < 0 .
+    Value neg = rewriter.create<arith::SubIOp>(loc, zeroCstInt, res);
+
+    rewriter.replaceOpWithNewOp<arith::SelectOp>(op, isNeg, neg, res);
+    return success();
+  }
+};
+
+//===----------------------------------------------------------------------===//
+// ConvertFPToUI
+//===----------------------------------------------------------------------===//
+
+struct ConvertFPToUI final : OpConversionPattern<arith::FPToUIOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(arith::FPToUIOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    // Get the input float type.
+    Value inFp = adaptor.getIn();
+    Type fpTy = inFp.getType();
+
+    Type intTy = op.getType();
+    auto newTy = getTypeConverter()->convertType<VectorType>(intTy);
+    if (!newTy)
+      return rewriter.notifyMatchFailure(
+          loc, llvm::formatv("unsupported type: {}", intTy));
+    unsigned newBitWidth = newTy.getElementTypeBitWidth();
+
+    Type newHalfType = IntegerType::get(inFp.getContext(), newBitWidth);
+    if (auto vecType = dyn_cast<VectorType>(fpTy))
+      newHalfType = VectorType::get(vecType.getShape(), newHalfType);
+
+    // The resulting integer has the upper part and the lower part. This would
+    // be interpreted as 2^N * high + low, where N is the bitwidth. Therefore,
+    // to calculate the higher part, we emit resHigh = fptoui(fp/2^N). For the
+    // lower part, we emit fptoui(fp - resHigh * 2^N). The special cases of
+    // overflows including +-inf, NaNs and negative numbers are UB.
+
+    const llvm::fltSemantics &fSemantics =
+        cast<FloatType>(getElementTypeOrSelf(fpTy)).getFloatSemantics();
+
+    auto powBitwidth = llvm::APFloat(fSemantics);
+    // If the integer does not fit the floating point number, we set the
+    // powBitwidth to inf. This ensures that the upper part is set
+    // correctly to 0. The opStatus inexact here only occurs when we have an
+    // overflow, since the number is always a power of two.
+    if (powBitwidth.convertFromAPInt(APInt(newBitWidth * 2, 1).shl(newBitWidth),
+                                     false, llvm::RoundingMode::TowardZero) ==
+        llvm::detail::opStatus::opInexact)
+      powBitwidth = llvm::APFloat::getInf(fSemantics);
+
+    TypedAttr powBitwidthAttr =
+        FloatAttr::get(getElementTypeOrSelf(fpTy), powBitwidth);
+    if (auto vecType = dyn_cast<VectorType>(fpTy))
+      powBitwidthAttr = SplatElementsAttr::get(vecType, powBitwidthAttr);
+    Value powBitwidthFloatCst =
+        rewriter.create<arith::ConstantOp>(loc, powBitwidthAttr);
+
+    Value fpDivPowBitwidth =
+        rewriter.create<arith::DivFOp>(loc, inFp, powBitwidthFloatCst);
+    Value resHigh =
+        rewriter.create<arith::FPToUIOp>(loc, newHalfType, fpDivPowBitwidth);
+    // Calculate fp - resHigh * 2^N by getting the remainder of the division
+    Value remainder =
+        rewriter.create<arith::RemFOp>(loc, inFp, powBitwidthFloatCst);
+    Value resLow =
+        rewriter.create<arith::FPToUIOp>(loc, newHalfType, remainder);
+
+    Value high = appendX1Dim(rewriter, loc, resHigh);
+    Value low = appendX1Dim(rewriter, loc, resLow);
+
+    Value resultVec = constructResultVector(rewriter, loc, newTy, {low, high});
+
+    rewriter.replaceOp(op, resultVec);
+    return success();
+  }
+};
+
 //===----------------------------------------------------------------------===//
 // ConvertTruncI
 //===----------------------------------------------------------------------===//
@@ -1184,5 +1307,6 @@ void arith::populateArithWideIntEmulationPatterns(
       ConvertIndexCastIntToIndex<arith::IndexCastUIOp>,
       ConvertIndexCastIndexToInt<arith::IndexCastOp, arith::ExtSIOp>,
       ConvertIndexCastIndexToInt<arith::IndexCastUIOp, arith::ExtUIOp>,
-      ConvertSIToFP, ConvertUIToFP>(typeConverter, patterns.getContext());
+      ConvertSIToFP, ConvertUIToFP, ConvertFPToUI, ConvertFPToSI>(
+      typeConverter, patterns.getContext());
 }

diff  --git a/mlir/test/Dialect/Arith/emulate-wide-int.mlir b/mlir/test/Dialect/Arith/emulate-wide-int.mlir
index 52da80ce26a73..936050cddb676 100644
--- a/mlir/test/Dialect/Arith/emulate-wide-int.mlir
+++ b/mlir/test/Dialect/Arith/emulate-wide-int.mlir
@@ -1046,3 +1046,112 @@ func.func @sitofp_i64_f64_vector(%a : vector<3xi64>) -> vector<3xf64> {
     %r = arith.sitofp %a : vector<3xi64> to vector<3xf64>
     return %r : vector<3xf64>
 }
+
+// CHECK-LABEL:   func @fptoui_i64_f64
+// CHECK-SAME:      ([[ARG:%.+]]: f64) -> vector<2xi32>
+// CHECK-NEXT:      [[POW:%.+]] = arith.constant 0x41F0000000000000 : f64
+// CHECK-NEXT:      [[DIV:%.+]] = arith.divf [[ARG]], [[POW]] : f64
+// CHECK-NEXT:      [[HIGHHALF:%.+]] = arith.fptoui [[DIV]] : f64 to i32
+// CHECK-NEXT:      [[REM:%.+]] = arith.remf [[ARG]], [[POW]] : f64
+// CHECK-NEXT:      [[LOWHALF:%.+]] = arith.fptoui [[REM]] : f64 to i32
+// CHECK:           %{{.+}} = vector.insert [[LOWHALF]], %{{.+}} [0]
+// CHECK-NEXT:      [[RESVEC:%.+]] = vector.insert [[HIGHHALF]], %{{.+}} [1]
+// CHECK:           return [[RESVEC]] : vector<2xi32>
+func.func @fptoui_i64_f64(%a : f64) -> i64 {
+    %r = arith.fptoui %a : f64 to i64
+    return %r : i64
+}
+
+// CHECK-LABEL:   func @fptoui_i64_f64_vector
+// CHECK-SAME:      ([[ARG:%.+]]: vector<3xf64>) -> vector<3x2xi32>
+// CHECK-NEXT:      [[POW:%.+]] = arith.constant dense<0x41F0000000000000> : vector<3xf64>
+// CHECK-NEXT:      [[DIV:%.+]] = arith.divf [[ARG]], [[POW]] : vector<3xf64>
+// CHECK-NEXT:      [[HIGHHALF:%.+]] = arith.fptoui [[DIV]] : vector<3xf64> to vector<3xi32>
+// CHECK-NEXT:      [[REM:%.+]] = arith.remf [[ARG]], [[POW]] : vector<3xf64>
+// CHECK-NEXT:      [[LOWHALF:%.+]] = arith.fptoui [[REM]] : vector<3xf64> to vector<3xi32>
+// CHECK-DAG:       [[HIGHHALFX1:%.+]] = vector.shape_cast [[HIGHHALF]] : vector<3xi32> to vector<3x1xi32>
+// CHECK-DAG:       [[LOWHALFX1:%.+]] = vector.shape_cast [[LOWHALF]] : vector<3xi32> to vector<3x1xi32>
+// CHECK:           %{{.+}} = vector.insert_strided_slice [[LOWHALFX1]], %{{.+}} {offsets = [0, 0], strides = [1, 1]}
+// CHECK-NEXT:      [[RESVEC:%.+]] = vector.insert_strided_slice [[HIGHHALFX1]], %{{.+}} {offsets = [0, 1], strides = [1, 1]}
+// CHECK:           return [[RESVEC]] : vector<3x2xi32>
+func.func @fptoui_i64_f64_vector(%a : vector<3xf64>) -> vector<3xi64> {
+    %r = arith.fptoui %a : vector<3xf64> to vector<3xi64>
+    return %r : vector<3xi64>
+}
+
+// This generates lines that are already verified by other patterns.
+// We do not re-verify these and just check for the wrapper around fptoui by following its low part.
+// CHECK-LABEL:   func @fptosi_i64_f64
+// CHECK-SAME:      ([[ARG:%.+]]: f64) -> vector<2xi32>
+// CHECK:           [[ZEROCST:%.+]] = arith.constant 0.000000e+00 : f64
+// CHECK:           [[ZEROCSTINT:%.+]] = arith.constant dense<0> : vector<2xi32>
+// CHECK-NEXT:      [[ISNEGATIVE:%.+]] = arith.cmpf olt, [[ARG]], [[ZEROCST]] : f64
+// CHECK-NEXT:      [[NEGATED:%.+]] = arith.negf [[ARG]] : f64
+// CHECK-NEXT:      [[ABSVALUE:%.+]] = arith.select [[ISNEGATIVE]], [[NEGATED]], [[ARG]] : f64
+// CHECK-NEXT:      [[POW:%.+]] = arith.constant 0x41F0000000000000 : f64
+// CHECK-NEXT:      [[DIV:%.+]] = arith.divf [[ABSVALUE]], [[POW]] : f64
+// CHECK-NEXT:      [[HIGHHALF:%.+]] = arith.fptoui [[DIV]] : f64 to i32
+// CHECK-NEXT:      [[REM:%.+]] = arith.remf [[ABSVALUE]], [[POW]] : f64
+// CHECK-NEXT:      [[LOWHALF:%.+]] = arith.fptoui [[REM]] : f64 to i32
+// CHECK:           vector.insert [[LOWHALF]], %{{.+}} [0] : i32 into vector<2xi32>
+// CHECK-NEXT:      [[FPTOUIRESVEC:%.+]] = vector.insert [[HIGHHALF]]
+// CHECK:           [[ZEROCSTINTHALF:%.+]] = vector.extract [[ZEROCSTINT]][0] : i32 from vector<2xi32>
+// CHECK:           [[SUB:%.+]] = arith.subi [[ZEROCSTINTHALF]], %{{.+}} : i32
+// CHECK-NEXT:      arith.cmpi ult, [[ZEROCSTINTHALF]], %{{.+}} : i32
+// CHECK-NEXT:      arith.extui
+// CHECK-NEXT:      arith.subi
+// CHECK-NEXT:      arith.subi
+// CHECK:           vector.insert [[SUB]]
+// CHECK:           [[SUBVEC:%.+]] = vector.insert
+// CHECK:           [[SUB:%.+]] = vector.extract [[SUBVEC]][0] : i32 from vector<2xi32>
+// CHECK:           [[LOWRES:%.+]] = vector.extract [[FPTOUIRESVEC]][0] : i32 from vector<2xi32>
+// CHECK:           [[ABSRES:%.+]] = arith.select [[ISNEGATIVE]], [[SUB]], [[LOWRES]] : i32
+// CHECK-NEXT:      arith.select [[ISNEGATIVE]]
+// CHECK:           vector.insert [[ABSRES]]
+// CHECK-NEXT:      [[ABSRESVEC:%.+]] = vector.insert
+// CHECK-NEXT:      return [[ABSRESVEC]] : vector<2xi32>
+func.func @fptosi_i64_f64(%a : f64) -> i64 {
+    %r = arith.fptosi %a : f64 to i64
+    return %r : i64
+}
+
+// Same as the non-vector one, we don't re-verify.
+// CHECK-LABEL:   func @fptosi_i64_f64_vector
+// CHECK-SAME:      ([[ARG:%.+]]: vector<3xf64>) -> vector<3x2xi32>
+// CHECK-NEXT:      [[ZEROCST:%.+]] = arith.constant dense<0.000000e+00> : vector<3xf64>
+// CHECK-NEXT:      [[ZEROCSTINT:%.+]] = arith.constant dense<0> : vector<3x2xi32>
+// CHECK-NEXT:      [[ISNEGATIVE:%.+]] = arith.cmpf olt, [[ARG]], [[ZEROCST]] : vector<3xf64>
+// CHECK-NEXT:      [[NEGATED:%.+]] = arith.negf [[ARG]] : vector<3xf64>
+// CHECK-NEXT:      [[ABSVALUE:%.+]] = arith.select [[ISNEGATIVE]], [[NEGATED]], [[ARG]] : vector<3xi1>, vector<3xf64>
+// CHECK-NEXT:      [[POW:%.+]] = arith.constant dense<0x41F0000000000000> : vector<3xf64>
+// CHECK-NEXT:      [[DIV:%.+]] = arith.divf [[ABSVALUE]], [[POW]] : vector<3xf64>
+// CHECK-NEXT:      [[HIGHHALF:%.+]] = arith.fptoui [[DIV]] : vector<3xf64> to vector<3xi32>
+// CHECK-NEXT:      [[REM:%.+]] = arith.remf [[ABSVALUE]], [[POW]] : vector<3xf64>
+// CHECK-NEXT:      [[LOWHALF:%.+]] = arith.fptoui [[REM]] : vector<3xf64> to vector<3xi32>
+// CHECK-NEXT:      [[HIGHHALFX1:%.+]] = vector.shape_cast [[HIGHHALF]] : vector<3xi32> to vector<3x1xi32>
+// CHECK-NEXT:      [[LOWHALFX1:%.+]] = vector.shape_cast [[LOWHALF]] : vector<3xi32> to vector<3x1xi32>
+// CHECK:           vector.insert_strided_slice [[LOWHALFX1]], %{{.+}} {offsets = [0, 0], strides = [1, 1]} : vector<3x1xi32> into vector<3x2xi32>
+// CHECK-NEXT:      [[FPTOUIRESVEC:%.+]] = vector.insert_strided_slice [[HIGHHALFX1]]
+// CHECK:           [[ZEROCSTINTHALF:%.+]] = vector.extract_strided_slice [[ZEROCSTINT]]
+// CHECK-SAME:      {offsets = [0, 0], sizes = [3, 1], strides = [1, 1]} : vector<3x2xi32> to vector<3x1xi32>
+// CHECK:           [[SUB:%.+]] = arith.subi [[ZEROCSTINTHALF]], %{{.+}} : vector<3x1xi32>
+// CHECK-NEXT:      arith.cmpi ult, [[ZEROCSTINTHALF]], %{{.+}} : vector<3x1xi32>
+// CHECK-NEXT:      arith.extui
+// CHECK-NEXT:      arith.subi
+// CHECK-NEXT:      arith.subi
+// CHECK:           vector.insert_strided_slice [[SUB]]
+// CHECK-NEXT:      [[SUBVEC:%.+]] = vector.insert_strided_slice
+// CHECK:           [[SUB:%.+]] = vector.extract_strided_slice [[SUBVEC]]
+// CHECK-SAME:      {offsets = [0, 0], sizes = [3, 1], strides = [1, 1]} : vector<3x2xi32> to vector<3x1xi32>
+// CHECK:           [[LOWRES:%.+]] = vector.extract_strided_slice [[FPTOUIRESVEC]]
+// CHECK-SAME:      {offsets = [0, 0], sizes = [3, 1], strides = [1, 1]} : vector<3x2xi32> to vector<3x1xi32>
+// CHECK:           [[ISNEGATIVEX1:%.+]] = vector.shape_cast [[ISNEGATIVE]] : vector<3xi1> to vector<3x1xi1>
+// CHECK:           [[ABSRES:%.+]] = arith.select [[ISNEGATIVEX1]], [[SUB]], [[LOWRES]] : vector<3x1xi1>, vector<3x1xi32>
+// CHECK-NEXT:      arith.select [[ISNEGATIVEX1]]
+// CHECK:           vector.insert_strided_slice [[ABSRES]]
+// CHECK-NEXT:      [[ABSRESVEC:%.+]] = vector.insert_strided_slice
+// CHECK-NEXT:      return [[ABSRESVEC]] : vector<3x2xi32>
+func.func @fptosi_i64_f64_vector(%a : vector<3xf64>) -> vector<3xi64> {
+    %r = arith.fptosi %a : vector<3xf64> to vector<3xi64>
+    return %r : vector<3xi64>
+}

diff  --git a/mlir/test/Integration/Dialect/Arith/CPU/test-wide-int-emulation-fptosi-i64.mlir b/mlir/test/Integration/Dialect/Arith/CPU/test-wide-int-emulation-fptosi-i64.mlir
new file mode 100644
index 0000000000000..d93b834c8f919
--- /dev/null
+++ b/mlir/test/Integration/Dialect/Arith/CPU/test-wide-int-emulation-fptosi-i64.mlir
@@ -0,0 +1,89 @@
+// Check that the wide integer `arith.fptosi` emulation produces the same result as wide
+// `arith.fptosi`. Emulate i64 ops with i32 ops.
+
+// RUN: mlir-opt %s --convert-scf-to-cf --convert-cf-to-llvm --convert-vector-to-llvm \
+// RUN:             --convert-func-to-llvm --convert-arith-to-llvm | \
+// RUN:   mlir-runner -e entry -entry-point-result=void \
+// RUN:                   --shared-libs=%mlir_c_runner_utils | \
+// RUN:   FileCheck %s --match-full-lines
+
+// RUN: mlir-opt %s --test-arith-emulate-wide-int="widest-int-supported=32" \
+// RUN:             --convert-scf-to-cf --convert-cf-to-llvm --convert-vector-to-llvm \
+// RUN:             --convert-func-to-llvm --convert-arith-to-llvm | \
+// RUN:   mlir-runner -e entry -entry-point-result=void \
+// RUN:                   --shared-libs=%mlir_c_runner_utils | \
+// RUN:   FileCheck %s --match-full-lines
+
+// Ops in this function *only* will be emulated using i32 types.
+func.func @emulate_fptosi(%arg: f64) -> i64 {
+  %res = arith.fptosi %arg : f64 to i64
+  return %res : i64
+}
+
+func.func @check_fptosi(%arg : f64) -> () {
+  %res = func.call @emulate_fptosi(%arg) : (f64) -> (i64)
+  vector.print %res : i64
+  return
+}
+
+func.func @entry() {
+  %cst0 = arith.constant 0.0 : f64
+  %cst_nzero = arith.constant 0x8000000000000000 : f64
+  %cst1 = arith.constant 1.0 : f64
+  %cst_n1 = arith.constant -1.0 : f64
+  %cst_n1_5 = arith.constant -1.5 : f64
+
+  %cstpow20 = arith.constant 1048576.0 : f64
+  %cstnpow20 = arith.constant -1048576.0 : f64
+  
+  %cst_i32_max = arith.constant 4294967295.0 : f64
+  %cst_i32_min = arith.constant -4294967296.0 : f64
+  %cst_i32_overflow = arith.constant 4294967296.0 : f64
+  %cst_i32_noverflow = arith.constant -4294967297.0 : f64
+
+
+  %cstpow40 = arith.constant 1099511627776.0 : f64
+  %cstnpow40 = arith.constant -1099511627776.0 : f64
+  %cst_pow40ppow20 = arith.constant 1099512676352.0 : f64
+  %cst_npow40ppow20 = arith.constant -1099512676352.0 : f64
+
+  %cst_max = arith.constant 9007199254740992.0
+  %cst_min = arith.constant -9007199254740992.0
+  
+  // CHECK:         0
+  func.call @check_fptosi(%cst0) : (f64) -> ()
+  // CHECK-NEXT:    0
+  func.call @check_fptosi(%cst_nzero) : (f64) -> ()
+  // CHECK-NEXT:    1
+  func.call @check_fptosi(%cst1) : (f64) -> ()
+  // CHECK-NEXT:    -1
+  func.call @check_fptosi(%cst_n1) : (f64) -> ()
+  // CHECK-NEXT:    -1
+  func.call @check_fptosi(%cst_n1_5) : (f64) -> ()
+  // CHECK-NEXT:    1048576
+  func.call @check_fptosi(%cstpow20) : (f64) -> ()
+  // CHECK-NEXT:    -1048576
+  func.call @check_fptosi(%cstnpow20) : (f64) -> ()
+  // CHECK-NEXT:    4294967295
+  func.call @check_fptosi(%cst_i32_max) : (f64) -> ()
+  // CHECK-NEXT:    -4294967296
+  func.call @check_fptosi(%cst_i32_min) : (f64) -> ()
+  // CHECK-NEXT:    4294967296
+  func.call @check_fptosi(%cst_i32_overflow) : (f64) -> ()
+  // CHECK-NEXT:    -4294967297
+  func.call @check_fptosi(%cst_i32_noverflow) : (f64) -> ()
+  // CHECK-NEXT:    1099511627776
+  func.call @check_fptosi(%cstpow40) : (f64) -> ()
+  // CHECK-NEXT:    -1099511627776
+  func.call @check_fptosi(%cstnpow40) : (f64) -> ()
+  // CHECK-NEXT:    1099512676352
+  func.call @check_fptosi(%cst_pow40ppow20) : (f64) -> ()
+  // CHECK-NEXT:    -1099512676352
+  func.call @check_fptosi(%cst_npow40ppow20) : (f64) -> ()
+  // CHECK-NEXT:    9007199254740992
+  func.call @check_fptosi(%cst_max) : (f64) -> ()
+  // CHECK-NEXT:    -9007199254740992
+  func.call @check_fptosi(%cst_min) : (f64) -> ()
+
+  return
+}

diff  --git a/mlir/test/Integration/Dialect/Arith/CPU/test-wide-int-emulation-fptoui-i64.mlir b/mlir/test/Integration/Dialect/Arith/CPU/test-wide-int-emulation-fptoui-i64.mlir
new file mode 100644
index 0000000000000..81283ee9fdfd8
--- /dev/null
+++ b/mlir/test/Integration/Dialect/Arith/CPU/test-wide-int-emulation-fptoui-i64.mlir
@@ -0,0 +1,64 @@
+// Check that the wide integer `arith.fptoui` emulation produces the same result as wide
+// `arith.fptoui`. Emulate i64 ops with i32 ops.
+
+// RUN: mlir-opt %s --convert-scf-to-cf --convert-cf-to-llvm --convert-vector-to-llvm \
+// RUN:             --convert-func-to-llvm --convert-arith-to-llvm | \
+// RUN:   mlir-runner -e entry -entry-point-result=void \
+// RUN:                   --shared-libs=%mlir_c_runner_utils | \
+// RUN:   FileCheck %s --match-full-lines
+
+// RUN: mlir-opt %s --test-arith-emulate-wide-int="widest-int-supported=32" \
+// RUN:             --convert-scf-to-cf --convert-cf-to-llvm --convert-vector-to-llvm \
+// RUN:             --convert-func-to-llvm --convert-arith-to-llvm | \
+// RUN:   mlir-runner -e entry -entry-point-result=void \
+// RUN:                   --shared-libs=%mlir_c_runner_utils | \
+// RUN:   FileCheck %s --match-full-lines
+
+// Ops in this function *only* will be emulated using i32 types.
+func.func @emulate_fptoui(%arg: f64) -> i64 {
+  %res = arith.fptoui %arg : f64 to i64
+  return %res : i64
+}
+
+func.func @check_fptoui(%arg : f64) -> () {
+  %res = func.call @emulate_fptoui(%arg) : (f64) -> (i64)
+  vector.print %res : i64
+  return
+}
+
+func.func @entry() {
+  %cst0 = arith.constant 0.0 : f64
+  %cst1 = arith.constant 1.0 : f64
+  %cst1_5 = arith.constant 1.5 : f64
+
+  %cstpow20 = arith.constant 1048576.0 : f64
+  %cst_i32_max = arith.constant 4294967295.0 : f64
+  %cst_i32_overflow = arith.constant 4294967296.0 : f64
+
+
+  %cstpow40 = arith.constant 1099511627776.0 : f64
+  %cst_pow40ppow20 = arith.constant 1099512676352.0 : f64
+
+  %cst_nzero = arith.constant 0x8000000000000000 : f64
+  
+  // CHECK:         0
+  func.call @check_fptoui(%cst0) : (f64) -> ()
+  // CHECK-NEXT:    1
+  func.call @check_fptoui(%cst1) : (f64) -> ()
+  // CHECK-NEXT:    1
+  func.call @check_fptoui(%cst1_5) : (f64) -> ()
+  // CHECK-NEXT:    1048576
+  func.call @check_fptoui(%cstpow20) : (f64) -> ()
+  // CHECK-NEXT:    4294967295
+  func.call @check_fptoui(%cst_i32_max) : (f64) -> ()
+  // CHECK-NEXT:    4294967296
+  func.call @check_fptoui(%cst_i32_overflow) : (f64) -> ()
+  // CHECK-NEXT:    1099511627776
+  func.call @check_fptoui(%cstpow40) : (f64) -> ()
+  // CHECK-NEXT:    1099512676352
+  func.call @check_fptoui(%cst_pow40ppow20) : (f64) -> ()
+  // CHECK-NEXT:    0
+  func.call @check_fptoui(%cst_nzero) : (f64) -> ()
+
+  return
+}


        


More information about the Mlir-commits mailing list