[ConstantFolding] Add constant folding for smul.fix and smul.fix.sat

[bjope]

Summary

This patch teaches ConstantFolding to constant fold
both scalar and vector variants of llvm.smul.fix and
llvm.smul.fix.sat.

As described in the LangRef rounding is unspecified for
these instrinsics. If the result cannot be represented
exactly the default behavior in ConstantFolding is to
round down towards negative infinity. If a target has a
preferred rounding that is different some kind of target
hook would be needed (same strategy as used by the
SelectionDAG legalizer).

[nikic]

llvm/trunk/lib/Analysis/ConstantFolding.cpp

2205

As a possible follow-up cleanup, this can be merged into the fshl/fshr code below. The smul_fix case could bail out if C0, C1 or C2 are null (or handle them, if there is interest to support undef operands).

llvm/trunk/lib/Analysis/ConstantFolding.cpp

bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {
   case Intrinsic::ssub_with_overflow:
   case Intrinsic::usub_with_overflow:
   case Intrinsic::smul_with_overflow:
   case Intrinsic::umul_with_overflow:
   case Intrinsic::sadd_sat:
   case Intrinsic::uadd_sat:
   case Intrinsic::ssub_sat:
   case Intrinsic::usub_sat:
+  case Intrinsic::smul_fix:
+  case Intrinsic::smul_fix_sat:
   case Intrinsic::convert_from_fp16:
   case Intrinsic::convert_to_fp16:
   case Intrinsic::bitreverse:
   case Intrinsic::x86_sse_cvtss2si:
   case Intrinsic::x86_sse_cvtss2si64:
   case Intrinsic::x86_sse_cvttss2si:
   case Intrinsic::x86_sse_cvttss2si64:
   case Intrinsic::x86_sse2_cvtsd2si:
if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
 
           return nullptr;
         }
         }
       }
     }
   }
 
+  if (const auto *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
+    if (const auto *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
+      if (const auto *Op3 = dyn_cast<ConstantInt>(Operands[2])) {
+        switch (IntrinsicID) {
+        default: break;
+        case Intrinsic::smul_fix:
+        case Intrinsic::smul_fix_sat: {
+          // This code performs rounding towards negative infinity in case the
+          // result cannot be represented exactly for the given scale. Targets
+          // that do care about rounding should use a target hook for specifying
+          // how rounding should be done, and provide their own folding to be
+          // consistent with rounding. This is the same approach as used by
+          // DAGTypeLegalizer::ExpandIntRes_MULFIX.
+          APInt Lhs = Op1->getValue();
+          APInt Rhs = Op2->getValue();
+          unsigned Scale = Op3->getValue().getZExtValue();
+          unsigned Width = Lhs.getBitWidth();
+          assert(Scale < Width && "Illegal scale.");
+          unsigned ExtendedWidth = Width * 2;
+          APInt Product = (Lhs.sextOrSelf(ExtendedWidth) *
+                           Rhs.sextOrSelf(ExtendedWidth)).ashr(Scale);
+          if (IntrinsicID == Intrinsic::smul_fix_sat) {
+            APInt MaxValue =
+              APInt::getSignedMaxValue(Width).sextOrSelf(ExtendedWidth);
+            APInt MinValue =
+              APInt::getSignedMinValue(Width).sextOrSelf(ExtendedWidth);
+            Product = APIntOps::smin(Product, MaxValue);
+            Product = APIntOps::smax(Product, MinValue);
+          }
+          return ConstantInt::get(Ty->getContext(),
+                                  Product.sextOrTrunc(Width));
+        }
+        }
+      }
+    }
+  }
+
   if (IntrinsicID == Intrinsic::fshl || IntrinsicID == Intrinsic::fshr) {
     const APInt *C0, *C1, *C2;
     if (!getConstIntOrUndef(Operands[0], C0) ||
         !getConstIntOrUndef(Operands[1], C1) ||
         !getConstIntOrUndef(Operands[2], C2))
       return nullptr;
 
     bool IsRight = IntrinsicID == Intrinsic::fshr;
for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
     for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
       // These intrinsics use a scalar type for their second argument.
       if (J == 1 &&
           (IntrinsicID == Intrinsic::cttz || IntrinsicID == Intrinsic::ctlz ||
            IntrinsicID == Intrinsic::powi)) {
         Lane[J] = Operands[J];
         continue;
       }
+      // These intrinsics use a scalar type for their third argument.
+      if (J == 2 &&
+          (IntrinsicID == Intrinsic::smul_fix ||
+           IntrinsicID == Intrinsic::smul_fix_sat)) {
+        Lane[J] = Operands[J];
+        continue;
+      }
 
       Constant *Agg = Operands[J]->getAggregateElement(I);
       if (!Agg)
         return nullptr;
 
       Lane[J] = Agg;
     }
 

llvm/trunk/test/Analysis/ConstantFolding/smul-fix-sat.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -constprop -S | FileCheck %s
+
+;-----------------------------------------------------------------------------
+; Simple test using scalar layout.
+;-----------------------------------------------------------------------------
+
+declare i32 @llvm.smul.fix.sat.i32(i32, i32, i32)
+
+define i32 @test_smul_fix_sat_i32_0() {
+; CHECK-LABEL: @test_smul_fix_sat_i32_0(
+; CHECK-NEXT:    ret i32 536870912
+;
+  %r = call i32 @llvm.smul.fix.sat.i32(i32 1073741824, i32 1073741824, i32 31) ; 0.5 * 0.5
+  ret i32 %r
+}
+
+;-----------------------------------------------------------------------------
+; More extensive tests based on vectors (basically using the scalar fold
+; for each index).
+;-----------------------------------------------------------------------------
+
+declare <8 x i3> @llvm.smul.fix.sat.v8i3(<8 x i3>, <8 x i3>, i32)
+
+define <8 x i3> @test_smul_fix_sat_v8i3_0() {
+; CHECK-LABEL: @test_smul_fix_sat_v8i3_0(
+; CHECK-NEXT:    ret <8 x i3> <i3 3, i3 3, i3 3, i3 3, i3 0, i3 -4, i3 -4, i3 -4>
+;
+  %r = call <8 x i3> @llvm.smul.fix.sat.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4>,
+  i32 0)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_sat_v8i3_1() {
+; CHECK-LABEL: @test_smul_fix_sat_v8i3_1(
+; CHECK-NEXT:    ret <8 x i3> <i3 3, i3 3, i3 3, i3 2, i3 0, i3 -2, i3 -4, i3 -4>
+;
+  %r = call <8 x i3> @llvm.smul.fix.sat.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4>,
+  i32 1)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_sat_v8i3_2() {
+; CHECK-LABEL: @test_smul_fix_sat_v8i3_2(
+; CHECK-NEXT:    ret <8 x i3> <i3 3, i3 3, i3 2, i3 1, i3 0, i3 -1, i3 -2, i3 -3>
+;
+  %r = call <8 x i3> @llvm.smul.fix.sat.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4>,
+  i32 2)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_sat_v8i3_3() {
+; CHECK-LABEL: @test_smul_fix_sat_v8i3_3(
+; CHECK-NEXT:    ret <8 x i3> <i3 3, i3 3, i3 2, i3 1, i3 0, i3 -1, i3 -2, i3 -3>
+;
+  %r = call <8 x i3> @llvm.smul.fix.sat.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1>,
+  i32 0)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_sat_v8i3_4() {
+; CHECK-LABEL: @test_smul_fix_sat_v8i3_4(
+; CHECK-NEXT:    ret <8 x i3> <i3 2, i3 1, i3 1, i3 0, i3 0, i3 -1, i3 -1, i3 -2>
+;
+  %r = call <8 x i3> @llvm.smul.fix.sat.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1>,
+  i32 1)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_sat_v8i3_5() {
+; CHECK-LABEL: @test_smul_fix_sat_v8i3_5(
+; CHECK-NEXT:    ret <8 x i3> <i3 1, i3 0, i3 0, i3 0, i3 0, i3 -1, i3 -1, i3 -1>
+;
+  %r = call <8 x i3> @llvm.smul.fix.sat.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1>,
+  i32 2)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_sat_v8i3_6() {
+; CHECK-LABEL: @test_smul_fix_sat_v8i3_6(
+; CHECK-NEXT:    ret <8 x i3> <i3 -4, i3 -4, i3 -4, i3 -3, i3 0, i3 3, i3 3, i3 3>
+;
+  %r = call <8 x i3> @llvm.smul.fix.sat.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3>,
+  i32 0)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_sat_v8i3_7() {
+; CHECK-LABEL: @test_smul_fix_sat_v8i3_7(
+; CHECK-NEXT:    ret <8 x i3> <i3 -4, i3 -4, i3 -3, i3 -2, i3 0, i3 1, i3 3, i3 3>
+;
+  %r = call <8 x i3> @llvm.smul.fix.sat.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3>,
+  i32 1)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_sat_v8i3_8() {
+; CHECK-LABEL: @test_smul_fix_sat_v8i3_8(
+; CHECK-NEXT:    ret <8 x i3> <i3 -3, i3 -3, i3 -2, i3 -1, i3 0, i3 0, i3 1, i3 2>
+;
+  %r = call <8 x i3> @llvm.smul.fix.sat.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3>,
+  i32 2)
+  ret <8 x i3> %r
+}

llvm/trunk/test/Analysis/ConstantFolding/smul-fix.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -constprop -S | FileCheck %s
+
+;-----------------------------------------------------------------------------
+; Simple test using scalar layout.
+;-----------------------------------------------------------------------------
+
+declare i32 @llvm.smul.fix.i32(i32, i32, i32)
+
+define i32 @test_smul_fix_i32_0() {
+; CHECK-LABEL: @test_smul_fix_i32_0(
+; CHECK-NEXT:    ret i32 536870912
+;
+  %r = call i32 @llvm.smul.fix.i32(i32 1073741824, i32 1073741824, i32 31) ; 0.5 * 0.5
+  ret i32 %r
+}
+
+;-----------------------------------------------------------------------------
+; More extensive tests based on vectors (basically using the scalar fold
+; for each index).
+;-----------------------------------------------------------------------------
+
+declare <8 x i3> @llvm.smul.fix.v8i3(<8 x i3>, <8 x i3>, i32)
+
+define <8 x i3> @test_smul_fix_v8i3_0() {
+; CHECK-LABEL: @test_smul_fix_v8i3_0(
+; CHECK-NEXT:    ret <8 x i3> <i3 0, i3 -4, i3 0, i3 -4, i3 0, i3 -4, i3 0, i3 -4>
+;
+  %r = call <8 x i3> @llvm.smul.fix.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4>,
+  i32 0)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_v8i3_1() {
+; CHECK-LABEL: @test_smul_fix_v8i3_1(
+; CHECK-NEXT:    ret <8 x i3> <i3 0, i3 -2, i3 -4, i3 2, i3 0, i3 -2, i3 -4, i3 2>
+;
+  %r = call <8 x i3> @llvm.smul.fix.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4>,
+  i32 1)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_v8i3_2() {
+; CHECK-LABEL: @test_smul_fix_v8i3_2(
+; CHECK-NEXT:    ret <8 x i3> <i3 -4, i3 3, i3 2, i3 1, i3 0, i3 -1, i3 -2, i3 -3>
+;
+  %r = call <8 x i3> @llvm.smul.fix.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4, i3 -4>,
+  i32 2)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_v8i3_3() {
+; CHECK-LABEL: @test_smul_fix_v8i3_3(
+; CHECK-NEXT:    ret <8 x i3> <i3 -4, i3 3, i3 2, i3 1, i3 0, i3 -1, i3 -2, i3 -3>
+;
+  %r = call <8 x i3> @llvm.smul.fix.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1>,
+  i32 0)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_v8i3_4() {
+; CHECK-LABEL: @test_smul_fix_v8i3_4(
+; CHECK-NEXT:    ret <8 x i3> <i3 2, i3 1, i3 1, i3 0, i3 0, i3 -1, i3 -1, i3 -2>
+;
+  %r = call <8 x i3> @llvm.smul.fix.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1>,
+  i32 1)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_v8i3_5() {
+; CHECK-LABEL: @test_smul_fix_v8i3_5(
+; CHECK-NEXT:    ret <8 x i3> <i3 1, i3 0, i3 0, i3 0, i3 0, i3 -1, i3 -1, i3 -1>
+;
+  %r = call <8 x i3> @llvm.smul.fix.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1, i3 -1>,
+  i32 2)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_v8i3_6() {
+; CHECK-LABEL: @test_smul_fix_v8i3_6(
+; CHECK-NEXT:    ret <8 x i3> <i3 -4, i3 -1, i3 2, i3 -3, i3 0, i3 3, i3 -2, i3 1>
+;
+  %r = call <8 x i3> @llvm.smul.fix.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3>,
+  i32 0)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_v8i3_7() {
+; CHECK-LABEL: @test_smul_fix_v8i3_7(
+; CHECK-NEXT:    ret <8 x i3> <i3 2, i3 3, i3 -3, i3 -2, i3 0, i3 1, i3 3, i3 -4>
+;
+  %r = call <8 x i3> @llvm.smul.fix.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3>,
+  i32 1)
+  ret <8 x i3> %r
+}
+
+define <8 x i3> @test_smul_fix_v8i3_8() {
+; CHECK-LABEL: @test_smul_fix_v8i3_8(
+; CHECK-NEXT:    ret <8 x i3> <i3 -3, i3 -3, i3 -2, i3 -1, i3 0, i3 0, i3 1, i3 2>
+;
+  %r = call <8 x i3> @llvm.smul.fix.v8i3(
+  <8 x i3> <i3 -4, i3 -3, i3 -2, i3 -1, i3 0, i3 1, i3 2, i3 3>,
+  <8 x i3> <i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3, i3 3>,
+  i32 2)
+  ret <8 x i3> %r
+}