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

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).

Reviewers: nikic, leonardchan, RKSimon

Reviewed By: leonardchan

Subscribers: hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63385

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363811 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Analysis/ConstantFolding.cpp b/lib/Analysis/ConstantFolding.cpp
index 88d003864ebd4576261332a02c5c8a54f4012b37..9372904f769d1b180c359581cef7cc4789e5ac9e 100644 (file)
--- a/lib/Analysis/ConstantFolding.cpp
+++ b/lib/Analysis/ConstantFolding.cpp
@@ -1422,6 +1422,8 @@ bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {
   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:
@@ -2198,6 +2200,43 @@ static Constant *ConstantFoldScalarCall3(StringRef Name, unsigned IntrinsicID,
     }
   }
 
+  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) ||
@@ -2307,6 +2346,13 @@ static Constant *ConstantFoldVectorCall(StringRef Name, unsigned IntrinsicID,
         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)

diff --git a/test/Analysis/ConstantFolding/smul-fix-sat.ll b/test/Analysis/ConstantFolding/smul-fix-sat.ll
new file mode 100644 (file)
index 0000000..3caae7c
--- /dev/null
+++ b/test/Analysis/ConstantFolding/smul-fix-sat.ll
@@ -0,0 +1,122 @@
+; 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
+}

diff --git a/test/Analysis/ConstantFolding/smul-fix.ll b/test/Analysis/ConstantFolding/smul-fix.ll
new file mode 100644 (file)
index 0000000..9fcf4a9
--- /dev/null
+++ b/test/Analysis/ConstantFolding/smul-fix.ll
@@ -0,0 +1,122 @@
+; 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
+}