[InstSimplify] Simplify smul.fix and smul.fix.sat

[bjope]

Summary

Add simplification of smul.fix and smul.fix.sat according to

X * 0 -> 0
X * undef -> 0
X * (1 << scale) -> X

This includes the commuted patterns and splatted vectors.

[nikic]

LGTM

[nagisa]

llvm/lib/Analysis/InstructionSimplify.cpp

5767

Wouldn't this fail to trigger if the intrinsic call starts as follows?

%res = call i16 @llvm.smul.fix.i16(i16 42, i16 undef, i32 ...)

If I understand it correctly, the canonicalization step above would swap the two operands, making the Op0 = undef and Op1 = 42, thus preventing this simplification from happening.

(This does not appear to be an issue for 0, it seems there's code somewhere else that const-evaluates this operation if both operands are constant)

[nikic]

llvm/lib/Analysis/InstructionSimplify.cpp

5767

Folding a call with all constant arguments is the responsibility of ConstantFolding rather than InstSimplify. Possibly the current constant folding implementation for fix-point multiplication doesn't handle undefs though.

[bjope]

llvm/lib/Analysis/InstructionSimplify.cpp

5767

@nagisa : Nice finding!
@nikic : Yes, that was a limitation in ConstantFolding. I've already started to prepare a fix for that but need to create some more tests. I can make that a predecessor for this patch (or maybe I'll just squash them together).

[bjope]

Update code comments (those comments depends on landing D98410 first).

llvm/lib/Analysis/InstructionSimplify.cpp

static Value *simplifyIntrinsic(CallBase *Call, const SimplifyQuery &Q) {
   case Intrinsic::fmuladd: {
     Value *Op0 = Call->getArgOperand(0);
     Value *Op1 = Call->getArgOperand(1);
     Value *Op2 = Call->getArgOperand(2);
     if (Value *V = simplifyFPOp({ Op0, Op1, Op2 }, {}, Q))
       return V;
     return nullptr;
   }
+  case Intrinsic::smul_fix:
+  case Intrinsic::smul_fix_sat: {
+    Value *Op0 = Call->getArgOperand(0);
+    Value *Op1 = Call->getArgOperand(1);
+    Value *Op2 = Call->getArgOperand(2);
+    Type *ReturnType = F->getReturnType();
+
+    // Canonicalize constant operand as Op1 (ConstantFolding handles the case
+    // when both Op0 and Op1 are constant so we do not care about that special
+    // case here).
+    if (isa<Constant>(Op0))
+      std::swap(Op0, Op1);
+
+    // X * 0 -> 0
+    if (match(Op1, m_Zero()))
+      return Constant::getNullValue(ReturnType);
+
+    // X * undef -> 0
+    if (Q.isUndefValue(Op1))
+      return Constant::getNullValue(ReturnType);
+
+    // X * (1 << Scale) -> X
+    APInt ScaledOne =
+        APInt::getOneBitSet(ReturnType->getScalarSizeInBits(),
+                            cast<ConstantInt>(Op2)->getZExtValue());
+    if (ScaledOne.isNonNegative() && match(Op1, m_SpecificInt(ScaledOne)))
+      return Op0;
+
+    return nullptr;
+  }
   default:
     return nullptr;
   }
 }
 
 static Value *tryConstantFoldCall(CallBase *Call, const SimplifyQuery &Q) {
   auto *F = dyn_cast<Function>(Call->getCalledOperand());
   if (!F || !canConstantFoldCallTo(Call, F))

llvm/test/Transforms/InstSimplify/smul_fix.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -S -instsimplify | FileCheck %s
+
+declare i16 @llvm.smul.fix.i16(i16, i16, i32)
+declare i16 @llvm.smul.fix.sat.i16(i16, i16, i32)
+declare <2 x i16> @llvm.smul.fix.v2i16(<2 x i16>, <2 x i16>, i32)
+declare <2 x i16> @llvm.smul.fix.sat.v2i16(<2 x i16>, <2 x i16>, i32)
+
+;;
+;; llvm.smul.fix (scalar)
+;;
+
+; X * 0 -> X
+define i16 @smul_fix_0(i16 %arg) {
+; CHECK-LABEL: @smul_fix_0(
+; CHECK-NEXT:    ret i16 0
+;
+  %res = call i16 @llvm.smul.fix.i16(i16 %arg, i16 0, i32 15)
+  ret i16 %res
+}
+
+; 0 * X -> X
+define i16 @smul_fix_1(i16 %arg) {
+; CHECK-LABEL: @smul_fix_1(
+; CHECK-NEXT:    ret i16 0
+;
+  %res = call i16 @llvm.smul.fix.i16(i16 0, i16 %arg, i32 15)
+  ret i16 %res
+}
+
+; X * undef -> undef
+define i16 @smul_fix_2(i16 %arg) {
+; CHECK-LABEL: @smul_fix_2(
+; CHECK-NEXT:    ret i16 0
+;
+  %res = call i16 @llvm.smul.fix.i16(i16 %arg, i16 undef, i32 15)
+  ret i16 %res
+}
+
+; undef * X -> undef
+define i16 @smul_fix_3(i16 %arg) {
+; CHECK-LABEL: @smul_fix_3(
+; CHECK-NEXT:    ret i16 0
+;
+  %res = call i16 @llvm.smul.fix.i16(i16 undef, i16 %arg, i32 15)
+  ret i16 %res
+}
+
+; X * 1 -> X
+define i16 @smul_fix_4(i16 %arg) {
+; CHECK-LABEL: @smul_fix_4(
+; CHECK-NEXT:    ret i16 [[ARG:%.*]]
+;
+  %res = call i16 @llvm.smul.fix.i16(i16 %arg, i16 16384, i32 14)
+  ret i16 %res
+}
+
+; 1 * X -> X
+define i16 @smul_fix_5(i16 %arg) {
+; CHECK-LABEL: @smul_fix_5(
+; CHECK-NEXT:    ret i16 [[ARG:%.*]]
+;
+  %res = call i16 @llvm.smul.fix.i16(i16 2, i16 %arg, i32 1)
+  ret i16 %res
+}
+
+;;
+;; llvm.smul.fix.sat (scalar)
+;;
+
+; X * 0 -> X
+define i16 @smul_fix_sat_0(i16 %arg) {
+; CHECK-LABEL: @smul_fix_sat_0(
+; CHECK-NEXT:    ret i16 0
+;
+  %res = call i16 @llvm.smul.fix.sat.i16(i16 %arg, i16 0, i32 15)
+  ret i16 %res
+}
+
+; 0 * X -> X
+define i16 @smul_fix_sat_1(i16 %arg) {
+; CHECK-LABEL: @smul_fix_sat_1(
+; CHECK-NEXT:    ret i16 0
+;
+  %res = call i16 @llvm.smul.fix.sat.i16(i16 0, i16 %arg, i32 15)
+  ret i16 %res
+}
+
+; X * undef -> undef
+define i16 @smul_fix_sat_2(i16 %arg) {
+; CHECK-LABEL: @smul_fix_sat_2(
+; CHECK-NEXT:    ret i16 0
+;
+  %res = call i16 @llvm.smul.fix.sat.i16(i16 %arg, i16 undef, i32 15)
+  ret i16 %res
+}
+
+; undef * X -> undef
+define i16 @smul_fix_sat_3(i16 %arg) {
+; CHECK-LABEL: @smul_fix_sat_3(
+; CHECK-NEXT:    ret i16 0
+;
+  %res = call i16 @llvm.smul.fix.sat.i16(i16 undef, i16 %arg, i32 15)
+  ret i16 %res
+}
+
+; X * 1 -> X
+define i16 @smul_fix_sat_4(i16 %arg) {
+; CHECK-LABEL: @smul_fix_sat_4(
+; CHECK-NEXT:    ret i16 [[ARG:%.*]]
+;
+  %res = call i16 @llvm.smul.fix.sat.i16(i16 %arg, i16 16384, i32 14)
+  ret i16 %res
+}
+
+; 1 * X -> X
+define i16 @smul_fix_sat_5(i16 %arg) {
+; CHECK-LABEL: @smul_fix_sat_5(
+; CHECK-NEXT:    ret i16 [[ARG:%.*]]
+;
+  %res = call i16 @llvm.smul.fix.sat.i16(i16 2, i16 %arg, i32 1)
+  ret i16 %res
+}
+
+;;
+;; llvm.smul.fix (vector)
+;;
+
+; X * 0 -> X
+define <2 x i16> @smul_fix_vec_0(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_vec_0(
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
+;
+  %res = call <2 x i16> @llvm.smul.fix.v2i16(<2 x i16> %arg, <2 x i16> zeroinitializer, i32 15)
+  ret <2 x i16> %res
+}
+
+; 0 * X -> X
+define <2 x i16> @smul_fix_vec_1(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_vec_1(
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
+;
+  %res = call <2 x i16> @llvm.smul.fix.v2i16(<2 x i16> zeroinitializer, <2 x i16> %arg, i32 15)
+  ret <2 x i16> %res
+}
+
+; X * undef -> undef
+define <2 x i16> @smul_fix_vec_2(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_vec_2(
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
+;
+  %res = call <2 x i16> @llvm.smul.fix.v2i16(<2 x i16> %arg, <2 x i16> undef, i32 15)
+  ret <2 x i16> %res
+}
+
+; undef * X -> undef
+define <2 x i16> @smul_fix_vec_3(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_vec_3(
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
+;
+  %res = call <2 x i16> @llvm.smul.fix.v2i16(<2 x i16> undef, <2 x i16> %arg, i32 15)
+  ret <2 x i16> %res
+}
+
+; X * 1 -> X
+define <2 x i16> @smul_fix_vec_4(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_vec_4(
+; CHECK-NEXT:    ret <2 x i16> [[ARG:%.*]]
+;
+  %res = call <2 x i16> @llvm.smul.fix.v2i16(<2 x i16> %arg, <2 x i16> <i16 16384, i16 16384>, i32 14)
+  ret <2 x i16> %res
+}
+
+; 1 * X -> X
+define <2 x i16> @smul_fix_vec_5(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_vec_5(
+; CHECK-NEXT:    ret <2 x i16> [[ARG:%.*]]
+;
+  %res = call <2 x i16> @llvm.smul.fix.v2i16(<2 x i16> <i16 2, i16 2>, <2 x i16> %arg, i32 1)
+  ret <2 x i16> %res
+}
+
+;;
+;; llvm.smul.fix.sat (vector)
+;;
+
+; X * 0 -> X
+define <2 x i16> @smul_fix_sat_vec_0(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_sat_vec_0(
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
+;
+  %res = call <2 x i16> @llvm.smul.fix.sat.v2i16(<2 x i16> %arg, <2 x i16> zeroinitializer, i32 15)
+  ret <2 x i16> %res
+}
+
+; 0 * X -> X
+define <2 x i16> @smul_fix_sat_vec_1(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_sat_vec_1(
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
+;
+  %res = call <2 x i16> @llvm.smul.fix.sat.v2i16(<2 x i16> zeroinitializer, <2 x i16> %arg, i32 15)
+  ret <2 x i16> %res
+}
+
+; X * undef -> undef
+define <2 x i16> @smul_fix_sat_vec_2(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_sat_vec_2(
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
+;
+  %res = call <2 x i16> @llvm.smul.fix.sat.v2i16(<2 x i16> %arg, <2 x i16> undef, i32 15)
+  ret <2 x i16> %res
+}
+
+; undef * X -> undef
+define <2 x i16> @smul_fix_sat_vec_3(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_sat_vec_3(
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
+;
+  %res = call <2 x i16> @llvm.smul.fix.sat.v2i16(<2 x i16> undef, <2 x i16> %arg, i32 15)
+  ret <2 x i16> %res
+}
+
+; X * 1 -> X
+define <2 x i16> @smul_fix_sat_vec_4(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_sat_vec_4(
+; CHECK-NEXT:    ret <2 x i16> [[ARG:%.*]]
+;
+  %res = call <2 x i16> @llvm.smul.fix.sat.v2i16(<2 x i16> %arg, <2 x i16> <i16 16384, i16 16384>, i32 14)
+  ret <2 x i16> %res
+}
+
+; 1 * X -> X
+define <2 x i16> @smul_fix_sat_vec_5(<2 x i16> %arg) {
+; CHECK-LABEL: @smul_fix_sat_vec_5(
+; CHECK-NEXT:    ret <2 x i16> [[ARG:%.*]]
+;
+  %res = call <2 x i16> @llvm.smul.fix.sat.v2i16(<2 x i16> <i16 2, i16 2>, <2 x i16> %arg, i32 1)
+  ret <2 x i16> %res
+}