[InstCombine] Fold usub_sat((sub nuw C1, A), C2) to usub_sat(C1 - C2, A) or 0

Author: elhewaty

llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp

@@ -2139,8 +2139,22 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
       }
     }
 
+    // usub_sat((sub nuw C, A), C1) -> usub_sat(usub_sat(C, C1), A)
+    // which after that:
+    // usub_sat((sub nuw C, A), C1) -> usub_sat(C - C1, A) if C1 u< C
+    // usub_sat((sub nuw C, A), C1) -> 0 otherwise
+    Constant *C, *C1;
+    Value *A;
+    if (IID == Intrinsic::usub_sat &&
+        match(Arg0, m_NUWSub(m_ImmConstant(C), m_Value(A))) &&
+        match(Arg1, m_ImmConstant(C1))) {
+      auto *NewC = Builder.CreateBinaryIntrinsic(Intrinsic::usub_sat, C, C1);
+      auto *NewSub =
+          Builder.CreateBinaryIntrinsic(Intrinsic::usub_sat, NewC, A);
+      return replaceInstUsesWith(*SI, NewSub);
+    }
+
     // ssub.sat(X, C) -> sadd.sat(X, -C) if C != MIN
-    Constant *C;
     if (IID == Intrinsic::ssub_sat && match(Arg1, m_Constant(C)) &&
         C->isNotMinSignedValue()) {
       Value *NegVal = ConstantExpr::getNeg(C);

llvm/test/Transforms/InstCombine/unsigned_saturated_sub.ll

@@ -11,8 +11,7 @@ declare void @usei1(i1)
 ; usub_sat((sub nuw C1, A), C2) to usub_sat(usub_sat(C1 - C2), A)
 define i32 @usub_sat_C1_C2(i32 %a){
 ; CHECK-LABEL: @usub_sat_C1_C2(
-; CHECK-NEXT:    [[ADD:%.*]] = sub nuw i32 64, [[A:%.*]]
-; CHECK-NEXT:    [[COND:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[ADD]], i32 14)
+; CHECK-NEXT:    [[COND:%.*]] = call i32 @llvm.usub.sat.i32(i32 50, i32 [[A:%.*]])
 ; CHECK-NEXT:    ret i32 [[COND]]
 ;
   %add = sub nuw i32 64, %a
@@ -22,9 +21,7 @@ define i32 @usub_sat_C1_C2(i32 %a){
 
 define i32 @usub_sat_C1_C2_produce_0(i32 %a){
 ; CHECK-LABEL: @usub_sat_C1_C2_produce_0(
-; CHECK-NEXT:    [[ADD:%.*]] = sub nuw i32 14, [[A:%.*]]
-; CHECK-NEXT:    [[COND:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[ADD]], i32 14)
-; CHECK-NEXT:    ret i32 [[COND]]
+; CHECK-NEXT:    ret i32 0
 ;
   %add = sub nuw i32 14, %a
   %cond = call i32 @llvm.usub.sat.i32(i32 %add, i32 14)
@@ -33,9 +30,7 @@ define i32 @usub_sat_C1_C2_produce_0(i32 %a){
 
 define i32 @usub_sat_C1_C2_produce_0_too(i32 %a){
 ; CHECK-LABEL: @usub_sat_C1_C2_produce_0_too(
-; CHECK-NEXT:    [[ADD:%.*]] = sub nuw i32 12, [[A:%.*]]
-; CHECK-NEXT:    [[COND:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[ADD]], i32 14)
-; CHECK-NEXT:    ret i32 [[COND]]
+; CHECK-NEXT:    ret i32 0
 ;
   %add = sub nuw i32 12, %a
   %cond = call i32 @llvm.usub.sat.i32(i32 %add, i32 14)
@@ -45,8 +40,7 @@ define i32 @usub_sat_C1_C2_produce_0_too(i32 %a){
 ; vector tests
 define <2 x i16> @usub_sat_C1_C2_splat(<2 x i16> %a) {
 ; CHECK-LABEL: @usub_sat_C1_C2_splat(
-; CHECK-NEXT:    [[ADD:%.*]] = sub nuw <2 x i16> <i16 64, i16 64>, [[A:%.*]]
-; CHECK-NEXT:    [[COND:%.*]] = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> [[ADD]], <2 x i16> <i16 14, i16 14>)
+; CHECK-NEXT:    [[COND:%.*]] = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> <i16 50, i16 50>, <2 x i16> [[A:%.*]])
 ; CHECK-NEXT:    ret <2 x i16> [[COND]]
 ;
   %add = sub nuw <2 x i16> <i16 64, i16 64>, %a
@@ -56,8 +50,7 @@ define <2 x i16> @usub_sat_C1_C2_splat(<2 x i16> %a) {
 
 define <2 x i16> @usub_sat_C1_C2_non_splat(<2 x i16> %a) {
 ; CHECK-LABEL: @usub_sat_C1_C2_non_splat(
-; CHECK-NEXT:    [[ADD:%.*]] = sub nuw <2 x i16> <i16 50, i16 64>, [[A:%.*]]
-; CHECK-NEXT:    [[COND:%.*]] = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> [[ADD]], <2 x i16> <i16 20, i16 14>)
+; CHECK-NEXT:    [[COND:%.*]] = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> <i16 30, i16 50>, <2 x i16> [[A:%.*]])
 ; CHECK-NEXT:    ret <2 x i16> [[COND]]
 ;
   %add = sub nuw <2 x i16> <i16 50, i16 64>, %a
@@ -67,9 +60,7 @@ define <2 x i16> @usub_sat_C1_C2_non_splat(<2 x i16> %a) {
 
 define <2 x i16> @usub_sat_C1_C2_splat_produce_0(<2 x i16> %a){
 ; CHECK-LABEL: @usub_sat_C1_C2_splat_produce_0(
-; CHECK-NEXT:    [[ADD:%.*]] = sub nuw <2 x i16> <i16 14, i16 14>, [[A:%.*]]
-; CHECK-NEXT:    [[COND:%.*]] = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> [[ADD]], <2 x i16> <i16 14, i16 14>)
-; CHECK-NEXT:    ret <2 x i16> [[COND]]
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
 ;
   %add = sub nuw <2 x i16> <i16 14, i16 14>, %a
   %cond = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> %add, <2 x i16> <i16 14, i16 14>)
@@ -78,9 +69,7 @@ define <2 x i16> @usub_sat_C1_C2_splat_produce_0(<2 x i16> %a){
 
 define <2 x i16> @usub_sat_C1_C2_splat_produce_0_too(<2 x i16> %a){
 ; CHECK-LABEL: @usub_sat_C1_C2_splat_produce_0_too(
-; CHECK-NEXT:    [[ADD:%.*]] = sub nuw <2 x i16> <i16 12, i16 12>, [[A:%.*]]
-; CHECK-NEXT:    [[COND:%.*]] = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> [[ADD]], <2 x i16> <i16 14, i16 14>)
-; CHECK-NEXT:    ret <2 x i16> [[COND]]
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
 ;
   %add = sub nuw <2 x i16> <i16 12, i16 12>, %a
   %cond = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> %add, <2 x i16> <i16 14, i16 14>)
@@ -89,9 +78,7 @@ define <2 x i16> @usub_sat_C1_C2_splat_produce_0_too(<2 x i16> %a){
 
 define <2 x i16> @usub_sat_C1_C2_non_splat_produce_0_too(<2 x i16> %a){
 ; CHECK-LABEL: @usub_sat_C1_C2_non_splat_produce_0_too(
-; CHECK-NEXT:    [[ADD:%.*]] = sub nuw <2 x i16> <i16 12, i16 13>, [[A:%.*]]
-; CHECK-NEXT:    [[COND:%.*]] = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> [[ADD]], <2 x i16> <i16 14, i16 15>)
-; CHECK-NEXT:    ret <2 x i16> [[COND]]
+; CHECK-NEXT:    ret <2 x i16> zeroinitializer
 ;
   %add = sub nuw <2 x i16> <i16 12, i16 13>, %a
   %cond = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> %add, <2 x i16> <i16 14, i16 15>)