[InstCombine] fold reassociative fadd with negated operand

We manage to iteratively achieve this result with no extra
uses, and the reassociate pass can also do this, but this
pattern falls through the cracks in the example from
issue #57053.

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index 13a0386..528c74a 100644 (file)
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -1665,6 +1665,11 @@ Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
         return BinaryOperator::CreateFMulFMF(X, NewMulC, &I);
     }
 
+    // (-X - Y) + (X + Z) --> Z - Y
+    if (match(&I, m_c_FAdd(m_FSub(m_FNeg(m_Value(X)), m_Value(Y)),
+                           m_c_FAdd(m_Deferred(X), m_Value(Z)))))
+      return BinaryOperator::CreateFSubFMF(Z, Y, &I);
+
     if (Value *V = FAddCombine(Builder).simplify(&I))
       return replaceInstUsesWith(I, V);
   }

diff --git a/llvm/test/Transforms/InstCombine/fadd.ll b/llvm/test/Transforms/InstCombine/fadd.ll
index 275cecd..32ecd8c 100644 (file)
--- a/llvm/test/Transforms/InstCombine/fadd.ll
+++ b/llvm/test/Transforms/InstCombine/fadd.ll
@@ -534,9 +534,7 @@ define float @fadd_fneg_reass_commute0(float %x, float %y, float %z) {
 ; CHECK-LABEL: @fadd_fneg_reass_commute0(
 ; CHECK-NEXT:    [[N:%.*]] = fneg reassoc nsz float [[X:%.*]]
 ; CHECK-NEXT:    call void @use(float [[N]])
-; CHECK-NEXT:    [[S:%.*]] = fsub reassoc nsz float [[N]], [[Y:%.*]]
-; CHECK-NEXT:    [[A:%.*]] = fadd reassoc nsz float [[X]], [[Z:%.*]]
-; CHECK-NEXT:    [[R:%.*]] = fadd reassoc nsz float [[S]], [[A]]
+; CHECK-NEXT:    [[R:%.*]] = fsub reassoc nsz float [[Z:%.*]], [[Y:%.*]]
 ; CHECK-NEXT:    ret float [[R]]
 ;
   %n = fneg reassoc nsz float %x
@@ -553,8 +551,7 @@ define float @fadd_fneg_reass_commute1(float %x, float %y, float %z) {
 ; CHECK-NEXT:    call void @use(float [[N]])
 ; CHECK-NEXT:    [[S:%.*]] = fsub float [[N]], [[Y:%.*]]
 ; CHECK-NEXT:    call void @use(float [[S]])
-; CHECK-NEXT:    [[A:%.*]] = fadd float [[X]], [[Z:%.*]]
-; CHECK-NEXT:    [[R:%.*]] = fadd reassoc nsz float [[A]], [[S]]
+; CHECK-NEXT:    [[R:%.*]] = fsub reassoc nsz float [[Z:%.*]], [[Y]]
 ; CHECK-NEXT:    ret float [[R]]
 ;
   %n = fneg float %x
@@ -574,7 +571,7 @@ define float @fadd_fneg_reass_commute2(float %x, float %y, float %z) {
 ; CHECK-NEXT:    call void @use(float [[S]])
 ; CHECK-NEXT:    [[A:%.*]] = fadd float [[Z:%.*]], [[X]]
 ; CHECK-NEXT:    call void @use(float [[A]])
-; CHECK-NEXT:    [[R:%.*]] = fadd fast float [[S]], [[A]]
+; CHECK-NEXT:    [[R:%.*]] = fsub fast float [[Z]], [[Y]]
 ; CHECK-NEXT:    ret float [[R]]
 ;
   %n = fneg float %x
@@ -591,9 +588,7 @@ define <2 x float> @fadd_fneg_reass_commute3(<2 x float> %x, <2 x float> %y, <2
 ; CHECK-LABEL: @fadd_fneg_reass_commute3(
 ; CHECK-NEXT:    [[N:%.*]] = fneg reassoc nsz <2 x float> [[X:%.*]]
 ; CHECK-NEXT:    call void @use_vec(<2 x float> [[N]])
-; CHECK-NEXT:    [[S:%.*]] = fsub reassoc nsz <2 x float> [[N]], [[Y:%.*]]
-; CHECK-NEXT:    [[A:%.*]] = fadd reassoc nsz <2 x float> [[Z:%.*]], [[X]]
-; CHECK-NEXT:    [[R:%.*]] = fadd reassoc nsz <2 x float> [[A]], [[S]]
+; CHECK-NEXT:    [[R:%.*]] = fsub reassoc nsz <2 x float> [[Z:%.*]], [[Y:%.*]]
 ; CHECK-NEXT:    ret <2 x float> [[R]]
 ;
   %n = fneg reassoc nsz <2 x float> %x
@@ -604,6 +599,8 @@ define <2 x float> @fadd_fneg_reass_commute3(<2 x float> %x, <2 x float> %y, <2
   ret <2 x float> %r
 }
 
+; negative test - need reassoc (+ nsz)
+
 define float @fadd_fneg_commute0(float %x, float %y, float %z) {
 ; CHECK-LABEL: @fadd_fneg_commute0(
 ; CHECK-NEXT:    [[N:%.*]] = fneg float [[X:%.*]]