[ValueTracking] Add rotate idiom to haveNoCommonBitsSet special cases (#122165)

An occasional idiom for rotation is "(A << B) + (A >> (BitWidth - B))".
Currently this is not well handled on targets with native
funnel-shift/rotate support. Add a special case to haveNoCommonBitsSet
to ensure that the addition is converted to a disjoint or in InstCombine
so during instruction selection the idiom can be converted to an
efficient rotation implementation.

Proof: https://alive2.llvm.org/ce/z/WdCZsN

Author: AlexMaclean

llvm/lib/Analysis/ValueTracking.cpp

@@ -229,6 +229,19 @@ static bool haveNoCommonBitsSetSpecialCases(const Value *LHS, const Value *RHS,
       return true;
   }
 
+  // Look for: (X << V) op (Y >> (BitWidth - V))
+  // or        (X >> V) op (Y << (BitWidth - V))
+  {
+    const Value *V;
+    const APInt *R;
+    if (((match(RHS, m_Shl(m_Value(), m_Sub(m_APInt(R), m_Value(V)))) &&
+          match(LHS, m_LShr(m_Value(), m_Specific(V)))) ||
+         (match(RHS, m_LShr(m_Value(), m_Sub(m_APInt(R), m_Value(V)))) &&
+          match(LHS, m_Shl(m_Value(), m_Specific(V))))) &&
+        R->uge(LHS->getType()->getScalarSizeInBits()))
+      return true;
+  }
+
   return false;
 }
 

llvm/test/Transforms/InstCombine/rotate.ll

@@ -191,7 +191,7 @@ define i32 @rotl_i32(i32 %x, i32 %y) {
 ; CHECK-NEXT:    [[SUB:%.*]] = sub i32 32, [[Y:%.*]]
 ; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[X:%.*]], [[Y]]
 ; CHECK-NEXT:    [[SHR:%.*]] = lshr i32 [[X]], [[SUB]]
-; CHECK-NEXT:    [[R:%.*]] = or i32 [[SHR]], [[SHL]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i32 [[SHR]], [[SHL]]
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %sub = sub i32 32, %y
@@ -208,7 +208,7 @@ define i37 @rotr_i37(i37 %x, i37 %y) {
 ; CHECK-NEXT:    [[SUB:%.*]] = sub i37 37, [[Y:%.*]]
 ; CHECK-NEXT:    [[SHL:%.*]] = shl i37 [[X:%.*]], [[SUB]]
 ; CHECK-NEXT:    [[SHR:%.*]] = lshr i37 [[X]], [[Y]]
-; CHECK-NEXT:    [[R:%.*]] = or i37 [[SHR]], [[SHL]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i37 [[SHR]], [[SHL]]
 ; CHECK-NEXT:    ret i37 [[R]]
 ;
   %sub = sub i37 37, %y
@@ -225,7 +225,7 @@ define i8 @rotr_i8_commute(i8 %x, i8 %y) {
 ; CHECK-NEXT:    [[SUB:%.*]] = sub i8 8, [[Y:%.*]]
 ; CHECK-NEXT:    [[SHL:%.*]] = shl i8 [[X:%.*]], [[SUB]]
 ; CHECK-NEXT:    [[SHR:%.*]] = lshr i8 [[X]], [[Y]]
-; CHECK-NEXT:    [[R:%.*]] = or i8 [[SHL]], [[SHR]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i8 [[SHL]], [[SHR]]
 ; CHECK-NEXT:    ret i8 [[R]]
 ;
   %sub = sub i8 8, %y
@@ -242,7 +242,7 @@ define <4 x i32> @rotl_v4i32(<4 x i32> %x, <4 x i32> %y) {
 ; CHECK-NEXT:    [[SUB:%.*]] = sub <4 x i32> splat (i32 32), [[Y:%.*]]
 ; CHECK-NEXT:    [[SHL:%.*]] = shl <4 x i32> [[X:%.*]], [[Y]]
 ; CHECK-NEXT:    [[SHR:%.*]] = lshr <4 x i32> [[X]], [[SUB]]
-; CHECK-NEXT:    [[R:%.*]] = or <4 x i32> [[SHL]], [[SHR]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint <4 x i32> [[SHL]], [[SHR]]
 ; CHECK-NEXT:    ret <4 x i32> [[R]]
 ;
   %sub = sub <4 x i32> <i32 32, i32 32, i32 32, i32 32>, %y
@@ -259,7 +259,7 @@ define <3 x i42> @rotr_v3i42(<3 x i42> %x, <3 x i42> %y) {
 ; CHECK-NEXT:    [[SUB:%.*]] = sub <3 x i42> splat (i42 42), [[Y:%.*]]
 ; CHECK-NEXT:    [[SHL:%.*]] = shl <3 x i42> [[X:%.*]], [[SUB]]
 ; CHECK-NEXT:    [[SHR:%.*]] = lshr <3 x i42> [[X]], [[Y]]
-; CHECK-NEXT:    [[R:%.*]] = or <3 x i42> [[SHR]], [[SHL]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint <3 x i42> [[SHR]], [[SHL]]
 ; CHECK-NEXT:    ret <3 x i42> [[R]]
 ;
   %sub = sub <3 x i42> <i42 42, i42 42, i42 42>, %y
@@ -838,7 +838,7 @@ define i24 @rotl_select_weird_type(i24 %x, i24 %shamt) {
 ; CHECK-NEXT:    [[SUB:%.*]] = sub i24 24, [[SHAMT]]
 ; CHECK-NEXT:    [[SHR:%.*]] = lshr i24 [[X:%.*]], [[SUB]]
 ; CHECK-NEXT:    [[SHL:%.*]] = shl i24 [[X]], [[SHAMT]]
-; CHECK-NEXT:    [[OR:%.*]] = or i24 [[SHL]], [[SHR]]
+; CHECK-NEXT:    [[OR:%.*]] = or disjoint i24 [[SHL]], [[SHR]]
 ; CHECK-NEXT:    [[R:%.*]] = select i1 [[CMP]], i24 [[X]], i24 [[OR]]
 ; CHECK-NEXT:    ret i24 [[R]]
 ;
@@ -981,3 +981,108 @@ define i16 @check_rotate_masked_16bit(i8 %shamt, i32 %cond) {
   %trunc = trunc i32 %or to i16
   ret i16 %trunc
 }
+
+define i32 @rotl_i32_add(i32 %x, i32 %y) {
+; CHECK-LABEL: @rotl_i32_add(
+; CHECK-NEXT:    [[SUB:%.*]] = sub i32 32, [[Y:%.*]]
+; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[X:%.*]], [[Y]]
+; CHECK-NEXT:    [[SHR:%.*]] = lshr i32 [[X]], [[SUB]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i32 [[SHR]], [[SHL]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %sub = sub i32 32, %y
+  %shl = shl i32 %x, %y
+  %shr = lshr i32 %x, %sub
+  %r = add i32 %shr, %shl
+  ret i32 %r
+}
+
+define i32 @rotr_i32_add(i32 %x, i32 %y) {
+; CHECK-LABEL: @rotr_i32_add(
+; CHECK-NEXT:    [[SUB:%.*]] = sub i32 32, [[Y:%.*]]
+; CHECK-NEXT:    [[SHL:%.*]] = lshr i32 [[X:%.*]], [[Y]]
+; CHECK-NEXT:    [[SHR:%.*]] = shl i32 [[X]], [[SUB]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i32 [[SHR]], [[SHL]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %sub = sub i32 32, %y
+  %shl = lshr i32 %x, %y
+  %shr = shl i32 %x, %sub
+  %r = add i32 %shr, %shl
+  ret i32 %r
+}
+
+define i32 @fshr_i32_add(i32 %x, i32 %y, i32 %z) {
+; CHECK-LABEL: @fshr_i32_add(
+; CHECK-NEXT:    [[SUB:%.*]] = sub i32 32, [[Z:%.*]]
+; CHECK-NEXT:    [[SHL:%.*]] = lshr i32 [[X:%.*]], [[Z]]
+; CHECK-NEXT:    [[SHR:%.*]] = shl i32 [[Y:%.*]], [[SUB]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i32 [[SHR]], [[SHL]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %sub = sub i32 32, %z
+  %shl = lshr i32 %x, %z
+  %shr = shl i32 %y, %sub
+  %r = add i32 %shr, %shl
+  ret i32 %r
+}
+
+define i32 @fshl_i32_add(i32 %x, i32 %y, i32 %z) {
+; CHECK-LABEL: @fshl_i32_add(
+; CHECK-NEXT:    [[SUB:%.*]] = sub i32 32, [[Z:%.*]]
+; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[Y:%.*]], [[Z]]
+; CHECK-NEXT:    [[SHR:%.*]] = lshr i32 [[X:%.*]], [[SUB]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i32 [[SHR]], [[SHL]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %sub = sub i32 32, %z
+  %shl = shl i32 %y, %z
+  %shr = lshr i32 %x, %sub
+  %r = add i32 %shr, %shl
+  ret i32 %r
+}
+
+define i32 @rotl_i32_add_greater(i32 %x, i32 %y) {
+; CHECK-LABEL: @rotl_i32_add_greater(
+; CHECK-NEXT:    [[SUB:%.*]] = sub i32 33, [[Y:%.*]]
+; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[X:%.*]], [[Y]]
+; CHECK-NEXT:    [[SHR:%.*]] = lshr i32 [[X]], [[SUB]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i32 [[SHR]], [[SHL]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %sub = sub i32 33, %y
+  %shl = shl i32 %x, %y
+  %shr = lshr i32 %x, %sub
+  %r = add i32 %shr, %shl
+  ret i32 %r
+}
+
+define i32 @rotr_i32_add_greater(i32 %x, i32 %y) {
+; CHECK-LABEL: @rotr_i32_add_greater(
+; CHECK-NEXT:    [[SUB:%.*]] = sub i32 34, [[Y:%.*]]
+; CHECK-NEXT:    [[SHL:%.*]] = lshr i32 [[X:%.*]], [[Y]]
+; CHECK-NEXT:    [[SHR:%.*]] = shl i32 [[X]], [[SUB]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i32 [[SHR]], [[SHL]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %sub = sub i32 34, %y
+  %shl = lshr i32 %x, %y
+  %shr = shl i32 %x, %sub
+  %r = add i32 %shr, %shl
+  ret i32 %r
+}
+
+define i32 @not_rotl_i32_add_less(i32 %x, i32 %y) {
+; CHECK-LABEL: @not_rotl_i32_add_less(
+; CHECK-NEXT:    [[SUB:%.*]] = sub i32 31, [[Y:%.*]]
+; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[X:%.*]], [[Y]]
+; CHECK-NEXT:    [[SHR:%.*]] = lshr i32 [[X]], [[SUB]]
+; CHECK-NEXT:    [[R:%.*]] = add i32 [[SHR]], [[SHL]]
+; CHECK-NEXT:    ret i32 [[R]]
+;
+  %sub = sub i32 31, %y
+  %shl = shl i32 %x, %y
+  %shr = lshr i32 %x, %sub
+  %r = add i32 %shr, %shl
+  ret i32 %r
+}