[InstSimplify] Fold and A, (zext (icmp eq A, 0)) into 0
#66676
Conversation
|
@llvm/pr-subscribers-llvm-transforms @llvm/pr-subscribers-llvm-analysis ChangesThis patch folds the pattern Full diff: https://github.com/llvm/llvm-project/pull/66676.diff 2 Files Affected:
diff --git a/llvm/lib/Analysis/InstructionSimplify.cpp b/llvm/lib/Analysis/InstructionSimplify.cpp
index e8f96e9f681f2d5..a835ca763e6565d 100644
--- a/llvm/lib/Analysis/InstructionSimplify.cpp
+++ b/llvm/lib/Analysis/InstructionSimplify.cpp
@@ -2079,6 +2079,13 @@ static Value *simplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
if (match(Op0, m_Not(m_Specific(Op1))) || match(Op1, m_Not(m_Specific(Op0))))
return Constant::getNullValue(Op0->getType());
+ // A & !A = !A & A = 0
+ ICmpInst::Predicate EqPred;
+ if ((match(Op0, m_ZExt(m_ICmp(EqPred, m_Specific(Op1), m_Zero()))) ||
+ match(Op1, m_ZExt(m_ICmp(EqPred, m_Specific(Op0), m_Zero())))) &&
+ EqPred == ICmpInst::ICMP_EQ)
+ return Constant::getNullValue(Op0->getType());
+
// (A | ?) & A = A
if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
return Op1;
diff --git a/llvm/test/Transforms/InstSimplify/and-or-icmp-zero.ll b/llvm/test/Transforms/InstSimplify/and-or-icmp-zero.ll
index 3e50a5968b46087..022add2c4717419 100644
--- a/llvm/test/Transforms/InstSimplify/and-or-icmp-zero.ll
+++ b/llvm/test/Transforms/InstSimplify/and-or-icmp-zero.ll
@@ -265,10 +265,7 @@ define i1 @and_cmps_ptr_eq_zero_with_mask_commute4(ptr %p, i64 %y) {
; tests from PR66606
define i32 @and_zext_eq_zero(i32 %a) {
; CHECK-LABEL: @and_zext_eq_zero(
-; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[A:%.*]], 0
-; CHECK-NEXT: [[NOT:%.*]] = zext i1 [[COND]] to i32
-; CHECK-NEXT: [[R:%.*]] = and i32 [[A]], [[NOT]]
-; CHECK-NEXT: ret i32 [[R]]
+; CHECK-NEXT: ret i32 0
;
%cond = icmp eq i32 %a, 0
%not = zext i1 %cond to i32
@@ -278,10 +275,7 @@ define i32 @and_zext_eq_zero(i32 %a) {
define i32 @and_zext_eq_zero_commuted(i32 %a) {
; CHECK-LABEL: @and_zext_eq_zero_commuted(
-; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[A:%.*]], 0
-; CHECK-NEXT: [[NOT:%.*]] = zext i1 [[COND]] to i32
-; CHECK-NEXT: [[R:%.*]] = and i32 [[NOT]], [[A]]
-; CHECK-NEXT: ret i32 [[R]]
+; CHECK-NEXT: ret i32 0
;
%cond = icmp eq i32 %a, 0
%not = zext i1 %cond to i32
|
There was a problem hiding this comment.
The general pattern here would be zext(c) & a, which is c ? (a & 1) : 0.
For the special case of c := a == C, if C is even then this reduces to 0, if C is odd it reduces to zext(c).
Possibly we shouldn't special case the a == C case though and just generally canonicalize to the select form, on which all possible folds can then be performed?
In any case, this needs to at least handle arbitrary constants.
|
|
The canonicalization is posted as #66740. |
#66740) This patch canonicalizes the pattern `and(zext(A), B)` into `select A, B & 1, 0`. Thus, we can reuse transforms `select B == even, B & 1, 0 -> 0` and `select B == odd, B & 1, 0 -> zext(B == odd)` in `InstCombine`. It is an alternative to #66676. Alive2: https://alive2.llvm.org/ce/z/598phE Fixes #66733. Fixes #66606. Fixes #28612.
llvm#66740) This patch canonicalizes the pattern `and(zext(A), B)` into `select A, B & 1, 0`. Thus, we can reuse transforms `select B == even, B & 1, 0 -> 0` and `select B == odd, B & 1, 0 -> zext(B == odd)` in `InstCombine`. It is an alternative to llvm#66676. Alive2: https://alive2.llvm.org/ce/z/598phE Fixes llvm#66733. Fixes llvm#66606. Fixes llvm#28612.
This patch folds the pattern
and A, (zext (icmp eq A, 0))into 0.Fixes #66606.