[InstCombine] Simplify and/or of icmp eq with op replacement #70335

nikic · 2023-10-26T14:05:21Z

and/or in logical (select) form benefit from generic simplifications via simplifyWithOpReplaced(). However, the corresponding fold for plain and/or currently does not exist.

Similar to selects, there are two general cases for this fold (illustrated with and, but there are or conjugates).

The basic case is something like (a == b) & c, where the replacement of a with b or b with a inside c allows it to fold to true or false. Then the whole operation will fold to either false or a == b.

The second case is something like (a != b) & c, where the replacement inside c allows it to fold to false. In that case, the operand can be replaced with c, because in the case where a == b (and thus the icmp is false), c itself will already be false.

As the test diffs show, this catches quite a lot of patterns in existing test coverage. This also obsoletes quite a few existing special-case and/or of icmp folds we have (e.g. simplifyAndOrOfICmpsWithLimitConst), but I haven't removed anything as part of this patch in the interest of risk mitigation.

Fixes #69050.
Fixes #69091.

llvmbot · 2023-10-26T14:06:34Z

@llvm/pr-subscribers-pgo
@llvm/pr-subscribers-llvm-analysis

@llvm/pr-subscribers-llvm-transforms

Author: Nikita Popov (nikic)

Changes

and/or in logical (select) form benefit from generic simplifications via simplifyWithOpReplaced(). However, the corresponding fold for plain and/or currently does not exist.

Similar to selects, there are two general cases for this fold (illustrated with and, but there are or conjugates).

The basic case is something like (a == b) & c, where the replacement of a with b or b with a inside c allows it to fold to true or false. Then the whole operation will fold to either false or a == b.

The second case is something like (a != b) & c, where the replacement inside c allows it to fold to false. In that case, the operand can be replaced with c, because in the case where a == b (and thus the icmp is false), c itself will already be false.

As the test diffs show, this catches quite a lot of patterns in existing test coverage. I believe this also obsoletes quite a few existing special-case and/or of icmp special case folds we have, but I haven't removed anything as part of this patch in the interest of risk mitigation.

Patch is 70.25 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/70335.diff

9 Files Affected:

(modified) llvm/lib/Analysis/InstructionSimplify.cpp (+58)
(modified) llvm/test/Transforms/InstCombine/div-by-0-guard-before-smul_ov.ll (+2-9)
(modified) llvm/test/Transforms/InstCombine/div-by-0-guard-before-umul_ov.ll (+2-9)
(modified) llvm/test/Transforms/InstCombine/ispow2.ll (+5-19)
(modified) llvm/test/Transforms/InstSimplify/and-or-icmp-ctpop.ll (+2-8)
(modified) llvm/test/Transforms/InstSimplify/and-or-icmp-min-max.ll (+150-467)
(modified) llvm/test/Transforms/InstSimplify/div-by-0-guard-before-smul_ov.ll (+2-9)
(modified) llvm/test/Transforms/InstSimplify/div-by-0-guard-before-umul_ov.ll (+2-9)
(modified) llvm/test/Transforms/InstSimplify/result-of-add-of-negative-is-non-zero-and-no-underflow.ll (+4-16)

diff --git a/llvm/lib/Analysis/InstructionSimplify.cpp b/llvm/lib/Analysis/InstructionSimplify.cpp
index 3d192d0759a1e67..ad3f89a3e93127f 100644
--- a/llvm/lib/Analysis/InstructionSimplify.cpp
+++ b/llvm/lib/Analysis/InstructionSimplify.cpp
@@ -2025,6 +2025,50 @@ static Value *simplifyAndOrOfCmps(const SimplifyQuery &Q, Value *Op0,
   return nullptr;
 }
 
+static Value *simplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
+                                     const SimplifyQuery &Q,
+                                     bool AllowRefinement,
+                                     SmallVectorImpl<Instruction *> *DropFlags,
+                                     unsigned MaxRecurse);
+
+static Value *simplifyAndOrWithICmpEq(unsigned Opcode, Value *Op0, Value *Op1,
+                                      const SimplifyQuery &Q,
+                                      unsigned MaxRecurse) {
+  assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
+         "Must be and/or");
+  ICmpInst::Predicate Pred;
+  Value *A, *B;
+  if (!match(Op0, m_ICmp(Pred, m_Value(A), m_Value(B))) ||
+      !ICmpInst::isEquality(Pred) || !MaxRecurse--)
+    return nullptr;
+
+  auto Simplify = [&](Value *Res) -> Value * {
+    // and (icmp eq a, b), x implies (a==b) inside x.
+    // or (icmp ne a, b), x implies (a==b) inside x.
+    // If x simplifies to true/false, we can simplify the and/or.
+    if (Pred ==
+        (Opcode == Instruction::And ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE))
+      return simplifyBinOp(Opcode, Op0, Res, Q, MaxRecurse);
+    // If we have and (icmp ne a, b), x and for a==b we can simplify x to false,
+    // then we can drop the icmp, as x will already be false in the case where
+    // the icmp is false. Similar for or and true.
+    if (Res == ConstantExpr::getBinOpAbsorber(Opcode, Res->getType()))
+      return Op1;
+    return nullptr;
+  };
+
+  if (Value *Res =
+          simplifyWithOpReplaced(Op1, A, B, Q, /* AllowRefinement */ true,
+                                 /* DropFlags */ nullptr, MaxRecurse))
+    return Simplify(Res);
+  if (Value *Res =
+          simplifyWithOpReplaced(Op1, B, A, Q, /* AllowRefinement */ true,
+                                 /* DropFlags */ nullptr, MaxRecurse))
+    return Simplify(Res);
+
+  return nullptr;
+}
+
 /// Given a bitwise logic op, check if the operands are add/sub with a common
 /// source value and inverted constant (identity: C - X -> ~(X + ~C)).
 static Value *simplifyLogicOfAddSub(Value *Op0, Value *Op1,
@@ -2159,6 +2203,13 @@ static Value *simplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
       isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT))
     return Constant::getNullValue(Op0->getType());
 
+  if (Value *V = simplifyAndOrWithICmpEq(Instruction::And, Op0, Op1, Q,
+                                         MaxRecurse))
+    return V;
+  if (Value *V = simplifyAndOrWithICmpEq(Instruction::And, Op1, Op0, Q,
+                                         MaxRecurse))
+    return V;
+
   if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
     return V;
 
@@ -2435,6 +2486,13 @@ static Value *simplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
       match(Op0, m_LShr(m_Specific(X), m_Specific(Y))))
     return Op1;
 
+  if (Value *V =
+          simplifyAndOrWithICmpEq(Instruction::Or, Op0, Op1, Q, MaxRecurse))
+    return V;
+  if (Value *V =
+          simplifyAndOrWithICmpEq(Instruction::Or, Op1, Op0, Q, MaxRecurse))
+    return V;
+
   if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
     return V;
 
diff --git a/llvm/test/Transforms/InstCombine/div-by-0-guard-before-smul_ov.ll b/llvm/test/Transforms/InstCombine/div-by-0-guard-before-smul_ov.ll
index 23bfc75b945ba10..08eefbebb736340 100644
--- a/llvm/test/Transforms/InstCombine/div-by-0-guard-before-smul_ov.ll
+++ b/llvm/test/Transforms/InstCombine/div-by-0-guard-before-smul_ov.ll
@@ -47,11 +47,7 @@ define i1 @n2_wrong_size(i4 %size0, i4 %size1, i4 %nmemb) {
 
 define i1 @n3_wrong_pred(i4 %size, i4 %nmemb) {
 ; CHECK-LABEL: @n3_wrong_pred(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i4 [[SIZE:%.*]], 0
-; CHECK-NEXT:    [[SMUL:%.*]] = tail call { i4, i1 } @llvm.smul.with.overflow.i4(i4 [[SIZE]], i4 [[NMEMB:%.*]])
-; CHECK-NEXT:    [[SMUL_OV:%.*]] = extractvalue { i4, i1 } [[SMUL]], 1
-; CHECK-NEXT:    [[AND:%.*]] = and i1 [[SMUL_OV]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[AND]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp eq i4 %size, 0 ; not 'ne'
   %smul = tail call { i4, i1 } @llvm.smul.with.overflow.i4(i4 %size, i4 %nmemb)
@@ -63,10 +59,7 @@ define i1 @n3_wrong_pred(i4 %size, i4 %nmemb) {
 define i1 @n4_not_and(i4 %size, i4 %nmemb) {
 ; CHECK-LABEL: @n4_not_and(
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i4 [[SIZE:%.*]], 0
-; CHECK-NEXT:    [[SMUL:%.*]] = tail call { i4, i1 } @llvm.smul.with.overflow.i4(i4 [[SIZE]], i4 [[NMEMB:%.*]])
-; CHECK-NEXT:    [[SMUL_OV:%.*]] = extractvalue { i4, i1 } [[SMUL]], 1
-; CHECK-NEXT:    [[AND:%.*]] = or i1 [[SMUL_OV]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[AND]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %cmp = icmp ne i4 %size, 0
   %smul = tail call { i4, i1 } @llvm.smul.with.overflow.i4(i4 %size, i4 %nmemb)
diff --git a/llvm/test/Transforms/InstCombine/div-by-0-guard-before-umul_ov.ll b/llvm/test/Transforms/InstCombine/div-by-0-guard-before-umul_ov.ll
index dbc3b5e7a25be38..047f8855fe5cb81 100644
--- a/llvm/test/Transforms/InstCombine/div-by-0-guard-before-umul_ov.ll
+++ b/llvm/test/Transforms/InstCombine/div-by-0-guard-before-umul_ov.ll
@@ -47,11 +47,7 @@ define i1 @n2_wrong_size(i4 %size0, i4 %size1, i4 %nmemb) {
 
 define i1 @n3_wrong_pred(i4 %size, i4 %nmemb) {
 ; CHECK-LABEL: @n3_wrong_pred(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i4 [[SIZE:%.*]], 0
-; CHECK-NEXT:    [[UMUL:%.*]] = tail call { i4, i1 } @llvm.umul.with.overflow.i4(i4 [[SIZE]], i4 [[NMEMB:%.*]])
-; CHECK-NEXT:    [[UMUL_OV:%.*]] = extractvalue { i4, i1 } [[UMUL]], 1
-; CHECK-NEXT:    [[AND:%.*]] = and i1 [[UMUL_OV]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[AND]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp eq i4 %size, 0 ; not 'ne'
   %umul = tail call { i4, i1 } @llvm.umul.with.overflow.i4(i4 %size, i4 %nmemb)
@@ -63,10 +59,7 @@ define i1 @n3_wrong_pred(i4 %size, i4 %nmemb) {
 define i1 @n4_not_and(i4 %size, i4 %nmemb) {
 ; CHECK-LABEL: @n4_not_and(
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i4 [[SIZE:%.*]], 0
-; CHECK-NEXT:    [[UMUL:%.*]] = tail call { i4, i1 } @llvm.umul.with.overflow.i4(i4 [[SIZE]], i4 [[NMEMB:%.*]])
-; CHECK-NEXT:    [[UMUL_OV:%.*]] = extractvalue { i4, i1 } [[UMUL]], 1
-; CHECK-NEXT:    [[AND:%.*]] = or i1 [[UMUL_OV]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[AND]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %cmp = icmp ne i4 %size, 0
   %umul = tail call { i4, i1 } @llvm.umul.with.overflow.i4(i4 %size, i4 %nmemb)
diff --git a/llvm/test/Transforms/InstCombine/ispow2.ll b/llvm/test/Transforms/InstCombine/ispow2.ll
index 740f79cd32b39e8..cc50c5cd1e6680a 100644
--- a/llvm/test/Transforms/InstCombine/ispow2.ll
+++ b/llvm/test/Transforms/InstCombine/ispow2.ll
@@ -392,9 +392,7 @@ define i1 @is_pow2_ctpop_wrong_pred1(i32 %x) {
 ; CHECK-LABEL: @is_pow2_ctpop_wrong_pred1(
 ; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i32 [[T0]], 2
-; CHECK-NEXT:    [[NOTZERO:%.*]] = icmp ne i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[NOTZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp ugt i32 %t0, 2
@@ -946,9 +944,7 @@ define i1 @is_pow2or0_ctpop_wrong_pred1(i32 %x) {
 ; CHECK-LABEL: @is_pow2or0_ctpop_wrong_pred1(
 ; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[T0]], 1
-; CHECK-NEXT:    [[ISZERO:%.*]] = icmp eq i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[ISZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp ne i32 %t0, 1
@@ -959,11 +955,7 @@ define i1 @is_pow2or0_ctpop_wrong_pred1(i32 %x) {
 
 define i1 @is_pow2or0_ctpop_wrong_pred2(i32 %x) {
 ; CHECK-LABEL: @is_pow2or0_ctpop_wrong_pred2(
-; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[T0]], 1
-; CHECK-NEXT:    [[ISZERO:%.*]] = icmp ne i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[ISZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp ne i32 %t0, 1
@@ -1149,9 +1141,7 @@ define i1 @isnot_pow2nor0_ctpop_wrong_pred1(i32 %x) {
 ; CHECK-LABEL: @isnot_pow2nor0_ctpop_wrong_pred1(
 ; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[T0]], 1
-; CHECK-NEXT:    [[NOTZERO:%.*]] = icmp ne i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[NOTZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp eq i32 %t0, 1
@@ -1162,11 +1152,7 @@ define i1 @isnot_pow2nor0_ctpop_wrong_pred1(i32 %x) {
 
 define i1 @isnot_pow2nor0_ctpop_wrong_pred2(i32 %x) {
 ; CHECK-LABEL: @isnot_pow2nor0_ctpop_wrong_pred2(
-; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[T0]], 1
-; CHECK-NEXT:    [[NOTZERO:%.*]] = icmp eq i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[NOTZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp eq i32 %t0, 1
diff --git a/llvm/test/Transforms/InstSimplify/and-or-icmp-ctpop.ll b/llvm/test/Transforms/InstSimplify/and-or-icmp-ctpop.ll
index 6de97c3a7a76deb..6fe8d29bd10bf5f 100644
--- a/llvm/test/Transforms/InstSimplify/and-or-icmp-ctpop.ll
+++ b/llvm/test/Transforms/InstSimplify/and-or-icmp-ctpop.ll
@@ -40,11 +40,7 @@ define <2 x i1> @eq_or_non_0_commute(<2 x i32> %x) {
 
 define i1 @eq_or_non_0_wrong_pred1(i32 %x) {
 ; CHECK-LABEL: @eq_or_non_0_wrong_pred1(
-; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]])
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[T0]], 10
-; CHECK-NEXT:    [[NOTZERO:%.*]] = icmp ne i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[NOTZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp ne i32 %t0, 10
@@ -90,9 +86,7 @@ define i1 @ne_and_is_0_wrong_pred1(i32 %x) {
 ; CHECK-LABEL: @ne_and_is_0_wrong_pred1(
 ; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]])
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[T0]], 10
-; CHECK-NEXT:    [[ISZERO:%.*]] = icmp eq i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[ISZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp ne i32 %t0, 10
diff --git a/llvm/test/Transforms/InstSimplify/and-or-icmp-min-max.ll b/llvm/test/Transforms/InstSimplify/and-or-icmp-min-max.ll
index 7ea1797c99898fd..4e3832f31e5a4c6 100644
--- a/llvm/test/Transforms/InstSimplify/and-or-icmp-min-max.ll
+++ b/llvm/test/Transforms/InstSimplify/and-or-icmp-min-max.ll
@@ -16,10 +16,7 @@
 
 define i1 @slt_and_max(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @slt_and_max(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i8 [[X]], 127
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMP]], [[CMPEQ]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp slt i8 %x, %y
   %cmpeq = icmp eq i8 %x, 127
@@ -29,10 +26,7 @@ define i1 @slt_and_max(i8 %x, i8 %y)  {
 
 define <2 x i1> @slt_and_max_commute(<2 x i8> %x, <2 x i8> %y)  {
 ; CHECK-LABEL: @slt_and_max_commute(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt <2 x i8> [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq <2 x i8> [[X]], <i8 127, i8 127>
-; CHECK-NEXT:    [[R:%.*]] = and <2 x i1> [[CMPEQ]], [[CMP]]
-; CHECK-NEXT:    ret <2 x i1> [[R]]
+; CHECK-NEXT:    ret <2 x i1> zeroinitializer
 ;
   %cmp = icmp slt <2 x i8> %x, %y
   %cmpeq = icmp eq <2 x i8> %x, <i8 127, i8 127>
@@ -42,10 +36,7 @@ define <2 x i1> @slt_and_max_commute(<2 x i8> %x, <2 x i8> %y)  {
 
 define i1 @slt_swap_and_max(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @slt_swap_and_max(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i8 [[X]], 127
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMP]], [[CMPEQ]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp sgt i8 %y, %x
   %cmpeq = icmp eq i8 %x, 127
@@ -55,10 +46,7 @@ define i1 @slt_swap_and_max(i8 %x, i8 %y)  {
 
 define i1 @slt_swap_and_max_commute(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @slt_swap_and_max_commute(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i8 [[X]], 127
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMPEQ]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp sgt i8 %y, %x
   %cmpeq = icmp eq i8 %x, 127
@@ -68,10 +56,7 @@ define i1 @slt_swap_and_max_commute(i8 %x, i8 %y)  {
 
 define i1 @ult_and_max(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @ult_and_max(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i8 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i8 [[X]], -1
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMP]], [[CMPEQ]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp ult i8 %x, %y
   %cmpeq = icmp eq i8 %x, 255
@@ -81,10 +66,7 @@ define i1 @ult_and_max(i8 %x, i8 %y)  {
 
 define i1 @ult_and_max_commute(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @ult_and_max_commute(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i8 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i8 [[X]], -1
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMPEQ]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp ult i8 %x, %y
   %cmpeq = icmp eq i8 %x, 255
@@ -94,10 +76,7 @@ define i1 @ult_and_max_commute(i8 %x, i8 %y)  {
 
 define i1 @ult_swap_and_max(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @ult_swap_and_max(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i8 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i8 [[X]], -1
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMP]], [[CMPEQ]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp ugt i8 %y, %x
   %cmpeq = icmp eq i8 %x, 255
@@ -107,10 +86,7 @@ define i1 @ult_swap_and_max(i8 %x, i8 %y)  {
 
 define i1 @ult_swap_and_max_commute(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @ult_swap_and_max_commute(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i8 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i8 [[X]], -1
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMPEQ]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp ugt i8 %y, %x
   %cmpeq = icmp eq i8 %x, 255
@@ -126,10 +102,7 @@ define i1 @ult_swap_and_max_commute(i8 %x, i8 %y)  {
 
 define i1 @sgt_and_min(i9 %x, i9 %y)  {
 ; CHECK-LABEL: @sgt_and_min(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i9 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i9 [[X]], -256
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMP]], [[CMPEQ]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp sgt i9 %x, %y
   %cmpeq = icmp eq i9 %x, 256
@@ -139,10 +112,7 @@ define i1 @sgt_and_min(i9 %x, i9 %y)  {
 
 define i1 @sgt_and_min_commute(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @sgt_and_min_commute(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i8 [[X]], -128
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMPEQ]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp sgt i8 %x, %y
   %cmpeq = icmp eq i8 %x, 128
@@ -152,10 +122,7 @@ define i1 @sgt_and_min_commute(i8 %x, i8 %y)  {
 
 define i1 @sgt_swap_and_min(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @sgt_swap_and_min(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i8 [[X]], -128
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMP]], [[CMPEQ]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp slt i8 %y, %x
   %cmpeq = icmp eq i8 %x, 128
@@ -165,10 +132,7 @@ define i1 @sgt_swap_and_min(i8 %x, i8 %y)  {
 
 define i1 @sgt_swap_and_min_commute(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @sgt_swap_and_min_commute(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i8 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp eq i8 [[X]], -128
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[CMPEQ]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp slt i8 %y, %x
   %cmpeq = icmp eq i8 %x, 128
@@ -224,10 +188,7 @@ define i1 @ugt_swap_and_min_commute(i8 %x, i8 %y)  {
 
 define i1 @sge_or_not_max(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @sge_or_not_max(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sge i8 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp ne i8 [[X]], 127
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[CMP]], [[CMPEQ]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %cmp = icmp sge i8 %x, %y
   %cmpeq = icmp ne i8 %x, 127
@@ -237,10 +198,7 @@ define i1 @sge_or_not_max(i8 %x, i8 %y)  {
 
 define i1 @sge_or_not_max_commute(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @sge_or_not_max_commute(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sge i8 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp ne i8 [[X]], 127
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[CMPEQ]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %cmp = icmp sge i8 %x, %y
   %cmpeq = icmp ne i8 %x, 127
@@ -250,10 +208,7 @@ define i1 @sge_or_not_max_commute(i8 %x, i8 %y)  {
 
 define i1 @sge_swap_or_not_max(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @sge_swap_or_not_max(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sle i8 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp ne i8 [[X]], 127
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[CMP]], [[CMPEQ]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %cmp = icmp sle i8 %y, %x
   %cmpeq = icmp ne i8 %x, 127
@@ -263,10 +218,7 @@ define i1 @sge_swap_or_not_max(i8 %x, i8 %y)  {
 
 define i1 @sge_swap_or_not_max_commute(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @sge_swap_or_not_max_commute(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sle i8 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp ne i8 [[X]], 127
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[CMPEQ]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %cmp = icmp sle i8 %y, %x
   %cmpeq = icmp ne i8 %x, 127
@@ -276,10 +228,7 @@ define i1 @sge_swap_or_not_max_commute(i8 %x, i8 %y)  {
 
 define i1 @uge_or_not_max(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @uge_or_not_max(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp uge i8 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp ne i8 [[X]], -1
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[CMP]], [[CMPEQ]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %cmp = icmp uge i8 %x, %y
   %cmpeq = icmp ne i8 %x, 255
@@ -289,10 +238,7 @@ define i1 @uge_or_not_max(i8 %x, i8 %y)  {
 
 define i1 @uge_or_not_max_commute(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @uge_or_not_max_commute(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp uge i8 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp ne i8 [[X]], -1
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[CMPEQ]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %cmp = icmp uge i8 %x, %y
   %cmpeq = icmp ne i8 %x, 255
@@ -302,10 +248,7 @@ define i1 @uge_or_not_max_commute(i8 %x, i8 %y)  {
 
 define i1 @uge_swap_or_not_max(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @uge_swap_or_not_max(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ule i8 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp ne i8 [[X]], -1
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[CMP]], [[CMPEQ]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %cmp = icmp ule i8 %y, %x
   %cmpeq = icmp ne i8 %x, 255
@@ -315,10 +258,7 @@ define i1 @uge_swap_or_not_max(i8 %x, i8 %y)  {
 
 define i1 @uge_swap_or_not_max_commute(i8 %x, i8 %y)  {
 ; CHECK-LABEL: @uge_swap_or_not_max_commute(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ule i8 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[CMPEQ:%.*]] = icmp ne i8 [[X]], -1
-; CHECK-NE...
[truncated]

github-actions · 2023-10-26T14:18:03Z

✅ With the latest revision this PR passed the C/C++ code formatter.

llvm/lib/Analysis/InstructionSimplify.cpp

dtcxzyw · 2023-10-26T18:27:23Z

It looks like your PR does similar optimizations as #69840. I will rebase it after landing this PR to see other missed optimizations.

dtcxzyw · 2023-10-26T18:31:28Z

Could you please check whether this PR fixes #68799, #69038, #69050 and #69091?

nikic · 2023-10-27T07:42:51Z

Could you please check whether this PR fixes #68799, #69038

These two are not fixed. They would need an extension like this one:

llvm-project/llvm/lib/Analysis/InstructionSimplify.cpp

Lines 4692 to 4716 in fb619b3

    
           Value *X; 
        
           Value *Y; 
        
           // select((X | Y) == 0 ?  X : 0) --> 0 (commuted 2 ways) 
        
           if (match(CmpLHS, m_Or(m_Value(X), m_Value(Y))) && 
        
               match(CmpRHS, m_Zero())) { 
        
             // (X | Y) == 0 implies X == 0 and Y == 0. 
        
             if (Value *V = simplifySelectWithICmpEq(X, CmpRHS, TrueVal, FalseVal, Q, 
        
                                                     MaxRecurse)) 
        
               return V; 
        
             if (Value *V = simplifySelectWithICmpEq(Y, CmpRHS, TrueVal, FalseVal, Q, 
        
                                                     MaxRecurse)) 
        
               return V; 
        
           } 
        
           // select((X & Y) == -1 ?  X : -1) --> -1 (commuted 2 ways) 
        
           if (match(CmpLHS, m_And(m_Value(X), m_Value(Y))) && 
        
               match(CmpRHS, m_AllOnes())) { 
        
             // (X & Y) == -1 implies X == -1 and Y == -1. 
        
             if (Value *V = simplifySelectWithICmpEq(X, CmpRHS, TrueVal, FalseVal, Q, 
        
                                                     MaxRecurse)) 
        
               return V; 
        
             if (Value *V = simplifySelectWithICmpEq(Y, CmpRHS, TrueVal, FalseVal, Q, 
        
                                                     MaxRecurse)) 
        
               return V; 
        
           }

, #69050 and #69091?

These are also not fixed, but only because of the recursion limit. I think the way I implemented it may be a bit too strict, as we'll decrement MaxRecurse in simplifyAndOrWithICmpEq() and again in simplifyWithOpReplaced().

nikic · 2023-10-27T09:07:15Z

I've slightly adjusted the handling of the recursion limit and added test cases from the two issues.

and/or in logical (select) form benefit from generic simplifications via simplifyWithOpReplaced(). However, the corresponding fold for plain and/or currently does not exist. Similar to selects, there are two general cases for this fold (illustrated with `and`, but there are `or` conjugates). The basic case is something like `(a == b) & c`, where the replacement of a with b or b with a inside c allows it to fold to true or false. Then the whole operation will fold to either false or `a == b`. The second case is something like `(a != b) & c`, where the replacement inside c allows it to fold to false. In that case, the operand can be replaced with c, because in the case where a == b (and thus the icmp is false), c itself will already be false. As the test diffs show, this catches quite a lot of patterns in existing test coverage. I believe this also obsoletes quite a few existing special-case and/or of icmp special case folds we have, but I haven't removed anything as part of this patch in the interest of risk mitigation.

nikic · 2023-10-27T09:12:17Z

llvm/test/Transforms/PGOProfile/chr.ll

 ; CHECK-NEXT:    [[V0:%.*]] = icmp eq i32 [[Z_FR]], 0
-; CHECK-NEXT:    [[V3_NONCHR:%.*]] = and i1 [[V0]], [[PRED_FR]]
+; CHECK-NEXT:    [[V3_NONCHR:%.*]] = and i1 [[V0]], [[PRED:%.*]]


What happens here is that now a frozen condition gets optimized away earlier, before freeze is propagated into it.

dtcxzyw

LGTM. Thanks!

goldsteinn · 2023-10-27T15:26:25Z

llvm/lib/Analysis/InstructionSimplify.cpp

+
+  auto Simplify = [&](Value *Res) -> Value * {
+    // and (icmp eq a, b), x implies (a==b) inside x.
+    // or (icmp ne a, b), x implies (a==b) inside x.


Do you get any benefit (and icmp ne (a, b), x) and use a != b inside x? (Likewise for eq + or?

If we had a way to say "simplify x under the assumption that a != b", I would expect that to be beneficial. But we can only do this for equalities right now, as in that case we can just replace operands.

I think we need some generic way to say "evaluate this predicate under the assumption that v has range CR" -- some of @dtcxzyw's recent PRs go in that direction, but I feel like there must be a more generic solution to this, which integrates with existing range propagation logic, instead of reimplementing it in parts. This seems like something that LVI/CVP should be doing, but I don't see a good way to integrate it there.

goldsteinn · 2023-10-27T15:28:52Z

llvm/lib/Analysis/InstructionSimplify.cpp

+  ICmpInst::Predicate Pred;
+  Value *A, *B;
+  if (!match(Op0, m_ICmp(Pred, m_Value(A), m_Value(B))) ||
+      !ICmpInst::isEquality(Pred) || !MaxRecurse--)


nit: Would have recursive check first.

…70335)" This reverts commit 1770a2e. Stage 2 llvm-tblgen crashes when generating X86GenAsmWriter.inc and other files.

nikic · 2023-10-31T15:55:20Z

I reverted this because I'm seeing segfaults in stage2 llvm-tblgen on a build server, but for some reason I can't reproduce this in local stage2 builds :/ Not sure what the difference is, my best guess is the libstdc++ version.

mikaelholmen · 2023-11-01T14:13:59Z

I reverted this because I'm seeing segfaults in stage2 llvm-tblgen on a build server, but for some reason I can't reproduce this in local stage2 builds :/ Not sure what the difference is, my best guess is the libstdc++ version.

@nikic
We saw miscompiles with this patch when testing our out-of-tree target. I don't have any reproducer to share at the moment but if you're struggling to reproduce yourself I can try to see if I might be able to extract something that shows the problem on trunk as well. Should I try?

nikic · 2023-11-01T15:21:41Z

I reverted this because I'm seeing segfaults in stage2 llvm-tblgen on a build server, but for some reason I can't reproduce this in local stage2 builds :/ Not sure what the difference is, my best guess is the libstdc++ version.

@nikic We saw miscompiles with this patch when testing our out-of-tree target. I don't have any reproducer to share at the moment but if you're struggling to reproduce yourself I can try to see if I might be able to extract something that shows the problem on trunk as well. Should I try?

If it's not too much effort, I'd really appreciate that!

mikaelholmen · 2023-11-02T05:22:15Z

Ok. I think this exposes a miscompile with this patch:
opt -passes=instcombine ic.ll -o - -S
With the patch we get

land.end:                                         ; preds = %for.body
  br i1 poison, label %cond.false, label %cond.end

which looks broken to me.
It can surely be further reduced but I didn't want to risk messing something up now.
ic.ll.gz

nikic · 2023-11-02T10:35:43Z

@mikaelholmen Thank you! Here is a reduced version:

declare void @barrier()

define i1 @test1(i32 %x) {
  %cmp1 = icmp ne i32 %x, 0
  call void @barrier()
  %div = udiv i32 2147483647, %x
  %cmp2 = icmp ugt i32 %x, %div
  %or = or i1 %cmp1, %cmp2
  ret i1 %or
}

What happens is that we try to evaluate %cmp2 with %x==0, in which case the division is UB and folds to poison. We then fold or with a poison operand, resulting in an overall poison result.

Another slight variant of this would be:

define i1 @test2(i32 %x) {
  %cmp1 = icmp ne i32 %x, 32 
  %shl = shl i32 1, %x
  %cmp2 = icmp ugt i32 %x, %shl 
  %or = or i1 %cmp1, %cmp2
  ret i1 %or
}

Same basic issue, but without UB involvement, just plain poison.

These were miscompiled with the initial version of the patch.

nikic · 2023-11-02T11:49:15Z

I've tried the following variant that only checks for true/false results (d8ea7f2). Unfortunately, this still causes stage2 llvm-tblgen crashes, so this wasn't the only issue :(

mikaelholmen · 2023-11-02T11:54:59Z

@mikaelholmen Thank you! Here is a reduced version:
declare void @barrier()

define i1 @test1(i32 %x) {
  %cmp1 = icmp ne i32 %x, 0
  call void @barrier()
  %div = udiv i32 2147483647, %x
  %cmp2 = icmp ugt i32 %x, %div
  %or = or i1 %cmp1, %cmp2
  ret i1 %or
}
What happens is that we try to evaluate %cmp2 with %x==0, in which case the division is UB and folds to poison. We then fold or with a poison operand, resulting in an overall poison result.

But... the @test1 example is indeed UB in case %x is 0, right? Since the div is always carried out, even if %x is 0, and then the result is used to calculate the return value %or (via icmp and or)? Or what am I missing here?

In the original ic.ll example the udiv was avoided if the denominator was 0 and UB was avoided.

nikic · 2023-11-02T11:57:56Z

@mikaelholmen Thank you! Here is a reduced version:
declare void @barrier()

define i1 @test1(i32 %x) {
  %cmp1 = icmp ne i32 %x, 0
  call void @barrier()
  %div = udiv i32 2147483647, %x
  %cmp2 = icmp ugt i32 %x, %div
  %or = or i1 %cmp1, %cmp2
  ret i1 %or
}
What happens is that we try to evaluate %cmp2 with %x==0, in which case the division is UB and folds to poison. We then fold or with a poison operand, resulting in an overall poison result.
But... the @test1 example is indeed UB in case %x is 0, right? Since the div is always carried out, even if %x is 0, and then the result is used to calculate the return value %or (via icmp and or)? Or what am I missing here?

In the original ic.ll example the udiv was avoided if the denominator was 0 and UB was avoided.

If %x is 0 there is indeed UB. But if %x is not zero, then the IR is well-defined and will always return true, so the fold to poison is incorrect.

mikaelholmen · 2023-11-02T12:00:11Z

@mikaelholmen Thank you! Here is a reduced version:
declare void @barrier()

define i1 @test1(i32 %x) {
  %cmp1 = icmp ne i32 %x, 0
  call void @barrier()
  %div = udiv i32 2147483647, %x
  %cmp2 = icmp ugt i32 %x, %div
  %or = or i1 %cmp1, %cmp2
  ret i1 %or
}
What happens is that we try to evaluate %cmp2 with %x==0, in which case the division is UB and folds to poison. We then fold or with a poison operand, resulting in an overall poison result.
But... the @test1 example is indeed UB in case %x is 0, right? Since the div is always carried out, even if %x is 0, and then the result is used to calculate the return value %or (via icmp and or)? Or what am I missing here?
In the original ic.ll example the udiv was avoided if the denominator was 0 and UB was avoided.
If %x is 0 there is indeed UB. But if %x is not zero, then the IR is well-defined and will always return true, so the fold to poison is incorrect.

Ah, ye it folded the whole thing to poison regardless of %x. I follow. Thanks!

mikaelholmen · 2023-11-02T12:09:26Z

I've tried the following variant that only checks for true/false results (d8ea7f2). Unfortunately, this still causes stage2 llvm-tblgen crashes, so this wasn't the only issue :(

I applied nikic@d8ea7f2 locally and reran the original testcase I had that failed, and now it passes so unfortunately that isn't of any use anymore. :(

nikic · 2023-11-02T15:08:47Z

I believe the problem is this:

define i1 @test(ptr %arg, ptr %arg2) {
  %icmp = icmp eq ptr %arg, %arg2
  %call = call i1 @llvm.is.constant.i1(i1 %icmp)
  %and = and i1 %call, %icmp
  ret i1 %and
}

declare i1 @llvm.is.constant.i1(i1)

The is.constant check gets folded away. I'd normally say this is just a QoI issue, but it seems that the libstdc++ implementation relies on this in some way I don't fully understand (there is a specialization of std::distance on std::list for getting the size of the full list, and then that is used in _Finalize_merge where we are currently in the process of updating list sizes, or something like that).

So I guess we should exclude llvm.is.constant from being folded by simplifyWithOpReplaced.

Test the interaction with llvm.is.constant.

) and/or in logical (select) form benefit from generic simplifications via simplifyWithOpReplaced(). However, the corresponding fold for plain and/or currently does not exist. Similar to selects, there are two general cases for this fold (illustrated with `and`, but there are `or` conjugates). The basic case is something like `(a == b) & c`, where the replacement of a with b or b with a inside c allows it to fold to true or false. Then the whole operation will fold to either false or `a == b`. The second case is something like `(a != b) & c`, where the replacement inside c allows it to fold to false. In that case, the operand can be replaced with c, because in the case where a == b (and thus the icmp is false), c itself will already be false. As the test diffs show, this catches quite a lot of patterns in existing test coverage. This also obsoletes quite a few existing special-case and/or of icmp folds we have (e.g. simplifyAndOrOfICmpsWithLimitConst), but I haven't removed anything as part of this patch in the interest of risk mitigation. Fixes llvm#69050. Fixes llvm#69091.

nikic · 2023-11-02T16:30:16Z

Looks like nikic@6343a80 does fix the llvm-tblgen issue.

…70335) Relative to the first attempt, this contains two changes: First, we only handle the case where one side simplifies to true or false, instead of calling simplification recursively. The previous approach would return poison if one operand simplified to poison (under the equality assumption), which is incorrect. Second, we do not fold llvm.is.constant in simplifyWithOpReplaced(). We may be assuming that a value is constant, if the equality holds, but it may not actually be constant. This is nominally just a QoI issue, but the std::list implementation in libstdc++ relies on the precise behavior in a way that causes miscompiles. ----- and/or in logical (select) form benefit from generic simplifications via simplifyWithOpReplaced(). However, the corresponding fold for plain and/or currently does not exist. Similar to selects, there are two general cases for this fold (illustrated with `and`, but there are `or` conjugates). The basic case is something like `(a == b) & c`, where the replacement of a with b or b with a inside c allows it to fold to true or false. Then the whole operation will fold to either false or `a == b`. The second case is something like `(a != b) & c`, where the replacement inside c allows it to fold to false. In that case, the operand can be replaced with c, because in the case where a == b (and thus the icmp is false), c itself will already be false. As the test diffs show, this catches quite a lot of patterns in existing test coverage. This also obsoletes quite a few existing special-case and/or of icmp folds we have (e.g. simplifyAndOrOfICmpsWithLimitConst), but I haven't removed anything as part of this patch in the interest of risk mitigation. Fixes #69050. Fixes #69091.

amykhuang · 2023-11-03T17:58:08Z

Might be already addressed in previous discussion on this bug, but in chrome we saw a binary size increase from this patch (crbug.com/1498542)

nikic · 2023-11-03T21:01:47Z

@amykhuang Do you still see a size increase with the reapplied version of the patch?

I think it's plausible that the llvm.is.constant issue could have resulted in a binary size increase, in which case that should be resolved now. If it's not, I'd expect that this is some kind of second order effect not directly related to this patch.

nikic requested review from dtcxzyw and goldsteinn October 26, 2023 14:05

llvmbot added llvm:analysis llvm:transforms labels Oct 26, 2023

nikic force-pushed the simplify-and-or-with-op-replaced branch from 5877211 to bcd36d9 Compare October 26, 2023 14:19

dtcxzyw reviewed Oct 26, 2023

View reviewed changes

llvm/lib/Analysis/InstructionSimplify.cpp Outdated Show resolved Hide resolved

nikic force-pushed the simplify-and-or-with-op-replaced branch from bcd36d9 to 3d05eb3 Compare October 27, 2023 08:32

llvmbot added the PGO Profile Guided Optimizations label Oct 27, 2023

nikic force-pushed the simplify-and-or-with-op-replaced branch from 3d05eb3 to 8a1d0a6 Compare October 27, 2023 09:10

nikic commented Oct 27, 2023

View reviewed changes

dtcxzyw approved these changes Oct 27, 2023

View reviewed changes

goldsteinn reviewed Oct 27, 2023

View reviewed changes

nikic merged commit 1770a2e into llvm:main Oct 30, 2023
3 checks passed

nikic added a commit that referenced this pull request Oct 30, 2023

Revert "[InstCombine] Simplify and/or of icmp eq with op replacement (#…

e46dd6f

…70335)" This reverts commit 1770a2e. Stage 2 llvm-tblgen crashes when generating X86GenAsmWriter.inc and other files.

nikic added a commit that referenced this pull request Nov 2, 2023

[InstSimplify] Add extra tests for #70335 (NFC)

e331ca5

These were miscompiled with the initial version of the patch.

nikic added a commit that referenced this pull request Nov 2, 2023

[InstSimplify] Add additional test for #70335 (NFC)

2a18217

Test the interaction with llvm.is.constant.

This was referenced Nov 12, 2023

Simplification for (A ==/!= B) &/| (((cast)A) CMP C) #68514

Closed

[ValueTracking] Improve isImpliedCondCommonOperandWithConstants to handle truncated LHS #69829

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[InstCombine] Simplify and/or of icmp eq with op replacement #70335

[InstCombine] Simplify and/or of icmp eq with op replacement #70335

nikic commented Oct 26, 2023 •

edited

llvmbot commented Oct 26, 2023 •

edited

github-actions bot commented Oct 26, 2023 •

edited

dtcxzyw commented Oct 26, 2023

dtcxzyw commented Oct 26, 2023

nikic commented Oct 27, 2023

nikic commented Oct 27, 2023

nikic Oct 27, 2023

dtcxzyw left a comment

goldsteinn Oct 27, 2023

nikic Oct 27, 2023

goldsteinn Oct 27, 2023

nikic commented Oct 31, 2023

mikaelholmen commented Nov 1, 2023

nikic commented Nov 1, 2023

mikaelholmen commented Nov 2, 2023

nikic commented Nov 2, 2023 •

edited

nikic commented Nov 2, 2023

mikaelholmen commented Nov 2, 2023 •

edited

nikic commented Nov 2, 2023

mikaelholmen commented Nov 2, 2023

mikaelholmen commented Nov 2, 2023

nikic commented Nov 2, 2023

nikic commented Nov 2, 2023

amykhuang commented Nov 3, 2023

nikic commented Nov 3, 2023

[InstCombine] Simplify and/or of icmp eq with op replacement #70335

[InstCombine] Simplify and/or of icmp eq with op replacement #70335

Conversation

nikic commented Oct 26, 2023 • edited

llvmbot commented Oct 26, 2023 • edited

github-actions bot commented Oct 26, 2023 • edited

dtcxzyw commented Oct 26, 2023

dtcxzyw commented Oct 26, 2023

nikic commented Oct 27, 2023

nikic commented Oct 27, 2023

nikic Oct 27, 2023

Choose a reason for hiding this comment

dtcxzyw left a comment

Choose a reason for hiding this comment

goldsteinn Oct 27, 2023

Choose a reason for hiding this comment

nikic Oct 27, 2023

Choose a reason for hiding this comment

goldsteinn Oct 27, 2023

Choose a reason for hiding this comment

nikic commented Oct 31, 2023

mikaelholmen commented Nov 1, 2023

nikic commented Nov 1, 2023

mikaelholmen commented Nov 2, 2023

nikic commented Nov 2, 2023 • edited

nikic commented Nov 2, 2023

mikaelholmen commented Nov 2, 2023 • edited

nikic commented Nov 2, 2023

mikaelholmen commented Nov 2, 2023

mikaelholmen commented Nov 2, 2023

nikic commented Nov 2, 2023

nikic commented Nov 2, 2023

amykhuang commented Nov 3, 2023

nikic commented Nov 3, 2023

nikic commented Oct 26, 2023 •

edited

llvmbot commented Oct 26, 2023 •

edited

github-actions bot commented Oct 26, 2023 •

edited

nikic commented Nov 2, 2023 •

edited

mikaelholmen commented Nov 2, 2023 •

edited