[InstCombine] Improve select equiv fold for plain condition #83405

vfdff · 2024-02-29T10:48:48Z

Fold select if the user of its select user indicates the condition

Fix #83225

llvmbot · 2024-02-29T10:50:11Z

@llvm/pr-subscribers-llvm-transforms

Author: Allen (vfdff)

Changes

Fold select if the user of its select user indicates the condition

Fix #83225

Full diff: https://github.com/llvm/llvm-project/pull/83405.diff

2 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp (+26)
(modified) llvm/test/Transforms/InstCombine/select.ll (+30)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
index 71fa9b9ba41ebb..1485fe8674b720 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
@@ -500,6 +500,29 @@ static bool isSelect01(const APInt &C1I, const APInt &C2I) {
   return C1I.isOne() || C1I.isAllOnes() || C2I.isOne() || C2I.isAllOnes();
 }
 
+/// Try to simplify a select instruction when the user of its select user
+/// indicates the condition.
+static bool simplifySeqSelectWithSameCond(Value *Cond, Value *F) {
+  Value *FalseVal = F;
+  Value *CondNext;
+  Value *FalseNext;
+  while (match(FalseVal, m_Select(m_Value(CondNext), m_Value(),
+                                  m_OneUse(m_Value(FalseNext))))) {
+    if (CondNext == Cond) {
+      auto *SI = cast<SelectInst>(FalseVal);
+      LLVM_DEBUG(dbgs() << "IC: Fold " << SI << " with " << *FalseNext << '\n');
+
+      SI->replaceAllUsesWith(FalseNext);
+      SI->eraseFromParent();
+      return true;
+    }
+
+    FalseVal = FalseNext;
+  }
+
+  return false;
+}
+
 /// Try to fold the select into one of the operands to allow further
 /// optimization.
 Instruction *InstCombinerImpl::foldSelectIntoOp(SelectInst &SI, Value *TrueVal,
@@ -557,6 +580,9 @@ Instruction *InstCombinerImpl::foldSelectIntoOp(SelectInst &SI, Value *TrueVal,
   if (Instruction *R = TryFoldSelectIntoOp(SI, FalseVal, TrueVal, true))
     return R;
 
+  if (simplifySeqSelectWithSameCond(SI.getCondition(), FalseVal))
+    return &SI;
+
   return nullptr;
 }
 
diff --git a/llvm/test/Transforms/InstCombine/select.ll b/llvm/test/Transforms/InstCombine/select.ll
index c5f1b77c6d7404..068d168876a328 100644
--- a/llvm/test/Transforms/InstCombine/select.ll
+++ b/llvm/test/Transforms/InstCombine/select.ll
@@ -3672,3 +3672,33 @@ define i32 @src_select_xxory_eq0_xorxy_y(i32 %x, i32 %y) {
   %cond = select i1 %xor0, i32 %xor, i32 %y
   ret i32 %cond
 }
+
+define i32 @sequence_select_with_same_cond (i1 %c1, i1 %c2){
+; CHECK-LABEL: @sequence_select_with_same_cond(
+; CHECK-NEXT:    [[S1:%.*]] = select i1 [[C1:%.*]], i32 23, i32 45
+; CHECK-NEXT:    [[S2:%.*]] = select i1 [[C2:%.*]], i32 666, i32 [[S1]]
+; CHECK-NEXT:    [[S3:%.*]] = select i1 [[C1]], i32 789, i32 [[S2]]
+; CHECK-NEXT:    ret i32 [[S3]]
+;
+  %s1 = select i1 %c1, i32 23, i32 45
+  %s2 = select i1 %c2, i32 666, i32 %s1
+  %s3 = select i1 %c1, i32 789, i32 %s2
+  ret i32 %s3
+}
+
+declare void @use32(i32)
+
+define i32 @sequence_select_with_same_cond_extra_use (i1 %c1, i1 %c2){
+; CHECK-LABEL: @sequence_select_with_same_cond_extra_use(
+; CHECK-NEXT:    [[S1:%.*]] = select i1 [[C1:%.*]], i32 23, i32 45
+; CHECK-NEXT:    call void @use32(i32 [[S1]])
+; CHECK-NEXT:    [[S2:%.*]] = select i1 [[C2:%.*]], i32 666, i32 [[S1]]
+; CHECK-NEXT:    [[S3:%.*]] = select i1 [[C1]], i32 789, i32 [[S2]]
+; CHECK-NEXT:    ret i32 [[S3]]
+;
+  %s1 = select i1 %c1, i32 23, i32 45
+  call void @use32(i32 %s1)
+  %s2 = select i1 %c2, i32 666, i32 %s1
+  %s3 = select i1 %c1, i32 789, i32 %s2
+  ret i32 %s3
+}

PR Link: llvm/llvm-project#83405

llvm/test/Transforms/InstCombine/select.ll

nikic · 2024-02-29T12:24:46Z

I think this should be a minor generalization of some existing simplifyWithOpReplaced or foldSelectValueEquivalence style fold.

vfdff · 2024-03-01T11:47:42Z

I think this should be a minor generalization of some existing simplifyWithOpReplaced or foldSelectValueEquivalence style fold.
Thanks, refactor with simplifyWithOpReplaced. （foldSelectValueEquivalence is used for select + cmp）

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

goldsteinn · 2024-03-18T16:35:12Z

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

+    while (match(ValOp, m_Select(m_Value(CondNext), m_Value(), m_Value()))) {
+      if (CondNext == CondVal)
+        if (simplifyWithOpReplaced(ValOp, CondVal, RepOp, SQ,
+                                   /* AllowRefinement */ true)) {


I don't understand what the simplifyWithOpReplaced is doing anymore.

deleted the redundant call simplifyWithOpReplaced , thanks

The select equivalence fold takes a select like "X == Y ? A : B" and then tries to simplify A based on the known equality. This patch also uses it for the case were we have just "C ? A : B" by treating the condition as either "C == 1" or "C != 1". This is intended as an alternative to llvm#83405 for fixing llvm#83225.

nikic · 2024-03-18T16:57:37Z

FWIW, what I had in mind with my comment was more something along the lines of #85663, which generalizes the existing select equivalence fold. The disadvantage is that it does not handle the multi-use example out of the box (but could be generalized).

vfdff · 2024-03-22T02:56:38Z

FWIW, what I had in mind with my comment was more something along the lines of #85663, which generalizes the existing select equivalence fold. The disadvantage is that it does not handle the multi-use example out of the box (but could be generalized).

hi @nikic, I saw #85663 is closed because the compile time regression, so would you suggest I continue with current solution?

vfdff · 2024-04-03T02:05:40Z

ping ？

goldsteinn · 2024-04-03T18:09:22Z

Still LGTM, ping @dtcxzyw @nikic

dtcxzyw · 2024-04-04T07:18:46Z

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

+                                          InstCombinerImpl &IC) {
+  Value *CondVal = SI.getCondition();
+  if (match(CondVal, m_ImmConstant()))
+    return false;


Isn't it handled by simplifySelectInst?

Yes, it isn't.
#85663 is a more general improvement by nikic, and it has some compile time regression, so I continue on this solution.

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

Lines 4767 to 4771 in a612524

if (auto *CondC = dyn_cast<Constant>(Cond)) {

if (auto *TrueC = dyn_cast<Constant>(TrueVal))

if (auto *FalseC = dyn_cast<Constant>(FalseVal))

if (Constant *C = ConstantFoldSelectInstruction(CondC, TrueC, FalseC))

return C;

Can you provide a test to show that the early exit here is necessary?

Thanks, I delete the early exit checking.
I added it just because I thought it was an optimized expression and didn't need to be optimized.

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

+    Value *CondNext;
+    while (match(ValOp, m_Select(m_Value(CondNext), m_Value(), m_Value()))) {
+      if (CondNext == CondVal) {
+        IC.replaceOperand(*SINext, OpIndex,


llvm/test/Transforms/InstCombine/select.ll

+; CHECK-NEXT:    [[SEL3:%.*]] = select i1 [[TMP1]], i8 [[SEL1_NEW]], i8 [[SEL4]]
+; CHECK-NEXT:    ret i8 [[SEL3]]
+;
+  %sel0 = select i1 %cond0, i8 %a, i8 %b


arsenm · 2024-05-08T15:18:16Z

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

+/// Try to simplify a select instruction when the user of its select user
+/// indicates the condition.


I don't know how to parse this comment "when the user of its select user". Can you rephrase this whole thing?

Add comment with a example.

arsenm · 2024-05-08T15:18:54Z

llvm/test/Transforms/InstCombine/select.ll

+  %sel2 = select i1 %cond1, i8 %sel1, i8 3
+  %sel3 = select i1 %cond0, i8 %sel1, i8 %sel2
+  ret i8 %sel3
+}


Can you add some tests that show flag preservation behavior?

Added sequence_select_with_same_cond_float_and_fmf_flag1 and sequence_select_with_same_cond_float_and_fmf_flag2, thanks

Better to have tests with different flags (and more specific than fast). Need to see the union / intersect behavior

Fold select if the user of its select user indicates the condition Fix llvm#83225

Refactor with a lambda function to address both false and true value.

vfdff · 2024-05-13T09:50:36Z

According performace considerations, I deleted the support with multi-use select node and rebase to top upstream as the code conflicts.

vfdff · 2024-05-24T02:59:49Z

ping @dtcxzyw, @arsenm

vfdff · 2024-06-01T01:57:12Z

close to avoid blocking other solutions

arsenm · 2024-06-01T06:01:36Z

close to avoid blocking other solutions

What do you mean? If there's an alternative patch, it should be mentioned in the closing message. Otherwise, why close it

vfdff · 2024-06-01T07:03:17Z

Oh，sorry, I reopen it now.
I close it because the solution is limit to single-use select node, which seems not perfect :)

arsenm · 2024-08-16T19:07:15Z

Oh，sorry, I reopen it now. I close it because the solution is limit to single-use select node, which seems not perfect :)

That is the solution to many combine problems. It's a common and reasonable restriction

goldsteinn · 2024-08-18T21:57:53Z

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

+    // directly.
+    // It is not profitable to build a new select for SINext with multi-arms.
+    while (match(ValOp, m_Select(m_Value(CondNext), m_Value(), m_Value()))) {
+      if (CondNext == CondVal && SINext->hasOneUse()) {


It should be match(ValOp, m_OneUse(m_Select(....)) then drop the SINext->hasOneUse(), otherwise this won't fold the base select if it has multiple uses.

For the following case, it can be optimized because %sel1 is a sinle use node.
If we adjust according above change, then it will not optimize because the operand of %sel1 （%sel0）is multi used.

define i8 @sequence_select_with_same_cond_multi_arms_src(i1 %cond0, i1 %cond1, i8 %a, i8 %b) { %sel0 = select i1 %cond0, i8 %a, i8 %b %sel1 = select i1 %cond1, i8 %sel0, i8 2; %sel1 used in single node %sel2 = select i1 %cond1, i8 %sel0, i8 3 %sel3 = select i1 %cond0, i8 %sel1, i8 %sel2 ret i8 %sel3 } define i8 @sequence_select_with_same_cond_multi_arms_tgt(i1 %cond0, i1 %cond1, i8 %a, i8 %b) { %sel0 = select i1 %cond0, i8 %a, i8 %b %sel1 = select i1 %cond1, i8 %a, i8 2; %sel1 used in single node %sel2 = select i1 %cond1, i8 %sel0, i8 3 %sel3 = select i1 %cond0, i8 %sel1, i8 %sel2 ret i8 %sel3 }

Makes sense, what about V == SINext || SINext->hasOneUse() then?

do you mean ValOp == SINext || SINext->hasOneUse() ? the condition ValOp == SINext is protected by ValOp = SINext->getOperand(OpIndex), and the CondNext == CondVal is also need, consider case. https://alive2.llvm.org/ce/z/TKLmMG

define i32 @src(i32 %a, i1 %c1, i1 %c2, i1 %c3){ %s1 = select i1 %c1, i32 23, i32 45 %s2 = select i1 %c2, i32 666, i32 %s1 ; this node can be optimized iff %c1 == %c3 %s3 = select i1 %c3, i32 789, i32 %s2 ret i32 %s3 } define i32 @tgt(i32 %a, i1 %c1, i1 %c2, i1 %c3){ %s1 = select i1 %c1, i32 23, i32 45 %s2 = select i1 %c2, i32 666, i32 45 %s3 = select i1 %c3, i32 789, i32 %s2 ret i32 %s3 }

dtcxzyw · 2024-12-15T15:20:08Z

Ping. I need this patch to unblock #119884.
I noticed this patch when drafting an alternative with replaceInInstruction.

vfdff · 2024-12-16T11:48:15Z

hi @dtcxzyw I don't work on this issue now, please update your MR at your convenience

vfdff requested a review from nikic as a code owner February 29, 2024 10:48

llvmbot added the llvm:transforms label Feb 29, 2024

dtcxzyw added a commit to dtcxzyw/llvm-opt-benchmark that referenced this pull request Feb 29, 2024

pre-commit: test PR83405

613d5ca

PR Link: llvm/llvm-project#83405

dtcxzyw mentioned this pull request Feb 29, 2024

pre-commit: test PR83405 dtcxzyw/llvm-opt-benchmark#282

Closed

vfdff requested a review from dtcxzyw February 29, 2024 11:19

dtcxzyw reviewed Feb 29, 2024

View reviewed changes

llvm/test/Transforms/InstCombine/select.ll Outdated Show resolved Hide resolved

vfdff force-pushed the PR83225 branch from 207b6a1 to 2b792fe Compare March 1, 2024 11:43

vfdff requested review from dtcxzyw, arsenm and goldsteinn March 8, 2024 10:28

goldsteinn reviewed Mar 8, 2024

View reviewed changes