[InstCombine] Implement `fcmp (fadd x, 0.0), y` => `fcmp x, y` optimization #88476

vmishelcs · 2024-04-12T06:32:10Z

This PR addresses issue #88168. I have implemented an optimization for the case of

define i1 @fcmp_fadd_zero_ugt(float %x, float %y) {
  %add = fadd float %x, 0.000000e+00
  %cmp = fcmp ugt float %add, %y
  ret i1 %cmp
}

=>

define i1 @fcmp_fadd_zero_ugt(float %x, float %y) {
  %cmp = fcmp ugt float %x, %y
  ret i1 %cmp
}

and all other types of fcmp instructions (uge, ogt, etc). I have also added tests to llvm/test/Transforms/InstCombine/fcmp.ll to check this behaviour.

Proofs:
fadd x, 0.0 https://alive2.llvm.org/ce/z/7FzNnM
fsub x, 0.0 https://alive2.llvm.org/ce/z/puUxLK

github-actions · 2024-04-12T06:32:28Z

Thank you for submitting a Pull Request (PR) to the LLVM Project!

This PR will be automatically labeled and the relevant teams will be
notified.

If you wish to, you can add reviewers by using the "Reviewers" section on this page.

If this is not working for you, it is probably because you do not have write
permissions for the repository. In which case you can instead tag reviewers by
name in a comment by using @ followed by their GitHub username.

If you have received no comments on your PR for a week, you can request a review
by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate
is once a week. Please remember that you are asking for valuable time from other developers.

If you have further questions, they may be answered by the LLVM GitHub User Guide.

You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums.

vmishelcs · 2024-04-12T06:38:18Z

@dtcxzyw Please take a look whenever you have time :) Let me know if I should make any changes.

Question: Should I also add tests for the case of fcmp (fsub x, 0.0), y? From what I can tell my implementation already covers this case. Is this because somewhere in the code we transform fsub x, 0.0 => fadd x, 0.0?

llvmbot · 2024-04-12T06:38:48Z

@llvm/pr-subscribers-llvm-transforms

Author: Vlad Mishel (vmishelcs)

Changes

This PR addresses issue #88168. I have implemented an optimization for the case of

define i1 @<!-- -->fcmp_fadd_zero_ugt(float %x, float %y) {
  %add = fadd float %x, 0.000000e+00
  %cmp = fcmp ugt float %add, %y
  ret i1 %cmp
}

=>

define i1 @<!-- -->fcmp_fadd_zero_ugt(float %x, float %y) {
  %cmp = fcmp ugt float %x, %y
  ret i1 %cmp
}

and all other types of fcmp instructions (uge, ogt, etc). I have also added tests to llvm/test/Transforms/InstCombine/fcmp.ll to check this behaviour.

Proofs:
fadd x, 0.0 https://alive2.llvm.org/ce/z/7FzNnM
fsub x, 0.0 https://alive2.llvm.org/ce/z/puUxLK

Full diff: https://github.com/llvm/llvm-project/pull/88476.diff

2 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp (+7)
(modified) llvm/test/Transforms/InstCombine/fcmp.ll (+142)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index 7292bb62702aaa..67f3e0386402b1 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -8100,6 +8100,13 @@ Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
         return new FCmpInst(I.getSwappedPredicate(), X, NegC, "", &I);
   }
 
+  // fcmp (fadd X, 0.0), Y --> fcmp X, Y
+  if (match(Op0, m_FAdd(m_Value(X), m_APFloat(C))) && C->isZero()) {
+    if (match(Op1, m_Value(Y))) {
+      return new FCmpInst(Pred, X, Y, "", &I);
+    }
+  }
+
   if (match(Op0, m_FPExt(m_Value(X)))) {
     // fcmp (fpext X), (fpext Y) -> fcmp X, Y
     if (match(Op1, m_FPExt(m_Value(Y))) && X->getType() == Y->getType())
diff --git a/llvm/test/Transforms/InstCombine/fcmp.ll b/llvm/test/Transforms/InstCombine/fcmp.ll
index f2701d16d0f3d1..003483c217771c 100644
--- a/llvm/test/Transforms/InstCombine/fcmp.ll
+++ b/llvm/test/Transforms/InstCombine/fcmp.ll
@@ -1284,3 +1284,145 @@ define <1 x i1> @bitcast_1vec_eq0(i32 %x) {
   %cmp = fcmp oeq <1 x float> %f, zeroinitializer
   ret <1 x i1> %cmp
 }
+
+; Simplify fcmp (x + 0.0), y => fcmp x, y
+
+define i1 @fcmp_fadd_zero_ugt(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_ugt(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ugt float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp ugt float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_uge(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_uge(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp uge float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp uge float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_ogt(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_ogt(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ogt float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp ogt float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_oge(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_oge(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp oge float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp oge float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_ult(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_ult(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ult float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp ult float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_ule(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_ule(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ule float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp ule float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_olt(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_olt(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp olt float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp olt float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_ole(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_ole(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ole float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp ole float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_oeq(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_oeq(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp oeq float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp oeq float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_one(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_one(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp one float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp one float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_ueq(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_ueq(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ueq float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp ueq float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_une(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_une(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp une float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp une float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_ord(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_ord(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ord float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp ord float %add, %y
+  ret i1 %cmp
+}
+
+define i1 @fcmp_fadd_zero_uno(float %x, float %y) {
+; CHECK-LABEL: @fcmp_fadd_zero_uno(
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp uno float [[ADD:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %add = fadd float %x, 0.000000e+00
+  %cmp = fcmp uno float %add, %y
+  ret i1 %cmp
+}

dtcxzyw · 2024-04-12T07:15:28Z

Question: Should I also add tests for the case of fcmp (fsub x, 0.0), y? From what I can tell my implementation already covers this case. Is this because somewhere in the code we transform fsub x, 0.0 => fadd x, 0.0?

Yeah, fsub x, +/-0.0 are not in canonical form. I don't think these tests are necessary.

dtcxzyw · 2024-04-12T07:20:28Z

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

+  if (match(Op0, m_FAdd(m_Value(X), m_APFloat(C))) && C->isZero()) {
+    if (match(Op1, m_Value(Y))) {
+      return new FCmpInst(Pred, X, Y, "", &I);
+    }
+  }


Suggested change

if (match(Op0, m_FAdd(m_Value(X), m_APFloat(C))) && C->isZero()) {

if (match(Op1, m_Value(Y))) {

return new FCmpInst(Pred, X, Y, "", &I);

}

}

if (match(Op0, m_FAdd(m_Value(X), m_AnyZeroFP())))

return new FCmpInst(Pred, X, Op1, "", &I);

if (match(Op1, m_FAdd(m_Value(X), m_AnyZeroFP())))

return new FCmpInst(Pred, Op0, X, "", &I);

Add a test for fcmp X, (fadd Y, 0.0)

Add a vector test for fcmp (fadd X, <0.0, -0.0>), Y

Add a test for fast-math flag preservation

Add a test for fcmp X, (fadd Y, 0.0)

Add a vector test for fcmp (fadd X, <0.0, -0.0>), Y

Add a test for fast-math flag preservation

Should I write a separate test case for each type of fcmp instruction (ugt, ogt, etc) for each of these? Or is one test case sufficient for each?

Just one test each is fine.

arsenm · 2024-04-15T18:52:45Z

llvm/test/Transforms/InstCombine/fcmp.ll

+  ret <2 x i1> %cmp
+}
+
+define i1 @fcmp_fadd_zero_fast(float %x, float %y) {


Test cases where each instruction has flags, and the other doesn't. Also add a scalar case with negative 0

Specifically you have to make sure the flags don't come from the fadd, they cannot be preserved

Specifically you have to make sure the flags don't come from the fadd, they cannot be preserved

Added! The behaviour seems to be correct from what I can tell.

I have a quick question: How come for some of the tests, the CHECK lines generate as

define i1 @fcmp_fadd_zero_ugt(float %x, float %y) { ; CHECK-LABEL: @fcmp_fadd_zero_ugt( ; CHECK-NEXT: [[CMP:%.*]] = fcmp ugt float [[ADD:%.*]], [[Y:%.*]] ; CHECK-NEXT: ret i1 [[CMP]] ; %add = fadd float %x, 0.000000e+00 %cmp = fcmp ugt float %add, %y ret i1 %cmp }

and for the tests which I added later, the CHECK lines are

define i1 @fcmp_fadd_neg_zero(float %x, float %y) { ; CHECK-LABEL: @fcmp_fadd_neg_zero( ; CHECK-NEXT: [[CMP:%.*]] = fcmp ugt float [[X:%.*]], [[Y:%.*]] ; CHECK-NEXT: ret i1 [[CMP]] ; %add = fadd float %x, -0.000000e+00 %cmp = fcmp ugt float %add, %y ret i1 %cmp }

?
In particular, why is it that for the former test, the CHECK line uses [[ADD:%.*]] but the latter uses [[X:%.*]]? When I ran the opt binary, the result is correct in both cases, but I was wondering why the slight difference.

I'm not sure exactly where update_test_checks pulls the name from. It looks like it tried the source in the first place?

Sounds like a regression from preserving variable names in UTC, cc @nhaehnle.

PR Link: llvm/llvm-project#88476

github-actions · 2024-04-16T03:12:05Z

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

PR Link: llvm/llvm-project#88476

vmishelcs · 2024-04-16T03:29:47Z

@dtcxzyw I see there are some issues with the formatting. Is there a script I can run and commit the changes with fixed formatting?

dtcxzyw · 2024-04-16T03:44:01Z

@dtcxzyw I see there are some issues with the formatting. Is there a script I can run and commit the changes with fixed formatting?

git clang-format HEAD~1

dtcxzyw

LGTM with the formatting fixed.

vmishelcs · 2024-04-16T04:32:45Z

Formatting should be fixed now.

github-actions · 2024-04-16T05:05:46Z

@vmishelcs Congratulations on having your first Pull Request (PR) merged into the LLVM Project!

Your changes will be combined with recent changes from other authors, then tested
by our build bots. If there is a problem with a build, you may receive a report in an email or a comment on this PR.

Please check whether problems have been caused by your change specifically, as
the builds can include changes from many authors. It is not uncommon for your
change to be included in a build that fails due to someone else's changes, or
infrastructure issues.

How to do this, and the rest of the post-merge process, is covered in detail here.

If your change does cause a problem, it may be reverted, or you can revert it yourself.
This is a normal part of LLVM development. You can fix your changes and open a new PR to merge them again.

If you don't get any reports, no action is required from you. Your changes are working as expected, well done!

vmishelcs added 2 commits April 11, 2024 23:00

[InstCombine] Add tests for fcmp combine

0812766

[InstCombine] Implement fcmp (fadd x, 0.0), y -> fcmp x, y

f7092dd

vmishelcs marked this pull request as ready for review April 12, 2024 06:38

vmishelcs requested a review from nikic as a code owner April 12, 2024 06:38

llvmbot added the llvm:transforms label Apr 12, 2024

XChy requested review from dtcxzyw, arsenm and goldsteinn April 12, 2024 06:41

dtcxzyw requested changes Apr 12, 2024

View reviewed changes

vmishelcs added 2 commits April 12, 2024 19:34

[InstCombine] Implement dtcxzyw's suggested changes

9fa945d

[InstCombine] Added tests for swap, vectors, and fast flag

daca19f

vmishelcs requested a review from dtcxzyw April 13, 2024 02:49

arsenm reviewed Apr 15, 2024

View reviewed changes

[InstCombine] Added tests as instructed by arsenm

a460fb0

vmishelcs requested a review from arsenm April 15, 2024 19:47

arsenm approved these changes Apr 15, 2024

View reviewed changes

dtcxzyw added a commit to dtcxzyw/llvm-codegen-benchmark that referenced this pull request Apr 16, 2024

pre-commit: test PR88476

66b58a0

PR Link: llvm/llvm-project#88476

dtcxzyw mentioned this pull request Apr 16, 2024

pre-commit: test PR88476 dtcxzyw/llvm-codegen-benchmark#5

Closed

dtcxzyw added a commit to dtcxzyw/llvm-opt-benchmark that referenced this pull request Apr 16, 2024

pre-commit: test PR88476

f5087ce

PR Link: llvm/llvm-project#88476

dtcxzyw mentioned this pull request Apr 16, 2024

pre-commit: test PR88476 dtcxzyw/llvm-opt-benchmark#519

Closed

dtcxzyw approved these changes Apr 16, 2024

View reviewed changes

[InstCombine] Format fix

dc54243

dtcxzyw merged commit edb0708 into llvm:main Apr 16, 2024
3 of 4 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[InstCombine] Implement `fcmp (fadd x, 0.0), y` => `fcmp x, y` optimization #88476

[InstCombine] Implement `fcmp (fadd x, 0.0), y` => `fcmp x, y` optimization #88476

vmishelcs commented Apr 12, 2024 •

edited

github-actions bot commented Apr 12, 2024

vmishelcs commented Apr 12, 2024

llvmbot commented Apr 12, 2024

dtcxzyw commented Apr 12, 2024

dtcxzyw Apr 12, 2024

vmishelcs Apr 13, 2024 •

edited

nikic Apr 13, 2024

arsenm Apr 15, 2024

arsenm Apr 15, 2024

vmishelcs Apr 15, 2024

arsenm Apr 15, 2024

nikic Apr 16, 2024

github-actions bot commented Apr 16, 2024 •

edited

vmishelcs commented Apr 16, 2024

dtcxzyw commented Apr 16, 2024

dtcxzyw left a comment

vmishelcs commented Apr 16, 2024

github-actions bot commented Apr 16, 2024

-  if (match(Op0, m_FAdd(m_Value(X), m_APFloat(C))) && C->isZero()) {
-    if (match(Op1, m_Value(Y))) {
-      return new FCmpInst(Pred, X, Y, "", &I);
-    }
-  }
+  if (match(Op0, m_FAdd(m_Value(X), m_AnyZeroFP())))
+      return new FCmpInst(Pred, X, Op1, "", &I);
+  if (match(Op1, m_FAdd(m_Value(X), m_AnyZeroFP())))
+      return new FCmpInst(Pred, Op0, X, "", &I);

[InstCombine] Implement fcmp (fadd x, 0.0), y => fcmp x, y optimization #88476

[InstCombine] Implement fcmp (fadd x, 0.0), y => fcmp x, y optimization #88476

Conversation

vmishelcs commented Apr 12, 2024 • edited

github-actions bot commented Apr 12, 2024

vmishelcs commented Apr 12, 2024

llvmbot commented Apr 12, 2024

dtcxzyw commented Apr 12, 2024

dtcxzyw Apr 12, 2024

Choose a reason for hiding this comment

vmishelcs Apr 13, 2024 • edited

Choose a reason for hiding this comment

nikic Apr 13, 2024

Choose a reason for hiding this comment

arsenm Apr 15, 2024

Choose a reason for hiding this comment

arsenm Apr 15, 2024

Choose a reason for hiding this comment

vmishelcs Apr 15, 2024

Choose a reason for hiding this comment

arsenm Apr 15, 2024

Choose a reason for hiding this comment

nikic Apr 16, 2024

Choose a reason for hiding this comment

github-actions bot commented Apr 16, 2024 • edited

vmishelcs commented Apr 16, 2024

dtcxzyw commented Apr 16, 2024

dtcxzyw left a comment

Choose a reason for hiding this comment

vmishelcs commented Apr 16, 2024

github-actions bot commented Apr 16, 2024

[InstCombine] Implement `fcmp (fadd x, 0.0), y` => `fcmp x, y` optimization #88476

[InstCombine] Implement `fcmp (fadd x, 0.0), y` => `fcmp x, y` optimization #88476

vmishelcs commented Apr 12, 2024 •

edited

vmishelcs Apr 13, 2024 •

edited

github-actions bot commented Apr 16, 2024 •

edited