[ValueTracking] Compute knownbits from `(icmp upred (add/sub nuw X, Y), C)` #87180

goldsteinn · 2024-03-31T03:42:57Z

[ValueTracking] Add tests for computing knownbits from (icmp upred (add/sub nuw X, Y), C); NFC
[ValueTracking] Compute knownbits from (icmp upred (add/sub nuw X, Y), C)

(icmp ule/ult (add nuw X, Y), C) implies both (icmp ule/ult X, C) and
(icmp ule/ult Y, C). We can use this to deduce leading zeros in X/Y.

(icmp uge/ugt (sub nuw X, Y), C) implies (icmp uge/uge X, C) . We
can use this to deduce leading ones in X.

Proofs: https://alive2.llvm.org/ce/z/sc5k22

llvmbot · 2024-03-31T03:43:27Z

@llvm/pr-subscribers-llvm-analysis

@llvm/pr-subscribers-llvm-transforms

Author: None (goldsteinn)

Changes

[ValueTracking] Add tests for computing knownbits from (icmp ult/ule (add nuw X, Y), C); NFC
[ValueTracking] Compute knownbits from (icmp ult/ule (add nuw X, Y), C)

Based on: #87180

Full diff: https://github.com/llvm/llvm-project/pull/87180.diff

2 Files Affected:

(modified) llvm/lib/Analysis/ValueTracking.cpp (+28-12)
(modified) llvm/test/Transforms/InstCombine/known-bits.ll (+98-5)

diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index b5e8a1d22f264b..5bc2d605d3a1b1 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -698,13 +698,19 @@ static void computeKnownBitsFromCmp(const Value *V, CmpInst::Predicate Pred,
     break;
   }
   default:
-    const APInt *Offset = nullptr;
-    if (match(LHS, m_CombineOr(m_V, m_AddLike(m_V, m_APInt(Offset)))) &&
-        match(RHS, m_APInt(C))) {
-      ConstantRange LHSRange = ConstantRange::makeAllowedICmpRegion(Pred, *C);
-      if (Offset)
-        LHSRange = LHSRange.sub(*Offset);
-      Known = Known.unionWith(LHSRange.toKnownBits());
+    if (match(RHS, m_APInt(C))) {
+      const APInt *Offset = nullptr;
+      if (match(LHS, m_CombineOr(m_V, m_AddLike(m_V, m_APInt(Offset))))) {
+        ConstantRange LHSRange = ConstantRange::makeAllowedICmpRegion(Pred, *C);
+        if (Offset)
+          LHSRange = LHSRange.sub(*Offset);
+        Known = Known.unionWith(LHSRange.toKnownBits());
+      }
+      if ((Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) &&
+          match(LHS, m_c_NUWAdd(m_V, m_Value()))) {
+        Known.Zero.setHighBits(
+            (*C - (Pred == ICmpInst::ICMP_ULT)).countLeadingZeros());
+      }
     }
     break;
   }
@@ -9283,11 +9289,21 @@ void llvm::findValuesAffectedByCondition(
             AddAffected(X);
         }
       } else {
-        // Handle (A + C1) u< C2, which is the canonical form of
-        // A > C3 && A < C4.
-        if (match(A, m_AddLike(m_Value(X), m_ConstantInt())) &&
-            match(B, m_ConstantInt()))
-          AddAffected(X);
+        if (match(B, m_ConstantInt())) {
+          // Handle (A + C1) u< C2, which is the canonical form of
+          // A > C3 && A < C4.
+          if (match(A, m_AddLike(m_Value(X), m_ConstantInt())))
+            AddAffected(X);
+
+          Value *Y;
+          // X & Y u> C -> X >u C && Y >u C
+          // X | Y u< C -> X u< C && Y u< C
+          if (ICmpInst::isUnsigned(Pred) &&
+              match(A, m_NUWAdd(m_Value(X), m_Value(Y)))) {
+            AddAffected(X);
+            AddAffected(Y);
+          }
+        }
 
         // Handle icmp slt/sgt (bitcast X to int), 0/-1, which is supported
         // by computeKnownFPClass().
diff --git a/llvm/test/Transforms/InstCombine/known-bits.ll b/llvm/test/Transforms/InstCombine/known-bits.ll
index 5305c78f691231..0ed63512f8cb49 100644
--- a/llvm/test/Transforms/InstCombine/known-bits.ll
+++ b/llvm/test/Transforms/InstCombine/known-bits.ll
@@ -124,7 +124,6 @@ exit:
   ret i8 %or2
 }
 
-
 define i8 @test_cond_and_bothways(i8 %x) {
 ; CHECK-LABEL: @test_cond_and_bothways(
 ; CHECK-NEXT:    [[AND:%.*]] = and i8 [[X:%.*]], 91
@@ -181,8 +180,6 @@ exit:
   ret i8 %or2
 }
 
-
-
 define i8 @test_cond_and_commuted(i8 %x, i1 %c1, i1 %c2) {
 ; CHECK-LABEL: @test_cond_and_commuted(
 ; CHECK-NEXT:    [[AND:%.*]] = and i8 [[X:%.*]], 3
@@ -343,7 +340,7 @@ exit:
   ret i8 %or2
 }
 
-define i32 @test_icmp_trunc1(i32 %x){
+define i32 @test_icmp_trunc1(i32 %x) {
 ; CHECK-LABEL: @test_icmp_trunc1(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[Y:%.*]] = trunc i32 [[X:%.*]] to i16
@@ -365,7 +362,7 @@ else:
   ret i32 0
 }
 
-define i32 @test_icmp_trunc_assume(i32 %x){
+define i32 @test_icmp_trunc_assume(i32 %x) {
 ; CHECK-LABEL: @test_icmp_trunc_assume(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[Y:%.*]] = trunc i32 [[X:%.*]] to i16
@@ -532,7 +529,103 @@ if.else:
   ret i1 %other
 }
 
+define i8 @test_icmp_add(i8 %n, i8 %n2, i8 %other) {
+; CHECK-LABEL: @test_icmp_add(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[N_ADD:%.*]] = add nuw i8 [[N:%.*]], [[N2:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i8 [[N_ADD]], 32
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    ret i8 0
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i8 [[OTHER:%.*]]
+;
+entry:
+  %n_add = add nuw i8 %n, %n2
+  %cmp = icmp ult i8 %n_add, 32
+  br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+  %r = and i8 %n, 32
+  ret i8 %r
+
+if.else:
+  ret i8 %other
+}
+
+define i8 @test_icmp_add2(i8 %n, i8 %n2, i8 %other) {
+; CHECK-LABEL: @test_icmp_add2(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[N_ADD:%.*]] = add nuw i8 [[N:%.*]], [[N2:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i8 [[N_ADD]], 14
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    ret i8 [[OTHER:%.*]]
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i8 0
+;
+entry:
+  %n_add = add nuw i8 %n, %n2
+  %cmp = icmp uge i8 %n_add, 15
+  br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+  ret i8 %other
+if.else:
+  %r = and i8 %n, 32
+  ret i8 %r
 
+}
+
+define i8 @test_icmp_add_fail_bad_range(i8 %n, i8 %n2, i8 %other) {
+; CHECK-LABEL: @test_icmp_add_fail_bad_range(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[N_ADD:%.*]] = add nuw i8 [[N:%.*]], [[N2:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i8 [[N_ADD]], 33
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    [[R:%.*]] = and i8 [[N]], 32
+; CHECK-NEXT:    ret i8 [[R]]
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i8 [[OTHER:%.*]]
+;
+entry:
+  %n_add = add nuw i8 %n, %n2
+  %cmp = icmp ule i8 %n_add, 32
+  br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+  %r = and i8 %n, 32
+  ret i8 %r
+
+if.else:
+  ret i8 %other
+}
+
+define i8 @test_icmp_add_fail_bad_pred(i8 %n, i8 %n2, i8 %other) {
+; CHECK-LABEL: @test_icmp_add_fail_bad_pred(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[N_ADD:%.*]] = add nuw i8 [[N:%.*]], [[N2:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i8 [[N_ADD]], 32
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    [[R:%.*]] = and i8 [[N]], 32
+; CHECK-NEXT:    ret i8 [[R]]
+; CHECK:       if.else:
+; CHECK-NEXT:    ret i8 [[OTHER:%.*]]
+;
+entry:
+  %n_add = add nuw i8 %n, %n2
+  %cmp = icmp ugt i8 %n_add, 32
+  br i1 %cmp, label %if.then, label %if.else
+
+if.then:
+  %r = and i8 %n, 32
+  ret i8 %r
+
+if.else:
+  ret i8 %other
+}
 
 declare void @use(i1)
 declare void @sink(i8)

llvm/lib/Analysis/ValueTracking.cpp

PR Link: llvm/llvm-project#87180

dtcxzyw · 2024-04-01T06:56:39Z

Can you have a look at the regression dtcxzyw/llvm-opt-benchmark#465 (comment)?

goldsteinn · 2024-04-01T18:42:57Z

Can you have a look at the regression dtcxzyw/llvm-opt-benchmark#465 (comment)?

Okay, reduced test:

define void @foo(i32 %x, i32 %y) {
  %z = add nuw i32 %x, %y
  %cmp0 = icmp ult i32 %z, 255
  br i1 %cmp0, label %L0, label %L2

L0:
  %sub_x = sub i32 255, %x
  br label %L1

L1:
  %phi = phi i32 [ 0, %L0 ], [ 1, %L1 ]
  %cmp1 = icmp ult i32 %phi, %sub_x
  br i1 %cmp1, label %L1, label %L2

L2:
  ret void
}

The issue occurs due to the InstCombine fold at:
https://github.com/llvm/llvm-project/blob/main/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp#L2280-L2288

We produce:

Bad:

define void @foo(i32 %x, i32 %y) local_unnamed_addr {
  %z = add nuw i32 %x, %y
  %cmp0 = icmp ult i32 %z, 255
  br i1 %cmp0, label %L0, label %L2

L0:                                               ; preds = %0
  %sub_x = xor i32 %x, 255
  br label %L1

L1:                                               ; preds = %L1, %L0
  %phi = phi i32 [ 0, %L0 ], [ 1, %L1 ]
  %cmp1 = icmp ult i32 %phi, %sub_x
  br i1 %cmp1, label %L1, label %L2

L2:                                               ; preds = %L1, %0
  ret void
}

vs
Good:

define void @foo(i32 %x, i32 %y) local_unnamed_addr {
  %z = add nuw i32 %x, %y
  %cmp0 = icmp ult i32 %z, 255
  br i1 %cmp0, label %L0, label %L2

L0:                                               ; preds = %0
  %sub_x = sub i32 255, %x
  br label %L1

L1:                                               ; preds = %L1, %L0
  %phi = phi i32 [ 0, %L0 ], [ 1, %L1 ]
  %cmp1 = icmp ult i32 %phi, %sub_x
  br i1 %cmp1, label %L1, label %L2

L2:                                               ; preds = %L1, %0
  ret void
}

The issue then arises because we fold (icmp eq/ne (sub C0, x), C1) -> (icmp eq/ne x, C2) even if the sub has multi-use but we don't do the same for xor:
https://github.com/llvm/llvm-project/blob/main/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp#L3463-L3474

Ill add a patch for the xor case.

Two folds unlocked: `(icmp eq/ne (xor x, C0), C1)` -> `(icmp eq/ne x, C2)` `(icmp eq/ne (xor x, y), 0)` -> `(icmp eq/ne x, y)` This fixes regressions assosiated with llvm#87180

goldsteinn · 2024-04-04T16:46:59Z

My preference would be to not have this blocked by the missing multi-use xor fold. It seems pretty unrelated that we end up with that pattern w/ this patch but not w/ (in that one case) and handling multi-use xor seems independently valuable.

Two folds unlocked: `(icmp eq/ne (xor x, C0), C1)` -> `(icmp eq/ne x, C2)` `(icmp eq/ne (xor x, y), 0)` -> `(icmp eq/ne x, y)` This fixes regressions assosiated with llvm#87180

goldsteinn · 2024-04-05T01:56:16Z

@dtcxzyw i update with support for sub nuw as well.

goldsteinn · 2024-04-11T18:09:16Z

rebased

nikic · 2024-04-25T03:15:49Z

Needs another rebase.

@dtcxzyw i update with support for sub nuw as well.

Why?

goldsteinn · 2024-04-25T16:01:51Z

Needs another rebase.

@dtcxzyw i update with support for sub nuw as well.

Why?

I guess the add nuw mirrors the or case and the sub nuw mirrors the and case. I had meant to leave it as a folllowup todo but it seemed to fit well enough.

goldsteinn · 2024-05-05T00:24:24Z

ping

goldsteinn · 2024-05-10T19:08:25Z

ping

nikic

LGTM

nikic · 2024-05-16T03:29:51Z

llvm/test/Transforms/InstCombine/known-bits.ll

+
+if.else:
+  ret i8 %other
+}    


Negative tests without nuw?

done + rebased/

…add/sub nuw X, Y), C)`; NFC

…), C)` `(icmp ule/ult (add nuw X, Y), C)` implies both `(icmp ule/ult X, C)` and `(icmp ule/ult Y, C)`. We can use this to deduce leading zeros in `X`/`Y`. `(icmp uge/ugt (sub nuw X, Y), C)` implies `(icmp uge/uge X, C)` . We can use this to deduce leading ones in `X`. Proofs: https://alive2.llvm.org/ce/z/sc5k22

Two folds unlocked: `(icmp eq/ne (xor x, C0), C1)` -> `(icmp eq/ne x, C2)` `(icmp eq/ne (xor x, y), 0)` -> `(icmp eq/ne x, y)` This fixes regressions assosiated with llvm#87180

Two folds unlocked: `(icmp eq/ne (xor x, C0), C1)` -> `(icmp eq/ne x, C2)` `(icmp eq/ne (xor x, y), 0)` -> `(icmp eq/ne x, y)` This fixes regressions assosiated with #87180 Closes #87275

goldsteinn requested a review from nikic as a code owner March 31, 2024 03:42

llvmbot added llvm:analysis llvm:transforms labels Mar 31, 2024

goldsteinn changed the title ~~goldsteinn/kb nuw add~~ [ValueTracking] Compute knownbits from (icmp ult/ule (add nuw X, Y), C) Mar 31, 2024

goldsteinn requested a review from dtcxzyw March 31, 2024 03:43

dtcxzyw requested changes Mar 31, 2024

View reviewed changes

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

goldsteinn force-pushed the goldsteinn/kb-nuw-add branch from 0dd2548 to 0341e89 Compare March 31, 2024 16:01

dtcxzyw added a commit to dtcxzyw/llvm-opt-benchmark that referenced this pull request Mar 31, 2024

pre-commit: test PR87180

cecfae9

PR Link: llvm/llvm-project#87180

dtcxzyw mentioned this pull request Mar 31, 2024

pre-commit: test PR87180 dtcxzyw/llvm-opt-benchmark#465

Open

dtcxzyw added a commit to dtcxzyw/llvm-opt-benchmark that referenced this pull request Mar 31, 2024

pre-commit: test PR87180

b487025

PR Link: llvm/llvm-project#87180

goldsteinn mentioned this pull request Apr 1, 2024

[InstCombine] Fold (icmp eq/ne (xor x, y), C1) even if multiuse #87275

Closed

goldsteinn changed the title ~~[ValueTracking] Compute knownbits from (icmp ult/ule (add nuw X, Y), C)~~ [ValueTracking] Compute knownbits from (icmp upred (add/sub nuw X, Y), C) Apr 5, 2024

goldsteinn force-pushed the goldsteinn/kb-nuw-add branch from 0341e89 to 6d6bb29 Compare April 5, 2024 01:55

goldsteinn force-pushed the goldsteinn/kb-nuw-add branch from 6d6bb29 to 7d54e14 Compare April 11, 2024 18:07

goldsteinn force-pushed the goldsteinn/kb-nuw-add branch from 7d54e14 to 88fa1fa Compare April 25, 2024 19:00

nikic approved these changes May 16, 2024

View reviewed changes

goldsteinn added 2 commits May 16, 2024 12:23

[ValueTracking] Add tests for computing knownbits from `(icmp upred (…

1690b9a

…add/sub nuw X, Y), C)`; NFC

goldsteinn force-pushed the goldsteinn/kb-nuw-add branch from 88fa1fa to 1adbd10 Compare May 16, 2024 17:25

goldsteinn closed this in 05347f8 May 16, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[ValueTracking] Compute knownbits from `(icmp upred (add/sub nuw X, Y), C)` #87180

[ValueTracking] Compute knownbits from `(icmp upred (add/sub nuw X, Y), C)` #87180

goldsteinn commented Mar 31, 2024 •

edited

llvmbot commented Mar 31, 2024 •

edited

dtcxzyw commented Apr 1, 2024

goldsteinn commented Apr 1, 2024

goldsteinn commented Apr 4, 2024

goldsteinn commented Apr 5, 2024

goldsteinn commented Apr 11, 2024

nikic commented Apr 25, 2024

goldsteinn commented Apr 25, 2024

goldsteinn commented May 5, 2024

goldsteinn commented May 10, 2024

nikic left a comment

nikic May 16, 2024

goldsteinn May 16, 2024

[ValueTracking] Compute knownbits from (icmp upred (add/sub nuw X, Y), C) #87180

[ValueTracking] Compute knownbits from (icmp upred (add/sub nuw X, Y), C) #87180

Conversation

goldsteinn commented Mar 31, 2024 • edited

llvmbot commented Mar 31, 2024 • edited

dtcxzyw commented Apr 1, 2024

goldsteinn commented Apr 1, 2024

goldsteinn commented Apr 4, 2024

goldsteinn commented Apr 5, 2024

goldsteinn commented Apr 11, 2024

nikic commented Apr 25, 2024

goldsteinn commented Apr 25, 2024

goldsteinn commented May 5, 2024

goldsteinn commented May 10, 2024

nikic left a comment

Choose a reason for hiding this comment

nikic May 16, 2024

Choose a reason for hiding this comment

goldsteinn May 16, 2024

Choose a reason for hiding this comment

[ValueTracking] Compute knownbits from `(icmp upred (add/sub nuw X, Y), C)` #87180

[ValueTracking] Compute knownbits from `(icmp upred (add/sub nuw X, Y), C)` #87180

goldsteinn commented Mar 31, 2024 •

edited

llvmbot commented Mar 31, 2024 •

edited