[llvm] [CVP]: Fold `icmp eq X, C` to `trunc X to i1` if C=2k+1 and X in [2k, 2k+1] (PR #83829)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Mar 4 03:42:01 PST 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-transforms
Author: Monad (YanWQ-monad)
<details>
<summary>Changes</summary>
For
```
define i1 @<!-- -->src(i8 %x) {
%1 = icmp ult i8 %x, 2
br i1 %1, label %bb1, label %bb2
bb2:
%2 = tail call i1 @<!-- -->dynamic()
br label %bb3
bb1:
%3 = icmp eq i8 %x, 1 ; <===
br label %bb3
bb3:
%4 = phi i1 [ %3, %bb1 ], [ %2, %bb2 ]
ret i1 %4
}
```
we can fold `icmp eq i8 %x, 1` to `trunc i8 %x to i1` since x is in [0, 1].
The alive2 proof is [https://alive2.llvm.org/ce/z/sGig3-](https://alive2.llvm.org/ce/z/sGig3-).
Generally, for
- `icmp eq X, C` if C = 2k+1 and X is in [2k, 2k+1]
- or `icmp ne X, C` if C = 2k and X is in [2k, 2k+1]
we can fold it to `trunc X to i1`.
With this fold, RISC-V can eliminate two instructions, while ARM can eliminate one instruction on the hot path.
The real-world case: https://github.com/rust-lang/rust/issues/121673
---
Full diff: https://github.com/llvm/llvm-project/pull/83829.diff
3 Files Affected:
- (modified) llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp (+36)
- (modified) llvm/test/Transforms/CorrelatedValuePropagation/icmp.ll (+62-2)
- (modified) llvm/test/Transforms/JumpThreading/pr33917.ll (+4-4)
``````````diff
diff --git a/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp b/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
index 490cb7e528eb6f..73be5f0b016603 100644
--- a/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
+++ b/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
@@ -332,6 +332,39 @@ static bool constantFoldCmp(CmpInst *Cmp, LazyValueInfo *LVI) {
return true;
}
+/// Given an icmp `icmp eq X, C`,
+/// if we already know that C is 2k+1 and X is in [2k, 2k+1],
+/// then we can fold it to `trunc X to i1`.
+static bool processEqualityICmp(CmpInst *Cmp, LazyValueInfo *LVI) {
+ if (Cmp->getType()->isVectorTy() ||
+ !Cmp->getOperand(0)->getType()->isIntegerTy() || !Cmp->isEquality())
+ return false;
+
+ Value *Op0 = Cmp->getOperand(0);
+ Value *Op1 = Cmp->getOperand(1);
+ ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
+ if (!CI)
+ return false;
+
+ ConstantRange Range =
+ LVI->getConstantRangeAtUse(Cmp->getOperandUse(0), /*UndefAllowed*/ true);
+ APInt RangeSize = Range.getUpper() - Range.getLower();
+ APInt Value = CI->getValue();
+ if (RangeSize != 2 || !Range.contains(Value))
+ return false;
+
+ bool ShouldBeOdd = Cmp->getPredicate() == ICmpInst::Predicate::ICMP_EQ;
+ if ((CI->getValue() & 1) == ShouldBeOdd) {
+ IRBuilder<> B{Cmp};
+ auto *Trunc = B.CreateTruncOrBitCast(Op0, Cmp->getType());
+ Cmp->replaceAllUsesWith(Trunc);
+ Cmp->eraseFromParent();
+ return true;
+ }
+
+ return false;
+}
+
static bool processCmp(CmpInst *Cmp, LazyValueInfo *LVI) {
if (constantFoldCmp(Cmp, LVI))
return true;
@@ -340,6 +373,9 @@ static bool processCmp(CmpInst *Cmp, LazyValueInfo *LVI) {
if (processICmp(ICmp, LVI))
return true;
+ if (processEqualityICmp(Cmp, LVI))
+ return true;
+
return false;
}
diff --git a/llvm/test/Transforms/CorrelatedValuePropagation/icmp.ll b/llvm/test/Transforms/CorrelatedValuePropagation/icmp.ll
index 101820a4c65f23..ccfbc274d570ce 100644
--- a/llvm/test/Transforms/CorrelatedValuePropagation/icmp.ll
+++ b/llvm/test/Transforms/CorrelatedValuePropagation/icmp.ll
@@ -594,10 +594,10 @@ define void @test_cmp_phi(i8 %a) {
; CHECK-NEXT: br i1 [[C0]], label [[LOOP:%.*]], label [[EXIT:%.*]]
; CHECK: loop:
; CHECK-NEXT: [[P:%.*]] = phi i8 [ [[A]], [[ENTRY:%.*]] ], [ [[B:%.*]], [[LOOP]] ]
-; CHECK-NEXT: [[C1:%.*]] = icmp ne i8 [[P]], 0
+; CHECK-NEXT: [[TMP0:%.*]] = trunc i8 [[P]] to i1
; CHECK-NEXT: [[C4:%.*]] = call i1 @get_bool()
; CHECK-NEXT: [[B]] = zext i1 [[C4]] to i8
-; CHECK-NEXT: br i1 [[C1]], label [[LOOP]], label [[EXIT]]
+; CHECK-NEXT: br i1 [[TMP0]], label [[LOOP]], label [[EXIT]]
; CHECK: exit:
; CHECK-NEXT: ret void
;
@@ -1455,3 +1455,63 @@ entry:
%select = select i1 %cmp1, i1 %cmp2, i1 false
ret i1 %select
}
+
+define i1 @test_icmp_eq_on_valid_bool_range(i8 %x) {
+; CHECK-LABEL: @test_icmp_eq_on_valid_bool_range(
+; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i8 [[X:%.*]], 2
+; CHECK-NEXT: br i1 [[TMP1]], label [[BB1:%.*]], label [[BB2:%.*]]
+; CHECK: bb2:
+; CHECK-NEXT: [[TMP2:%.*]] = tail call i1 @get_bool()
+; CHECK-NEXT: br label [[BB3:%.*]]
+; CHECK: bb1:
+; CHECK-NEXT: [[TMP3:%.*]] = trunc i8 [[X]] to i1
+; CHECK-NEXT: br label [[BB3]]
+; CHECK: bb3:
+; CHECK-NEXT: [[TMP4:%.*]] = phi i1 [ [[TMP3]], [[BB1]] ], [ [[TMP2]], [[BB2]] ]
+; CHECK-NEXT: ret i1 [[TMP4]]
+;
+ %1 = icmp ult i8 %x, 2
+ br i1 %1, label %bb1, label %bb2
+
+bb2:
+ %2 = tail call i1 @get_bool()
+ br label %bb3
+
+bb1:
+ %3 = icmp eq i8 %x, 1
+ br label %bb3
+
+bb3:
+ %4 = phi i1 [ %3, %bb1 ], [ %2, %bb2 ]
+ ret i1 %4
+}
+
+define i1 @test_icmp_ne_on_valid_bool_range(i8 %x) {
+; CHECK-LABEL: @test_icmp_ne_on_valid_bool_range(
+; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i8 [[X:%.*]], 2
+; CHECK-NEXT: br i1 [[TMP1]], label [[BB1:%.*]], label [[BB2:%.*]]
+; CHECK: bb2:
+; CHECK-NEXT: [[TMP2:%.*]] = tail call i1 @get_bool()
+; CHECK-NEXT: br label [[BB3:%.*]]
+; CHECK: bb1:
+; CHECK-NEXT: [[TMP3:%.*]] = trunc i8 [[X]] to i1
+; CHECK-NEXT: br label [[BB3]]
+; CHECK: bb3:
+; CHECK-NEXT: [[TMP4:%.*]] = phi i1 [ [[TMP3]], [[BB1]] ], [ [[TMP2]], [[BB2]] ]
+; CHECK-NEXT: ret i1 [[TMP4]]
+;
+ %1 = icmp ult i8 %x, 2
+ br i1 %1, label %bb1, label %bb2
+
+bb2:
+ %2 = tail call i1 @get_bool()
+ br label %bb3
+
+bb1:
+ %3 = icmp ne i8 %x, 0
+ br label %bb3
+
+bb3:
+ %4 = phi i1 [ %3, %bb1 ], [ %2, %bb2 ]
+ ret i1 %4
+}
diff --git a/llvm/test/Transforms/JumpThreading/pr33917.ll b/llvm/test/Transforms/JumpThreading/pr33917.ll
index 7d21a4e1781519..20380c769bf173 100644
--- a/llvm/test/Transforms/JumpThreading/pr33917.ll
+++ b/llvm/test/Transforms/JumpThreading/pr33917.ll
@@ -15,16 +15,16 @@ define void @patatino() personality ptr @rust_eh_personality {
; CHECK-LABEL: @patatino(
; CHECK-NEXT: bb9:
; CHECK-NEXT: [[T9:%.*]] = invoke ptr @foo()
-; CHECK-NEXT: to label [[GOOD:%.*]] unwind label [[BAD:%.*]]
+; CHECK-NEXT: to label [[GOOD:%.*]] unwind label [[BAD:%.*]]
; CHECK: bad:
; CHECK-NEXT: [[T10:%.*]] = landingpad { ptr, i32 }
-; CHECK-NEXT: cleanup
+; CHECK-NEXT: cleanup
; CHECK-NEXT: resume { ptr, i32 } [[T10]]
; CHECK: good:
; CHECK-NEXT: [[T11:%.*]] = icmp ne ptr [[T9]], null
; CHECK-NEXT: [[T12:%.*]] = zext i1 [[T11]] to i64
-; CHECK-NEXT: [[COND:%.*]] = icmp eq i64 [[T12]], 1
-; CHECK-NEXT: br i1 [[COND]], label [[IF_TRUE:%.*]], label [[DONE:%.*]]
+; CHECK-NEXT: [[TMP0:%.*]] = trunc i64 [[T12]] to i1
+; CHECK-NEXT: br i1 [[TMP0]], label [[IF_TRUE:%.*]], label [[DONE:%.*]]
; CHECK: if_true:
; CHECK-NEXT: call void @llvm.assume(i1 [[T11]])
; CHECK-NEXT: br label [[DONE]]
``````````
</details>
https://github.com/llvm/llvm-project/pull/83829
More information about the llvm-commits
mailing list