[llvm] [CVP] Refactor `processMinMaxIntrinsic` to use context-sensitive information (PR #82596)

Yingwei Zheng via llvm-commits llvm-commits at lists.llvm.org
Wed Feb 21 23:55:18 PST 2024


https://github.com/dtcxzyw created https://github.com/llvm/llvm-project/pull/82596

This patch uses `getConstantRangeAtUse` in `processMinMaxIntrinsic` to address the comment https://github.com/llvm/llvm-project/pull/82478#discussion_r1497300920. After this patch we can reuse the range result in https://github.com/llvm/llvm-project/pull/82478.

To my surprise this patch enables a lot of optimizations in my benchmark. I am not sure whether it causes miscompilation :(


>From 970df17d688e5c68a7483a5ae946546a26d20e6f Mon Sep 17 00:00:00 2001
From: Yingwei Zheng <dtcxzyw2333 at gmail.com>
Date: Thu, 22 Feb 2024 15:41:33 +0800
Subject: [PATCH 1/2] [CVP] Add pre-commit tests. NFC.

---
 .../CorrelatedValuePropagation/min-max.ll     | 31 +++++++++++++++++++
 1 file changed, 31 insertions(+)

diff --git a/llvm/test/Transforms/CorrelatedValuePropagation/min-max.ll b/llvm/test/Transforms/CorrelatedValuePropagation/min-max.ll
index 705b6e96fe9e36..05a716748e5d36 100644
--- a/llvm/test/Transforms/CorrelatedValuePropagation/min-max.ll
+++ b/llvm/test/Transforms/CorrelatedValuePropagation/min-max.ll
@@ -235,3 +235,34 @@ define i8 @test19(i8 %x) {
   %r = call i8 @llvm.smax(i8 %x, i8 42)
   ret i8 %r
 }
+
+declare void @body(i32)
+
+define void @test() {
+; CHECK-LABEL: @test(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
+; CHECK:       for.body:
+; CHECK-NEXT:    [[INDVAR:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[INDVAR]], i32 65535)
+; CHECK-NEXT:    call void @body(i32 [[SMAX]])
+; CHECK-NEXT:    [[INC]] = add nsw i32 [[INDVAR]], 1
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i32 [[INDVAR]], 65535
+; CHECK-NEXT:    br i1 [[CMP]], label [[FOR_BODY]], label [[EXIT:%.*]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %for.body
+
+for.body:
+  %indvar = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %smax = call i32 @llvm.smax.i32(i32 %indvar, i32 65535)
+  call void @body(i32 %smax)
+  %inc = add nsw i32 %indvar, 1
+  %cmp = icmp slt i32 %indvar, 65535
+  br i1 %cmp, label %for.body, label %exit
+
+exit:
+  ret void
+}

>From 64e36b88752339d1853abcf16da295e125b82121 Mon Sep 17 00:00:00 2001
From: Yingwei Zheng <dtcxzyw2333 at gmail.com>
Date: Thu, 22 Feb 2024 15:42:07 +0800
Subject: [PATCH 2/2] [CVP] Refactor `processMinMaxIntrinsic` to use
 context-sensitive information

---
 .../Scalar/CorrelatedValuePropagation.cpp     | 26 ++++++++++++-------
 .../CorrelatedValuePropagation/min-max.ll     |  7 +----
 2 files changed, 18 insertions(+), 15 deletions(-)

diff --git a/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp b/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
index 9235850de92f3e..f1efb3109aebe6 100644
--- a/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
+++ b/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
@@ -530,15 +530,23 @@ static bool processAbsIntrinsic(IntrinsicInst *II, LazyValueInfo *LVI) {
 // See if this min/max intrinsic always picks it's one specific operand.
 static bool processMinMaxIntrinsic(MinMaxIntrinsic *MM, LazyValueInfo *LVI) {
   CmpInst::Predicate Pred = CmpInst::getNonStrictPredicate(MM->getPredicate());
-  LazyValueInfo::Tristate Result = LVI->getPredicateAt(
-      Pred, MM->getLHS(), MM->getRHS(), MM, /*UseBlockValue=*/true);
-  if (Result == LazyValueInfo::Unknown)
-    return false;
-
-  ++NumMinMax;
-  MM->replaceAllUsesWith(MM->getOperand(!Result));
-  MM->eraseFromParent();
-  return true;
+  ConstantRange LHS_CR = LVI->getConstantRangeAtUse(MM->getOperandUse(0),
+                                                    /*UndefAllowed*/ true);
+  ConstantRange RHS_CR = LVI->getConstantRangeAtUse(MM->getOperandUse(1),
+                                                    /*UndefAllowed*/ true);
+  if (LHS_CR.icmp(Pred, RHS_CR)) {
+    ++NumMinMax;
+    MM->replaceAllUsesWith(MM->getLHS());
+    MM->eraseFromParent();
+    return true;
+  }
+  if (RHS_CR.icmp(Pred, LHS_CR)) {
+    ++NumMinMax;
+    MM->replaceAllUsesWith(MM->getRHS());
+    MM->eraseFromParent();
+    return true;
+  }
+  return false;
 }
 
 // Rewrite this with.overflow intrinsic as non-overflowing.
diff --git a/llvm/test/Transforms/CorrelatedValuePropagation/min-max.ll b/llvm/test/Transforms/CorrelatedValuePropagation/min-max.ll
index 05a716748e5d36..3e9f1874a6e7d4 100644
--- a/llvm/test/Transforms/CorrelatedValuePropagation/min-max.ll
+++ b/llvm/test/Transforms/CorrelatedValuePropagation/min-max.ll
@@ -71,7 +71,6 @@ define i8 @test6(i8 %x) {
 ; CHECK-LABEL: @test6(
 ; CHECK-NEXT:    [[LIM:%.*]] = icmp uge i8 [[X:%.*]], 42
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[LIM]])
-; CHECK-NEXT:    [[R:%.*]] = call i8 @llvm.umin.i8(i8 [[X]], i8 42)
 ; CHECK-NEXT:    ret i8 42
 ;
   %lim = icmp uge i8 %x, 42
@@ -119,7 +118,6 @@ define i8 @test10(i8 %x) {
 ; CHECK-LABEL: @test10(
 ; CHECK-NEXT:    [[LIM:%.*]] = icmp ule i8 [[X:%.*]], 42
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[LIM]])
-; CHECK-NEXT:    [[R:%.*]] = call i8 @llvm.umax.i8(i8 [[X]], i8 42)
 ; CHECK-NEXT:    ret i8 42
 ;
   %lim = icmp ule i8 %x, 42
@@ -167,7 +165,6 @@ define i8 @test14(i8 %x) {
 ; CHECK-LABEL: @test14(
 ; CHECK-NEXT:    [[LIM:%.*]] = icmp sge i8 [[X:%.*]], 42
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[LIM]])
-; CHECK-NEXT:    [[R:%.*]] = call i8 @llvm.smin.i8(i8 [[X]], i8 42)
 ; CHECK-NEXT:    ret i8 42
 ;
   %lim = icmp sge i8 %x, 42
@@ -215,7 +212,6 @@ define i8 @test18(i8 %x) {
 ; CHECK-LABEL: @test18(
 ; CHECK-NEXT:    [[LIM:%.*]] = icmp sle i8 [[X:%.*]], 42
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[LIM]])
-; CHECK-NEXT:    [[R:%.*]] = call i8 @llvm.smax.i8(i8 [[X]], i8 42)
 ; CHECK-NEXT:    ret i8 42
 ;
   %lim = icmp sle i8 %x, 42
@@ -244,8 +240,7 @@ define void @test() {
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    [[INDVAR:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
-; CHECK-NEXT:    [[SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[INDVAR]], i32 65535)
-; CHECK-NEXT:    call void @body(i32 [[SMAX]])
+; CHECK-NEXT:    call void @body(i32 65535)
 ; CHECK-NEXT:    [[INC]] = add nsw i32 [[INDVAR]], 1
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i32 [[INDVAR]], 65535
 ; CHECK-NEXT:    br i1 [[CMP]], label [[FOR_BODY]], label [[EXIT:%.*]]



More information about the llvm-commits mailing list