[llvm] [FuncSpec] Improve handling of Comparison Instructions (PR #114073)
Hari Limaye via llvm-commits
llvm-commits at lists.llvm.org
Tue Oct 29 08:40:33 PDT 2024
https://github.com/hazzlim created https://github.com/llvm/llvm-project/pull/114073
When visiting comparison instructions during computation of a specializations's bonus, make use of information from the lattice value of the other operand in the case where we have not found this to have a specific constant value.
>From 3c03c26aae3752bf8f66323a0dffb2c0781304e6 Mon Sep 17 00:00:00 2001
From: Hari Limaye <hari.limaye at arm.com>
Date: Tue, 29 Oct 2024 12:11:52 +0000
Subject: [PATCH] [FuncSpec] Improve handling of Comparison Instructions
When visiting comparison instructions during computation of a
specializations's bonus, make use of information from the lattice value
of the other operand in the case where we have not found this to have a
specific constant value.
---
.../Transforms/IPO/FunctionSpecialization.cpp | 22 ++++--
.../FunctionSpecialization/cmp-with-range.ll | 74 +++++++++++++++++++
2 files changed, 90 insertions(+), 6 deletions(-)
create mode 100644 llvm/test/Transforms/FunctionSpecialization/cmp-with-range.ll
diff --git a/llvm/lib/Transforms/IPO/FunctionSpecialization.cpp b/llvm/lib/Transforms/IPO/FunctionSpecialization.cpp
index 919d3143a13f7e..2c1f29d719c5b8 100644
--- a/llvm/lib/Transforms/IPO/FunctionSpecialization.cpp
+++ b/llvm/lib/Transforms/IPO/FunctionSpecialization.cpp
@@ -468,16 +468,26 @@ Constant *InstCostVisitor::visitCastInst(CastInst &I) {
Constant *InstCostVisitor::visitCmpInst(CmpInst &I) {
assert(LastVisited != KnownConstants.end() && "Invalid iterator!");
+ Constant *Const = LastVisited->second;
bool Swap = I.getOperand(1) == LastVisited->first;
Value *V = Swap ? I.getOperand(0) : I.getOperand(1);
Constant *Other = findConstantFor(V, KnownConstants);
- if (!Other)
- return nullptr;
- Constant *Const = LastVisited->second;
- return Swap ?
- ConstantFoldCompareInstOperands(I.getPredicate(), Other, Const, DL)
- : ConstantFoldCompareInstOperands(I.getPredicate(), Const, Other, DL);
+ if (Other)
+ return Swap ? ConstantFoldCompareInstOperands(I.getPredicate(), Other, Const,
+ DL)
+ : ConstantFoldCompareInstOperands(I.getPredicate(), Const, Other,
+ DL);
+
+ // If we haven't found Other to be a specific constant value, we may still be
+ // able constant fold the comparison using information from the lattice value.
+ ValueLatticeElement ConstLV = ValueLatticeElement();
+ ConstLV.markConstant(Const);
+ const ValueLatticeElement &OtherLV = Solver.getLatticeValueFor(V);
+ auto &V1State = Swap ? OtherLV : ConstLV;
+ auto &V2State = Swap ? ConstLV : OtherLV;
+ return V1State.getCompare(I.getPredicate(), I.getType(), V2State, DL);
+
}
Constant *InstCostVisitor::visitUnaryOperator(UnaryOperator &I) {
diff --git a/llvm/test/Transforms/FunctionSpecialization/cmp-with-range.ll b/llvm/test/Transforms/FunctionSpecialization/cmp-with-range.ll
new file mode 100644
index 00000000000000..bea044a9904ff5
--- /dev/null
+++ b/llvm/test/Transforms/FunctionSpecialization/cmp-with-range.ll
@@ -0,0 +1,74 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --include-generated-funcs --version 5
+; RUN: opt -passes="ipsccp<func-spec>" -funcspec-min-function-size=1 \
+; RUN: -funcspec-for-literal-constant=true \
+; RUN: -funcspec-min-codesize-savings=50 \
+; RUN: -funcspec-min-latency-savings=0 \
+; RUN: -S < %s | FileCheck %s
+
+; Verify that we are able to estimate the codesize savings arising from a branch
+; based on a comparison with a value found to have a constant range by IPSCCP.
+define i32 @main() {
+ %notspec = call i32 @test(i32 8)
+ %spec = call i32 @test(i32 0)
+ %sum = add i32 %notspec, %spec
+ ret i32 %sum
+}
+
+define i32 @test(i32 %x) {
+entry:
+ %range = call i32 @foo(), !range !{ i32 1, i32 0 }
+ %bound = shl nsw nuw i32 %range, 3
+ %cmp = icmp uge i32 %x, %bound
+ br i1 %cmp, label %if.then, label %if.end
+
+if.then:
+ call void @do_something()
+ call void @do_something()
+ call void @do_something()
+ call void @do_something()
+ br label %if.end
+
+if.end:
+ %res = phi i32 [ 0, %entry ], [ 1, %if.then]
+ ret i32 %res
+}
+
+declare i32 @foo()
+declare void @do_something()
+; CHECK-LABEL: define range(i32 0, 2) i32 @main() {
+; CHECK-NEXT: [[NOTSPEC:%.*]] = call i32 @test(i32 8)
+; CHECK-NEXT: [[SPEC:%.*]] = call i32 @test.specialized.1(i32 0)
+; CHECK-NEXT: [[SUM:%.*]] = add nuw nsw i32 [[NOTSPEC]], 0
+; CHECK-NEXT: ret i32 [[SUM]]
+;
+;
+; CHECK-LABEL: define range(i32 0, 2) i32 @test(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*]]:
+; CHECK-NEXT: [[RANGE:%.*]] = call i32 @foo(), !range [[RNG0:![0-9]+]]
+; CHECK-NEXT: [[BOUND:%.*]] = shl nuw nsw i32 [[RANGE]], 3
+; CHECK-NEXT: [[CMP:%.*]] = icmp uge i32 [[X]], [[BOUND]]
+; CHECK-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
+; CHECK: [[IF_THEN]]:
+; CHECK-NEXT: call void @do_something()
+; CHECK-NEXT: call void @do_something()
+; CHECK-NEXT: call void @do_something()
+; CHECK-NEXT: call void @do_something()
+; CHECK-NEXT: br label %[[IF_END]]
+; CHECK: [[IF_END]]:
+; CHECK-NEXT: [[RES:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ 1, %[[IF_THEN]] ]
+; CHECK-NEXT: ret i32 [[RES]]
+;
+;
+; CHECK-LABEL: define internal i32 @test.specialized.1(
+; CHECK-SAME: i32 [[X:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[RANGE:%.*]] = call i32 @foo(), !range [[RNG0]]
+; CHECK-NEXT: [[BOUND:%.*]] = shl nuw nsw i32 [[RANGE]], 3
+; CHECK-NEXT: br label %[[IF_END:.*]]
+; CHECK: [[IF_END]]:
+; CHECK-NEXT: ret i32 poison
+;
+;.
+; CHECK: [[RNG0]] = !{i32 1, i32 0}
+;.
More information about the llvm-commits
mailing list