[Mlir-commits] [llvm] [mlir] [lld] [clang] [NFC][IRCE] Add unit test to show room for improvement (PR #71506)

[Aleksandr Popov]

https://github.com/aleks-tmb updated https://github.com/llvm/llvm-project/pull/71506

>From 1c946c72f19bb553ad7af2e9b4edb1796eeaaaae Mon Sep 17 00:00:00 2001
From: Aleksander Popov <apopov at azul.com>
Date: Tue, 7 Nov 2023 05:16:45 +0100
Subject: [PATCH] [NFC][IRCE] Add unit test to show room for improvement

Add tests for compound loop bounds where IRCE is possible

if (K > 0 && M > 0)
  for (i = 0; i < min(K, M); i++) {...}

if (K > 0 && M > 0)
  for (i = min(K, M); i >= 0; i--) {...}
---
 .../Transforms/IRCE/compound-loop-bound.ll    | 132 ++++++++++++++++++
 1 file changed, 132 insertions(+)
 create mode 100644 llvm/test/Transforms/IRCE/compound-loop-bound.ll

diff --git a/llvm/test/Transforms/IRCE/compound-loop-bound.ll b/llvm/test/Transforms/IRCE/compound-loop-bound.ll
new file mode 100644
index 000000000000000..0930d19e22154fc
--- /dev/null
+++ b/llvm/test/Transforms/IRCE/compound-loop-bound.ll
@@ -0,0 +1,132 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 3
+; RUN: opt -passes=irce < %s -S | FileCheck %s
+
+; if (K > 0 && M > 0)
+;   for (i = 0; i < min(K, M); i++) {...}
+;
+; TODO: Loop bounds are safe according to loop guards. IRCE is allowed.
+define void @incrementing_loop(ptr %arr, ptr %len_ptr, i32 %K, i32 %M) {
+; CHECK-LABEL: define void @incrementing_loop(
+; CHECK-SAME: ptr [[ARR:%.*]], ptr [[LEN_PTR:%.*]], i32 [[K:%.*]], i32 [[M:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[LEN:%.*]] = load i32, ptr [[LEN_PTR]], align 4, !range [[RNG0:![0-9]+]]
+; CHECK-NEXT:    [[CHECK0:%.*]] = icmp sgt i32 [[K]], 0
+; CHECK-NEXT:    [[CHECK1:%.*]] = icmp sgt i32 [[M]], 0
+; CHECK-NEXT:    [[AND:%.*]] = and i1 [[CHECK0]], [[CHECK1]]
+; CHECK-NEXT:    br i1 [[AND]], label [[PREHEADER:%.*]], label [[EXIT:%.*]]
+; CHECK:       preheader:
+; CHECK-NEXT:    [[SMIN:%.*]] = call i32 @llvm.smin.i32(i32 [[K]], i32 [[M]])
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[IDX:%.*]] = phi i32 [ 0, [[PREHEADER]] ], [ [[IDX_NEXT:%.*]], [[IN_BOUNDS:%.*]] ]
+; CHECK-NEXT:    [[IDX_NEXT]] = add i32 [[IDX]], 1
+; CHECK-NEXT:    [[GUARD:%.*]] = icmp slt i32 [[IDX]], [[LEN]]
+; CHECK-NEXT:    br i1 [[GUARD]], label [[IN_BOUNDS]], label [[OUT_OF_BOUNDS:%.*]]
+; CHECK:       in.bounds:
+; CHECK-NEXT:    [[ADDR:%.*]] = getelementptr i32, ptr [[ARR]], i32 [[IDX]]
+; CHECK-NEXT:    store i32 0, ptr [[ADDR]], align 4
+; CHECK-NEXT:    [[NEXT:%.*]] = icmp slt i32 [[IDX_NEXT]], [[SMIN]]
+; CHECK-NEXT:    br i1 [[NEXT]], label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
+; CHECK:       out.of.bounds:
+; CHECK-NEXT:    ret void
+; CHECK:       exit.loopexit:
+; CHECK-NEXT:    br label [[EXIT]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %len = load i32, ptr %len_ptr, !range !0
+  %check0 = icmp sgt i32 %K, 0
+  %check1 = icmp sgt i32 %M, 0
+  %and = and i1 %check0, %check1
+  br i1 %and, label %preheader, label %exit
+
+preheader:
+  %smin = call i32 @llvm.smin.i32(i32 %K, i32 %M)
+  br label %loop
+
+loop:
+  %idx = phi i32 [ 0, %preheader ], [ %idx.next, %in.bounds ]
+  %idx.next = add i32 %idx, 1
+  %guard = icmp slt i32 %idx, %len
+  br i1 %guard, label %in.bounds, label %out.of.bounds
+
+in.bounds:
+  %addr = getelementptr i32, ptr %arr, i32 %idx
+  store i32 0, ptr %addr
+  %next = icmp slt i32 %idx.next, %smin
+  br i1 %next, label %loop, label %exit
+
+out.of.bounds:
+  ret void
+
+exit:
+  ret void
+}
+
+; if (K > 0 && M > 0)
+;   for (i = min(K, M); i >= 0; i--) {...}
+;
+; TODO: Loop bounds are safe according to loop guards. IRCE is allowed.
+define void @decrementing_loop(ptr %arr, ptr %len_ptr, i32 %K, i32 %M) {
+; CHECK-LABEL: define void @decrementing_loop(
+; CHECK-SAME: ptr [[ARR:%.*]], ptr [[LEN_PTR:%.*]], i32 [[K:%.*]], i32 [[M:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[LEN:%.*]] = load i32, ptr [[LEN_PTR]], align 4, !range [[RNG0]]
+; CHECK-NEXT:    [[CHECK0:%.*]] = icmp sgt i32 [[K]], 0
+; CHECK-NEXT:    [[CHECK1:%.*]] = icmp sgt i32 [[M]], 0
+; CHECK-NEXT:    [[AND:%.*]] = and i1 [[CHECK0]], [[CHECK1]]
+; CHECK-NEXT:    br i1 [[AND]], label [[PREHEADER:%.*]], label [[EXIT:%.*]]
+; CHECK:       preheader:
+; CHECK-NEXT:    [[SMIN:%.*]] = call i32 @llvm.smin.i32(i32 [[K]], i32 [[M]])
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[IDX:%.*]] = phi i32 [ [[SMIN]], [[PREHEADER]] ], [ [[IDX_DEC:%.*]], [[IN_BOUNDS:%.*]] ]
+; CHECK-NEXT:    [[IDX_DEC]] = sub i32 [[IDX]], 1
+; CHECK-NEXT:    [[GUARD:%.*]] = icmp slt i32 [[IDX]], [[LEN]]
+; CHECK-NEXT:    br i1 [[GUARD]], label [[IN_BOUNDS]], label [[OUT_OF_BOUNDS:%.*]]
+; CHECK:       in.bounds:
+; CHECK-NEXT:    [[ADDR:%.*]] = getelementptr i32, ptr [[ARR]], i32 [[IDX]]
+; CHECK-NEXT:    store i32 0, ptr [[ADDR]], align 4
+; CHECK-NEXT:    [[NEXT:%.*]] = icmp sgt i32 [[IDX_DEC]], -1
+; CHECK-NEXT:    br i1 [[NEXT]], label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
+; CHECK:       out.of.bounds:
+; CHECK-NEXT:    ret void
+; CHECK:       exit.loopexit:
+; CHECK-NEXT:    br label [[EXIT]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+  entry:
+  %len = load i32, ptr %len_ptr, !range !0
+  %check0 = icmp sgt i32 %K, 0
+  %check1 = icmp sgt i32 %M, 0
+  %and = and i1 %check0, %check1
+  br i1 %and, label %preheader, label %exit
+
+  preheader:
+  %smin = call i32 @llvm.smin.i32(i32 %K, i32 %M)
+  br label %loop
+
+  loop:
+  %idx = phi i32 [ %smin, %preheader ] , [ %idx.dec, %in.bounds ]
+  %idx.dec = sub i32 %idx, 1
+  %guard = icmp slt i32 %idx, %len
+  br i1 %guard, label %in.bounds, label %out.of.bounds
+
+  in.bounds:
+  %addr = getelementptr i32, ptr %arr, i32 %idx
+  store i32 0, ptr %addr
+  %next = icmp sgt i32 %idx.dec, -1
+  br i1 %next, label %loop, label %exit
+
+  out.of.bounds:
+  ret void
+
+  exit:
+  ret void
+}
+
+declare i32 @llvm.smin.i32(i32, i32)
+
+!0 = !{i32 0, i32 2147483647}