[llvm] [ARM] Handle debug instructions in LowOverheadLoop (PR #221699)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-arm

Author: David Green (davemgreen)

<details>
<summary>Changes</summary>

The last instruction in the block might be a debug instruction, in which case we should get the last non-debug instruction.

Fixes #<!-- -->221633.

---
Full diff: https://github.com/llvm/llvm-project/pull/221699.diff


2 Files Affected:

- (modified) llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp (+3-2) 
- (added) llvm/test/CodeGen/Thumb2/LowOverheadLoops/dbg-verify.ll (+81) 


``````````diff
diff --git a/llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp b/llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp
index 1719165fb6717..f89f209d6ecf8 100644
--- a/llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp
+++ b/llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp
@@ -685,10 +685,11 @@ bool LowOverheadLoop::ValidateTailPredicate() {
     // before entering the loop.
     auto CannotProvideElements = [this](MachineBasicBlock *MBB,
                                         MCRegister NumElements) {
-      if (MBB->empty())
+      auto Back = MBB->getLastNonDebugInstr();
+      if (Back == MBB->end())
         return false;
       // NumElements is redefined in this block.
-      if (RDI.hasLocalDefBefore(&MBB->back(), NumElements))
+      if (RDI.hasLocalDefBefore(&*Back, NumElements))
         return true;
 
       // Don't continue searching up through multiple predecessors.
diff --git a/llvm/test/CodeGen/Thumb2/LowOverheadLoops/dbg-verify.ll b/llvm/test/CodeGen/Thumb2/LowOverheadLoops/dbg-verify.ll
new file mode 100644
index 0000000000000..a13c65e28be67
--- /dev/null
+++ b/llvm/test/CodeGen/Thumb2/LowOverheadLoops/dbg-verify.ll
@@ -0,0 +1,81 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve %s -o - | FileCheck %s
+
+ at ccmdv2_carrier_address = constant ptr blockaddress(@ccmdv2_source_carrier, %carrier)
+
+define arm_aapcs_vfpcc <4 x float> @ccmdv2_source_carrier(ptr %pSrc, i32 %blockSize) !dbg !7 {
+; CHECK-LABEL: ccmdv2_source_carrier:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    .save {r7, lr}
+; CHECK-NEXT:    push {r7, lr}
+; CHECK-NEXT:    .cfi_def_cfa_offset 8
+; CHECK-NEXT:    .cfi_offset lr, -4
+; CHECK-NEXT:    .cfi_offset r7, -8
+; CHECK-NEXT:    .loc 1 1 0 prologue_end
+; CHECK-NEXT:    cbz r1, .LBB0_5
+; CHECK-NEXT:  @ %bb.1: @ %start.block
+; CHECK-NEXT:    dlstp.32 lr, r1
+; CHECK-NEXT:  .Ltmp0: @ Block address taken
+; CHECK-NEXT:  @ %bb.2: @ %carrier
+; CHECK-NEXT:    @APP
+; CHECK-NEXT:    @NO_APP
+; CHECK-NEXT:    @DEBUG_VALUE: ccmdv2_source_carrier:count <- $r1
+; CHECK-NEXT:  .LBB0_3: @ %do.body
+; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    vldrw.u32 q0, [r0], #16
+; CHECK-NEXT:    letp lr, .LBB0_3
+; CHECK-NEXT:  @ %bb.4: @ %common.ret
+; CHECK-NEXT:    pop {r7, pc}
+; CHECK-NEXT:  .LBB0_5:
+; CHECK-NEXT:    vmov.i32 q0, #0x0
+; CHECK-NEXT:    pop {r7, pc}
+entry:
+  %is.zero = icmp eq i32 %blockSize, 0
+  %rnd.up = add i32 %blockSize, 3
+  %iterations = lshr i32 %rnd.up, 2
+  br i1 %is.zero, label %cleanup, label %start.block
+
+start.block:
+  %loop.start = call i32 @llvm.start.loop.iterations.i32(i32 %iterations)
+  br label %carrier
+
+carrier:
+  %ccmdv2_count = call i32 asm sideeffect "", "={r1},0"(i32 %blockSize)
+  call void @llvm.dbg.value(metadata i32 %ccmdv2_count, metadata !6, metadata !DIExpression()), !dbg !9
+  br label %do.body
+
+do.body:
+  %hw.counter = phi i32 [ %loop.start, %carrier ], [ %hw.next, %do.body ]
+  %blockSize.addr.0 = phi i32 [ %ccmdv2_count, %carrier ], [ %sub, %do.body ]
+  %pSrc.addr.0 = phi ptr [ %pSrc, %carrier ], [ %add.ptr, %do.body ]
+  %pred = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %blockSize.addr.0)
+  %value = tail call <4 x float> @llvm.masked.load.v4f32.p0(ptr %pSrc.addr.0, i32 4, <4 x i1> %pred, <4 x float> zeroinitializer)
+  %add.ptr = getelementptr inbounds float, ptr %pSrc.addr.0, i32 4
+  %sub = add i32 %blockSize.addr.0, -4
+  %hw.next = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 %hw.counter, i32 1)
+  %continue = icmp ne i32 %hw.next, 0
+  br i1 %continue, label %do.body, label %do.end
+
+do.end:
+  ret <4 x float> %value
+
+cleanup:
+  ret <4 x float> zeroinitializer
+}
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!3, !4}
+!llvm.ident = !{!5}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C11, file: !1, producer: "ccmdv2", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, enums: !2)
+!1 = !DIFile(filename: "llvm-arm-loloop-query-window-source-v1.c", directory: "/")
+!2 = !{}
+!3 = !{i32 2, !"Dwarf Version", i32 4}
+!4 = !{i32 2, !"Debug Info Version", i32 3}
+!5 = !{!"ccmdv2"}
+!6 = !DILocalVariable(name: "count", scope: !7, file: !1, line: 2, type: !8)
+!7 = distinct !DISubprogram(name: "ccmdv2_source_carrier", scope: !1, file: !1, line: 1, type: !10, scopeLine: 1, spFlags: DISPFlagDefinition, unit: !0, retainedNodes: !2)
+!8 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
+!9 = !DILocation(line: 2, column: 3, scope: !7)
+!10 = !DISubroutineType(types: !2)
+

``````````

</details>


https://github.com/llvm/llvm-project/pull/221699