[DebugInfo][AMDGPU] Convert a debug-intrinsic method to debug records

[jmorse]

CC AMDGPU folks -- I've recently run into this function when stripping out support for debug intrinsics in LLVM. It appears this function wasn't handled in the initial migration to debug records a year ago, probably because it didn't lead to any test failures, and it's been dormant/unexercised since. I've updated the code to deal with debug records, but I've no idea whether it's doing the right thing or not.

The test update is done by the automatic-updates script, so that's just a mechanical change (that could be wrong).

Note that we drop a call to copyMetadata: debug records can't carry additional metadata like instructions, I don't believe anything debug-info related should be carrying such metadata.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Jeremy Morse (jmorse)

Changes

CC AMDGPU folks -- I've recently run into this function when stripping out support for debug intrinsics in LLVM. It appears this function wasn't handled in the initial migration to debug records a year ago, probably because it didn't lead to any test failures, and it's been dormant/unexercised since. I've updated the code to deal with debug records, but I've no idea whether it's doing the right thing or not.

The test update is done by the automatic-updates script, so that's just a mechanical change (that could be wrong).

Note that we drop a call to copyMetadata: debug records can't carry additional metadata like instructions, I don't believe anything debug-info related should be carrying such metadata.

---

Full diff: https://github.com/llvm/llvm-project/pull/149505.diff

2 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/AMDGPULowerBufferFatPointers.cpp (+7-8)
• (modified) llvm/test/CodeGen/AMDGPU/lower-buffer-fat-pointers-unoptimized-debug-data.ll (+12-6)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPULowerBufferFatPointers.cpp b/llvm/lib/Target/AMDGPU/AMDGPULowerBufferFatPointers.cpp
index fa8af68817dfc..359e992d61580 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULowerBufferFatPointers.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULowerBufferFatPointers.cpp
@@ -1583,15 +1583,14 @@ void SplitPtrStructs::killAndReplaceSplitInstructions(
     if (!SplitUsers.contains(I))
       continue;
 
-    SmallVector<DbgValueInst *> Dbgs;
-    findDbgValues(Dbgs, I);
-    for (auto *Dbg : Dbgs) {
-      IRB.SetInsertPoint(Dbg);
+    SmallVector<DbgValueInst *> DIs;
+    SmallVector<DbgVariableRecord *> Dbgs;
+    findDbgValues(DIs, I, &Dbgs);
+    for (DbgVariableRecord *Dbg : Dbgs) {
       auto &DL = I->getDataLayout();
       assert(isSplitFatPtr(I->getType()) &&
              "We should've RAUW'd away loads, stores, etc. at this point");
-      auto *OffDbg = cast<DbgValueInst>(Dbg->clone());
-      copyMetadata(OffDbg, Dbg);
+      DbgVariableRecord *OffDbg = Dbg->clone();
       auto [Rsrc, Off] = getPtrParts(I);
 
       int64_t RsrcSz = DL.getTypeSizeInBits(Rsrc->getType());
@@ -1606,9 +1605,9 @@ void SplitPtrStructs::killAndReplaceSplitInstructions(
       if (OffExpr) {
         OffDbg->setExpression(*OffExpr);
         OffDbg->replaceVariableLocationOp(I, Off);
-        IRB.Insert(OffDbg);
+        OffDbg->insertBefore(Dbg);
       } else {
-        OffDbg->deleteValue();
+        OffDbg->eraseFromParent();
       }
       if (RsrcExpr) {
         Dbg->setExpression(*RsrcExpr);
diff --git a/llvm/test/CodeGen/AMDGPU/lower-buffer-fat-pointers-unoptimized-debug-data.ll b/llvm/test/CodeGen/AMDGPU/lower-buffer-fat-pointers-unoptimized-debug-data.ll
index 047bdde40ee57..82813200c75ec 100644
--- a/llvm/test/CodeGen/AMDGPU/lower-buffer-fat-pointers-unoptimized-debug-data.ll
+++ b/llvm/test/CodeGen/AMDGPU/lower-buffer-fat-pointers-unoptimized-debug-data.ll
@@ -11,11 +11,13 @@ define float @debug_stash_pointer(ptr addrspace(8) %buf, i32 %idx, ptr addrspace
 ; CHECK-NEXT:      #dbg_value(ptr addrspace(5) [[BUF_PTR_VAR]], [[META10:![0-9]+]], !DIExpression(), [[DBG21]])
 ; CHECK-NEXT:    [[AUX_PTR_VAR:%.*]] = alloca i160, align 32, addrspace(5), !dbg [[DBG22:![0-9]+]]
 ; CHECK-NEXT:      #dbg_value(ptr addrspace(5) [[AUX_PTR_VAR]], [[META12:![0-9]+]], !DIExpression(), [[DBG22]])
-; CHECK-NEXT:      #dbg_value({ ptr addrspace(8), i32 } poison, [[META13:![0-9]+]], !DIExpression(), [[META23:![0-9]+]])
+; CHECK-NEXT:      #dbg_value(i32 0, [[META13:![0-9]+]], !DIExpression(DW_OP_LLVM_fragment, 128, 32), [[META23:![0-9]+]])
+; CHECK-NEXT:      #dbg_value(ptr addrspace(8) [[BUF]], [[META13]], !DIExpression(DW_OP_LLVM_fragment, 0, 128), [[META23]])
 ; CHECK-NEXT:    [[BUF_PTR_INT_RSRC:%.*]] = ptrtoint ptr addrspace(8) [[BUF]] to i160, !dbg [[DBG24:![0-9]+]]
 ; CHECK-NEXT:    [[BUF_PTR_INT:%.*]] = shl nuw i160 [[BUF_PTR_INT_RSRC]], 32, !dbg [[DBG24]]
 ; CHECK-NEXT:    store i160 [[BUF_PTR_INT]], ptr addrspace(5) [[BUF_PTR_VAR]], align 32, !dbg [[DBG24]]
-; CHECK-NEXT:      #dbg_value({ ptr addrspace(8), i32 } poison, [[META15:![0-9]+]], !DIExpression(), [[META25:![0-9]+]])
+; CHECK-NEXT:      #dbg_value(i32 0, [[META15:![0-9]+]], !DIExpression(DW_OP_LLVM_fragment, 128, 32), [[META25:![0-9]+]])
+; CHECK-NEXT:      #dbg_value(ptr addrspace(8) [[AUX]], [[META15]], !DIExpression(DW_OP_LLVM_fragment, 0, 128), [[META25]])
 ; CHECK-NEXT:    [[AUX_PTR_INT_RSRC:%.*]] = ptrtoint ptr addrspace(8) [[AUX]] to i160, !dbg [[DBG26:![0-9]+]]
 ; CHECK-NEXT:    [[AUX_PTR_INT:%.*]] = shl nuw i160 [[AUX_PTR_INT_RSRC]], 32, !dbg [[DBG26]]
 ; CHECK-NEXT:    store i160 [[AUX_PTR_INT]], ptr addrspace(5) [[AUX_PTR_VAR]], align 32, !dbg [[DBG26]]
@@ -24,10 +26,12 @@ define float @debug_stash_pointer(ptr addrspace(8) %buf, i32 %idx, ptr addrspace
 ; CHECK-NEXT:    [[TMP2:%.*]] = trunc i160 [[TMP1]] to i128, !dbg [[DBG27]]
 ; CHECK-NEXT:    [[BUF_PTR_2_PTR_RSRC:%.*]] = inttoptr i128 [[TMP2]] to ptr addrspace(8), !dbg [[DBG27]]
 ; CHECK-NEXT:    [[BUF_PTR_2_PTR_OFF:%.*]] = trunc i160 [[BUF_PTR_2]] to i32, !dbg [[DBG27]]
-; CHECK-NEXT:      #dbg_value({ ptr addrspace(8), i32 } poison, [[META16:![0-9]+]], !DIExpression(), [[DBG27]])
+; CHECK-NEXT:      #dbg_value(i32 [[BUF_PTR_2_PTR_OFF]], [[META16:![0-9]+]], !DIExpression(DW_OP_LLVM_fragment, 128, 32), [[DBG27]])
+; CHECK-NEXT:      #dbg_value(ptr addrspace(8) [[BUF_PTR_2_PTR_RSRC]], [[META16]], !DIExpression(DW_OP_LLVM_fragment, 0, 128), [[DBG27]])
 ; CHECK-NEXT:    [[BUF_PTR_3_IDX:%.*]] = mul i32 [[IDX]], 4, !dbg [[DBG28:![0-9]+]]
 ; CHECK-NEXT:    [[BUF_PTR_3:%.*]] = add i32 [[BUF_PTR_2_PTR_OFF]], [[BUF_PTR_3_IDX]], !dbg [[DBG28]]
-; CHECK-NEXT:      #dbg_value({ ptr addrspace(8), i32 } poison, [[META17:![0-9]+]], !DIExpression(), [[DBG28]])
+; CHECK-NEXT:      #dbg_value(i32 [[BUF_PTR_3]], [[META17:![0-9]+]], !DIExpression(DW_OP_LLVM_fragment, 128, 32), [[DBG28]])
+; CHECK-NEXT:      #dbg_value(ptr addrspace(8) [[BUF_PTR_2_PTR_RSRC]], [[META17]], !DIExpression(DW_OP_LLVM_fragment, 0, 128), [[DBG28]])
 ; CHECK-NEXT:    [[BUF_PTR_3_INT_RSRC:%.*]] = ptrtoint ptr addrspace(8) [[BUF_PTR_2_PTR_RSRC]] to i160, !dbg [[DBG29:![0-9]+]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = shl nuw i160 [[BUF_PTR_3_INT_RSRC]], 32, !dbg [[DBG29]]
 ; CHECK-NEXT:    [[BUF_PTR_3_INT_OFF:%.*]] = zext i32 [[BUF_PTR_3]] to i160, !dbg [[DBG29]]
@@ -38,7 +42,8 @@ define float @debug_stash_pointer(ptr addrspace(8) %buf, i32 %idx, ptr addrspace
 ; CHECK-NEXT:    [[TMP5:%.*]] = trunc i160 [[TMP4]] to i128, !dbg [[DBG30]]
 ; CHECK-NEXT:    [[BUF_PTR_4_PTR_RSRC:%.*]] = inttoptr i128 [[TMP5]] to ptr addrspace(8), !dbg [[DBG30]]
 ; CHECK-NEXT:    [[BUF_PTR_4_PTR_OFF:%.*]] = trunc i160 [[BUF_PTR_4]] to i32, !dbg [[DBG30]]
-; CHECK-NEXT:      #dbg_value({ ptr addrspace(8), i32 } poison, [[META18:![0-9]+]], !DIExpression(), [[DBG30]])
+; CHECK-NEXT:      #dbg_value(i32 [[BUF_PTR_4_PTR_OFF]], [[META18:![0-9]+]], !DIExpression(DW_OP_LLVM_fragment, 128, 32), [[DBG30]])
+; CHECK-NEXT:      #dbg_value(ptr addrspace(8) [[BUF_PTR_4_PTR_RSRC]], [[META18]], !DIExpression(DW_OP_LLVM_fragment, 0, 128), [[DBG30]])
 ; CHECK-NEXT:    [[RET:%.*]] = call float @llvm.amdgcn.raw.ptr.buffer.load.f32(ptr addrspace(8) align 4 [[BUF_PTR_4_PTR_RSRC]], i32 [[BUF_PTR_4_PTR_OFF]], i32 0, i32 0), !dbg [[DBG31:![0-9]+]]
 ; CHECK-NEXT:      #dbg_value(float [[RET]], [[META19:![0-9]+]], !DIExpression(), [[DBG31]])
 ; CHECK-NEXT:    [[AUX_PTR_2:%.*]] = load i160, ptr addrspace(5) [[AUX_PTR_VAR]], align 32, !dbg [[DBG32:![0-9]+]]
@@ -46,7 +51,8 @@ define float @debug_stash_pointer(ptr addrspace(8) %buf, i32 %idx, ptr addrspace
 ; CHECK-NEXT:    [[TMP7:%.*]] = trunc i160 [[TMP6]] to i128, !dbg [[DBG32]]
 ; CHECK-NEXT:    [[AUX_PTR_2_PTR_RSRC:%.*]] = inttoptr i128 [[TMP7]] to ptr addrspace(8), !dbg [[DBG32]]
 ; CHECK-NEXT:    [[AUX_PTR_2_PTR_OFF:%.*]] = trunc i160 [[AUX_PTR_2]] to i32, !dbg [[DBG32]]
-; CHECK-NEXT:      #dbg_value({ ptr addrspace(8), i32 } poison, [[META20:![0-9]+]], !DIExpression(), [[DBG32]])
+; CHECK-NEXT:      #dbg_value(i32 [[AUX_PTR_2_PTR_OFF]], [[META20:![0-9]+]], !DIExpression(DW_OP_LLVM_fragment, 128, 32), [[DBG32]])
+; CHECK-NEXT:      #dbg_value(ptr addrspace(8) [[AUX_PTR_2_PTR_RSRC]], [[META20]], !DIExpression(DW_OP_LLVM_fragment, 0, 128), [[DBG32]])
 ; CHECK-NEXT:    [[BUF_PTR_4_LEGAL:%.*]] = bitcast i160 [[BUF_PTR_4]] to <5 x i32>, !dbg [[DBG33:![0-9]+]]
 ; CHECK-NEXT:    [[BUF_PTR_4_SLICE_0:%.*]] = shufflevector <5 x i32> [[BUF_PTR_4_LEGAL]], <5 x i32> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>, !dbg [[DBG33]]
 ; CHECK-NEXT:    call void @llvm.amdgcn.raw.ptr.buffer.store.v4i32(<4 x i32> [[BUF_PTR_4_SLICE_0]], ptr addrspace(8) align 32 [[AUX_PTR_2_PTR_RSRC]], i32 [[AUX_PTR_2_PTR_OFF]], i32 0, i32 0), !dbg [[DBG33]]

[krzysz00]

Lgtm, thanks for keeping up with this