[llvm] AMDGPU: Avoid report_fatal_error on ds ordered intrinsics (PR #145202)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

<details>
<summary>Changes</summary>



---
Full diff: https://github.com/llvm/llvm-project/pull/145202.diff


3 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp (+15-6) 
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+17-6) 
- (added) llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add-errors.ll (+30) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index 672520390c8bf..b632b16f5c198 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -1767,8 +1767,12 @@ bool AMDGPUInstructionSelector::selectDSOrderedIntrinsic(
   bool WaveRelease = MI.getOperand(8).getImm() != 0;
   bool WaveDone = MI.getOperand(9).getImm() != 0;
 
-  if (WaveDone && !WaveRelease)
-    report_fatal_error("ds_ordered_count: wave_done requires wave_release");
+  if (WaveDone && !WaveRelease) {
+    // TODO: Move this to IR verifier
+    const Function &Fn = MF->getFunction();
+    Fn.getContext().diagnose(DiagnosticInfoUnsupported(
+        Fn, "ds_ordered_count: wave_done requires wave_release", DL));
+  }
 
   unsigned OrderedCountIndex = IndexOperand & 0x3f;
   IndexOperand &= ~0x3f;
@@ -1779,13 +1783,18 @@ bool AMDGPUInstructionSelector::selectDSOrderedIntrinsic(
     IndexOperand &= ~(0xf << 24);
 
     if (CountDw < 1 || CountDw > 4) {
-      report_fatal_error(
-        "ds_ordered_count: dword count must be between 1 and 4");
+      const Function &Fn = MF->getFunction();
+      Fn.getContext().diagnose(DiagnosticInfoUnsupported(
+          Fn, "ds_ordered_count: dword count must be between 1 and 4", DL));
+      CountDw = 1;
     }
   }
 
-  if (IndexOperand)
-    report_fatal_error("ds_ordered_count: bad index operand");
+  if (IndexOperand) {
+    const Function &Fn = MF->getFunction();
+    Fn.getContext().diagnose(DiagnosticInfoUnsupported(
+        Fn, "ds_ordered_count: bad index operand", DL));
+  }
 
   unsigned Instruction = IntrID == Intrinsic::amdgcn_ds_ordered_add ? 0 : 1;
   unsigned ShaderType = SIInstrInfo::getDSShaderTypeValue(*MF);
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index b9023b6d7a3a6..672896146c781 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -9330,16 +9330,27 @@ SDValue SITargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
       IndexOperand &= ~(0xf << 24);
 
       if (CountDw < 1 || CountDw > 4) {
-        report_fatal_error(
-            "ds_ordered_count: dword count must be between 1 and 4");
+        const Function &Fn = DAG.getMachineFunction().getFunction();
+        DAG.getContext()->diagnose(DiagnosticInfoUnsupported(
+            Fn, "ds_ordered_count: dword count must be between 1 and 4",
+            DL.getDebugLoc()));
+        CountDw = 1;
       }
     }
 
-    if (IndexOperand)
-      report_fatal_error("ds_ordered_count: bad index operand");
+    if (IndexOperand) {
+      const Function &Fn = DAG.getMachineFunction().getFunction();
+      DAG.getContext()->diagnose(DiagnosticInfoUnsupported(
+          Fn, "ds_ordered_count: bad index operand", DL.getDebugLoc()));
+    }
 
-    if (WaveDone && !WaveRelease)
-      report_fatal_error("ds_ordered_count: wave_done requires wave_release");
+    if (WaveDone && !WaveRelease) {
+      // TODO: Move this to IR verifier
+      const Function &Fn = DAG.getMachineFunction().getFunction();
+      DAG.getContext()->diagnose(DiagnosticInfoUnsupported(
+          Fn, "ds_ordered_count: wave_done requires wave_release",
+          DL.getDebugLoc()));
+    }
 
     unsigned Instruction = IntrID == Intrinsic::amdgcn_ds_ordered_add ? 0 : 1;
     unsigned ShaderType =
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add-errors.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add-errors.ll
new file mode 100644
index 0000000000000..36b13e3b16cc6
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add-errors.ll
@@ -0,0 +1,30 @@
+; RUN: not llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1010 -filetype=null %s 2>&1 | FileCheck %s
+; RUN: not llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1010 -filetype=null %s 2>&1 | FileCheck %s
+
+; CHECK: error: <unknown>:0:0: in function ds_ordered_add_dword_count_too_low void (ptr addrspace(2), ptr addrspace(1)): ds_ordered_count: dword count must be between 1 and 4
+define amdgpu_kernel void @ds_ordered_add_dword_count_too_low(ptr addrspace(2) inreg %gds, ptr addrspace(1) %out) {
+  %val = call i32 @llvm.amdgcn.ds.ordered.add(ptr addrspace(2) %gds, i32 31, i32 0, i32 0, i1 false, i32 0, i1 true, i1 true)
+  store i32 %val, ptr addrspace(1) %out
+  ret void
+}
+
+; CHECK: error: <unknown>:0:0: in function ds_ordered_add_dword_count_too_high void (ptr addrspace(2), ptr addrspace(1)): ds_ordered_count: dword count must be between 1 and 4
+define amdgpu_kernel void @ds_ordered_add_dword_count_too_high(ptr addrspace(2) inreg %gds, ptr addrspace(1) %out) {
+  %val = call i32 @llvm.amdgcn.ds.ordered.add(ptr addrspace(2) %gds, i32 31, i32 0, i32 0, i1 false, i32 5, i1 true, i1 true)
+  store i32 %val, ptr addrspace(1) %out
+  ret void
+}
+
+; CHECK: error: <unknown>:0:0: in function ds_ordered_add_bad_index_operand void (ptr addrspace(2), ptr addrspace(1)): ds_ordered_count: bad index operand
+define amdgpu_kernel void @ds_ordered_add_bad_index_operand(ptr addrspace(2) inreg %gds, ptr addrspace(1) %out) {
+  %val = call i32 @llvm.amdgcn.ds.ordered.add(ptr addrspace(2) %gds, i32 31, i32 0, i32 1, i1 false, i32 -1, i1 true, i1 true)
+  store i32 %val, ptr addrspace(1) %out
+  ret void
+}
+
+; CHECK: error: <unknown>:0:0: in function ds_ordered_add_dword_count_wave_done_without_wave_release void (ptr addrspace(2), ptr addrspace(1)): ds_ordered_count: wave_done requires wave_release
+define amdgpu_kernel void @ds_ordered_add_dword_count_wave_done_without_wave_release(ptr addrspace(2) inreg %gds, ptr addrspace(1) %out) {
+  %val = call i32 @llvm.amdgcn.ds.ordered.add(ptr addrspace(2) %gds, i32 31, i32 0, i32 0, i1 false, i32 1, i1 false, i1 true)
+  store i32 %val, ptr addrspace(1) %out
+  ret void
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/145202