[llvm] [AMDGPU] Fix ISel crash on llvm.amdgcn.class with bf16 operand (PR #210971)

Arseniy Obolenskiy via llvm-commits llvm-commits at lists.llvm.org
Tue Jul 21 07:52:40 PDT 2026


https://github.com/aobolensk updated https://github.com/llvm/llvm-project/pull/210971

>From 9822f768bfbf8f1b001cf34b45a97cc1a8f5ef6a Mon Sep 17 00:00:00 2001
From: Arseniy Obolenskiy <arseniy.obolenskiy at amd.com>
Date: Tue, 21 Jul 2026 14:40:34 +0200
Subject: [PATCH 1/3] [AMDGPU] Fix ISel crash on llvm.amdgcn.class with bf16
 operand

FP_CLASS has no bf16 pattern, so widen the source to f32 first (that is a free bitcast, since bf16 shares f32 exponent field)
---
 llvm/lib/Target/AMDGPU/SIISelLowering.cpp     | 11 +++--
 .../CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll  | 48 +++++++++++++++++++
 2 files changed, 56 insertions(+), 3 deletions(-)
 create mode 100644 llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index f92b1731c3f7d..4c22ac053e365 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -11185,9 +11185,14 @@ SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
   case Intrinsic::amdgcn_fract:
     return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
 
-  case Intrinsic::amdgcn_class:
-    return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT, Op.getOperand(1),
-                       Op.getOperand(2));
+  case Intrinsic::amdgcn_class: {
+    SDValue Src = Op.getOperand(1);
+    if (Src.getValueType() == MVT::bf16) {
+      SDValue Ext = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Src);
+      return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT, Ext, Op.getOperand(2));
+    }
+    return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT, Src, Op.getOperand(2));
+  }
   case Intrinsic::amdgcn_div_fmas:
     return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT, Op.getOperand(1),
                        Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll
new file mode 100644
index 0000000000000..4e67a92448a9c
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll
@@ -0,0 +1,48 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -global-isel=0 -amdgpu-scalarize-global-loads=false -mtriple=amdgpu8.03-amd-amdhsa -mattr=-flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI-SDAG %s
+; RUN: llc -global-isel=1 -global-isel-abort=2 -amdgpu-scalarize-global-loads=false -mtriple=amdgpu8.03-amd-amdhsa -mattr=-flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI-GISEL %s
+
+declare bfloat @llvm.fabs.bf16(bfloat %a)
+declare i1 @llvm.amdgcn.class.bf16(bfloat %a, i32 %b)
+
+define amdgpu_kernel void @class_bf16(
+; GCN-LABEL: class_bf16:
+; GCN:       ; %bb.0: ; %entry
+; GCN-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
+; GCN-NEXT:    s_load_dwordx2 s[8:9], s[8:9], 0x10
+; GCN-NEXT:    s_mov_b32 s7, 0x1100f000
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    s_mov_b32 s14, s6
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_mov_b32 s12, s2
+; GCN-NEXT:    s_mov_b32 s13, s3
+; GCN-NEXT:    s_mov_b32 s15, s7
+; GCN-NEXT:    s_mov_b32 s10, s6
+; GCN-NEXT:    s_mov_b32 s11, s7
+; GCN-NEXT:    buffer_load_ushort v0, off, s[12:15], 0
+; GCN-NEXT:    buffer_load_dword v1, off, s[8:11], 0
+; GCN-NEXT:    s_mov_b32 s4, s0
+; GCN-NEXT:    s_mov_b32 s5, s1
+; GCN-NEXT:    s_waitcnt vmcnt(1)
+; GCN-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GCN-NEXT:    s_waitcnt vmcnt(0)
+; GCN-NEXT:    v_cmp_class_f32_e32 vcc, v0, v1
+; GCN-NEXT:    s_and_b64 s[0:1], vcc, exec
+; GCN-NEXT:    s_cselect_b32 s0, -1, 0
+; GCN-NEXT:    v_mov_b32_e32 v0, s0
+; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GCN-NEXT:    s_endpgm
+    ptr addrspace(1) %r,
+    ptr addrspace(1) %a,
+    ptr addrspace(1) %b) {
+entry:
+  %a.val = load bfloat, ptr addrspace(1) %a
+  %b.val = load i32, ptr addrspace(1) %b
+  %r.val = call i1 @llvm.amdgcn.class.bf16(bfloat %a.val, i32 %b.val)
+  %r.val.sext = sext i1 %r.val to i32
+  store i32 %r.val.sext, ptr addrspace(1) %r
+  ret void
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; VI-GISEL: {{.*}}
+; VI-SDAG: {{.*}}

>From 85719455149466d0ea7c86978d37ff47610cc779 Mon Sep 17 00:00:00 2001
From: Arseniy Obolenskiy <arseniy.obolenskiy at amd.com>
Date: Tue, 21 Jul 2026 15:02:18 +0200
Subject: [PATCH 2/3] Address comments

---
 llvm/lib/Target/AMDGPU/SIISelLowering.cpp     |  6 ++--
 .../CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll  | 36 ++-----------------
 2 files changed, 7 insertions(+), 35 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 4c22ac053e365..e78256e94d122 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -11188,8 +11188,10 @@ SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
   case Intrinsic::amdgcn_class: {
     SDValue Src = Op.getOperand(1);
     if (Src.getValueType() == MVT::bf16) {
-      SDValue Ext = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Src);
-      return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT, Ext, Op.getOperand(2));
+      DAG.getContext()->diagnose(DiagnosticInfoUnsupported(
+          DAG.getMachineFunction().getFunction(),
+          "llvm.amdgcn.class does not support bf16", DL.getDebugLoc()));
+      return DAG.getPOISON(VT);
     }
     return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT, Src, Op.getOperand(2));
   }
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll
index 4e67a92448a9c..f6145905eac17 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll
@@ -1,37 +1,10 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
-; RUN: llc -global-isel=0 -amdgpu-scalarize-global-loads=false -mtriple=amdgpu8.03-amd-amdhsa -mattr=-flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI-SDAG %s
-; RUN: llc -global-isel=1 -global-isel-abort=2 -amdgpu-scalarize-global-loads=false -mtriple=amdgpu8.03-amd-amdhsa -mattr=-flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI-GISEL %s
+; RUN: not llc -mtriple=amdgpu9.50 < %s 2>&1 | FileCheck -check-prefix=ERROR %s
+
+; ERROR: error: <unknown>:0:0: in function class_bf16 void (ptr addrspace(1), ptr addrspace(1), ptr addrspace(1)): llvm.amdgcn.class does not support bf16
 
-declare bfloat @llvm.fabs.bf16(bfloat %a)
 declare i1 @llvm.amdgcn.class.bf16(bfloat %a, i32 %b)
 
 define amdgpu_kernel void @class_bf16(
-; GCN-LABEL: class_bf16:
-; GCN:       ; %bb.0: ; %entry
-; GCN-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
-; GCN-NEXT:    s_load_dwordx2 s[8:9], s[8:9], 0x10
-; GCN-NEXT:    s_mov_b32 s7, 0x1100f000
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s14, s6
-; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s12, s2
-; GCN-NEXT:    s_mov_b32 s13, s3
-; GCN-NEXT:    s_mov_b32 s15, s7
-; GCN-NEXT:    s_mov_b32 s10, s6
-; GCN-NEXT:    s_mov_b32 s11, s7
-; GCN-NEXT:    buffer_load_ushort v0, off, s[12:15], 0
-; GCN-NEXT:    buffer_load_dword v1, off, s[8:11], 0
-; GCN-NEXT:    s_mov_b32 s4, s0
-; GCN-NEXT:    s_mov_b32 s5, s1
-; GCN-NEXT:    s_waitcnt vmcnt(1)
-; GCN-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_cmp_class_f32_e32 vcc, v0, v1
-; GCN-NEXT:    s_and_b64 s[0:1], vcc, exec
-; GCN-NEXT:    s_cselect_b32 s0, -1, 0
-; GCN-NEXT:    v_mov_b32_e32 v0, s0
-; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
-; GCN-NEXT:    s_endpgm
     ptr addrspace(1) %r,
     ptr addrspace(1) %a,
     ptr addrspace(1) %b) {
@@ -43,6 +16,3 @@ entry:
   store i32 %r.val.sext, ptr addrspace(1) %r
   ret void
 }
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; VI-GISEL: {{.*}}
-; VI-SDAG: {{.*}}

>From 788fff03b9629aa20e8ca1264f1e8e0f7b310faa Mon Sep 17 00:00:00 2001
From: Arseniy Obolenskiy <arseniy.obolenskiy at amd.com>
Date: Tue, 21 Jul 2026 16:52:26 +0200
Subject: [PATCH 3/3] Address comments

---
 llvm/lib/Target/AMDGPU/SIISelLowering.cpp          | 6 ++++--
 llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll | 3 ++-
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index e78256e94d122..f5ad2567dbb3e 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -11187,10 +11187,12 @@ SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
 
   case Intrinsic::amdgcn_class: {
     SDValue Src = Op.getOperand(1);
-    if (Src.getValueType() == MVT::bf16) {
+    EVT SrcVT = Src.getValueType();
+    if (SrcVT != MVT::f16 && SrcVT != MVT::f32 && SrcVT != MVT::f64) {
       DAG.getContext()->diagnose(DiagnosticInfoUnsupported(
           DAG.getMachineFunction().getFunction(),
-          "llvm.amdgcn.class does not support bf16", DL.getDebugLoc()));
+          "llvm.amdgcn.class only supports f16, f32, and f64",
+          DL.getDebugLoc()));
       return DAG.getPOISON(VT);
     }
     return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT, Src, Op.getOperand(2));
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll
index f6145905eac17..46531efe31725 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.class.bf16.ll
@@ -1,6 +1,7 @@
 ; RUN: not llc -mtriple=amdgpu9.50 < %s 2>&1 | FileCheck -check-prefix=ERROR %s
+; RUN: not llc -mtriple=amdgpu6.00 < %s 2>&1 | FileCheck -check-prefix=ERROR %s
 
-; ERROR: error: <unknown>:0:0: in function class_bf16 void (ptr addrspace(1), ptr addrspace(1), ptr addrspace(1)): llvm.amdgcn.class does not support bf16
+; ERROR: error: <unknown>:0:0: in function class_bf16 void (ptr addrspace(1), ptr addrspace(1), ptr addrspace(1)): llvm.amdgcn.class only supports f16, f32, and f64
 
 declare i1 @llvm.amdgcn.class.bf16(bfloat %a, i32 %b)
 



More information about the llvm-commits mailing list