[PATCH] D131951: Use ISA versions instead of attributes to determine intrinsic legality.

[Leon Clark via Phabricator via llvm-commits]

Leonc created this revision.
Herald added subscribers: kosarev, foad, kerbowa, hiraditya, nhaehnle, jvesely, arsenm.
Herald added a project: All.
Leonc requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Subtarget attributes are not guaranteed to match those of the actual hardware. This can cause the compiler to crash in some circumstances.
Modify intrinsic lowering to be conditional on ISA version numbers instead of subtarget attributes.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D131951

Files:
  llvm/lib/Target/AMDGPU/SIISelLowering.cpp
  llvm/test/CodeGen/AMDGPU/llvm.amdgcn.atomic.fadd.ll


Index: llvm/test/CodeGen/AMDGPU/llvm.amdgcn.atomic.fadd.ll
===================================================================
--- llvm/test/CodeGen/AMDGPU/llvm.amdgcn.atomic.fadd.ll
+++ llvm/test/CodeGen/AMDGPU/llvm.amdgcn.atomic.fadd.ll
@@ -93,7 +93,8 @@
 ; Make sure this artificially selects with an incorrect subtarget, but
 ; the feature set.
 ; GCN-LABEL: {{^}}global_atomic_fadd_f32_wrong_subtarget:
-; GCN: global_atomic_add_f32 v{{[0-9]+}}, v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]$}}
+; GCN-NOT: global_atomic_add_f32
+; GCN: v_illegal
 define amdgpu_kernel void @global_atomic_fadd_f32_wrong_subtarget(float addrspace(1)* %ptr, float %data) #0 {
   %ret = call float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* %ptr, float %data)
   ret void
Index: llvm/lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -35,6 +35,7 @@
 #include "llvm/IR/IntrinsicsR600.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/KnownBits.h"
+#include "llvm/Support/TargetParser.h"
 
 using namespace llvm;
 
@@ -7821,10 +7822,17 @@
     };
     unsigned Opcode = 0;
     switch (IntrID) {
-    case Intrinsic::amdgcn_global_atomic_fadd:
-      if (!Subtarget->hasAtomicFaddNoRtnInsts())
+    case Intrinsic::amdgcn_global_atomic_fadd: {
+      auto IV = AMDGPU::getIsaVersion(Subtarget->getCPU());
+      auto IVTupl = std::make_tuple(IV.Major, IV.Minor, IV.Stepping);
+      using TuplTy = decltype(IVTupl);
+      // Emit `v_illegal` if ISA version is between gfx908 and gfx940
+      // (inclusive), or if ISA version is gfx11 or above.
+      if ((IVTupl < TuplTy(9u, 0u, 8u)) ||
+          (IVTupl > TuplTy(9u, 4u, 0u) && IVTupl < TuplTy(11u, 0u, 0u)))
         return makeV_ILLEGAL(Op, DAG);
       [[fallthrough]];
+    }
     case Intrinsic::amdgcn_flat_atomic_fadd: {
       EVT VT = Op.getOperand(3).getValueType();
       return DAG.getAtomic(ISD::ATOMIC_LOAD_FADD, DL, VT,


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D131951.452928.patch
Type: text/x-patch
Size: 2053 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220816/5e097858/attachment.bin>