[PATCH] D66223: AMDGPU/GlobalISel: Fix assert on load from constant address

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added reviewers: tstellar, nhaehnle.
Herald added subscribers: Petar.Avramovic, t-tye, tpr, dstuttard, rovka, yaxunl, wdng, jvesely, kzhuravl.

https://reviews.llvm.org/D66223

Files:
  lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
  test/CodeGen/AMDGPU/GlobalISel/inst-select-load-smrd.mir


Index: test/CodeGen/AMDGPU/GlobalISel/inst-select-load-smrd.mir
===================================================================
--- test/CodeGen/AMDGPU/GlobalISel/inst-select-load-smrd.mir
+++ test/CodeGen/AMDGPU/GlobalISel/inst-select-load-smrd.mir
@@ -5,6 +5,7 @@
 --- |
   define amdgpu_kernel void @smrd_imm(i32 addrspace(4)* %const0) { ret void }
   define amdgpu_kernel void @smrd_wide() { ret void }
+  define amdgpu_kernel void @constant_address_positive() { ret void }
 ...
 ---
 
@@ -185,3 +186,29 @@
     %5:sgpr(<16 x s32>) = G_LOAD %1 :: (load 64, addrspace 1)
     $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = COPY %5
 ...
+
+
+# Test a load of an offset from a constant base address
+# GCN-LABEL: name: constant_address_positive{{$}}
+# GCN: %4:sreg_32_xm0 = S_MOV_B32 44
+# GCN: %5:sreg_32_xm0 = S_MOV_B32 0
+# GCN: %0:sreg_64 = REG_SEQUENCE %4, %subreg.sub0, %5, %subreg.sub1
+
+# VI: %3:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0, 64, 0, 0 :: (dereferenceable invariant load 4, addrspace 4)
+# SICI: %3:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %0, 16, 0, 0 :: (dereferenceable invariant load 4, addrspace 4)
+
+---
+
+name:            constant_address_positive
+legalized:       true
+regBankSelected: true
+
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1, $vgpr2_vgpr3
+    %0:sgpr(p4) = G_CONSTANT i64 44
+    %1:sgpr(s64) = G_CONSTANT i64 64
+    %2:sgpr(p4) = G_GEP %0, %1
+    %3:sgpr(s32) = G_LOAD %2 :: (dereferenceable invariant load 4, align 4, addrspace 4)
+    S_ENDPGM 0, implicit %3
+...
Index: lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -1146,13 +1146,13 @@
 
   GEPInfo GEPInfo(*PtrMI);
 
-  for (unsigned i = 1, e = 3; i < e; ++i) {
+  for (unsigned i = 1; i != 3; ++i) {
     const MachineOperand &GEPOp = PtrMI->getOperand(i);
     const MachineInstr *OpDef = MRI.getUniqueVRegDef(GEPOp.getReg());
     assert(OpDef);
-    if (isConstant(*OpDef)) {
-      // FIXME: Is it possible to have multiple Imm parts?  Maybe if we
-      // are lacking other optimizations.
+    if (i == 2 && isConstant(*OpDef)) {
+      // TODO: Could handle constant base + variable offset, but a combine
+      // probably should have commuted it.
       assert(GEPInfo.Imm == 0);
       GEPInfo.Imm = OpDef->getOperand(1).getCImm()->getSExtValue();
       continue;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D66223.215132.patch
Type: text/x-patch
Size: 2516 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20190814/9d6a1bef/attachment.bin>