[llvm] a21544a - AMDGPU: Fix handling of target flags when commuting instruction

[Matt Arsenault via llvm-commits]

Author: Matt Arsenault
Date: 2020-07-01T13:59:13-04:00
New Revision: a21544ad11fe443f4c38161630d3f2fc90d098bf

URL: https://github.com/llvm/llvm-project/commit/a21544ad11fe443f4c38161630d3f2fc90d098bf
DIFF: https://github.com/llvm/llvm-project/commit/a21544ad11fe443f4c38161630d3f2fc90d098bf.diff

LOG: AMDGPU: Fix handling of target flags when commuting instruction

If the original register operand had a subregister, it wasn't getting
cleared. This resulted in reinterpreted the subreg index as
unrecognized target flags, which produced unparseable MIR.

Added: 
    llvm/test/CodeGen/AMDGPU/machine-cse-commute-target-flags.mir

Modified: 
    llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 78d96426b432..efdfca35da29 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -1860,6 +1860,9 @@ static MachineInstr *swapRegAndNonRegOperand(MachineInstr &MI,
   else
     return nullptr;
 
+  // Make sure we don't reinterpret a subreg index in the target flags.
+  RegOp.setTargetFlags(NonRegOp.getTargetFlags());
+
   NonRegOp.ChangeToRegister(Reg, false, false, IsKill, IsDead, IsUndef, IsDebug);
   NonRegOp.setSubReg(SubReg);
 

diff  --git a/llvm/test/CodeGen/AMDGPU/machine-cse-commute-target-flags.mir b/llvm/test/CodeGen/AMDGPU/machine-cse-commute-target-flags.mir
new file mode 100644
index 000000000000..22ba87b9572d
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/machine-cse-commute-target-flags.mir
@@ -0,0 +1,74 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -run-pass=machine-cse -o - %s | FileCheck %s
+
+--- |
+  define void @commute_instruction_subreg_target_flag() { ret void }
+  define void @commute_target_flag_frame_index() { ret void }
+  define void @commute_target_flag_global() { ret void }
+  declare void @func()
+
+...
+
+# Make sure the subreg index is cleared when commuting a register and immediate.
+
+---
+name: commute_instruction_subreg_target_flag
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CHECK-LABEL: name: commute_instruction_subreg_target_flag
+    ; CHECK: liveins: $vgpr0_vgpr1
+    ; CHECK: [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+    ; CHECK: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[COPY]].sub1, 64, 0, implicit $exec
+    ; CHECK: S_ENDPGM 0, implicit [[V_ADD_U32_e64_]], implicit [[V_ADD_U32_e64_]]
+    %0:vreg_64 = COPY $vgpr0_vgpr1
+    %1:vgpr_32 = V_ADD_U32_e64 %0.sub1, 64, 0, implicit $exec
+    %2:vgpr_32 = V_ADD_U32_e64 64, %0.sub1, 0, implicit $exec
+    S_ENDPGM 0, implicit %1, implicit %2
+
+...
+
+# FIXME: Why doesn't this CSE?
+---
+name: commute_target_flag_frame_index
+tracksRegLiveness: true
+stack:
+  - { id: 0, type: default, offset: 0, size: 4, alignment: 4 }
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; CHECK-LABEL: name: commute_target_flag_frame_index
+    ; CHECK: liveins: $vgpr0_vgpr1
+    ; CHECK: [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+    ; CHECK: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 %stack.0, [[COPY]].sub0, 0, implicit $exec
+    ; CHECK: S_ENDPGM 0, implicit [[V_ADD_U32_e64_]], implicit [[V_ADD_U32_e64_]]
+    %0:vreg_64 = COPY $vgpr0_vgpr1
+    %1:vgpr_32 = V_ADD_U32_e64 %0.sub0, %stack.0, 0, implicit $exec
+    %2:vgpr_32 = V_ADD_U32_e64 %stack.0, %0.sub0, 0, implicit $exec
+    S_ENDPGM 0, implicit %1, implicit %2
+
+...
+
+# FIXME: Handle commuting global variables
+---
+name: commute_target_flag_global
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1
+
+    ; CHECK-LABEL: name: commute_target_flag_global
+    ; CHECK: liveins: $sgpr0_sgpr1
+    ; CHECK: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; CHECK: [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY]].sub0, target-flags(amdgpu-rel32-lo) @func, implicit-def dead $scc
+    ; CHECK: [[S_ADD_U32_1:%[0-9]+]]:sreg_32 = S_ADD_U32 target-flags(amdgpu-rel32-lo) @func, [[COPY]].sub0, implicit-def dead $scc
+    ; CHECK: S_ENDPGM 0, implicit [[S_ADD_U32_]], implicit [[S_ADD_U32_1]]
+    %0:sreg_64 = COPY $sgpr0_sgpr1
+    %1:sreg_32 = S_ADD_U32 %0.sub0, target-flags(amdgpu-rel32-lo) @func, implicit-def dead $scc
+    %2:sreg_32 = S_ADD_U32 target-flags(amdgpu-rel32-lo) @func, %0.sub0, implicit-def dead $scc
+    S_ENDPGM 0, implicit %1, implicit %2
+
+...