RenameIndependentSubregs: Fix handling of undef tied operands

[arsenm]

Summary

If a tied source operand was undef, it would be replaced but not
update the other tied operand, which would end up using different
virtual registers.

[MatzeB]

Good catch, LGTM.

lib/CodeGen/RenameIndependentSubregs.cpp

246–247

Could you add a comment along the lines of `
/// Undef use operands are not tracked in the equivalence class but need to be update if they are tied.

test/CodeGen/AMDGPU/rename-independent-subregs-invalid-mac-operands.mir

2–23

Do you actually need the IR, see also: http://llvm.org/docs/MIRLangRef.html#simplifying-mir-files

[arsenm]

test/CodeGen/AMDGPU/rename-independent-subregs-invalid-mac-operands.mir

2–23

It complains if I try to remove it entirely. It seems to want the blocks at least

[MatzeB]

test/CodeGen/AMDGPU/rename-independent-subregs-invalid-mac-operands.mir

2–23

You have to use the bb.0 instead of bb.0.bb etc. (the .bb at the end is a reference to the IR block).

[arsenm]

r304747

lib/CodeGen/RenameIndependentSubregs.cpp

for (const SubRangeInfo &SRInfo : SubRangeInfos) {
       unsigned LocalID = SRInfo.ConEQ.getEqClass(VNI);
       // Global ID
       ID = Classes[LocalID + SRInfo.Index];
       break;
     }
 
     unsigned VReg = Intervals[ID]->reg;
     MO.setReg(VReg);
+    if (MO.isTied())
+      MO.getParent()->substituteRegister(Reg, VReg, 0, TRI);
   }
   // TODO: We could attempt to recompute new register classes while visiting
   // the operands: Some of the split register may be fine with less constraint
   // classes than the original vreg.
 }
 
 void RenameIndependentSubregs::distribute(const IntEqClasses &Classes,
     const SmallVectorImpl<SubRangeInfo> &SubRangeInfos,

test/CodeGen/AMDGPU/rename-independent-subregs-invalid-mac-operands.mir

+# RUN: llc -march=amdgcn -verify-machineinstrs -run-pass=simple-register-coalescing,rename-independent-subregs -o - %s | FileCheck -check-prefix=GCN %s
+--- |
+  define amdgpu_kernel void @mac_invalid_operands(float %arg) #0 {
+  bb:
+    %tmp = fcmp oeq float %arg, 0.000000e+00
+    br i1 %tmp, label %bb1, label %bb6
+
+  bb1:                                              ; preds = %bb
+    %tmp2 = insertelement <3 x float> undef, float %arg, i32 0
+    %tmp3 = shufflevector <3 x float> %tmp2, <3 x float> undef, <3 x i32> zeroinitializer
+    %tmp4 = fmul <3 x float> %tmp3, undef
+    %tmp5 = fadd <3 x float> %tmp4, undef
+    br label %bb6
+
+  bb6:                                              ; preds = %bb1, %bb
+    %tmp7 = phi <3 x float> [ <float fadd (float undef, float undef), float undef, float undef>, %bb ], [ %tmp5, %bb1 ]
+    store volatile <3 x float> %tmp7, <3 x float> addrspace(1)* undef, align 16
+    ret void
+  }
+
+  attributes #0 = { norecurse nounwind }
+
+...
+---
+
+# GCN-LABEL: name: mac_invalid_operands
+# GCN: undef %18.sub0 = V_MAC_F32_e32 undef %3, undef %9, undef %18.sub0, implicit %exec
+
+name:            mac_invalid_operands
+alignment:       0
+exposesReturnsTwice: false
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_128 }
+  - { id: 1, class: vreg_128 }
+  - { id: 2, class: sgpr_64 }
+  - { id: 3, class: vgpr_32 }
+  - { id: 4, class: vgpr_32 }
+  - { id: 5, class: vgpr_32 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: sreg_64 }
+  - { id: 8, class: vgpr_32 }
+  - { id: 9, class: vgpr_32 }
+  - { id: 10, class: vreg_64 }
+  - { id: 11, class: vreg_64 }
+  - { id: 12, class: vreg_128 }
+  - { id: 13, class: vreg_128 }
+  - { id: 14, class: vgpr_32 }
+  - { id: 15, class: vreg_64 }
+  - { id: 16, class: vgpr_32 }
+  - { id: 17, class: vreg_128 }
+body:             |
+  bb.0.bb:
+    successors: %bb.2.bb1(0x50000000), %bb.1(0x30000000)
+
+    %7 = V_CMP_NEQ_F32_e64 0, 0, 0, undef %3, 0, 0, implicit %exec
+    %vcc = COPY killed %7
+    S_CBRANCH_VCCZ %bb.2.bb1, implicit killed %vcc
+
+  bb.1:
+    successors: %bb.3.bb6(0x80000000)
+
+    %4 = V_ADD_F32_e32 undef %6, undef %5, implicit %exec
+    undef %12.sub0 = COPY killed %4
+    %17 = COPY killed %12
+    S_BRANCH %bb.3.bb6
+
+  bb.2.bb1:
+    successors: %bb.3.bb6(0x80000000)
+
+    %8 = V_MAC_F32_e32 undef %3, undef %9, undef %8, implicit %exec
+    undef %13.sub0 = COPY %8
+    %13.sub1 = COPY %8
+    %13.sub2 = COPY killed %8
+    %0 = COPY killed %13
+    %17 = COPY killed %0
+
+  bb.3.bb6:
+    %1 = COPY killed %17
+    FLAT_STORE_DWORD undef %10, %1.sub2, 0, 0, implicit %exec, implicit %flat_scr
+    %14 = COPY %1.sub1
+    %16 = COPY killed %1.sub0
+    undef %15.sub0 = COPY killed %16
+    %15.sub1 = COPY killed %14
+    FLAT_STORE_DWORDX2 undef %11, killed %15, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+
+...