AMDGPU/GlobalISel: Set insert point when emitting control flow pseudos

This was implicitly assuming the branch instruction was the next after
the pseudo. It's possible for another non-terminator instruction to be
inserted between the intrinsic and the branch, so adjust the insertion
point. Fixes a non-terminator after terminator verifier error (which
without the verifier, manifested itself as an infinite loop in
analyzeBranch much later on).

Author: arsenm

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

@@ -4144,6 +4144,7 @@ bool AMDGPULegalizerInfo::legalizeIntrinsic(MachineInstr &MI,
       Register Use = MI.getOperand(3).getReg();
 
       MachineBasicBlock *CondBrTarget = BrCond->getOperand(1).getMBB();
+      B.setInsertPt(B.getMBB(), BrCond->getIterator());
       if (IntrID == Intrinsic::amdgcn_if) {
         B.buildInstr(AMDGPU::SI_IF)
           .addDef(Def)
@@ -4184,6 +4185,8 @@ bool AMDGPULegalizerInfo::legalizeIntrinsic(MachineInstr &MI,
 
       MachineBasicBlock *CondBrTarget = BrCond->getOperand(1).getMBB();
       Register Reg = MI.getOperand(2).getReg();
+
+      B.setInsertPt(B.getMBB(), BrCond->getIterator());
       B.buildInstr(AMDGPU::SI_LOOP)
         .addUse(Reg)
         .addMBB(UncondBrTarget);

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-brcond.mir

@@ -1,6 +1,6 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -O0 -run-pass=legalizer %s -o - | FileCheck -check-prefix=WAVE64 %s
-# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 -mattr=+wavefrontsize32,-wavefrontsize64 -O0 -run-pass=legalizer %s -o - | FileCheck -check-prefix=WAVE32 %s
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -O0 -run-pass=legalizer -verify-machineinstrs %s -o - | FileCheck -check-prefix=WAVE64 %s
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 -mattr=+wavefrontsize32,-wavefrontsize64 -O0 -run-pass=legalizer -verify-machineinstrs %s -o - | FileCheck -check-prefix=WAVE32 %s
 
 ---
 name: legal_brcond_vcc
@@ -310,3 +310,99 @@ body:             |
 
   bb.2:
 ...
+
+# There's another instruction between the intrinsic and the
+# conditional branch, so we need to move the insert point.
+---
+name: brcond_si_if_need_insert_terminator_point
+body:             |
+  ; WAVE64-LABEL: name: brcond_si_if_need_insert_terminator_point
+  ; WAVE64: bb.0:
+  ; WAVE64:   successors: %bb.1(0x80000000)
+  ; WAVE64:   [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+  ; WAVE64:   [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
+  ; WAVE64:   [[ICMP:%[0-9]+]]:sreg_64_xexec(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
+  ; WAVE64:   [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr2
+  ; WAVE64:   [[SI_IF:%[0-9]+]]:sreg_64_xexec(s64) = SI_IF [[ICMP]](s1), %bb.1, implicit-def $exec, implicit-def $scc, implicit $exec
+  ; WAVE64:   G_BR %bb.1
+  ; WAVE64: bb.1:
+  ; WAVE64:   S_ENDPGM 0, implicit [[COPY2]](s32)
+  ; WAVE32-LABEL: name: brcond_si_if_need_insert_terminator_point
+  ; WAVE32: bb.0:
+  ; WAVE32:   successors: %bb.1(0x80000000)
+  ; WAVE32:   [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+  ; WAVE32:   [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
+  ; WAVE32:   [[ICMP:%[0-9]+]]:sreg_32_xm0_xexec(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[COPY1]]
+  ; WAVE32:   [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr2
+  ; WAVE32:   [[SI_IF:%[0-9]+]]:sreg_32_xm0_xexec(s64) = SI_IF [[ICMP]](s1), %bb.1, implicit-def $exec, implicit-def $scc, implicit $exec
+  ; WAVE32:   G_BR %bb.1
+  ; WAVE32: bb.1:
+  ; WAVE32:   S_ENDPGM 0, implicit [[COPY2]](s32)
+  bb.0:
+    successors: %bb.1
+    liveins: $vgpr0, $vgpr1, $vgpr2
+    %0:_(s32) = COPY $vgpr0
+    %1:_(s32) = COPY $vgpr1
+    %2:_(s1) = G_ICMP intpred(ne), %0, %1
+    %3:_(s1), %4:_(s64) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.if), %2
+    %5:_(s32) = COPY $vgpr2
+    G_BRCOND %3, %bb.1
+
+  bb.1:
+    S_ENDPGM 0, implicit %5
+...
+
+---
+name: brcond_si_loop_need_terminator_insert_point
+tracksRegLiveness: true
+body:             |
+  ; WAVE64-LABEL: name: brcond_si_loop_need_terminator_insert_point
+  ; WAVE64: bb.0:
+  ; WAVE64:   successors: %bb.1(0x80000000)
+  ; WAVE64:   liveins: $vgpr0, $vgpr1, $sgpr0_sgpr1
+  ; WAVE64:   [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+  ; WAVE64:   [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
+  ; WAVE64:   [[COPY2:%[0-9]+]]:sreg_64_xexec(s64) = COPY $sgpr0_sgpr1
+  ; WAVE64: bb.1:
+  ; WAVE64:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; WAVE64:   S_NOP 0
+  ; WAVE64:   S_NOP 0
+  ; WAVE64:   S_NOP 0
+  ; WAVE64:   SI_LOOP [[COPY2]](s64), %bb.1, implicit-def $exec, implicit-def $scc, implicit $exec
+  ; WAVE64:   G_BR %bb.2
+  ; WAVE64: bb.2:
+  ; WAVE64:   S_NOP 0
+  ; WAVE32-LABEL: name: brcond_si_loop_need_terminator_insert_point
+  ; WAVE32: bb.0:
+  ; WAVE32:   successors: %bb.1(0x80000000)
+  ; WAVE32:   liveins: $vgpr0, $vgpr1, $sgpr0_sgpr1
+  ; WAVE32:   [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+  ; WAVE32:   [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
+  ; WAVE32:   [[COPY2:%[0-9]+]]:sreg_32_xm0_xexec(s64) = COPY $sgpr0_sgpr1
+  ; WAVE32: bb.1:
+  ; WAVE32:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; WAVE32:   S_NOP 0
+  ; WAVE32:   S_NOP 0
+  ; WAVE32:   S_NOP 0
+  ; WAVE32:   SI_LOOP [[COPY2]](s64), %bb.1, implicit-def $exec, implicit-def $scc, implicit $exec
+  ; WAVE32:   G_BR %bb.2
+  ; WAVE32: bb.2:
+  ; WAVE32:   S_NOP 0
+  bb.0:
+    liveins: $vgpr0, $vgpr1, $sgpr0_sgpr1
+    %0:_(s32) = COPY $vgpr0
+    %1:_(s32) = COPY $vgpr1
+    %2:_(s64) = COPY $sgpr0_sgpr1
+
+  bb.1:
+    successors: %bb.1, %bb.2
+    S_NOP 0
+    %3:_(s1) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.loop), %2
+    S_NOP 0
+    S_NOP 0
+    G_BRCOND %3, %bb.2
+    G_BR %bb.1
+
+  bb.2:
+    S_NOP 0
+...