[AArch64][GlobalISel] Fold G_ASHR into TB(N)Z bit calculation

[paquette]

Summary

This implements walking over G_ASHR in the same way as getTestBitOperand in AArch64ISelLowering.

(tbz (ashr x, c), b) -> (tbz x, b+c) or (tbz x, msb) if b+c is > # bits in x

[aemerson]

LGTM with nit.

llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp

1076

Can you factor our these getType() calls here and in the G_SHL case into a variable?

llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp

case TargetOpcode::G_AND: {
         // same side at some point earlier?
         std::swap(ConstantReg, TestReg);
         VRegAndVal = getConstantVRegValWithLookThrough(ConstantReg, MRI);
       }
       if (VRegAndVal)
         C = VRegAndVal->Value;
       break;
     }
+    case TargetOpcode::G_ASHR:
     case TargetOpcode::G_SHL: {
       TestReg = MI->getOperand(1).getReg();
       auto VRegAndVal =
           getConstantVRegValWithLookThrough(MI->getOperand(2).getReg(), MRI);
       if (VRegAndVal)
         C = VRegAndVal->Value;
       break;
     }
while (MachineInstr *MI = getDefIgnoringCopies(Reg, MRI)) {
     case TargetOpcode::G_SHL:
       // (tbz (shl x, c), b) -> (tbz x, b-c) when b-c is positive and fits in
       // the type of the register.
       if (*C <= Bit && (Bit - *C) < MRI.getType(TestReg).getSizeInBits()) {
         NextReg = TestReg;
         Bit = Bit - *C;
       }
       break;
+    case TargetOpcode::G_ASHR:
+      // (tbz (ashr x, c), b) -> (tbz x, b+c) or (tbz x, msb) if b+c is > # bits
+      // in x
+      NextReg = TestReg;
+      Bit = Bit + *C;
+      if (Bit >= MRI.getType(TestReg).getSizeInBits())
+        Bit = MRI.getType(TestReg).getSizeInBits() - 1;
+      break;
     }
 
     // Check if we found anything worth folding.
     if (!NextReg.isValid() || !MRI.hasOneUse(NextReg))
       break;
     Reg = NextReg;
   }
 

llvm/test/CodeGen/AArch64/GlobalISel/opt-fold-shift-tbz-tbnz.mir

bb.0:
 
     %and:gpr(s64) = G_AND %fold_me, %bit
     %cmp:gpr(s32) = G_ICMP intpred(ne), %and(s64), %zero
     %cmp_trunc:gpr(s1) = G_TRUNC %cmp(s32)
     G_BRCOND %cmp_trunc(s1), %bb.1
     G_BR %bb.0
   bb.1:
     RET_ReallyLR
+...
+---
+name:            fold_ashr_in_range
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: fold_ashr_in_range
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr64all = COPY $x0
+  ; CHECK:   [[COPY:%[0-9]+]]:gpr32all = COPY %copy.sub_32
+  ; CHECK:   [[COPY1:%[0-9]+]]:gpr32 = COPY [[COPY]]
+  ; CHECK:   TBNZW [[COPY1]], 4, %bb.1
+  ; CHECK:   B %bb.0
+  ; CHECK: bb.1:
+  ; CHECK:   RET_ReallyLR
+  bb.0:
+    successors: %bb.0, %bb.1
+    liveins: $x0
+    %copy:gpr(s64) = COPY $x0
+    %bit:gpr(s64) = G_CONSTANT i64 8
+    %zero:gpr(s64) = G_CONSTANT i64 0
+
+    ; tb(n)z (ashr x, c), b == tbz(x, b + c) when b+c <= the size of the type.
+    ; In this case, we should get 1 + 3 = 4 as the bit number.
+    %fold_cst:gpr(s64) = G_CONSTANT i64 1
+    %fold_me:gpr(s64) = G_ASHR %copy, %fold_cst
+
+    %and:gpr(s64) = G_AND %fold_me, %bit
+    %cmp:gpr(s32) = G_ICMP intpred(ne), %and(s64), %zero
+    %cmp_trunc:gpr(s1) = G_TRUNC %cmp(s32)
+    G_BRCOND %cmp_trunc(s1), %bb.1
+    G_BR %bb.0
+  bb.1:
+    RET_ReallyLR
+
+...
+---
+name:            fold_ashr_msb_1
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: fold_ashr_msb_1
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr32 = COPY $w0
+  ; CHECK:   TBNZW %copy, 31, %bb.1
+  ; CHECK:   B %bb.0
+  ; CHECK: bb.1:
+  ; CHECK:   RET_ReallyLR
+  bb.0:
+    successors: %bb.0, %bb.1
+    liveins: $x0
+    %copy:gpr(s32) = COPY $w0
+    %bit:gpr(s32) = G_CONSTANT i32 8
+    %zero:gpr(s32) = G_CONSTANT i32 0
+
+    ; We should get a TBNZW with a 31 as the bit.
+    %fold_cst:gpr(s32) = G_CONSTANT i32 1234
+    %fold_me:gpr(s32) = G_ASHR %copy, %fold_cst
+
+    %and:gpr(s32) = G_AND %fold_me, %bit
+    %cmp:gpr(s32) = G_ICMP intpred(ne), %and(s32), %zero
+    %cmp_trunc:gpr(s1) = G_TRUNC %cmp(s32)
+    G_BRCOND %cmp_trunc(s1), %bb.1
+    G_BR %bb.0
+  bb.1:
+    RET_ReallyLR
+
+...
+---
+name:            fold_ashr_msb_2
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: fold_ashr_msb_2
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr64 = COPY $x0
+  ; CHECK:   TBNZX %copy, 63, %bb.1
+  ; CHECK:   B %bb.0
+  ; CHECK: bb.1:
+  ; CHECK:   RET_ReallyLR
+  bb.0:
+    successors: %bb.0, %bb.1
+    liveins: $x0
+    %copy:gpr(s64) = COPY $x0
+    %bit:gpr(s64) = G_CONSTANT i64 8
+    %zero:gpr(s64) = G_CONSTANT i64 0
+
+    ; We should get a TBNZX with a 63 as the bit.
+    %fold_cst:gpr(s64) = G_CONSTANT i64 1234
+    %fold_me:gpr(s64) = G_ASHR %copy, %fold_cst
+
+    %and:gpr(s64) = G_AND %fold_me, %bit
+    %cmp:gpr(s32) = G_ICMP intpred(ne), %and(s64), %zero
+    %cmp_trunc:gpr(s1) = G_TRUNC %cmp(s32)
+    G_BRCOND %cmp_trunc(s1), %bb.1
+    G_BR %bb.0
+  bb.1:
+    RET_ReallyLR