# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py # RUN: llc -mtriple=aarch64-apple-darwin -run-pass=regbankselect -verify-machineinstrs %s -o - | FileCheck %s ... --- name: select_f32 alignment: 4 legalized: true tracksRegLiveness: true machineFunctionInfo: {} body: | bb.0: liveins: $s0, $s1, $w0 ; CHECK-LABEL: name: select_f32 ; CHECK: liveins: $s0, $s1, $w0 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr(s32) = COPY $w0 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:fpr(s32) = COPY $s0 ; CHECK-NEXT: [[COPY2:%[0-9]+]]:fpr(s32) = COPY $s1 ; CHECK-NEXT: [[SELECT:%[0-9]+]]:fpr(s32) = G_SELECT [[COPY]](s32), [[COPY1]], [[COPY2]] ; CHECK-NEXT: $s0 = COPY [[SELECT]](s32) ; CHECK-NEXT: RET_ReallyLR implicit $s0 %3:_(s32) = COPY $w0 %1:_(s32) = COPY $s0 %2:_(s32) = COPY $s1 %4:_(s32) = G_SELECT %3, %1, %2 $s0 = COPY %4(s32) RET_ReallyLR implicit $s0 ... --- name: select_f64 alignment: 4 legalized: true tracksRegLiveness: true machineFunctionInfo: {} body: | bb.0: liveins: $d0, $d1, $w0 ; CHECK-LABEL: name: select_f64 ; CHECK: liveins: $d0, $d1, $w0 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr(s32) = COPY $w0 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:fpr(s64) = COPY $d0 ; CHECK-NEXT: [[COPY2:%[0-9]+]]:fpr(s64) = COPY $d1 ; CHECK-NEXT: [[SELECT:%[0-9]+]]:fpr(s64) = G_SELECT [[COPY]](s32), [[COPY1]], [[COPY2]] ; CHECK-NEXT: $d0 = COPY [[SELECT]](s64) ; CHECK-NEXT: RET_ReallyLR implicit $d0 %3:_(s32) = COPY $w0 %1:_(s64) = COPY $d0 %2:_(s64) = COPY $d1 %4:_(s64) = G_SELECT %3, %1, %2 $d0 = COPY %4(s64) RET_ReallyLR implicit $d0 ... --- name: two_fpr_inputs_gpr_output alignment: 4 legalized: true tracksRegLiveness: true machineFunctionInfo: {} body: | bb.0: liveins: $d0, $d1, $w0 ; Verify that the G_SELECT only has FPRs. ; The only difference between fcsel and csel are the register banks. So, ; if we have two FPR inputs and a GPR output, we should do a floating point ; select anyway. This will cost one copy for the output, but that's less ; than doing two to put the inputs on GPRs. ; CHECK-LABEL: name: two_fpr_inputs_gpr_output ; CHECK: liveins: $d0, $d1, $w0 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr(s32) = COPY $w0 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:fpr(s64) = COPY $d0 ; CHECK-NEXT: [[COPY2:%[0-9]+]]:fpr(s64) = COPY $d1 ; CHECK-NEXT: [[SELECT:%[0-9]+]]:fpr(s64) = G_SELECT [[COPY]](s32), [[COPY1]], [[COPY2]] ; CHECK-NEXT: $x0 = COPY [[SELECT]](s64) ; CHECK-NEXT: RET_ReallyLR implicit $x0 %3:_(s32) = COPY $w0 %1:_(s64) = COPY $d0 %2:_(s64) = COPY $d1 %4:_(s64) = G_SELECT %3, %1, %2 $x0 = COPY %4(s64) RET_ReallyLR implicit $x0 ... --- name: one_fpr_input_fpr_output alignment: 4 legalized: true tracksRegLiveness: true machineFunctionInfo: {} body: | bb.0: liveins: $d0, $x1, $w0 ; Same idea as the above test. If the output is an FPR, and one of the ; inputs is an FPR, then it's fewer copies to just do a FCSEL. ; CHECK-LABEL: name: one_fpr_input_fpr_output ; CHECK: liveins: $d0, $x1, $w0 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr(s32) = COPY $w0 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:fpr(s64) = COPY $d0 ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr(s64) = COPY $x1 ; CHECK-NEXT: [[COPY3:%[0-9]+]]:fpr(s64) = COPY [[COPY2]](s64) ; CHECK-NEXT: [[SELECT:%[0-9]+]]:fpr(s64) = G_SELECT [[COPY]](s32), [[COPY1]], [[COPY3]] ; CHECK-NEXT: $d0 = COPY [[SELECT]](s64) ; CHECK-NEXT: RET_ReallyLR implicit $d0 %3:_(s32) = COPY $w0 %1:_(s64) = COPY $d0 %2:_(s64) = COPY $x1 %4:_(s64) = G_SELECT %3, %1, %2 $d0 = COPY %4(s64) RET_ReallyLR implicit $d0 ... --- name: one_fpr_input_gpr_output alignment: 4 legalized: true tracksRegLiveness: true machineFunctionInfo: {} body: | bb.0: liveins: $d0, $x1, $w0 ; Now we have more GPR registers on the G_SELECT. It's cheaper here to put ; everything on GPR. ; CHECK-LABEL: name: one_fpr_input_gpr_output ; CHECK: liveins: $d0, $x1, $w0 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr(s32) = COPY $w0 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:fpr(s64) = COPY $d0 ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr(s64) = COPY $x1 ; CHECK-NEXT: [[COPY3:%[0-9]+]]:gpr(s64) = COPY [[COPY1]](s64) ; CHECK-NEXT: [[SELECT:%[0-9]+]]:gpr(s64) = G_SELECT [[COPY]](s32), [[COPY3]], [[COPY2]] ; CHECK-NEXT: $x0 = COPY [[SELECT]](s64) ; CHECK-NEXT: RET_ReallyLR implicit $x0 %3:_(s32) = COPY $w0 %1:_(s64) = COPY $d0 %2:_(s64) = COPY $x1 %4:_(s64) = G_SELECT %3, %1, %2 $x0 = COPY %4(s64) RET_ReallyLR implicit $x0 ... --- name: two_gpr_input_fpr_output alignment: 4 legalized: true tracksRegLiveness: true machineFunctionInfo: {} body: | bb.0: liveins: $x0, $x1, $w0 ; Same as above. The G_SELECT should get all GPRS. ; CHECK-LABEL: name: two_gpr_input_fpr_output ; CHECK: liveins: $x0, $x1, $w0 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr(s32) = COPY $w0 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr(s64) = COPY $x0 ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr(s64) = COPY $x1 ; CHECK-NEXT: [[SELECT:%[0-9]+]]:gpr(s64) = G_SELECT [[COPY]](s32), [[COPY1]], [[COPY2]] ; CHECK-NEXT: $d0 = COPY [[SELECT]](s64) ; CHECK-NEXT: RET_ReallyLR implicit $d0 %3:_(s32) = COPY $w0 %1:_(s64) = COPY $x0 %2:_(s64) = COPY $x1 %4:_(s64) = G_SELECT %3, %1, %2 $d0 = COPY %4(s64) RET_ReallyLR implicit $d0