[NFC] Tests for future commit in DAGCombiner

[arsenm]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

Why use mir for this?

24

Should also do some SGPR tests

[kmitropoulou]

Updating D153479: [NFC] Add tests for future commit in DAGCombiner

[kmitropoulou]

Updating D153479: [NFC] Add tests for future commit in DAGCombiner

[kmitropoulou]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

CSE changes my optimization. Therefore, I need to do the checking earlier.

For example, the following test:

define i1 @test1(i32 %arg1, i32 %arg2) #0 {

%cmp1 = icmp slt i32 %arg1, 1000
%cmp2 = icmp slt i32 %arg2, 1000
%or  = or i1 %cmp1, %cmp2
ret i1 %or

}

will be optimized as follows with my optimization:

bb.0 (%ir-block.0):
  liveins: $vgpr0, $vgpr1
  %1:vgpr_32 = COPY $vgpr1
  %0:vgpr_32 = COPY $vgpr0
  %2:vgpr_32 = V_MIN_I32_e64 %0, %1, implicit $exec
  %3:sreg_32 = S_MOV_B32 1000
  %4:sreg_32_xm0_xexec = V_CMP_LT_I32_e64 killed %2, killed %3, implicit $exec
  %5:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1, killed %4, implicit $exec
  $vgpr0 = COPY %5
  SI_RETURN implicit $vgpr0

This is the output after the instruction selection. After CSE, the predicate of the compare instruction changes:

; %bb.0:

s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
s_waitcnt_vscnt null, 0x0
v_min_i32_e32 v0, v0, v1
s_delay_alu instid0(VALU_DEP_1)
v_cmp_gt_i32_e32 vcc_lo, 0x3e8, v0
v_cndmask_b32_e64 v0, 0, 1, vcc_lo
s_setpc_b64 s[30:31]

[arsenm]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

I don't understand. I assume you mean MachineCSE? Is your patch not actually a DAG combine as the description states?

Can you stop somewhere after SIFixSGPRCopies instead?

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[kmitropoulou]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

I am sorry I meant MachineCSE. The patch will upload implements is in DAGCombiner.
The new checks are generated after amdgpu-isel .

[kmitropoulou]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

*The patch that I will upload implements the optimization in DAGCombiner.

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[kmitropoulou]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

I am sorry I did not understand your comment earlier :) I update the test.

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[arsenm]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

My original point still stands. Why can't you test the end ISA? In general optimization patches are better of testing end to end unless you specifically need to check some intermediate state

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[kmitropoulou]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

Lets' say we have the following test:

define i1 @test1(i32 %arg1, i32 %arg2) #0 {

%cmp1 = icmp slt i32 %arg1, 1000
%cmp2 = icmp slt i32 %arg2, 1000
%or  = or i1 %cmp1, %cmp2
ret i1 %or

}

The dump after SI Fix SGPR copies is:

bb.0 (%ir-block.0):
  liveins: $vgpr0, $vgpr1
  %1:vgpr_32 = COPY $vgpr1
  %0:vgpr_32 = COPY $vgpr0
  %2:vgpr_32 = V_MIN_I32_e64 %0, %1, implicit $exec
  %3:sreg_32 = S_MOV_B32 1000
  %4:sreg_32_xm0_xexec = V_CMP_LT_I32_e64 killed %2, killed %3, implicit $exec
  %5:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1, killed %4, implicit $exec
  $vgpr0 = COPY %5
  SI_RETURN implicit $vgpr0

The final output is:

%bb.0:

s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
s_waitcnt_vscnt null, 0x0
v_min_i32_e32 v0, v0, v1
s_delay_alu instid0(VALU_DEP_1)
v_cmp_gt_i32_e32 vcc_lo, 0x3e8, v0
v_cndmask_b32_e64 v0, 0, 1, vcc_lo
s_setpc_b64 s[30:31]

So, it is easier to check the correctness of the optimization after SI Fix SGPR copies.

If you do not like it, then I can change it.

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[arsenm]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

You don't actually care about the MIR here, so just go to ISA. You can disable the s_delay_salu insertion for gfx11 with the flag

950

amdgpu_gfx with inreg is less noisy

[kmitropoulou]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

Which flag?

[foad]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

3

-amdgpu-enable-delay-alu=0

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[kmitropoulou]

Thank you all :)

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[arsenm]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

19–20

Do you need tests where these use different constants?

275–276

Some tests where the compare types don't match?

1101

Don't need these attributes

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

275–276

In test62, the predicates are different and in test63, the compare types are different.

[kmitropoulou]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

19–20

I added test65.

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[arsenm]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

1875

Also test f16 and v2f16

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

[kmitropoulou]

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

1875

I added tests: test84, test85, test86, test87, test88, test89, tes90 and test91.

[kmitropoulou]

Updating D153479: [NFC] Tests for future commit in DAGCombiner

llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -amdgpu-enable-delay-alu=0 < %s | FileCheck %s
+
+; The tests check the following optimization of DAGCombiner:
+; CMP(A,C)||CMP(B,C) => CMP(MIN/MAX(A,B), C)
+; CMP(A,C)&&CMP(B,C) => CMP(MIN/MAX(A,B), C)
+
+define i1 @test1(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test1:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_i32_e32 vcc_lo, 0x3e8, v0
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp slt i32 %arg1, 1000
+  %cmp2 = icmp slt i32 %arg2, 1000
+  %or  = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test2(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test2:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 0x3e8, v0
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg1, 1000
+  %cmp2 = icmp ult i32 %arg2, 1000
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test3(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test3:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_i32_e32 vcc_lo, 0x3e9, v0
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, 0x3e9, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sle i32 %arg1, 1000
+  %cmp2 = icmp sle i32 %arg2, 1000
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test4(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test4:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 0x3e9, v0
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, 0x3e9, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ule i32 %arg1, 1000
+  %cmp2 = icmp ule i32 %arg2, 1000
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test5(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test5:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, 0x3e8, v0
+; CHECK-NEXT:    v_cmp_lt_i32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sgt i32 %arg1, 1000
+  %cmp2 = icmp sgt i32 %arg2, 1000
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test6(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test6:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, 0x3e8, v0
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ugt i32 %arg1, 1000
+  %cmp2 = icmp ugt i32 %arg2, 1000
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test7(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test7:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, 0x3e7, v0
+; CHECK-NEXT:    v_cmp_lt_i32_e64 s0, 0x3e7, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sge i32 %arg1, 1000
+  %cmp2 = icmp sge i32 %arg2, 1000
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test8(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test8:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, 0x3e7, v0
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s0, 0x3e7, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp uge i32 %arg1, 1000
+  %cmp2 = icmp uge i32 %arg2, 1000
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test9(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test9:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp slt i32 %arg1, %arg3
+  %cmp2 = icmp slt i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test10(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test10:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg1, %arg3
+  %cmp2 = icmp ult i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test11(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test11:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_le_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_le_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sle i32 %arg1, %arg3
+  %cmp2 = icmp sle i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test12(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test12:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_le_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_le_u32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ule i32 %arg1, %arg3
+  %cmp2 = icmp ule i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test13(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test13:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sgt i32 %arg1, %arg3
+  %cmp2 = icmp sgt i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test14(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test14:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ugt i32 %arg1, %arg3
+  %cmp2 = icmp ugt i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test15(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test15:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_ge_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sge i32 %arg1, %arg3
+  %cmp2 = icmp sge i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test16(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test16:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_u32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp uge i32 %arg1, %arg3
+  %cmp2 = icmp uge i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test17(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test17:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_i32_e32 vcc_lo, 0x3e8, v0
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp slt i32 %arg1, 1000
+  %cmp2 = icmp slt i32 %arg2, 1000
+  %and  = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test18(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test18:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 0x3e8, v0
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg1, 1000
+  %cmp2 = icmp ult i32 %arg2, 1000
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test19(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test19:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_i32_e32 vcc_lo, 0x3e9, v0
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, 0x3e9, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sle i32 %arg1, 1000
+  %cmp2 = icmp sle i32 %arg2, 1000
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test20(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test20:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 0x3e9, v0
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, 0x3e9, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ule i32 %arg1, 1000
+  %cmp2 = icmp ule i32 %arg2, 1000
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test21(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test21:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, 0x3e8, v0
+; CHECK-NEXT:    v_cmp_lt_i32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sgt i32 %arg1, 1000
+  %cmp2 = icmp sgt i32 %arg2, 1000
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test22(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test22:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, 0x3e8, v0
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ugt i32 %arg1, 1000
+  %cmp2 = icmp ugt i32 %arg2, 1000
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test23(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test23:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, 0x3e7, v0
+; CHECK-NEXT:    v_cmp_lt_i32_e64 s0, 0x3e7, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sge i32 %arg1, 1000
+  %cmp2 = icmp sge i32 %arg2, 1000
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test24(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test24:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, 0x3e7, v0
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s0, 0x3e7, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp uge i32 %arg1, 1000
+  %cmp2 = icmp uge i32 %arg2, 1000
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test25(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test25:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp slt i32 %arg1, %arg3
+  %cmp2 = icmp slt i32 %arg2, %arg3
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test26(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test26:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg1, %arg3
+  %cmp2 = icmp ult i32 %arg2, %arg3
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test27(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test27:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_le_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_le_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sle i32 %arg1, %arg3
+  %cmp2 = icmp sle i32 %arg2, %arg3
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test28(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test28:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_le_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_le_u32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ule i32 %arg1, %arg3
+  %cmp2 = icmp ule i32 %arg2, %arg3
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test29(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test29:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sgt i32 %arg1, %arg3
+  %cmp2 = icmp sgt i32 %arg2, %arg3
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test30(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test30:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ugt i32 %arg1, %arg3
+  %cmp2 = icmp ugt i32 %arg2, %arg3
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test31(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test31:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_ge_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sge i32 %arg1, %arg3
+  %cmp2 = icmp sge i32 %arg2, %arg3
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test32(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test32:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_u32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp uge i32 %arg1, %arg3
+  %cmp2 = icmp uge i32 %arg2, %arg3
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test33(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test33:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, v0, v1
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, 0x3e8, v0
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp slt i32 %arg1, %arg2
+  %cmp2 = icmp slt i32 %arg1, 1000
+  %or  = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define amdgpu_gfx void @test34(i32 inreg %arg1, i32 inreg %arg2) {
+; CHECK-LABEL: test34:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_cmpk_lt_i32 s4, 0x3e9
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_cselect_b32 s0, -1, 0
+; CHECK-NEXT:    s_cmpk_lt_i32 s5, 0x3e9
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_cselect_b32 s1, -1, 0
+; CHECK-NEXT:    s_or_b32 s0, s0, s1
+; CHECK-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
+; CHECK-NEXT:    global_store_b8 v[0:1], v2, off dlc
+; CHECK-NEXT:    s_waitcnt_vscnt null, 0x0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sle i32 %arg1, 1000
+  %cmp2 = icmp sle i32 %arg2, 1000
+  %or = or i1 %cmp1, %cmp2
+  store volatile i1 %or, ptr addrspace(1) null
+  ret void
+}
+
+define amdgpu_gfx void @test35(i32 inreg %arg1, i32 inreg %arg2) {
+; CHECK-LABEL: test35:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_cmpk_gt_i32 s4, 0x3e8
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_cselect_b32 s0, -1, 0
+; CHECK-NEXT:    s_cmpk_gt_i32 s5, 0x3e8
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_cselect_b32 s1, -1, 0
+; CHECK-NEXT:    s_or_b32 s0, s0, s1
+; CHECK-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
+; CHECK-NEXT:    global_store_b8 v[0:1], v2, off dlc
+; CHECK-NEXT:    s_waitcnt_vscnt null, 0x0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sgt i32 %arg1, 1000
+  %cmp2 = icmp sgt i32 %arg2, 1000
+  %or = or i1 %cmp1, %cmp2
+  store volatile i1 %or, ptr addrspace(1) null
+  ret void
+}
+
+define amdgpu_gfx void @test36(i32 inreg %arg1, i32 inreg %arg2, i32 inreg %arg3) {
+; CHECK-LABEL: test36:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_cmp_lt_u32 s4, s6
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_cselect_b32 s0, -1, 0
+; CHECK-NEXT:    s_cmp_lt_u32 s5, s6
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_cselect_b32 s1, -1, 0
+; CHECK-NEXT:    s_or_b32 s0, s0, s1
+; CHECK-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
+; CHECK-NEXT:    global_store_b8 v[0:1], v2, off dlc
+; CHECK-NEXT:    s_waitcnt_vscnt null, 0x0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg1, %arg3
+  %cmp2 = icmp ult i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  store volatile i1 %or, ptr addrspace(1) null
+  ret void
+}
+
+define amdgpu_gfx void @test37(i32 inreg %arg1, i32 inreg %arg2, i32 inreg %arg3) {
+; CHECK-LABEL: test37:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_cmp_ge_i32 s4, s6
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_cselect_b32 s0, -1, 0
+; CHECK-NEXT:    s_cmp_ge_i32 s5, s6
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_cselect_b32 s1, -1, 0
+; CHECK-NEXT:    s_or_b32 s0, s0, s1
+; CHECK-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
+; CHECK-NEXT:    global_store_b8 v[0:1], v2, off dlc
+; CHECK-NEXT:    s_waitcnt_vscnt null, 0x0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sge i32 %arg1, %arg3
+  %cmp2 = icmp sge i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  store volatile i1 %or, ptr addrspace(1) null
+  ret void
+}
+
+define amdgpu_gfx void @test38(i32 inreg %arg1, i32 inreg %arg2) {
+; CHECK-LABEL: test38:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_cmpk_lt_u32 s4, 0x3e9
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_cselect_b32 s0, -1, 0
+; CHECK-NEXT:    s_cmpk_lt_u32 s5, 0x3e9
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_cselect_b32 s1, -1, 0
+; CHECK-NEXT:    s_and_b32 s0, s0, s1
+; CHECK-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
+; CHECK-NEXT:    global_store_b8 v[0:1], v2, off dlc
+; CHECK-NEXT:    s_waitcnt_vscnt null, 0x0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ule i32 %arg1, 1000
+  %cmp2 = icmp ule i32 %arg2, 1000
+  %and = and i1 %cmp1, %cmp2
+  store volatile i1 %and, ptr addrspace(1) null
+  ret void
+}
+
+define amdgpu_gfx void @test39(i32 inreg %arg1, i32 inreg %arg2) {
+; CHECK-LABEL: test39:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_cmpk_gt_i32 s4, 0x3e7
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_cselect_b32 s0, -1, 0
+; CHECK-NEXT:    s_cmpk_gt_i32 s5, 0x3e7
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_cselect_b32 s1, -1, 0
+; CHECK-NEXT:    s_and_b32 s0, s0, s1
+; CHECK-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
+; CHECK-NEXT:    global_store_b8 v[0:1], v2, off dlc
+; CHECK-NEXT:    s_waitcnt_vscnt null, 0x0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sge i32 %arg1, 1000
+  %cmp2 = icmp sge i32 %arg2, 1000
+  %and = and i1 %cmp1, %cmp2
+  store volatile i1 %and, ptr addrspace(1) null
+  ret void
+}
+
+define amdgpu_gfx void @test40(i32 inreg %arg1, i32 inreg %arg2, i32 inreg %arg3) {
+; CHECK-LABEL: test40:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_cmp_le_i32 s4, s6
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_cselect_b32 s0, -1, 0
+; CHECK-NEXT:    s_cmp_le_i32 s5, s6
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_cselect_b32 s1, -1, 0
+; CHECK-NEXT:    s_and_b32 s0, s0, s1
+; CHECK-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
+; CHECK-NEXT:    global_store_b8 v[0:1], v2, off dlc
+; CHECK-NEXT:    s_waitcnt_vscnt null, 0x0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sle i32 %arg1, %arg3
+  %cmp2 = icmp sle i32 %arg2, %arg3
+  %and = and i1 %cmp1, %cmp2
+  store volatile i1 %and, ptr addrspace(1) null
+  ret void
+}
+
+define amdgpu_gfx void @test41(i32 inreg %arg1, i32 inreg %arg2, i32 inreg %arg3) {
+; CHECK-LABEL: test41:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_cmp_ge_u32 s4, s6
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_cselect_b32 s0, -1, 0
+; CHECK-NEXT:    s_cmp_ge_u32 s5, s6
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_cselect_b32 s1, -1, 0
+; CHECK-NEXT:    s_and_b32 s0, s0, s1
+; CHECK-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
+; CHECK-NEXT:    global_store_b8 v[0:1], v2, off dlc
+; CHECK-NEXT:    s_waitcnt_vscnt null, 0x0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp uge i32 %arg1, %arg3
+  %cmp2 = icmp uge i32 %arg2, %arg3
+  %and = and i1 %cmp1, %cmp2
+  store volatile i1 %and, ptr addrspace(1) null
+  ret void
+}
+
+define i1 @test42(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test42:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, v2, v0
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s0, v2, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg3, %arg1
+  %cmp2 = icmp ult i32 %arg3, %arg2
+  %or = and i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test43(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test43:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, v2, v0
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s0, v2, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg3, %arg1
+  %cmp2 = icmp ult i32 %arg3, %arg2
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test44(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test44:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_u32_e32 vcc_lo, v2, v0
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, v2, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ugt i32 %arg3, %arg1
+  %cmp2 = icmp ugt i32 %arg3, %arg2
+  %or = and i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test45(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test45:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_u32_e32 vcc_lo, v2, v0
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, v2, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ugt i32 %arg3, %arg1
+  %cmp2 = icmp ugt i32 %arg3, %arg2
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test46(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test46:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, v2, v0
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp slt i32 %arg3, %arg1
+  %cmp2 = icmp sgt i32 %arg2, %arg3
+  %or  = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test47(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test47:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_i32_e64 s0, v2, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sgt i32 %arg1, %arg3
+  %cmp2 = icmp slt i32 %arg3, %arg2
+  %or  = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test48(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test48:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, v2, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp slt i32 %arg1, %arg3
+  %cmp2 = icmp sgt i32 %arg3, %arg2
+  %or  = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test49(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test49:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_i32_e32 vcc_lo, v2, v0
+; CHECK-NEXT:    v_cmp_lt_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sgt i32 %arg3, %arg1
+  %cmp2 = icmp slt i32 %arg2, %arg3
+  %or  = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test50(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test50:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, v2, v0
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp slt i32 %arg3, %arg1
+  %cmp2 = icmp sgt i32 %arg2, %arg3
+  %and  = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test51(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test51:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_i32_e64 s0, v2, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sgt i32 %arg1, %arg3
+  %cmp2 = icmp slt i32 %arg3, %arg2
+  %and  = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test52(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test52:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_i32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, v2, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp slt i32 %arg1, %arg3
+  %cmp2 = icmp sgt i32 %arg3, %arg2
+  %and  = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test53(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test53:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_i32_e32 vcc_lo, v2, v0
+; CHECK-NEXT:    v_cmp_lt_i32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp sgt i32 %arg3, %arg1
+  %cmp2 = icmp slt i32 %arg2, %arg3
+  %and  = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @test54(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test54:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp olt float %arg1, %arg3
+  %cmp2 = fcmp olt float %arg2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test55(double %arg1, double %arg2, double %arg3) #0 {
+; CHECK-LABEL: test55:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_le_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_le_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ole double %arg1, %arg3
+  %cmp2 = fcmp ole double %arg2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test56(double %arg1, double %arg2, double %arg3) #0 {
+; CHECK-LABEL: test56:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_gt_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ogt double %arg1, %arg3
+  %cmp2 = fcmp ogt double %arg2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test57(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test57:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_ge_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp oge float %arg1, %arg3
+  %cmp2 = fcmp oge float %arg2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test58(double %arg1, double %arg2, double %arg3) #0 {
+; CHECK-LABEL: test58:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_nle_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_nle_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ugt double %arg1, %arg3
+  %cmp2 = fcmp ugt double %arg2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test59(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test59:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_nlt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nlt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp uge float %arg1, %arg3
+  %cmp2 = fcmp uge float %arg2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test60(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test60:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_ngt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ngt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ule float %arg1, %arg3
+  %cmp2 = fcmp ule float %arg2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test61(double %arg1, double %arg2, double %arg3) #0 {
+; CHECK-LABEL: test61:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_nge_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_nge_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ult double %arg1, %arg3
+  %cmp2 = fcmp ult double %arg2, %arg3
+  %and1 = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test62(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test62:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_add_f32 v0, 1.0, v0 :: v_dual_add_f32 v1, 2.0, v1
+; CHECK-NEXT:    v_cmp_lt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %add1 = fadd nnan float %arg1, 1.0
+  %add2 = fadd nnan float %arg2, 2.0
+  %cmp1 = fcmp nnan olt float %add1, %arg3
+  %cmp2 = fcmp nnan olt float %add2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test63(double %arg1, double %arg2, double %arg3) #0 {
+; CHECK-LABEL: test63:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; CHECK-NEXT:    v_add_f64 v[2:3], v[2:3], 2.0
+; CHECK-NEXT:    v_cmp_le_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_le_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %add1 = fadd nnan double %arg1, 1.0
+  %add2 = fadd nnan double %arg2, 2.0
+  %cmp1 = fcmp nnan ole double %add1, %arg3
+  %cmp2 = fcmp nnan ole double %add2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test64(double %arg1, double %arg2, double %arg3) #0 {
+; CHECK-LABEL: test64:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; CHECK-NEXT:    v_add_f64 v[2:3], v[2:3], 2.0
+; CHECK-NEXT:    v_cmp_gt_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_gt_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %add1 = fadd nnan double %arg1, 1.0
+  %add2 = fadd nnan double %arg2, 2.0
+  %cmp1 = fcmp nnan ogt double %add1, %arg3
+  %cmp2 = fcmp nnan ogt double %add2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test65(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test65:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_add_f32 v0, 1.0, v0 :: v_dual_add_f32 v1, 2.0, v1
+; CHECK-NEXT:    v_cmp_ge_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %add1 = fadd nnan float %arg1, 1.0
+  %add2 = fadd nnan float %arg2, 2.0
+  %cmp1 = fcmp nnan oge float %add1, %arg3
+  %cmp2 = fcmp nnan oge float %add2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test66(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test66:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; CHECK-NEXT:    v_add_f64 v[2:3], v[2:3], 2.0
+; CHECK-NEXT:    v_cmp_gt_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_gt_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %add1 = fadd nnan double %arg1, 1.0
+  %add2 = fadd nnan double %arg2, 2.0
+  %cmp1 = fcmp nnan ugt double %add1, %arg3
+  %cmp2 = fcmp nnan ugt double %add2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test67(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test67:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_add_f32 v0, 1.0, v0 :: v_dual_add_f32 v1, 2.0, v1
+; CHECK-NEXT:    v_cmp_ge_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %add1 = fadd nnan float %arg1, 1.0
+  %add2 = fadd nnan float %arg2, 2.0
+  %cmp1 = fcmp nnan uge float %add1, %arg3
+  %cmp2 = fcmp nnan uge float %add2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test68(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test68:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_add_f32 v0, 1.0, v0 :: v_dual_add_f32 v1, 2.0, v1
+; CHECK-NEXT:    v_cmp_le_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_le_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %add1 = fadd nnan float %arg1, 1.0
+  %add2 = fadd nnan float %arg2, 2.0
+  %cmp1 = fcmp nnan ule float %add1, %arg3
+  %cmp2 = fcmp nnan ule float %add2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test69(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test69:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; CHECK-NEXT:    v_add_f64 v[2:3], v[2:3], 2.0
+; CHECK-NEXT:    v_cmp_lt_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_lt_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %add1 = fadd nnan double %arg1, 1.0
+  %add2 = fadd nnan double %arg2, 2.0
+  %cmp1 = fcmp nnan ult double %add1, %arg3
+  %cmp2 = fcmp nnan ult double %add2, %arg3
+  %and1 = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test70(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test70:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_max_f32 v0, v0, v0 :: v_dual_max_f32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_lt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call float @llvm.canonicalize.f32(float %arg1)
+  %var2 = call float @llvm.canonicalize.f32(float %arg2)
+  %cmp1 = fcmp olt float %var1, %arg3
+  %cmp2 = fcmp olt float %var2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test71(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test71:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f64 v[0:1], v[0:1], v[0:1]
+; CHECK-NEXT:    v_max_f64 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_cmp_le_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_le_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call double @llvm.canonicalize.f64(double %arg1)
+  %var2 = call double @llvm.canonicalize.f64(double %arg2)
+  %cmp1 = fcmp ole double %var1, %arg3
+  %cmp2 = fcmp ole double %var2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test72(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test72:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f64 v[0:1], v[0:1], v[0:1]
+; CHECK-NEXT:    v_max_f64 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_cmp_gt_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_gt_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call double @llvm.canonicalize.f64(double %arg1)
+  %var2 = call double @llvm.canonicalize.f64(double %arg2)
+  %cmp1 = fcmp ogt double %var1, %arg3
+  %cmp2 = fcmp ogt double %var2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test73(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test73:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_max_f32 v0, v0, v0 :: v_dual_max_f32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_ge_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call float @llvm.canonicalize.f32(float %arg1)
+  %var2 = call float @llvm.canonicalize.f32(float %arg2)
+  %cmp1 = fcmp oge float %var1, %arg3
+  %cmp2 = fcmp oge float %var2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test74(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test74:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f64 v[0:1], v[0:1], v[0:1]
+; CHECK-NEXT:    v_max_f64 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_cmp_nle_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_nle_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call double @llvm.canonicalize.f64(double %arg1)
+  %var2 = call double @llvm.canonicalize.f64(double %arg2)
+  %cmp1 = fcmp ugt double %var1, %arg3
+  %cmp2 = fcmp ugt double %var2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test75(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test75:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_max_f32 v0, v0, v0 :: v_dual_max_f32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_nlt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nlt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call float @llvm.canonicalize.f32(float %arg1)
+  %var2 = call float @llvm.canonicalize.f32(float %arg2)
+  %cmp1 = fcmp uge float %var1, %arg3
+  %cmp2 = fcmp uge float %var2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test76(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test76:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_max_f32 v0, v0, v0 :: v_dual_max_f32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_ngt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ngt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call float @llvm.canonicalize.f32(float %arg1)
+  %var2 = call float @llvm.canonicalize.f32(float %arg2)
+  %cmp1 = fcmp ule float %var1, %arg3
+  %cmp2 = fcmp ule float %var2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test77(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test77:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f64 v[0:1], v[0:1], v[0:1]
+; CHECK-NEXT:    v_max_f64 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_cmp_nge_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_nge_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call double @llvm.canonicalize.f64(double %arg1)
+  %var2 = call double @llvm.canonicalize.f64(double %arg2)
+  %cmp1 = fcmp ult double %var1, %arg3
+  %cmp2 = fcmp ult double %var2, %arg3
+  %and1 = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test78(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test78:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_gt_f32_e64 s0, v2, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp olt float %arg1, %arg3
+  %cmp2 = fcmp ogt float %arg3, %arg2
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test79(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test79:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_nge_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nle_f32_e64 s0, v2, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ult float %arg1, %arg3
+  %cmp2 = fcmp ugt float %arg3, %arg2
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test80(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test80:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_add_f32 v0, 1.0, v0 :: v_dual_add_f32 v1, 2.0, v1
+; CHECK-NEXT:    v_cmp_ge_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_le_f32_e64 s0, v2, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %add1 = fadd nnan float %arg1, 1.0
+  %add2 = fadd nnan float %arg2, 2.0
+  %cmp1 = fcmp nnan oge float %add1, %arg3
+  %cmp2 = fcmp nnan ole float %arg3, %add2
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test81(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test81:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; CHECK-NEXT:    v_add_f64 v[2:3], v[2:3], 2.0
+; CHECK-NEXT:    v_cmp_gt_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_lt_f64_e64 s0, v[4:5], v[2:3]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %add1 = fadd nnan double %arg1, 1.0
+  %add2 = fadd nnan double %arg2, 2.0
+  %cmp1 = fcmp nnan ugt double %add1, %arg3
+  %cmp2 = fcmp nnan ult double %arg3, %add2
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test82(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test82:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f64 v[0:1], v[0:1], v[0:1]
+; CHECK-NEXT:    v_max_f64 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_cmp_le_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_ge_f64_e64 s0, v[4:5], v[2:3]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call double @llvm.canonicalize.f64(double %arg1)
+  %var2 = call double @llvm.canonicalize.f64(double %arg2)
+  %cmp1 = fcmp ole double %var1, %arg3
+  %cmp2 = fcmp oge double %arg3, %var2
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test83(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test83:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_max_f32 v0, v0, v0 :: v_dual_max_f32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_ngt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nlt_f32_e64 s0, v2, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call float @llvm.canonicalize.f32(float %arg1)
+  %var2 = call float @llvm.canonicalize.f32(float %arg2)
+  %cmp1 = fcmp ule float %var1, %arg3
+  %cmp2 = fcmp uge float %arg3, %var2
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test84(half %arg1, half %arg2, half %arg3) {
+; CHECK-LABEL: test84:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f16_e32 v0, v0, v0
+; CHECK-NEXT:    v_max_f16_e32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_lt_f16_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_f16_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call half @llvm.canonicalize.f16(half %arg1)
+  %var2 = call half @llvm.canonicalize.f16(half %arg2)
+  %cmp1 = fcmp olt half %var1, %arg3
+  %cmp2 = fcmp olt half %var2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define <2 x i1> @test85(<2 x half> %arg1, <2 x half> %arg2, <2 x half> %arg3) {
+; CHECK-LABEL: test85:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_pk_max_f16 v0, v0, v0
+; CHECK-NEXT:    v_pk_max_f16 v1, v1, v1
+; CHECK-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
+; CHECK-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
+; CHECK-NEXT:    v_lshrrev_b32_e32 v5, 16, v1
+; CHECK-NEXT:    v_cmp_le_f16_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_le_f16_e64 s0, v1, v2
+; CHECK-NEXT:    v_cmp_le_f16_e64 s1, v4, v3
+; CHECK-NEXT:    v_cmp_le_f16_e64 s2, v5, v3
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_or_b32 s0, s1, s2
+; CHECK-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %arg1)
+  %var2 = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %arg2)
+  %cmp1 = fcmp ole <2 x half> %var1, %arg3
+  %cmp2 = fcmp ole <2 x half> %var2, %arg3
+  %or1  = or <2 x i1> %cmp1, %cmp2
+  ret <2 x i1> %or1
+}
+
+define <2 x i1> @test86(<2 x half> %arg1, <2 x half> %arg2, <2 x half> %arg3) {
+; CHECK-LABEL: test86:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_pk_max_f16 v0, v0, v0
+; CHECK-NEXT:    v_pk_max_f16 v1, v1, v1
+; CHECK-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
+; CHECK-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
+; CHECK-NEXT:    v_lshrrev_b32_e32 v5, 16, v1
+; CHECK-NEXT:    v_cmp_gt_f16_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_gt_f16_e64 s0, v1, v2
+; CHECK-NEXT:    v_cmp_gt_f16_e64 s1, v4, v3
+; CHECK-NEXT:    v_cmp_gt_f16_e64 s2, v5, v3
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_or_b32 s0, s1, s2
+; CHECK-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %arg1)
+  %var2 = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %arg2)
+  %cmp1 = fcmp ogt <2 x half> %var1, %arg3
+  %cmp2 = fcmp ogt <2 x half> %var2, %arg3
+  %or1  = or <2 x i1> %cmp1, %cmp2
+  ret <2 x i1> %or1
+}
+
+define i1 @test87(half %arg1, half %arg2, half %arg3) {
+; CHECK-LABEL: test87:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f16_e32 v0, v0, v0
+; CHECK-NEXT:    v_max_f16_e32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_ge_f16_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_f16_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call half @llvm.canonicalize.f16(half %arg1)
+  %var2 = call half @llvm.canonicalize.f16(half %arg2)
+  %cmp1 = fcmp oge half %var1, %arg3
+  %cmp2 = fcmp oge half %var2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define <2 x i1> @test88(<2 x half> %arg1, <2 x half> %arg2, <2 x half> %arg3) {
+; CHECK-LABEL: test88:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_pk_max_f16 v0, v0, v0
+; CHECK-NEXT:    v_pk_max_f16 v1, v1, v1
+; CHECK-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
+; CHECK-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
+; CHECK-NEXT:    v_lshrrev_b32_e32 v5, 16, v1
+; CHECK-NEXT:    v_cmp_nle_f16_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nle_f16_e64 s0, v1, v2
+; CHECK-NEXT:    v_cmp_nle_f16_e64 s1, v4, v3
+; CHECK-NEXT:    v_cmp_nle_f16_e64 s2, v5, v3
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_and_b32 s0, s1, s2
+; CHECK-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %arg1)
+  %var2 = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %arg2)
+  %cmp1 = fcmp ugt <2 x half> %var1, %arg3
+  %cmp2 = fcmp ugt <2 x half> %var2, %arg3
+  %and1  = and <2 x i1> %cmp1, %cmp2
+  ret <2 x i1> %and1
+}
+
+define i1 @test89(half %arg1, half %arg2, half %arg3) {
+; CHECK-LABEL: test89:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f16_e32 v0, v0, v0
+; CHECK-NEXT:    v_max_f16_e32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_nlt_f16_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nlt_f16_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call half @llvm.canonicalize.f16(half %arg1)
+  %var2 = call half @llvm.canonicalize.f16(half %arg2)
+  %cmp1 = fcmp uge half %var1, %arg3
+  %cmp2 = fcmp uge half %var2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test90(half %arg1, half %arg2, half %arg3) {
+; CHECK-LABEL: test90:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f16_e32 v0, v0, v0
+; CHECK-NEXT:    v_max_f16_e32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_ngt_f16_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ngt_f16_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call half @llvm.canonicalize.f16(half %arg1)
+  %var2 = call half @llvm.canonicalize.f16(half %arg2)
+  %cmp1 = fcmp ule half %var1, %arg3
+  %cmp2 = fcmp ule half %var2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define <2 x i1> @test91(<2 x half> %arg1, <2 x half> %arg2, <2 x half> %arg3) {
+; CHECK-LABEL: test91:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_pk_max_f16 v0, v0, v0
+; CHECK-NEXT:    v_pk_max_f16 v1, v1, v1
+; CHECK-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
+; CHECK-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
+; CHECK-NEXT:    v_lshrrev_b32_e32 v5, 16, v1
+; CHECK-NEXT:    v_cmp_nge_f16_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nge_f16_e64 s0, v1, v2
+; CHECK-NEXT:    v_cmp_nge_f16_e64 s1, v4, v3
+; CHECK-NEXT:    v_cmp_nge_f16_e64 s2, v5, v3
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_and_b32 s0, s1, s2
+; CHECK-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %arg1)
+  %var2 = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %arg2)
+  %cmp1 = fcmp ult <2 x half> %var1, %arg3
+  %cmp2 = fcmp ult <2 x half> %var2, %arg3
+  %and1 = and <2 x i1> %cmp1, %cmp2
+  ret <2 x i1> %and1
+}
+
+; The optimization does not apply to the following tests.
+
+define i1 @test92(i32 %arg1, i64 %arg2) {
+; CHECK-LABEL: test92:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_mov_b64 s[0:1], 0x3e8
+; CHECK-NEXT:    v_cmp_gt_i64_e32 vcc_lo, s[0:1], v[1:2]
+; CHECK-NEXT:    v_cmp_gt_i32_e64 s0, 0x3e8, v0
+; CHECK-NEXT:    s_or_b32 s0, s0, vcc_lo
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp slt i32 %arg1, 1000
+  %cmp2 = icmp slt i64 %arg2, 1000
+  %or  = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test93(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test93:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0x3e8, v0
+; CHECK-NEXT:    v_cmp_eq_u32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp eq i32 %arg1, 1000
+  %cmp2 = icmp eq i32 %arg2, 1000
+  %or  = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test94(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test94:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0x3e8, v0
+; CHECK-NEXT:    v_cmp_ne_u32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ne i32 %arg1, 1000
+  %cmp2 = icmp ne i32 %arg2, 1000
+  %or  = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test95(i64 %arg1, i64 %arg2, i64 %arg3) {
+; CHECK-LABEL: test95:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_lt_u64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+   %cmp1 = icmp ult i64 %arg1, %arg3
+   %cmp2 = icmp ult i64 %arg2, %arg3
+   %or = or i1 %cmp1, %cmp2
+   ret i1 %or
+}
+
+define i1 @test96(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test96:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s0, v2, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg1, %arg3
+  %cmp2 = icmp ult i32 %arg3, %arg2
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test97(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test97:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_le_u32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg1, %arg3
+  %cmp2 = icmp ule i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test98(i32 %arg1, i32 %arg2, i32 %arg3) {
+; CHECK-LABEL: test98:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_le_u32_e32 vcc_lo, v2, v0
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ule i32 %arg3, %arg1
+  %cmp2 = icmp ugt i32 %arg2, %arg3
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test99(i16 %arg1, i32 %arg2) {
+; CHECK-LABEL: test99:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_u16_e32 vcc_lo, 10, v0
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, 10, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i16 %arg1, 10
+  %cmp2 = icmp ult i32 %arg2, 10
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test100(i32 %arg1, i32 %arg2, i32 %arg3, i32 %arg4) {
+; CHECK-LABEL: test100:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_u32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s0, v1, v2
+; CHECK-NEXT:    v_cmp_lt_u32_e64 s1, v0, v3
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    s_or_b32 s1, s1, vcc_lo
+; CHECK-NEXT:    s_or_b32 s0, s0, s1
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg1, %arg3
+  %cmp2 = icmp ult i32 %arg2, %arg3
+  %or1 = or i1 %cmp1, %cmp2
+  %cmp3 = icmp ult i32 %arg1, %arg4
+  %or2 = or i1 %cmp3, %cmp1
+  %or3 = or i1 %or1, %or2
+  ret i1 %or3
+}
+
+define i1 @test101(i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: test101:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_u32_e32 vcc_lo, 0x64, v0
+; CHECK-NEXT:    v_cmp_gt_u32_e64 s0, 0x3e8, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = icmp ult i32 %arg1, 100
+  %cmp2 = icmp ult i32 %arg2, 1000
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @test102(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test102:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_gt_f32_e64 s0, v2, v1
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp olt float %arg1, %arg3
+  %cmp2 = fcmp ogt float %arg3, %arg2
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test103(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test103:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_nge_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nle_f32_e64 s0, v2, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ult float %arg1, %arg3
+  %cmp2 = fcmp ugt float %arg3, %arg2
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test104(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test104:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f64 v[0:1], v[0:1], v[0:1]
+; CHECK-NEXT:    v_max_f64 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_cmp_le_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_ge_f64_e64 s0, v[4:5], v[2:3]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call double @llvm.canonicalize.f64(double %arg1)
+  %var2 = call double @llvm.canonicalize.f64(double %arg2)
+  %cmp1 = fcmp ole double %var1, %arg3
+  %cmp2 = fcmp oge double %arg3, %var2
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test105(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test105:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_max_f32 v0, v0, v0 :: v_dual_max_f32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_ngt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nlt_f32_e64 s0, v2, v1
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call float @llvm.canonicalize.f32(float %arg1)
+  %var2 = call float @llvm.canonicalize.f32(float %arg2)
+  %cmp1 = fcmp ule float %var1, %arg3
+  %cmp2 = fcmp uge float %arg3, %var2
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test106(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test106:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_lt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp olt float %arg1, %arg3
+  %cmp2 = fcmp olt float %arg2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test107(double %arg1, double %arg2, double %arg3) #0 {
+; CHECK-LABEL: test107:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_le_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_le_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ole double %arg1, %arg3
+  %cmp2 = fcmp ole double %arg2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test108(double %arg1, double %arg2, double %arg3) #0 {
+; CHECK-LABEL: test108:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_gt_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_gt_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ogt double %arg1, %arg3
+  %cmp2 = fcmp ogt double %arg2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test109(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test109:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_ge_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp oge float %arg1, %arg3
+  %cmp2 = fcmp oge float %arg2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test110(double %arg1, double %arg2, double %arg3) #0 {
+; CHECK-LABEL: test110:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_nle_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_nle_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ugt double %arg1, %arg3
+  %cmp2 = fcmp ugt double %arg2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test111(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test111:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_nlt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nlt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp uge float %arg1, %arg3
+  %cmp2 = fcmp uge float %arg2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test112(float %arg1, float %arg2, float %arg3) #0 {
+; CHECK-LABEL: test112:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_ngt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ngt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ule float %arg1, %arg3
+  %cmp2 = fcmp ule float %arg2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test113(double %arg1, double %arg2, double %arg3) #0 {
+; CHECK-LABEL: test113:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_nge_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_nge_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %cmp1 = fcmp ult double %arg1, %arg3
+  %cmp2 = fcmp ult double %arg2, %arg3
+  %or1 = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test114(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test114:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_max_f32 v0, v0, v0 :: v_dual_max_f32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_lt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_lt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call float @llvm.canonicalize.f32(float %arg1)
+  %var2 = call float @llvm.canonicalize.f32(float %arg2)
+  %cmp1 = fcmp olt float %var1, %arg3
+  %cmp2 = fcmp olt float %var2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test115(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test115:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f64 v[0:1], v[0:1], v[0:1]
+; CHECK-NEXT:    v_max_f64 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_cmp_le_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_le_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call double @llvm.canonicalize.f64(double %arg1)
+  %var2 = call double @llvm.canonicalize.f64(double %arg2)
+  %cmp1 = fcmp ole double %var1, %arg3
+  %cmp2 = fcmp ole double %var2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test116(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test116:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f64 v[0:1], v[0:1], v[0:1]
+; CHECK-NEXT:    v_max_f64 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_cmp_gt_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_gt_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call double @llvm.canonicalize.f64(double %arg1)
+  %var2 = call double @llvm.canonicalize.f64(double %arg2)
+  %cmp1 = fcmp ogt double %var1, %arg3
+  %cmp2 = fcmp ogt double %var2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test117(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test117:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_max_f32 v0, v0, v0 :: v_dual_max_f32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_ge_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ge_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call float @llvm.canonicalize.f32(float %arg1)
+  %var2 = call float @llvm.canonicalize.f32(float %arg2)
+  %cmp1 = fcmp oge float %var1, %arg3
+  %cmp2 = fcmp oge float %var2, %arg3
+  %and1  = and i1 %cmp1, %cmp2
+  ret i1 %and1
+}
+
+define i1 @test118(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test118:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f64 v[0:1], v[0:1], v[0:1]
+; CHECK-NEXT:    v_max_f64 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_cmp_nle_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_nle_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call double @llvm.canonicalize.f64(double %arg1)
+  %var2 = call double @llvm.canonicalize.f64(double %arg2)
+  %cmp1 = fcmp ugt double %var1, %arg3
+  %cmp2 = fcmp ugt double %var2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test119(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test119:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_max_f32 v0, v0, v0 :: v_dual_max_f32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_nlt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_nlt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call float @llvm.canonicalize.f32(float %arg1)
+  %var2 = call float @llvm.canonicalize.f32(float %arg2)
+  %cmp1 = fcmp uge float %var1, %arg3
+  %cmp2 = fcmp uge float %var2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test120(float %arg1, float %arg2, float %arg3) {
+; CHECK-LABEL: test120:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_dual_max_f32 v0, v0, v0 :: v_dual_max_f32 v1, v1, v1
+; CHECK-NEXT:    v_cmp_ngt_f32_e32 vcc_lo, v0, v2
+; CHECK-NEXT:    v_cmp_ngt_f32_e64 s0, v1, v2
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call float @llvm.canonicalize.f32(float %arg1)
+  %var2 = call float @llvm.canonicalize.f32(float %arg2)
+  %cmp1 = fcmp ule float %var1, %arg3
+  %cmp2 = fcmp ule float %var2, %arg3
+  %or1  = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+define i1 @test121(double %arg1, double %arg2, double %arg3) {
+; CHECK-LABEL: test121:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_max_f64 v[0:1], v[0:1], v[0:1]
+; CHECK-NEXT:    v_max_f64 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_cmp_nge_f64_e32 vcc_lo, v[0:1], v[4:5]
+; CHECK-NEXT:    v_cmp_nge_f64_e64 s0, v[2:3], v[4:5]
+; CHECK-NEXT:    s_or_b32 s0, vcc_lo, s0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %var1 = call double @llvm.canonicalize.f64(double %arg1)
+  %var2 = call double @llvm.canonicalize.f64(double %arg2)
+  %cmp1 = fcmp ult double %var1, %arg3
+  %cmp2 = fcmp ult double %var2, %arg3
+  %or1 = or i1 %cmp1, %cmp2
+  ret i1 %or1
+}
+
+declare double @llvm.canonicalize.f64(double)
+declare float @llvm.canonicalize.f32(float)
+declare half @llvm.canonicalize.f16(half)
+declare <2 x half> @llvm.canonicalize.v2f16(<2 x half>)
+
+attributes #0 = { nounwind "amdgpu-ieee"="false" }