AMDGPU/GlobalISel: Remove selection of MAD/MAC when not available

[critson]

Summary

Add code to respect mad-mac-f32-insts target feature.

[arsenm]

llvm/test/CodeGen/AMDGPU/GlobalISel/udiv.i64.ll

4

I think it's a bit much to have this giant extra test just for this. Can you just add a run line to legalize-fmad.* instead?

184

The check prefix being an actual word confused me. NOFMAD maybe?

[critson]

Simplify test changes.

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

getActionDefinitionsBuilder(G_FSUB)
       .legalFor({S32})
       // Must use fadd + fneg
       .lowerFor({S64, S16, V2S16})
       .scalarize(0)
       .clampScalar(0, S32, S64);
 
   // Whether this is legal depends on the floating point mode for the function.
   auto &FMad = getActionDefinitionsBuilder(G_FMAD);
-  if (ST.hasMadF16())
+  if (ST.hasMadF16() && ST.hasMadMacF32Insts())
     FMad.customFor({S32, S16});
-  else
+  else if (ST.hasMadMacF32Insts())
     FMad.customFor({S32});
+  else if (ST.hasMadF16())
+    FMad.customFor({S16});
   FMad.scalarize(0)
       .lower();
 
   // TODO: Do we need to clamp maximum bitwidth?
   getActionDefinitionsBuilder(G_TRUNC)
     .legalIf(isScalar(0))
     .legalFor({{V2S16, V2S32}})
     .clampMaxNumElements(0, S16, 2)

llvm/test/CodeGen/AMDGPU/GlobalISel/udiv.i64.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -mtriple=amdgcn-amd-amdhsa -denormal-fp-math-f32=preserve-sign < %s | FileCheck -check-prefixes=CHECK,GISEL %s
-; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=0 -mtriple=amdgcn-amd-amdhsa -denormal-fp-math-f32=preserve-sign < %s | FileCheck -check-prefixes=CHECK,CGP %s
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -mtriple=amdgcn-amd-amdhsa -denormal-fp-math-f32=preserve-sign -mattr=+mad-mac-f32-insts < %s | FileCheck -check-prefixes=CHECK,GISEL %s
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=0 -mtriple=amdgcn-amd-amdhsa -denormal-fp-math-f32=preserve-sign -mattr=+mad-mac-f32-insts < %s | FileCheck -check-prefixes=CHECK,CGP %s
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=0 -mtriple=amdgcn-amd-amdhsa -denormal-fp-math-f32=preserve-sign -mattr=-mad-mac-f32-insts < %s | FileCheck -check-prefixes=NOMAD %s
 
 ; The same 32-bit expansion is implemented in the legalizer and in AMDGPUCodeGenPrepare.
 
 define i64 @v_udiv_i64(i64 %num, i64 %den) {
 ; CHECK-LABEL: v_udiv_i64:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    v_or_b32_e32 v5, v1, v3

 ; CHECK-NEXT:    v_cndmask_b32_e64 v0, v1, v4, s[4:5]
 ; CHECK-NEXT:    v_cndmask_b32_e32 v4, v5, v0, vcc
 ; CHECK-NEXT:    v_mov_b32_e32 v5, 0
 ; CHECK-NEXT:  BB0_4:
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; CHECK-NEXT:    v_mov_b32_e32 v0, v4
 ; CHECK-NEXT:    v_mov_b32_e32 v1, v5
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_udiv_i64:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    v_or_b32_e32 v5, v1, v3
+; NOMAD-NEXT:    v_mov_b32_e32 v4, 0
+; NOMAD-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[4:5]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, 1
+; NOMAD-NEXT:    s_xor_b64 s[4:5], vcc, s[4:5]
+; NOMAD-NEXT:    ; implicit-def: $vgpr4_vgpr5
+; NOMAD-NEXT:    s_and_saveexec_b64 s[6:7], s[4:5]
+; NOMAD-NEXT:    s_xor_b64 s[6:7], exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB0_2
+; NOMAD-NEXT:  ; %bb.1:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, v2
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v5, v3
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, 0, v2
+; NOMAD-NEXT:    v_subb_u32_e32 v7, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x4f800000, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v5, v4
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v4
+; NOMAD-NEXT:    v_trunc_f32_e32 v5, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0xcf800000, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v7, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v5, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v5, v9
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v4, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v5, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v5, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v13, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v13, v12
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_addc_u32_e64 v9, s[4:5], v5, v8, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v7, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v6, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v9, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v4, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v9, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v4, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v9, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v4, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v6, v9, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v11, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v10, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v7, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v8, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v9, v7
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v10, v11
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v8, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v9, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v5, vcc, v5, v6, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v1, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v7, v0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v1, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v0, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v1, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v1, v5
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v9, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v2, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v2, v5
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, 1, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v11, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, 1, v10
+; NOMAD-NEXT:    v_addc_u32_e32 v12, vcc, 0, v11, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v9
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, v0, v7
+; NOMAD-NEXT:    v_subb_u32_e64 v9, s[4:5], v1, v6, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v6
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v9, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v1, vcc, v1, v3, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v9, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v6, v13, v6, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, v7, v2
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v7, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v9, v7, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v10, v8, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v11, v12, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v4, v1, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v5, v3, vcc
+; NOMAD-NEXT:  BB0_2: ; %Flow
+; NOMAD-NEXT:    s_or_saveexec_b64 s[6:7], s[6:7]
+; NOMAD-NEXT:    s_xor_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB0_4
+; NOMAD-NEXT:  ; %bb.3:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v1, v2
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v3, v1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v1, v2
+; NOMAD-NEXT:    v_sub_i32_e32 v5, vcc, 0, v3
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v3, v1
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v1, v3
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v1, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v3, v1, v2
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, 1, v1
+; NOMAD-NEXT:    v_subrev_i32_e32 v5, vcc, 1, v1
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, v0, v3
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v3
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v6, v2
+; NOMAD-NEXT:    s_and_b64 s[4:5], s[4:5], vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, v1, v4, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v5, v0, vcc
+; NOMAD-NEXT:    v_mov_b32_e32 v5, 0
+; NOMAD-NEXT:  BB0_4:
+; NOMAD-NEXT:    s_or_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    v_mov_b32_e32 v0, v4
+; NOMAD-NEXT:    v_mov_b32_e32 v1, v5
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = udiv i64 %num, %den
   ret i64 %result
 }
 
 ; FIXME: This is a workaround for not handling uniform VGPR case.
 declare i32 @llvm.amdgcn.readfirstlane(i32)
 
 define amdgpu_ps i64 @s_udiv_i64(i64 inreg %num, i64 inreg %den) {

 ; CHECK-NEXT:    v_cmp_le_u32_e64 s[0:1], s2, v4
 ; CHECK-NEXT:    s_and_b64 s[0:1], s[0:1], vcc
 ; CHECK-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
 ; CHECK-NEXT:  BB1_4:
 ; CHECK-NEXT:    v_readfirstlane_b32 s0, v0
 ; CHECK-NEXT:    s_mov_b32 s1, s0
 ; CHECK-NEXT:    ; return to shader part epilog
+;
+; NOMAD-LABEL: s_udiv_i64:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_or_b64 s[4:5], s[0:1], s[2:3]
+; NOMAD-NEXT:    s_mov_b32 s6, 0
+; NOMAD-NEXT:    s_mov_b32 s7, -1
+; NOMAD-NEXT:    s_and_b64 s[4:5], s[4:5], s[6:7]
+; NOMAD-NEXT:    v_cmp_eq_u64_e64 s[4:5], s[4:5], 0
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[6:7], 0, 1
+; NOMAD-NEXT:    s_xor_b64 vcc, s[4:5], s[6:7]
+; NOMAD-NEXT:    s_mov_b32 s4, 1
+; NOMAD-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; NOMAD-NEXT:    s_cbranch_vccz BB1_2
+; NOMAD-NEXT:  ; %bb.1:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v0, s2
+; NOMAD-NEXT:    v_mov_b32_e32 v1, s3
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v2, s3
+; NOMAD-NEXT:    s_sub_u32 s6, 0, s2
+; NOMAD-NEXT:    s_cselect_b32 s4, 1, 0
+; NOMAD-NEXT:    v_mov_b32_e32 v3, s1
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
+; NOMAD-NEXT:    s_and_b32 s4, s4, 1
+; NOMAD-NEXT:    v_add_f32_e32 v0, v2, v0
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; NOMAD-NEXT:    s_cmp_lg_u32 s4, 0
+; NOMAD-NEXT:    s_subb_u32 s7, 0, s3
+; NOMAD-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x2f800000, v0
+; NOMAD-NEXT:    v_trunc_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0xcf800000, v2
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; NOMAD-NEXT:    v_add_f32_e32 v0, v4, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s6, v2
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s6, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v6, s7, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s6, v0
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v2, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v2, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v0, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v2, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v9, v8
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v5
+; NOMAD-NEXT:    v_addc_u32_e64 v5, s[4:5], v2, v4, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s6, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v6, s7, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s6, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s6, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v5, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v5, v4
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v0, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v5, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v0, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v5, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v9, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v8, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v8, v9
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v6
+; NOMAD-NEXT:    v_addc_u32_e32 v2, vcc, v2, v5, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s1, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v5, s0, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v0, s1, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v6, s0, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v7, s1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v8, s0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, s1, v2
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v7, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s2, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v6, s3, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s2, v0
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, 1, v0
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, 1, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v2, s2, v2
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v6, v2
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v7
+; NOMAD-NEXT:    v_sub_i32_e32 v5, vcc, s0, v5
+; NOMAD-NEXT:    v_subb_u32_e64 v3, s[4:5], v3, v2, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v2, s[4:5], s1, v2
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s2, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s3, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v1, vcc, v2, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v7, v6, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v3, vcc, s2, v5
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s2, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v3, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s3, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v5, v3, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v8, v4, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
+; NOMAD-NEXT:    s_mov_b32 s4, 0
+; NOMAD-NEXT:  BB1_2: ; %Flow
+; NOMAD-NEXT:    s_and_b32 s1, s4, 1
+; NOMAD-NEXT:    s_cmp_lg_u32 s1, 0
+; NOMAD-NEXT:    s_cbranch_scc0 BB1_4
+; NOMAD-NEXT:  ; %bb.3:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v0, s2
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v1, v0, s2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v0, s2
+; NOMAD-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v2
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v1, v0
+; NOMAD-NEXT:    v_add_i32_e64 v2, s[4:5], v0, v1
+; NOMAD-NEXT:    v_sub_i32_e64 v0, s[4:5], v0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v0, v0, s0
+; NOMAD-NEXT:    v_mul_lo_u32 v1, v0, s2
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, 1, v0
+; NOMAD-NEXT:    v_subrev_i32_e32 v3, vcc, 1, v0
+; NOMAD-NEXT:    v_sub_i32_e32 v4, vcc, s0, v1
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, s0, v1
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[0:1], s2, v4
+; NOMAD-NEXT:    s_and_b64 s[0:1], s[0:1], vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, v0, v2, s[0:1]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; NOMAD-NEXT:  BB1_4:
+; NOMAD-NEXT:    v_readfirstlane_b32 s0, v0
+; NOMAD-NEXT:    s_mov_b32 s1, s0
+; NOMAD-NEXT:    ; return to shader part epilog
   %result = udiv i64 %num, %den
   %cast = bitcast i64 %result to <2 x i32>
   %elt.0 = extractelement <2 x i32> %cast, i32 0
   %elt.1 = extractelement <2 x i32> %cast, i32 1
   %res.0 = call i32 @llvm.amdgcn.readfirstlane(i32 %elt.0)
   %res.1 = call i32 @llvm.amdgcn.readfirstlane(i32 %elt.1)
   %ins.0 = insertelement <2 x i32> undef, i32 %res.0, i32 0
   %ins.1 = insertelement <2 x i32> %ins.0, i32 %res.0, i32 1

 ; CGP-NEXT:    v_cndmask_b32_e64 v2, v3, v5, s[4:5]
 ; CGP-NEXT:    v_cndmask_b32_e32 v4, v7, v2, vcc
 ; CGP-NEXT:    v_mov_b32_e32 v5, 0
 ; CGP-NEXT:  BB2_8:
 ; CGP-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; CGP-NEXT:    v_mov_b32_e32 v2, v4
 ; CGP-NEXT:    v_mov_b32_e32 v3, v5
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_udiv_v2i64:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    v_mov_b32_e32 v8, v0
+; NOMAD-NEXT:    v_mov_b32_e32 v9, v1
+; NOMAD-NEXT:    v_or_b32_e32 v1, v9, v5
+; NOMAD-NEXT:    v_mov_b32_e32 v0, 0
+; NOMAD-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[0:1]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, 1
+; NOMAD-NEXT:    s_xor_b64 s[4:5], vcc, s[4:5]
+; NOMAD-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; NOMAD-NEXT:    s_and_saveexec_b64 s[6:7], s[4:5]
+; NOMAD-NEXT:    s_xor_b64 s[6:7], exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB2_2
+; NOMAD-NEXT:  ; %bb.1:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v0, v4
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v1, v5
+; NOMAD-NEXT:    v_sub_i32_e32 v10, vcc, 0, v4
+; NOMAD-NEXT:    v_subb_u32_e32 v11, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; NOMAD-NEXT:    v_add_f32_e32 v0, v1, v0
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; NOMAD-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; NOMAD-NEXT:    v_trunc_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_f32_e32 v12, 0xcf800000, v1
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; NOMAD-NEXT:    v_add_f32_e32 v0, v12, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v10, v1
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v10, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v11, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v10, v0
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v14, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v1, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v0, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v1, v13
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v0, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v17, v1, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v0, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v1, v12
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v17, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v15, v14
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v17, v16
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v15, v14
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v14
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v13
+; NOMAD-NEXT:    v_addc_u32_e64 v13, s[4:5], v1, v12, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v1, s[4:5], v1, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v10, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v11, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v10, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v10, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v0, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v13, v12
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v11, v10
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v10, v14
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v0, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v13, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v0, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v13, v10
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v15, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[4:5], v14, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[4:5], v12, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v13, v11
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v14, v15
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v12, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[4:5], v13, v12
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v10, v12
+; NOMAD-NEXT:    v_addc_u32_e32 v1, vcc, v1, v10, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v11
+; NOMAD-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v9, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v8, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v0, v9, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v8, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v9, v1
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v8, v1
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v9, v1
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v13, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v12, v10
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v13, v11
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v4, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v5, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v4, v0
+; NOMAD-NEXT:    v_add_i32_e32 v1, vcc, v1, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v4, v1
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, 1, v0
+; NOMAD-NEXT:    v_addc_u32_e32 v15, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v12, v10
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, 1, v14
+; NOMAD-NEXT:    v_addc_u32_e32 v16, vcc, 0, v15, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v13
+; NOMAD-NEXT:    v_sub_i32_e32 v11, vcc, v8, v11
+; NOMAD-NEXT:    v_subb_u32_e64 v13, s[4:5], v9, v10, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v9, s[4:5], v9, v10
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v11, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v13, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v9, vcc, v9, v5, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v13, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v10, v17, v10, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v11, vcc, v11, v4
+; NOMAD-NEXT:    v_subbrev_u32_e32 v9, vcc, 0, v9, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v11, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v13, v11, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v14, v12, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v9, v15, v16, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v10
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc
+; NOMAD-NEXT:  BB2_2: ; %Flow2
+; NOMAD-NEXT:    s_or_saveexec_b64 s[6:7], s[6:7]
+; NOMAD-NEXT:    s_xor_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB2_4
+; NOMAD-NEXT:  ; %bb.3:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v0, v4
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v1, v0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v0, v4
+; NOMAD-NEXT:    v_sub_i32_e32 v9, vcc, 0, v1
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v1, v0
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v0, v1
+; NOMAD-NEXT:    v_sub_i32_e64 v0, s[4:5], v0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v0, v0, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v1, v0, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, 1, v0
+; NOMAD-NEXT:    v_subrev_i32_e32 v9, vcc, 1, v0
+; NOMAD-NEXT:    v_sub_i32_e32 v10, vcc, v8, v1
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v8, v1
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v10, v4
+; NOMAD-NEXT:    s_and_b64 s[4:5], s[4:5], vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, v0, v5, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc
+; NOMAD-NEXT:    v_mov_b32_e32 v1, 0
+; NOMAD-NEXT:  BB2_4:
+; NOMAD-NEXT:    s_or_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    v_or_b32_e32 v5, v3, v7
+; NOMAD-NEXT:    v_mov_b32_e32 v4, 0
+; NOMAD-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[4:5]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, 1
+; NOMAD-NEXT:    s_xor_b64 s[4:5], vcc, s[4:5]
+; NOMAD-NEXT:    ; implicit-def: $vgpr4_vgpr5
+; NOMAD-NEXT:    s_and_saveexec_b64 s[6:7], s[4:5]
+; NOMAD-NEXT:    s_xor_b64 s[6:7], exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB2_6
+; NOMAD-NEXT:  ; %bb.5:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, v6
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v5, v7
+; NOMAD-NEXT:    v_sub_i32_e32 v8, vcc, 0, v6
+; NOMAD-NEXT:    v_subb_u32_e32 v9, vcc, 0, v7, vcc
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x4f800000, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v5, v4
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v4
+; NOMAD-NEXT:    v_trunc_f32_e32 v5, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v10, 0xcf800000, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v10, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v8, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v9, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v8, v4
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v12, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v5, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v5, v11
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v4, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v5, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v4, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v5, v10
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v15, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v15, v14
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v11
+; NOMAD-NEXT:    v_addc_u32_e64 v11, s[4:5], v5, v10, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v9, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v8, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v11, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v11, v10
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v9, v8
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v8, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v4, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v11, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v4, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v11, v8
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v13, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v12, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v9, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v10, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v11, v9
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v12, v13
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v10, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v11, v10
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v8, v10
+; NOMAD-NEXT:    v_addc_u32_e32 v5, vcc, v5, v8, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v2, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v3, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v2, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v5
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v11, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v7, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v5
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, 1, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v13, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, 1, v12
+; NOMAD-NEXT:    v_addc_u32_e32 v14, vcc, 0, v13, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
+; NOMAD-NEXT:    v_sub_i32_e32 v9, vcc, v2, v9
+; NOMAD-NEXT:    v_subb_u32_e64 v11, s[4:5], v3, v8, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v3, s[4:5], v3, v8
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v9, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v11, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v3, vcc, v3, v7, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v11, v7
+; NOMAD-NEXT:    v_cndmask_b32_e32 v8, v15, v8, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v9, vcc, v9, v6
+; NOMAD-NEXT:    v_subbrev_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v9, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v3, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v7
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v11, v9, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v12, v10, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v7, v13, v14, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v4, v3, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v5, v7, vcc
+; NOMAD-NEXT:  BB2_6: ; %Flow
+; NOMAD-NEXT:    s_or_saveexec_b64 s[6:7], s[6:7]
+; NOMAD-NEXT:    s_xor_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB2_8
+; NOMAD-NEXT:  ; %bb.7:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v3, v6
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v3, v3
+; NOMAD-NEXT:    v_mul_f32_e32 v3, 0x4f800000, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v3, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v6
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, 0, v4
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v4, v7, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v4, v3
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v3, v4
+; NOMAD-NEXT:    v_sub_i32_e64 v3, s[4:5], v3, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v3, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v3, v6
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, 1, v3
+; NOMAD-NEXT:    v_subrev_i32_e32 v7, vcc, 1, v3
+; NOMAD-NEXT:    v_sub_i32_e32 v8, vcc, v2, v4
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v2, v4
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v8, v6
+; NOMAD-NEXT:    s_and_b64 s[4:5], s[4:5], vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, v3, v5, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v7, v2, vcc
+; NOMAD-NEXT:    v_mov_b32_e32 v5, 0
+; NOMAD-NEXT:  BB2_8:
+; NOMAD-NEXT:    s_or_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    v_mov_b32_e32 v2, v4
+; NOMAD-NEXT:    v_mov_b32_e32 v3, v5
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = udiv <2 x i64> %num, %den
   ret <2 x i64> %result
 }
 
 define i64 @v_udiv_i64_pow2k_denom(i64 %num) {
 ; CHECK-LABEL: v_udiv_i64_pow2k_denom:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
 ; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v8, v6, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e32 v1, v9, v10, vcc
 ; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_udiv_i64_pow2k_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_movk_i32 s6, 0x1000
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v2, 0
+; NOMAD-NEXT:    s_mov_b32 s7, 0xfffff000
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v3, s6
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
+; NOMAD-NEXT:    v_add_f32_e32 v2, v2, v3
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; NOMAD-NEXT:    v_trunc_f32_e32 v3, v3
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0xcf800000, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; NOMAD-NEXT:    v_add_f32_e32 v2, v4, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s7, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, -1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s7, v2
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v3, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v2, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v9, v8
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
+; NOMAD-NEXT:    v_addc_u32_e64 v5, s[4:5], v3, v4, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v3, s[4:5], v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, -1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s7, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v5, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v5, v4
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v5, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v2, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v5, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v9, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v8, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v8, v9
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v6
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, v3, v5, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v1, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v0, v3
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v1, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v0, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v1, v3
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v7, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s6, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, 0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s6, v2
+; NOMAD-NEXT:    v_add_i32_e32 v3, vcc, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s6, v3
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, 1, v2
+; NOMAD-NEXT:    v_addc_u32_e32 v9, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, 1, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v10, vcc, 0, v9, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v5
+; NOMAD-NEXT:    v_subb_u32_e64 v5, s[4:5], v1, v4, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v4
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s6, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v0, vcc, s6, v0
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s6, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v8, v6, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v9, v10, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = udiv i64 %num, 4096
   ret i64 %result
 }
 
 define <2 x i64> @v_udiv_v2i64_pow2k_denom(<2 x i64> %num) {
 ; GISEL-LABEL: v_udiv_v2i64_pow2k_denom:
 ; GISEL:       ; %bb.0:
 ; GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

 ; CGP-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
 ; CGP-NEXT:    v_cndmask_b32_e32 v2, v4, v1, vcc
 ; CGP-NEXT:    v_cndmask_b32_e64 v1, v19, v15, s[4:5]
 ; CGP-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v10
 ; CGP-NEXT:    v_cndmask_b32_e64 v0, v5, v0, s[4:5]
 ; CGP-NEXT:    v_cndmask_b32_e64 v1, v7, v1, s[4:5]
 ; CGP-NEXT:    v_cndmask_b32_e32 v3, v6, v3, vcc
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_udiv_v2i64_pow2k_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_movk_i32 s10, 0x1000
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, 0
+; NOMAD-NEXT:    s_mov_b32 s8, 0xfffff000
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v5, s10
+; NOMAD-NEXT:    v_mul_f32_e32 v6, 0x4f800000, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x4f800000, v4
+; NOMAD-NEXT:    v_add_f32_e32 v6, v6, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v4, v5
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v5, v6
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x5f7ffffc, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v6, 0x2f800000, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v7, 0x2f800000, v4
+; NOMAD-NEXT:    v_trunc_f32_e32 v6, v6
+; NOMAD-NEXT:    v_trunc_f32_e32 v7, v7
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0xcf800000, v6
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v6, v6
+; NOMAD-NEXT:    v_mul_f32_e32 v9, 0xcf800000, v7
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v7, v7
+; NOMAD-NEXT:    v_add_f32_e32 v5, v8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s8, v6
+; NOMAD-NEXT:    v_add_f32_e32 v4, v9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, s8, v7
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v11, -1, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v12, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v13, s8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, -1, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v15, s8, v4
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v11, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v6, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v5, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v6, v10
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v14, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v7, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v4, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v7, v13
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v6, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v5, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v6, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v19, v4, v9
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v17
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v7, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v7, v9
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v15, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v14, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v10, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v15, v16
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v19, v18
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v17
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v14, v15
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v12
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v10
+; NOMAD-NEXT:    v_addc_u32_e64 v10, s[4:5], v6, v8, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v11, -1, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v12, s8, v5
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v13
+; NOMAD-NEXT:    v_addc_u32_e64 v13, s[6:7], v7, v9, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[6:7], v7, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v9, s8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, -1, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v15, s8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v16, s8, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v17, v10, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v5, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v19, s8, v13
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v16
+; NOMAD-NEXT:    v_mul_lo_u32 v16, v13, v9
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[6:7], v14, v19
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v13, v9
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v12
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[6:7], v14, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v5, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v4, v12
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v16, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v15, v19
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v10, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v5, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v10, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v4, v12
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[8:9], v17, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[8:9], v15, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[8:9], v11, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[8:9], v14, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[8:9], v8, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[6:7], v9, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[6:7], v17, v14
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v15, v18
+; NOMAD-NEXT:    v_add_i32_e64 v16, s[6:7], v16, v19
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v12
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[6:7], v8, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[6:7], v9, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[6:7], v15, v12
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v14
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[6:7], v10, v12
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v13, v11
+; NOMAD-NEXT:    v_addc_u32_e32 v6, vcc, v6, v10, vcc
+; NOMAD-NEXT:    v_addc_u32_e64 v7, vcc, v7, v11, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v6, vcc, 0, v6, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v1, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v1, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v0, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v1, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v0, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v6, v1, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v2, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v16, v3, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v2, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v7, v3, v7
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v13, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v16, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v13, v10
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v15, v9
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v16, v11
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v10, s10, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v12, 0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v13, s10, v5
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v11, s10, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, 0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v15, s10, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v7, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s10, v6
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, 1, v5
+; NOMAD-NEXT:    v_addc_u32_e32 v16, vcc, 0, v6, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v17, s10, v7
+; NOMAD-NEXT:    v_add_i32_e32 v18, vcc, 1, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v19, vcc, 0, v7, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v14, v17
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, 1, v9
+; NOMAD-NEXT:    v_addc_u32_e32 v17, vcc, 0, v16, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v13
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v15
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, 1, v18
+; NOMAD-NEXT:    v_addc_u32_e32 v15, vcc, 0, v19, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v10
+; NOMAD-NEXT:    v_subb_u32_e64 v10, s[4:5], v1, v8, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v8
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s10, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v2, s[4:5], v2, v11
+; NOMAD-NEXT:    v_subb_u32_e64 v11, s[6:7], v3, v12, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v3, s[6:7], v3, v12
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[6:7], 0, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[6:7]
+; NOMAD-NEXT:    v_cmp_eq_u32_e64 s[6:7], 0, v10
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[8:9], s10, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[8:9]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, v12, v8, s[6:7]
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc
+; NOMAD-NEXT:    v_subbrev_u32_e64 v3, vcc, 0, v3, s[4:5]
+; NOMAD-NEXT:    v_subrev_i32_e32 v0, vcc, s10, v0
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s10, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v11
+; NOMAD-NEXT:    v_cndmask_b32_e32 v10, v12, v10, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v2, vcc, s10, v2
+; NOMAD-NEXT:    v_subbrev_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s10, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v11, v0, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v12, v2, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v9, v14, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v1, v18, v13, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v16, v17, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, v19, v15, s[4:5]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, v4, v1, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v6, v3, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v3, v7, v5, s[4:5]
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = udiv <2 x i64> %num, <i64 4096, i64 4096>
   ret <2 x i64> %result
 }
 
 define i64 @v_udiv_i64_oddk_denom(i64 %num) {
 ; CHECK-LABEL: v_udiv_i64_oddk_denom:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
 ; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v8, v6, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e32 v1, v9, v10, vcc
 ; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_udiv_i64_oddk_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_mov_b32 s6, 0x12d8fb
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v2, 0
+; NOMAD-NEXT:    s_mov_b32 s7, 0xffed2705
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v3, s6
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
+; NOMAD-NEXT:    v_add_f32_e32 v2, v2, v3
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; NOMAD-NEXT:    v_trunc_f32_e32 v3, v3
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0xcf800000, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; NOMAD-NEXT:    v_add_f32_e32 v2, v4, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s7, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, -1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s7, v2
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v3, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v2, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v9, v8
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
+; NOMAD-NEXT:    v_addc_u32_e64 v5, s[4:5], v3, v4, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v3, s[4:5], v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, -1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s7, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v5, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v5, v4
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v5, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v2, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v5, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v9, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v8, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v8, v9
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v6
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, v3, v5, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v1, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v0, v3
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v1, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v0, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v1, v3
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v7, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s6, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, 0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s6, v2
+; NOMAD-NEXT:    v_add_i32_e32 v3, vcc, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s6, v3
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, 1, v2
+; NOMAD-NEXT:    v_addc_u32_e32 v9, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, 1, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v10, vcc, 0, v9, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v5
+; NOMAD-NEXT:    v_subb_u32_e64 v5, s[4:5], v1, v4, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v4
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s6, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v0, vcc, s6, v0
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s6, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v8, v6, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v9, v10, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = udiv i64 %num, 1235195
   ret i64 %result
 }
 
 define <2 x i64> @v_udiv_v2i64_oddk_denom(<2 x i64> %num) {
 ; GISEL-LABEL: v_udiv_v2i64_oddk_denom:
 ; GISEL:       ; %bb.0:
 ; GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

 ; CGP-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
 ; CGP-NEXT:    v_cndmask_b32_e32 v2, v4, v1, vcc
 ; CGP-NEXT:    v_cndmask_b32_e64 v1, v19, v15, s[4:5]
 ; CGP-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v10
 ; CGP-NEXT:    v_cndmask_b32_e64 v0, v5, v0, s[4:5]
 ; CGP-NEXT:    v_cndmask_b32_e64 v1, v7, v1, s[4:5]
 ; CGP-NEXT:    v_cndmask_b32_e32 v3, v6, v3, vcc
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_udiv_v2i64_oddk_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_mov_b32 s10, 0x12d8fb
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, 0
+; NOMAD-NEXT:    s_mov_b32 s8, 0xffed2705
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v5, s10
+; NOMAD-NEXT:    v_mul_f32_e32 v6, 0x4f800000, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x4f800000, v4
+; NOMAD-NEXT:    v_add_f32_e32 v6, v6, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v4, v5
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v5, v6
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x5f7ffffc, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v6, 0x2f800000, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v7, 0x2f800000, v4
+; NOMAD-NEXT:    v_trunc_f32_e32 v6, v6
+; NOMAD-NEXT:    v_trunc_f32_e32 v7, v7
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0xcf800000, v6
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v6, v6
+; NOMAD-NEXT:    v_mul_f32_e32 v9, 0xcf800000, v7
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v7, v7
+; NOMAD-NEXT:    v_add_f32_e32 v5, v8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s8, v6
+; NOMAD-NEXT:    v_add_f32_e32 v4, v9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, s8, v7
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v11, -1, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v12, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v13, s8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, -1, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v15, s8, v4
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v11, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v6, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v5, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v6, v10
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v14, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v7, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v4, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v7, v13
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v6, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v5, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v6, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v19, v4, v9
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v17
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v7, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v7, v9
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v15, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v14, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v10, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v15, v16
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v19, v18
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v17
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v14, v15
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v12
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v10
+; NOMAD-NEXT:    v_addc_u32_e64 v10, s[4:5], v6, v8, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v11, -1, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v12, s8, v5
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v13
+; NOMAD-NEXT:    v_addc_u32_e64 v13, s[6:7], v7, v9, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[6:7], v7, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v9, s8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, -1, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v15, s8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v16, s8, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v17, v10, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v5, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v19, s8, v13
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v16
+; NOMAD-NEXT:    v_mul_lo_u32 v16, v13, v9
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[6:7], v14, v19
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v13, v9
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v12
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[6:7], v14, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v5, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v4, v12
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v16, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v15, v19
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v10, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v5, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v10, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v4, v12
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[8:9], v17, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[8:9], v15, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[8:9], v11, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[8:9], v14, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[8:9], v8, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[6:7], v9, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[6:7], v17, v14
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v15, v18
+; NOMAD-NEXT:    v_add_i32_e64 v16, s[6:7], v16, v19
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v12
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[6:7], v8, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[6:7], v9, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[6:7], v15, v12
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v14
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[6:7], v10, v12
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v13, v11
+; NOMAD-NEXT:    v_addc_u32_e32 v6, vcc, v6, v10, vcc
+; NOMAD-NEXT:    v_addc_u32_e64 v7, vcc, v7, v11, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v6, vcc, 0, v6, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v1, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v1, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v0, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v1, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v0, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v6, v1, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v2, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v16, v3, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v2, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v7, v3, v7
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v13, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v16, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v13, v10
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v15, v9
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v16, v11
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v10, s10, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v12, 0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v13, s10, v5
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v11, s10, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, 0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v15, s10, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v7, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s10, v6
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, 1, v5
+; NOMAD-NEXT:    v_addc_u32_e32 v16, vcc, 0, v6, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v17, s10, v7
+; NOMAD-NEXT:    v_add_i32_e32 v18, vcc, 1, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v19, vcc, 0, v7, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v14, v17
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, 1, v9
+; NOMAD-NEXT:    v_addc_u32_e32 v17, vcc, 0, v16, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v13
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v15
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, 1, v18
+; NOMAD-NEXT:    v_addc_u32_e32 v15, vcc, 0, v19, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v10
+; NOMAD-NEXT:    v_subb_u32_e64 v10, s[4:5], v1, v8, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v8
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s10, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v2, s[4:5], v2, v11
+; NOMAD-NEXT:    v_subb_u32_e64 v11, s[6:7], v3, v12, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v3, s[6:7], v3, v12
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[6:7], 0, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[6:7]
+; NOMAD-NEXT:    v_cmp_eq_u32_e64 s[6:7], 0, v10
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[8:9], s10, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[8:9]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, v12, v8, s[6:7]
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc
+; NOMAD-NEXT:    v_subbrev_u32_e64 v3, vcc, 0, v3, s[4:5]
+; NOMAD-NEXT:    v_subrev_i32_e32 v0, vcc, s10, v0
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s10, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v11
+; NOMAD-NEXT:    v_cndmask_b32_e32 v10, v12, v10, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v2, vcc, s10, v2
+; NOMAD-NEXT:    v_subbrev_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s10, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v11, v0, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v12, v2, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v9, v14, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v1, v18, v13, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v16, v17, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, v19, v15, s[4:5]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, v4, v1, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v6, v3, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v3, v7, v5, s[4:5]
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = udiv <2 x i64> %num, <i64 1235195, i64 1235195>
   ret <2 x i64> %result
 }
 
 define i64 @v_udiv_i64_pow2_shl_denom(i64 %x, i64 %y) {
 ; CHECK-LABEL: v_udiv_i64_pow2_shl_denom:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

 ; CHECK-NEXT:    v_cndmask_b32_e64 v0, v1, v3, s[4:5]
 ; CHECK-NEXT:    v_cndmask_b32_e32 v2, v5, v0, vcc
 ; CHECK-NEXT:    v_mov_b32_e32 v3, 0
 ; CHECK-NEXT:  BB7_4:
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; CHECK-NEXT:    v_mov_b32_e32 v0, v2
 ; CHECK-NEXT:    v_mov_b32_e32 v1, v3
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_udiv_i64_pow2_shl_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_movk_i32 s4, 0x1000
+; NOMAD-NEXT:    s_mov_b32 s5, 0
+; NOMAD-NEXT:    v_mov_b32_e32 v6, 0
+; NOMAD-NEXT:    v_lshl_b64 v[4:5], s[4:5], v2
+; NOMAD-NEXT:    v_or_b32_e32 v7, v1, v5
+; NOMAD-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[6:7]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, 1
+; NOMAD-NEXT:    s_xor_b64 s[4:5], vcc, s[4:5]
+; NOMAD-NEXT:    ; implicit-def: $vgpr2_vgpr3
+; NOMAD-NEXT:    s_and_saveexec_b64 s[6:7], s[4:5]
+; NOMAD-NEXT:    s_xor_b64 s[6:7], exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB7_2
+; NOMAD-NEXT:  ; %bb.1:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v2, v4
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v3, v5
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, 0, v4
+; NOMAD-NEXT:    v_subb_u32_e32 v7, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_mul_f32_e32 v3, 0x4f800000, v3
+; NOMAD-NEXT:    v_add_f32_e32 v2, v3, v2
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; NOMAD-NEXT:    v_trunc_f32_e32 v3, v3
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0xcf800000, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; NOMAD-NEXT:    v_add_f32_e32 v2, v8, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v6, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v7, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v6, v2
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v3, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v2, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v3, v9
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v2, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v3, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v2, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v3, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v13, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v13, v12
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v9
+; NOMAD-NEXT:    v_addc_u32_e64 v9, s[4:5], v3, v8, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v3, s[4:5], v3, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v7, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v6, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v6, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v9, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v2, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v9, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v9, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v2, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v6, v9, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v11, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v10, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v7, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v8, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v9, v7
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v10, v11
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v8, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v9, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, v3, v6, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v7
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, v0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v1, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v0, v3
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v1, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v1, v3
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v9, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v4, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v5, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v4, v2
+; NOMAD-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v4, v3
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, 1, v2
+; NOMAD-NEXT:    v_addc_u32_e32 v11, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, 1, v10
+; NOMAD-NEXT:    v_addc_u32_e32 v12, vcc, 0, v11, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v9
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, v0, v7
+; NOMAD-NEXT:    v_subb_u32_e64 v9, s[4:5], v1, v6, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v6
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v1, vcc, v1, v5, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v6, v13, v6, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, v7, v4
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v7, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v1, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v9, v7, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v10, v8, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v11, v12, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v2, v1, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; NOMAD-NEXT:  BB7_2: ; %Flow
+; NOMAD-NEXT:    s_or_saveexec_b64 s[6:7], s[6:7]
+; NOMAD-NEXT:    s_xor_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB7_4
+; NOMAD-NEXT:  ; %bb.3:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v1, v4
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v2, v1, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v1, v4
+; NOMAD-NEXT:    v_sub_i32_e32 v5, vcc, 0, v2
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v2, v1
+; NOMAD-NEXT:    v_add_i32_e64 v3, s[4:5], v1, v2
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v2
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v1, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v2, v1, v4
+; NOMAD-NEXT:    v_add_i32_e32 v3, vcc, 1, v1
+; NOMAD-NEXT:    v_subrev_i32_e32 v5, vcc, 1, v1
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, v0, v2
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v0, v2
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v6, v4
+; NOMAD-NEXT:    s_and_b64 s[4:5], s[4:5], vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, v1, v3, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v5, v0, vcc
+; NOMAD-NEXT:    v_mov_b32_e32 v3, 0
+; NOMAD-NEXT:  BB7_4:
+; NOMAD-NEXT:    s_or_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    v_mov_b32_e32 v0, v2
+; NOMAD-NEXT:    v_mov_b32_e32 v1, v3
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %shl.y = shl i64 4096, %y
   %r = udiv i64 %x, %shl.y
   ret i64 %r
 }
 
 define <2 x i64> @v_udiv_v2i64_pow2_shl_denom(<2 x i64> %x, <2 x i64> %y) {
 ; GISEL-LABEL: v_udiv_v2i64_pow2_shl_denom:
 ; GISEL:       ; %bb.0:

 ; CGP-NEXT:    v_cndmask_b32_e64 v2, v3, v5, s[4:5]
 ; CGP-NEXT:    v_cndmask_b32_e32 v4, v6, v2, vcc
 ; CGP-NEXT:    v_mov_b32_e32 v5, 0
 ; CGP-NEXT:  BB8_8:
 ; CGP-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; CGP-NEXT:    v_mov_b32_e32 v2, v4
 ; CGP-NEXT:    v_mov_b32_e32 v3, v5
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_udiv_v2i64_pow2_shl_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    v_mov_b32_e32 v5, v0
+; NOMAD-NEXT:    v_mov_b32_e32 v7, v1
+; NOMAD-NEXT:    s_movk_i32 s4, 0x1000
+; NOMAD-NEXT:    s_mov_b32 s5, 0
+; NOMAD-NEXT:    v_mov_b32_e32 v0, 0
+; NOMAD-NEXT:    v_lshl_b64 v[10:11], s[4:5], v4
+; NOMAD-NEXT:    v_lshl_b64 v[8:9], s[4:5], v6
+; NOMAD-NEXT:    v_or_b32_e32 v1, v7, v11
+; NOMAD-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[0:1]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, 1
+; NOMAD-NEXT:    s_xor_b64 s[4:5], vcc, s[4:5]
+; NOMAD-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; NOMAD-NEXT:    s_and_saveexec_b64 s[6:7], s[4:5]
+; NOMAD-NEXT:    s_xor_b64 s[6:7], exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB8_2
+; NOMAD-NEXT:  ; %bb.1:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v0, v10
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v1, v11
+; NOMAD-NEXT:    v_sub_i32_e32 v4, vcc, 0, v10
+; NOMAD-NEXT:    v_subb_u32_e32 v6, vcc, 0, v11, vcc
+; NOMAD-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; NOMAD-NEXT:    v_add_f32_e32 v0, v1, v0
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; NOMAD-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; NOMAD-NEXT:    v_trunc_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_f32_e32 v12, 0xcf800000, v1
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; NOMAD-NEXT:    v_add_f32_e32 v0, v12, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v4, v1
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v4, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v6, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v4, v0
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v14, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v1, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v0, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v1, v13
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v0, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v17, v1, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v0, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v1, v12
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v17, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v15, v14
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v17, v16
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v15, v14
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v14
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v13
+; NOMAD-NEXT:    v_addc_u32_e64 v13, s[4:5], v1, v12, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v1, s[4:5], v1, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v4, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v6, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v4, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v0, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v13, v12
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v6, v4
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v14
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v0, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v13, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v13, v4
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v15, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[4:5], v14, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[4:5], v12, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v13, v6
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v14, v15
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v12, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[4:5], v13, v12
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v12
+; NOMAD-NEXT:    v_addc_u32_e32 v1, vcc, v1, v4, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v6
+; NOMAD-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v7, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v6, v5, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v0, v7, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v5, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v7, v1
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v5, v1
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v7, v1
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v13, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v12, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v13, v6
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v10, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v11, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v10, v0
+; NOMAD-NEXT:    v_add_i32_e32 v1, vcc, v1, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v10, v1
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, 1, v0
+; NOMAD-NEXT:    v_addc_u32_e32 v15, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v12, v4
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, 1, v14
+; NOMAD-NEXT:    v_addc_u32_e32 v16, vcc, 0, v15, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v13
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, v5, v6
+; NOMAD-NEXT:    v_subb_u32_e64 v13, s[4:5], v7, v4, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v4, s[4:5], v7, v4
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v6, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v13, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v4, vcc, v4, v11, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v13, v11
+; NOMAD-NEXT:    v_cndmask_b32_e32 v7, v17, v7, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, v6, v10
+; NOMAD-NEXT:    v_subbrev_u32_e32 v4, vcc, 0, v4, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v6, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v4, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v11
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v13, v6, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v14, v12, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v6, v15, v16, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v7
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc
+; NOMAD-NEXT:  BB8_2: ; %Flow2
+; NOMAD-NEXT:    s_or_saveexec_b64 s[6:7], s[6:7]
+; NOMAD-NEXT:    s_xor_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB8_4
+; NOMAD-NEXT:  ; %bb.3:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v0, v10
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v1, v0, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v0, v10
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, 0, v1
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v1, v0
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v0, v1
+; NOMAD-NEXT:    v_sub_i32_e64 v0, s[4:5], v0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v0, v0, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v1, v0, v10
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, 1, v0
+; NOMAD-NEXT:    v_subrev_i32_e32 v6, vcc, 1, v0
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, v5, v1
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v5, v1
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v10
+; NOMAD-NEXT:    s_and_b64 s[4:5], s[4:5], vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
+; NOMAD-NEXT:    v_mov_b32_e32 v1, 0
+; NOMAD-NEXT:  BB8_4:
+; NOMAD-NEXT:    s_or_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    v_or_b32_e32 v5, v3, v9
+; NOMAD-NEXT:    v_mov_b32_e32 v4, 0
+; NOMAD-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[4:5]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, 1
+; NOMAD-NEXT:    s_xor_b64 s[4:5], vcc, s[4:5]
+; NOMAD-NEXT:    ; implicit-def: $vgpr4_vgpr5
+; NOMAD-NEXT:    s_and_saveexec_b64 s[6:7], s[4:5]
+; NOMAD-NEXT:    s_xor_b64 s[6:7], exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB8_6
+; NOMAD-NEXT:  ; %bb.5:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, v8
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v5, v9
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, 0, v8
+; NOMAD-NEXT:    v_subb_u32_e32 v7, vcc, 0, v9, vcc
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x4f800000, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v5, v4
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v4
+; NOMAD-NEXT:    v_trunc_f32_e32 v5, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v10, 0xcf800000, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v10, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v6, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v7, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v12, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v5, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v5, v11
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v4, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v5, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v4, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v5, v10
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v15, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v15, v14
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v11
+; NOMAD-NEXT:    v_addc_u32_e64 v11, s[4:5], v5, v10, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v7, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v6, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v11, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v11, v10
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v4, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v11, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v4, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v6, v11, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v13, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v12, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v7, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v10, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v11, v7
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v12, v13
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v10, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v11, v10
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_addc_u32_e32 v5, vcc, v5, v6, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v7, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v2, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v3, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v2, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v5
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v11, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v10, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v11, v7
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v9, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v8, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v8, v5
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, 1, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v13, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v10, v6
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, 1, v12
+; NOMAD-NEXT:    v_addc_u32_e32 v14, vcc, 0, v13, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v11
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, v2, v7
+; NOMAD-NEXT:    v_subb_u32_e64 v11, s[4:5], v3, v6, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v3, s[4:5], v3, v6
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v11, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v3, vcc, v3, v9, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v11, v9
+; NOMAD-NEXT:    v_cndmask_b32_e32 v6, v15, v6, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, v7, v8
+; NOMAD-NEXT:    v_subbrev_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v7, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v3, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v9
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v11, v7, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v12, v10, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v7, v13, v14, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v6
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v4, v3, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v5, v7, vcc
+; NOMAD-NEXT:  BB8_6: ; %Flow
+; NOMAD-NEXT:    s_or_saveexec_b64 s[6:7], s[6:7]
+; NOMAD-NEXT:    s_xor_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB8_8
+; NOMAD-NEXT:  ; %bb.7:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v3, v8
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v3, v3
+; NOMAD-NEXT:    v_mul_f32_e32 v3, 0x4f800000, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v3, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v8
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, 0, v4
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v4, v6, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v4, v3
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v3, v4
+; NOMAD-NEXT:    v_sub_i32_e64 v3, s[4:5], v3, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v3, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v3, v8
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, 1, v3
+; NOMAD-NEXT:    v_subrev_i32_e32 v6, vcc, 1, v3
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, v2, v4
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v2, v4
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v8
+; NOMAD-NEXT:    s_and_b64 s[4:5], s[4:5], vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, v3, v5, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v6, v2, vcc
+; NOMAD-NEXT:    v_mov_b32_e32 v5, 0
+; NOMAD-NEXT:  BB8_8:
+; NOMAD-NEXT:    s_or_b64 exec, exec, s[6:7]
+; NOMAD-NEXT:    v_mov_b32_e32 v2, v4
+; NOMAD-NEXT:    v_mov_b32_e32 v3, v5
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %shl.y = shl <2 x i64> <i64 4096, i64 4096>, %y
   %r = udiv <2 x i64> %x, %shl.y
   ret <2 x i64> %r
 }
 
 define i64 @v_udiv_i64_24bit(i64 %num, i64 %den) {
 ; GISEL-LABEL: v_udiv_i64_24bit:
 ; GISEL:       ; %bb.0:

 ; CGP-NEXT:    s_mov_b32 s4, 0xffffff
 ; CGP-NEXT:    v_and_b32_e32 v0, s4, v0
 ; CGP-NEXT:    v_and_b32_e32 v1, s4, v2
 ; CGP-NEXT:    v_cvt_f32_u32_e32 v0, v0
 ; CGP-NEXT:    v_cvt_f32_u32_e32 v1, v1
 ; CGP-NEXT:    v_rcp_f32_e32 v2, v1
 ; CGP-NEXT:    v_mul_f32_e32 v2, v0, v2
 ; CGP-NEXT:    v_trunc_f32_e32 v2, v2
-; CGP-NEXT:    v_fma_f32 v0, -v2, v1, v0
+; CGP-NEXT:    v_mad_f32 v0, -v2, v1, v0
 ; CGP-NEXT:    v_cvt_u32_f32_e32 v2, v2
 ; CGP-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v0|, v1
 ; CGP-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
 ; CGP-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
 ; CGP-NEXT:    v_and_b32_e32 v0, 0xffffff, v0
 ; CGP-NEXT:    v_mov_b32_e32 v1, 0
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_udiv_i64_24bit:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_mov_b32 s4, 0xffffff
+; NOMAD-NEXT:    v_and_b32_e32 v0, s4, v0
+; NOMAD-NEXT:    v_and_b32_e32 v1, s4, v2
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v0, v0
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v1, v1
+; NOMAD-NEXT:    v_rcp_f32_e32 v2, v1
+; NOMAD-NEXT:    v_mul_f32_e32 v2, v0, v2
+; NOMAD-NEXT:    v_trunc_f32_e32 v2, v2
+; NOMAD-NEXT:    v_fma_f32 v0, -v2, v1, v0
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; NOMAD-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v0|, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
+; NOMAD-NEXT:    v_and_b32_e32 v0, 0xffffff, v0
+; NOMAD-NEXT:    v_mov_b32_e32 v1, 0
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %num.mask = and i64 %num, 16777215
   %den.mask = and i64 %den, 16777215
   %result = udiv i64 %num.mask, %den.mask
   ret i64 %result
 }
 
 define <2 x i64> @v_udiv_v2i64_24bit(<2 x i64> %num, <2 x i64> %den) {
 ; GISEL-LABEL: v_udiv_v2i64_24bit:

 ; CGP-NEXT:    v_cvt_f32_u32_e32 v2, v2
 ; CGP-NEXT:    v_cvt_f32_u32_e32 v4, v4
 ; CGP-NEXT:    v_rcp_f32_e32 v5, v3
 ; CGP-NEXT:    v_rcp_f32_e32 v6, v4
 ; CGP-NEXT:    v_mul_f32_e32 v5, v0, v5
 ; CGP-NEXT:    v_mul_f32_e32 v6, v2, v6
 ; CGP-NEXT:    v_trunc_f32_e32 v5, v5
 ; CGP-NEXT:    v_trunc_f32_e32 v6, v6
-; CGP-NEXT:    v_fma_f32 v0, -v5, v3, v0
+; CGP-NEXT:    v_mad_f32 v0, -v5, v3, v0
 ; CGP-NEXT:    v_cvt_u32_f32_e32 v5, v5
-; CGP-NEXT:    v_fma_f32 v2, -v6, v4, v2
+; CGP-NEXT:    v_mad_f32 v2, -v6, v4, v2
 ; CGP-NEXT:    v_cvt_u32_f32_e32 v6, v6
 ; CGP-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v0|, v3
 ; CGP-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
 ; CGP-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v2|, v4
 ; CGP-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[4:5]
 ; CGP-NEXT:    v_add_i32_e32 v0, vcc, v5, v0
 ; CGP-NEXT:    v_add_i32_e32 v2, vcc, v6, v2
 ; CGP-NEXT:    v_and_b32_e32 v0, s6, v0
 ; CGP-NEXT:    v_and_b32_e32 v2, s6, v2
 ; CGP-NEXT:    v_mov_b32_e32 v3, v1
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_udiv_v2i64_24bit:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_mov_b32 s6, 0xffffff
+; NOMAD-NEXT:    v_mov_b32_e32 v1, 0
+; NOMAD-NEXT:    v_and_b32_e32 v0, s6, v0
+; NOMAD-NEXT:    v_and_b32_e32 v2, s6, v2
+; NOMAD-NEXT:    v_and_b32_e32 v3, s6, v4
+; NOMAD-NEXT:    v_and_b32_e32 v4, s6, v6
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v0, v0
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v3, v3
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v2, v2
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, v4
+; NOMAD-NEXT:    v_rcp_f32_e32 v5, v3
+; NOMAD-NEXT:    v_rcp_f32_e32 v6, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, v0, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v6, v2, v6
+; NOMAD-NEXT:    v_trunc_f32_e32 v5, v5
+; NOMAD-NEXT:    v_trunc_f32_e32 v6, v6
+; NOMAD-NEXT:    v_fma_f32 v0, -v5, v3, v0
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_fma_f32 v2, -v6, v4, v2
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v6, v6
+; NOMAD-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v0|, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v2|, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v5, v0
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v6, v2
+; NOMAD-NEXT:    v_and_b32_e32 v0, s6, v0
+; NOMAD-NEXT:    v_and_b32_e32 v2, s6, v2
+; NOMAD-NEXT:    v_mov_b32_e32 v3, v1
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %num.mask = and <2 x i64> %num, <i64 16777215, i64 16777215>
   %den.mask = and <2 x i64> %den, <i64 16777215, i64 16777215>
   %result = udiv <2 x i64> %num.mask, %den.mask
   ret <2 x i64> %result
 }

llvm/test/CodeGen/AMDGPU/GlobalISel/urem.i64.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -mtriple=amdgcn-amd-amdhsa -denormal-fp-math-f32=preserve-sign < %s | FileCheck -check-prefixes=CHECK,GISEL %s
-; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=0 -mtriple=amdgcn-amd-amdhsa -denormal-fp-math-f32=preserve-sign < %s | FileCheck -check-prefixes=CHECK,CGP %s
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -mtriple=amdgcn-amd-amdhsa -denormal-fp-math-f32=preserve-sign -mattr=+mad-mac-f32-insts < %s | FileCheck -check-prefixes=CHECK,GISEL %s
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=0 -mtriple=amdgcn-amd-amdhsa -denormal-fp-math-f32=preserve-sign -mattr=+mad-mac-f32-insts < %s | FileCheck -check-prefixes=CHECK,CGP %s
+; RUN: llc -global-isel -amdgpu-codegenprepare-disable-idiv-expansion=1 -mtriple=amdgcn-amd-amdhsa -denormal-fp-math-f32=preserve-sign -mattr=-mad-mac-f32-insts < %s | FileCheck -check-prefixes=NOMAD %s
 
 ; The same 32-bit expansion is implemented in the legalizer and in AMDGPUCodeGenPrepare.
 
 define i64 @v_urem_i64(i64 %num, i64 %den) {
 ; CHECK-LABEL: v_urem_i64:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    v_or_b32_e32 v5, v1, v3

 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e64 v4, v4, v0, s[4:5]
 ; CHECK-NEXT:    v_mov_b32_e32 v5, 0
 ; CHECK-NEXT:  BB0_4:
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[8:9]
 ; CHECK-NEXT:    v_mov_b32_e32 v0, v4
 ; CHECK-NEXT:    v_mov_b32_e32 v1, v5
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_urem_i64:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    v_or_b32_e32 v5, v1, v3
+; NOMAD-NEXT:    v_mov_b32_e32 v4, 0
+; NOMAD-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[4:5]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, 1
+; NOMAD-NEXT:    s_xor_b64 s[4:5], vcc, s[4:5]
+; NOMAD-NEXT:    ; implicit-def: $vgpr4_vgpr5
+; NOMAD-NEXT:    s_and_saveexec_b64 s[6:7], s[4:5]
+; NOMAD-NEXT:    s_xor_b64 s[6:7], exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB0_2
+; NOMAD-NEXT:  ; %bb.1:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, v2
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v5, v3
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, 0, v2
+; NOMAD-NEXT:    v_subb_u32_e32 v7, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x4f800000, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v5, v4
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v4
+; NOMAD-NEXT:    v_trunc_f32_e32 v5, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0xcf800000, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v7, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v5, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v5, v9
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v4, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v5, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v5, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v13, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v13, v12
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_addc_u32_e64 v9, s[4:5], v5, v8, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v7, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v6, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v9, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v4, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v9, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v4, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v9, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v4, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v6, v9, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v11, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v10, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v7, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v8, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v9, v7
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v10, v11
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v8, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v9, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v5, vcc, v5, v6, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v1, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v7, v0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v1, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v0, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v1, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v1, v5
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v9, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v2, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v5, v2, v5
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v8, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; NOMAD-NEXT:    v_sub_i32_e32 v5, vcc, v0, v7
+; NOMAD-NEXT:    v_subb_u32_e64 v6, s[4:5], v1, v4, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v4
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v5, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v6, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v1, vcc, v1, v3, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v6, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, v5, v2
+; NOMAD-NEXT:    v_subbrev_u32_e64 v8, s[4:5], 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v1, vcc, v1, v3, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v8, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v11, vcc, v7, v2
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v8, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v10, v9, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v7, v11, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v5, v3, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v6, v1, vcc
+; NOMAD-NEXT:  BB0_2: ; %Flow
+; NOMAD-NEXT:    s_or_saveexec_b64 s[8:9], s[6:7]
+; NOMAD-NEXT:    s_xor_b64 exec, exec, s[8:9]
+; NOMAD-NEXT:    s_cbranch_execz BB0_4
+; NOMAD-NEXT:  ; %bb.3:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v1, v2
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v3, v1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v1, v2
+; NOMAD-NEXT:    v_sub_i32_e32 v5, vcc, 0, v3
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v3, v1
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v1, v3
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v1, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v1, v1, v2
+; NOMAD-NEXT:    v_sub_i32_e32 v3, vcc, v0, v1
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v3, v2
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v3, v2
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v1
+; NOMAD-NEXT:    v_sub_i32_e64 v0, s[6:7], v3, v2
+; NOMAD-NEXT:    s_and_b64 vcc, vcc, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, v4, v0, s[4:5]
+; NOMAD-NEXT:    v_mov_b32_e32 v5, 0
+; NOMAD-NEXT:  BB0_4:
+; NOMAD-NEXT:    s_or_b64 exec, exec, s[8:9]
+; NOMAD-NEXT:    v_mov_b32_e32 v0, v4
+; NOMAD-NEXT:    v_mov_b32_e32 v1, v5
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = urem i64 %num, %den
   ret i64 %result
 }
 
 ; FIXME: This is a workaround for not handling uniform VGPR case.
 declare i32 @llvm.amdgcn.readfirstlane(i32)
 
 define amdgpu_ps i64 @s_urem_i64(i64 inreg %num, i64 inreg %den) {

 ; CHECK-NEXT:    v_subrev_i32_e64 v0, s[2:3], s2, v1
 ; CHECK-NEXT:    s_and_b64 vcc, vcc, s[0:1]
 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
 ; CHECK-NEXT:  BB1_4:
 ; CHECK-NEXT:    v_readfirstlane_b32 s0, v0
 ; CHECK-NEXT:    s_mov_b32 s1, s0
 ; CHECK-NEXT:    ; return to shader part epilog
+;
+; NOMAD-LABEL: s_urem_i64:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_or_b64 s[4:5], s[0:1], s[2:3]
+; NOMAD-NEXT:    s_mov_b32 s6, 0
+; NOMAD-NEXT:    s_mov_b32 s7, -1
+; NOMAD-NEXT:    s_and_b64 s[4:5], s[4:5], s[6:7]
+; NOMAD-NEXT:    v_cmp_eq_u64_e64 s[4:5], s[4:5], 0
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[6:7], 0, 1
+; NOMAD-NEXT:    s_xor_b64 vcc, s[4:5], s[6:7]
+; NOMAD-NEXT:    s_mov_b32 s4, 1
+; NOMAD-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; NOMAD-NEXT:    s_cbranch_vccz BB1_2
+; NOMAD-NEXT:  ; %bb.1:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v0, s2
+; NOMAD-NEXT:    v_mov_b32_e32 v1, s3
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v2, s3
+; NOMAD-NEXT:    s_sub_u32 s6, 0, s2
+; NOMAD-NEXT:    s_cselect_b32 s4, 1, 0
+; NOMAD-NEXT:    v_mov_b32_e32 v3, s1
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
+; NOMAD-NEXT:    s_and_b32 s4, s4, 1
+; NOMAD-NEXT:    v_add_f32_e32 v0, v2, v0
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; NOMAD-NEXT:    s_cmp_lg_u32 s4, 0
+; NOMAD-NEXT:    s_subb_u32 s7, 0, s3
+; NOMAD-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v0
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x2f800000, v0
+; NOMAD-NEXT:    v_trunc_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0xcf800000, v2
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; NOMAD-NEXT:    v_add_f32_e32 v0, v4, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s6, v2
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s6, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v6, s7, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s6, v0
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v2, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v2, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v0, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v2, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v9, v8
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v5
+; NOMAD-NEXT:    v_addc_u32_e64 v5, s[4:5], v2, v4, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v2, s[4:5], v2, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s6, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v6, s7, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s6, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s6, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v5, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v5, v4
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v0, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v5, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v0, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v5, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v9, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v8, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v8, v9
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v6
+; NOMAD-NEXT:    v_addc_u32_e32 v2, vcc, v2, v5, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s1, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v5, s0, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v0, s1, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v6, s0, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v7, s1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v8, s0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, s1, v2
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v7, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s2, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v6, s3, v0
+; NOMAD-NEXT:    v_mul_hi_u32 v0, s2, v0
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v2, s2, v2
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v6, v2
+; NOMAD-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
+; NOMAD-NEXT:    v_sub_i32_e32 v2, vcc, s0, v5
+; NOMAD-NEXT:    v_subb_u32_e64 v3, s[4:5], v3, v0, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v0, s[4:5], s1, v0
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s2, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s3, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v0, vcc, v0, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v5, v4, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v3, vcc, s2, v2
+; NOMAD-NEXT:    v_subbrev_u32_e32 v0, vcc, 0, v0, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s2, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v5, vcc, s2, v3
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s3, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, s3, v0
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v6, v4, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v3, v5, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
+; NOMAD-NEXT:    s_mov_b32 s4, 0
+; NOMAD-NEXT:  BB1_2: ; %Flow
+; NOMAD-NEXT:    s_and_b32 s1, s4, 1
+; NOMAD-NEXT:    s_cmp_lg_u32 s1, 0
+; NOMAD-NEXT:    s_cbranch_scc0 BB1_4
+; NOMAD-NEXT:  ; %bb.3:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v0, s2
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_f32_e32 v0, 0x4f800000, v0
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v1, v0, s2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v0, s2
+; NOMAD-NEXT:    v_sub_i32_e32 v3, vcc, 0, v1
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v2
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v1, v0
+; NOMAD-NEXT:    v_add_i32_e64 v2, s[4:5], v0, v1
+; NOMAD-NEXT:    v_sub_i32_e64 v0, s[4:5], v0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v0, v0, s0
+; NOMAD-NEXT:    v_mul_lo_u32 v0, v0, s2
+; NOMAD-NEXT:    v_sub_i32_e32 v1, vcc, s0, v0
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s2, v1
+; NOMAD-NEXT:    v_add_i32_e64 v2, s[4:5], s2, v1
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[0:1], s0, v0
+; NOMAD-NEXT:    v_subrev_i32_e64 v0, s[2:3], s2, v1
+; NOMAD-NEXT:    s_and_b64 vcc, vcc, s[0:1]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, v2, v0, s[0:1]
+; NOMAD-NEXT:  BB1_4:
+; NOMAD-NEXT:    v_readfirstlane_b32 s0, v0
+; NOMAD-NEXT:    s_mov_b32 s1, s0
+; NOMAD-NEXT:    ; return to shader part epilog
   %result = urem i64 %num, %den
   %cast = bitcast i64 %result to <2 x i32>
   %elt.0 = extractelement <2 x i32> %cast, i32 0
   %elt.1 = extractelement <2 x i32> %cast, i32 1
   %res.0 = call i32 @llvm.amdgcn.readfirstlane(i32 %elt.0)
   %res.1 = call i32 @llvm.amdgcn.readfirstlane(i32 %elt.1)
   %ins.0 = insertelement <2 x i32> undef, i32 %res.0, i32 0
   %ins.1 = insertelement <2 x i32> %ins.0, i32 %res.0, i32 1

 ; CGP-NEXT:    v_cndmask_b32_e32 v2, v4, v2, vcc
 ; CGP-NEXT:    v_cndmask_b32_e64 v4, v5, v2, s[4:5]
 ; CGP-NEXT:    v_mov_b32_e32 v5, 0
 ; CGP-NEXT:  BB2_8:
 ; CGP-NEXT:    s_or_b64 exec, exec, s[8:9]
 ; CGP-NEXT:    v_mov_b32_e32 v2, v4
 ; CGP-NEXT:    v_mov_b32_e32 v3, v5
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_urem_v2i64:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v8, v4
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v9, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v9, 0x4f800000, v9
+; NOMAD-NEXT:    v_add_f32_e32 v8, v9, v8
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v8, v8
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0x5f7ffffc, v8
+; NOMAD-NEXT:    v_mul_f32_e32 v9, 0x2f800000, v8
+; NOMAD-NEXT:    v_trunc_f32_e32 v9, v9
+; NOMAD-NEXT:    v_mul_f32_e32 v10, 0xcf800000, v9
+; NOMAD-NEXT:    v_add_f32_e32 v8, v10, v8
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v8, v8
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v9, v9
+; NOMAD-NEXT:    v_sub_i32_e32 v10, vcc, 0, v4
+; NOMAD-NEXT:    v_subb_u32_e32 v11, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v11, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v10, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v10, v8
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v14
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v9, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v8, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v8, v12
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v15, v14
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v9, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v9, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v8, v13
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v15, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v15, v16
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v15, v14
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v9, v13
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v14
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_addc_u32_e64 v12, s[4:5], v9, v13, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v9, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v11, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v10, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v10, v8
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v14
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v11, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v12, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v8, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v8, v13
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v14, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v12, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v12, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v8, v10
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v14, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v13, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[4:5], v14, v15
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v13, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v14, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v12, v10
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v10, v13
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v8, v11
+; NOMAD-NEXT:    v_addc_u32_e32 v9, vcc, v9, v10, vcc
+; NOMAD-NEXT:    v_addc_u32_e64 v9, vcc, 0, v9, s[4:5]
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v1, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v0, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v0, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v1, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v1, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v0, v9
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v11, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v12
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v1, v9
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v4, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v4, v8
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v10
+; NOMAD-NEXT:    v_subb_u32_e64 v9, s[4:5], v1, v8, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v8
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_eq_u32_e64 s[4:5], v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, v8, v10, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v10, s[4:5], v0, v4
+; NOMAD-NEXT:    v_subb_u32_e32 v1, vcc, v1, v5, vcc
+; NOMAD-NEXT:    v_subbrev_u32_e64 v11, vcc, 0, v1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v11, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v10, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v11, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v12, v12, v13, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v4, vcc, v10, v4
+; NOMAD-NEXT:    v_subb_u32_e64 v1, s[4:5], v1, v5, s[4:5]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v10, v4, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v11, v1, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, v6
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v5, v7
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x4f800000, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v5, v4
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v4
+; NOMAD-NEXT:    v_trunc_f32_e32 v5, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0xcf800000, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v8, v4
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_sub_i32_e32 v8, vcc, 0, v6
+; NOMAD-NEXT:    v_subb_u32_e32 v9, vcc, 0, v7, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v8, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v8, v4
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v12
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v5, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v4, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v10
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v5, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v5, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v11
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v13, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v14
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v5, v11
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v12
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v10
+; NOMAD-NEXT:    v_addc_u32_e64 v10, s[4:5], v5, v11, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v8, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v8, v4
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v9, v12
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v9, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v10, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v4, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v4, v11
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v9, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v9, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v12, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v10, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v10, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v4, v8
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v12, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[4:5], v12, v13
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v11, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v12, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v10, v8
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v8, v11
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v9
+; NOMAD-NEXT:    v_addc_u32_e32 v5, vcc, v5, v8, vcc
+; NOMAD-NEXT:    v_addc_u32_e64 v5, vcc, 0, v5, s[4:5]
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v2, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v4
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v9, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v5
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v7, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v5, v6, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; NOMAD-NEXT:    v_sub_i32_e32 v2, vcc, v2, v8
+; NOMAD-NEXT:    v_subb_u32_e64 v5, s[4:5], v3, v4, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v3, s[4:5], v3, v4
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v5, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v2, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_eq_u32_e64 s[4:5], v5, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, v4, v8, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v8, s[4:5], v2, v6
+; NOMAD-NEXT:    v_subb_u32_e32 v3, vcc, v3, v7, vcc
+; NOMAD-NEXT:    v_subbrev_u32_e64 v9, vcc, 0, v3, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v9, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v8, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v9, v7
+; NOMAD-NEXT:    v_cndmask_b32_e32 v10, v10, v11, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, v8, v6
+; NOMAD-NEXT:    v_subb_u32_e64 v3, s[4:5], v3, v7, s[4:5]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v10
+; NOMAD-NEXT:    v_cndmask_b32_e32 v6, v8, v6, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v9, v3, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v5, v3, vcc
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = urem <2 x i64> %num, %den
   ret <2 x i64> %result
 }
 
 define i64 @v_urem_i64_pow2k_denom(i64 %num) {
 ; CHECK-LABEL: v_urem_i64_pow2k_denom:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

 ; CHECK-NEXT:    v_cndmask_b32_e32 v5, v6, v5, vcc
 ; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v5
 ; CHECK-NEXT:    v_cndmask_b32_e32 v4, v4, v7, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e32 v1, v1, v8, vcc
 ; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_urem_i64_pow2k_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_movk_i32 s6, 0x1000
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v2, 0
+; NOMAD-NEXT:    s_mov_b32 s7, 0xfffff000
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v3, s6
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
+; NOMAD-NEXT:    v_add_f32_e32 v2, v2, v3
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; NOMAD-NEXT:    v_trunc_f32_e32 v3, v3
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0xcf800000, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; NOMAD-NEXT:    v_add_f32_e32 v2, v4, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s7, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, -1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s7, v2
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v3, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v2, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v9, v8
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
+; NOMAD-NEXT:    v_addc_u32_e64 v5, s[4:5], v3, v4, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v3, s[4:5], v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, -1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s7, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v5, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v5, v4
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v5, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v2, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v5, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v9, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v8, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v8, v9
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v6
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, v3, v5, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v1, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v0, v3
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v1, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v0, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v1, v3
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v7, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s6, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, 0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, s6, v2
+; NOMAD-NEXT:    v_add_i32_e32 v3, vcc, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v3, s6, v3
+; NOMAD-NEXT:    v_add_i32_e32 v3, vcc, v6, v3
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v5
+; NOMAD-NEXT:    v_subb_u32_e64 v3, s[4:5], v1, v2, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v2
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s6, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v4, v2, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v4, vcc, s6, v0
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s6, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v7, vcc, s6, v4
+; NOMAD-NEXT:    v_subbrev_u32_e32 v8, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v6, v5, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v4, v7, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v8, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = urem i64 %num, 4096
   ret i64 %result
 }
 
 define <2 x i64> @v_urem_v2i64_pow2k_denom(<2 x i64> %num) {
 ; GISEL-LABEL: v_urem_v2i64_pow2k_denom:
 ; GISEL:       ; %bb.0:
 ; GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

 ; CGP-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
 ; CGP-NEXT:    v_cndmask_b32_e32 v2, v2, v8, vcc
 ; CGP-NEXT:    v_cndmask_b32_e64 v1, v1, v17, s[4:5]
 ; CGP-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v5
 ; CGP-NEXT:    v_cndmask_b32_e64 v0, v0, v9, s[4:5]
 ; CGP-NEXT:    v_cndmask_b32_e64 v1, v7, v1, s[4:5]
 ; CGP-NEXT:    v_cndmask_b32_e32 v3, v6, v3, vcc
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_urem_v2i64_pow2k_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_movk_i32 s10, 0x1000
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, 0
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v5, s10
+; NOMAD-NEXT:    v_mul_f32_e32 v6, 0x4f800000, v4
+; NOMAD-NEXT:    s_sub_u32 s8, 0, s10
+; NOMAD-NEXT:    s_cselect_b32 s4, 1, 0
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x4f800000, v4
+; NOMAD-NEXT:    v_add_f32_e32 v6, v6, v5
+; NOMAD-NEXT:    s_and_b32 s4, s4, 1
+; NOMAD-NEXT:    v_add_f32_e32 v4, v4, v5
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v5, v6
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x5f7ffffc, v5
+; NOMAD-NEXT:    s_cmp_lg_u32 s4, 0
+; NOMAD-NEXT:    s_subb_u32 s6, 0, 0
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v6, 0x2f800000, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v7, 0x2f800000, v4
+; NOMAD-NEXT:    s_sub_u32 s9, 0, s10
+; NOMAD-NEXT:    s_cselect_b32 s4, 1, 0
+; NOMAD-NEXT:    v_trunc_f32_e32 v6, v6
+; NOMAD-NEXT:    v_trunc_f32_e32 v7, v7
+; NOMAD-NEXT:    s_and_b32 s4, s4, 1
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0xcf800000, v6
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v6, v6
+; NOMAD-NEXT:    v_mul_f32_e32 v9, 0xcf800000, v7
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v7, v7
+; NOMAD-NEXT:    v_add_f32_e32 v5, v8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s8, v6
+; NOMAD-NEXT:    v_add_f32_e32 v4, v9, v4
+; NOMAD-NEXT:    s_cmp_lg_u32 s4, 0
+; NOMAD-NEXT:    s_subb_u32 s11, 0, 0
+; NOMAD-NEXT:    v_mul_lo_u32 v9, s9, v7
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v11, s6, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v12, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v13, s9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, s11, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v15, s9, v4
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v11, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v6, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v5, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v6, v10
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v14, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v7, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v4, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v7, v13
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v6, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v5, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v6, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v19, v4, v9
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v17
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v7, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v7, v9
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v15, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v14, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v10, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v15, v16
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v19, v18
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v17
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v14, v15
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v12
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v10
+; NOMAD-NEXT:    v_addc_u32_e64 v10, s[4:5], v6, v8, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v11, s6, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v12, s8, v5
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v13
+; NOMAD-NEXT:    v_addc_u32_e64 v13, s[6:7], v7, v9, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[6:7], v7, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v9, s9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, s11, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v15, s9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v16, s8, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v17, v10, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v5, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v19, s9, v13
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v16
+; NOMAD-NEXT:    v_mul_lo_u32 v16, v13, v9
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[6:7], v14, v19
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v13, v9
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v12
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[6:7], v14, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v5, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v4, v12
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v16, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v15, v19
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v10, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v5, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v10, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v4, v12
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[8:9], v17, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[8:9], v15, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[8:9], v11, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[8:9], v14, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[8:9], v8, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[6:7], v9, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[6:7], v17, v14
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v15, v18
+; NOMAD-NEXT:    v_add_i32_e64 v16, s[6:7], v16, v19
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v12
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[6:7], v8, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[6:7], v9, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[6:7], v15, v12
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v14
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[6:7], v10, v12
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v13, v11
+; NOMAD-NEXT:    v_addc_u32_e32 v6, vcc, v6, v10, vcc
+; NOMAD-NEXT:    v_addc_u32_e64 v7, vcc, v7, v11, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v6, vcc, 0, v6, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v1, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v1, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v0, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v1, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v0, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v6, v1, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v2, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v16, v3, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v2, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v7, v3, v7
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v13, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v16, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v13, v10
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v15, v9
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v16, v11
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v10, s10, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v12, 0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, s10, v5
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v11, s10, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v13, 0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, s10, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v7, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v6, s10, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v7, s10, v7
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v12, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v13, v7
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v7, v4
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v10
+; NOMAD-NEXT:    v_subb_u32_e64 v6, s[4:5], v1, v5, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v5
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s10, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v2, s[4:5], v2, v11
+; NOMAD-NEXT:    v_subb_u32_e64 v7, s[6:7], v3, v4, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v3, s[6:7], v3, v4
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[6:7], s10, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[6:7]
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[6:7], 0, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[6:7]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, -1, vcc
+; NOMAD-NEXT:    v_subbrev_u32_e64 v3, vcc, 0, v3, s[4:5]
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v6
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v8, vcc, s10, v0
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s10, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v7
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v9, v4, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v9, vcc, s10, v2
+; NOMAD-NEXT:    v_subbrev_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s10, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v13, vcc, s10, v8
+; NOMAD-NEXT:    v_subbrev_u32_e32 v14, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, -1, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v16, vcc, s10, v9
+; NOMAD-NEXT:    v_subbrev_u32_e32 v17, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v10, v12, v10, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v11, v15, v11, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v10
+; NOMAD-NEXT:    v_cndmask_b32_e32 v8, v8, v13, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, v9, v16, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v14, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v3, v3, v17, s[4:5]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, v2, v9, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v3, v7, v3, s[4:5]
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = urem <2 x i64> %num, <i64 4096, i64 4096>
   ret <2 x i64> %result
 }
 
 define i64 @v_urem_i64_oddk_denom(i64 %num) {
 ; CHECK-LABEL: v_urem_i64_oddk_denom:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

 ; CHECK-NEXT:    v_cndmask_b32_e32 v5, v6, v5, vcc
 ; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v5
 ; CHECK-NEXT:    v_cndmask_b32_e32 v4, v4, v7, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e32 v1, v1, v8, vcc
 ; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_urem_i64_oddk_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_mov_b32 s6, 0x12d8fb
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v2, 0
+; NOMAD-NEXT:    s_mov_b32 s7, 0xffed2705
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v3, s6
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
+; NOMAD-NEXT:    v_add_f32_e32 v2, v2, v3
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; NOMAD-NEXT:    v_trunc_f32_e32 v3, v3
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0xcf800000, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; NOMAD-NEXT:    v_add_f32_e32 v2, v4, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s7, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, -1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s7, v2
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v3, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v2, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v9, v8
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v5
+; NOMAD-NEXT:    v_addc_u32_e64 v5, s[4:5], v3, v4, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v3, s[4:5], v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v4, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, -1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, s7, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s7, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v5, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v5, v4
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v5, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v2, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v5, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v9, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v8, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v8, v9
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v6
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, v3, v5, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v4, v1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v1, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v0, v3
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v1, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v0, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v1, v3
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v7, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v7, v5
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v5, s6, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, 0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, s6, v2
+; NOMAD-NEXT:    v_add_i32_e32 v3, vcc, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v3, s6, v3
+; NOMAD-NEXT:    v_add_i32_e32 v3, vcc, v6, v3
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v5
+; NOMAD-NEXT:    v_subb_u32_e64 v3, s[4:5], v1, v2, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v2
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s6, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v4, v2, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v4, vcc, s6, v0
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s6, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, -1, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v7, vcc, s6, v4
+; NOMAD-NEXT:    v_subbrev_u32_e32 v8, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v6, v5, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v4, v7, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v8, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = urem i64 %num, 1235195
   ret i64 %result
 }
 
 define <2 x i64> @v_urem_v2i64_oddk_denom(<2 x i64> %num) {
 ; GISEL-LABEL: v_urem_v2i64_oddk_denom:
 ; GISEL:       ; %bb.0:
 ; GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

 ; CGP-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
 ; CGP-NEXT:    v_cndmask_b32_e32 v2, v2, v8, vcc
 ; CGP-NEXT:    v_cndmask_b32_e64 v1, v1, v17, s[4:5]
 ; CGP-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v5
 ; CGP-NEXT:    v_cndmask_b32_e64 v0, v0, v9, s[4:5]
 ; CGP-NEXT:    v_cndmask_b32_e64 v1, v7, v1, s[4:5]
 ; CGP-NEXT:    v_cndmask_b32_e32 v3, v6, v3, vcc
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_urem_v2i64_oddk_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_mov_b32 s10, 0x12d8fb
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, 0
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v5, s10
+; NOMAD-NEXT:    v_mul_f32_e32 v6, 0x4f800000, v4
+; NOMAD-NEXT:    s_sub_u32 s8, 0, s10
+; NOMAD-NEXT:    s_cselect_b32 s4, 1, 0
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x4f800000, v4
+; NOMAD-NEXT:    v_add_f32_e32 v6, v6, v5
+; NOMAD-NEXT:    s_and_b32 s4, s4, 1
+; NOMAD-NEXT:    v_add_f32_e32 v4, v4, v5
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v5, v6
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x5f7ffffc, v5
+; NOMAD-NEXT:    s_cmp_lg_u32 s4, 0
+; NOMAD-NEXT:    s_subb_u32 s6, 0, 0
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v6, 0x2f800000, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v7, 0x2f800000, v4
+; NOMAD-NEXT:    s_sub_u32 s9, 0, s10
+; NOMAD-NEXT:    s_cselect_b32 s4, 1, 0
+; NOMAD-NEXT:    v_trunc_f32_e32 v6, v6
+; NOMAD-NEXT:    v_trunc_f32_e32 v7, v7
+; NOMAD-NEXT:    s_and_b32 s4, s4, 1
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0xcf800000, v6
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v6, v6
+; NOMAD-NEXT:    v_mul_f32_e32 v9, 0xcf800000, v7
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v7, v7
+; NOMAD-NEXT:    v_add_f32_e32 v5, v8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s8, v6
+; NOMAD-NEXT:    v_add_f32_e32 v4, v9, v4
+; NOMAD-NEXT:    s_cmp_lg_u32 s4, 0
+; NOMAD-NEXT:    s_subb_u32 s11, 0, 0
+; NOMAD-NEXT:    v_mul_lo_u32 v9, s9, v7
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v10, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v11, s6, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v12, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v13, s9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, s11, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v15, s9, v4
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v11, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v6, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v5, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v6, v10
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v14, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v7, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v4, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v7, v13
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v6, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v5, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v6, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v19, v4, v9
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v17
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v7, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v7, v9
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v15, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v14, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v10, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v15, v16
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v19, v18
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v17
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v12, v11
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v14, v15
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v12
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v10
+; NOMAD-NEXT:    v_addc_u32_e64 v10, s[4:5], v6, v8, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, s8, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v11, s6, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v12, s8, v5
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v13
+; NOMAD-NEXT:    v_addc_u32_e64 v13, s[6:7], v7, v9, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[6:7], v7, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v9, s9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v14, s11, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v15, s9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v16, s8, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v17, v10, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v5, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v19, s9, v13
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v16
+; NOMAD-NEXT:    v_mul_lo_u32 v16, v13, v9
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[6:7], v14, v19
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v13, v9
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v12
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[6:7], v14, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v5, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v4, v12
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v16, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v15, v19
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v10, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v5, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v10, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v4, v12
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[8:9], v17, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[8:9], v15, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[8:9], v11, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[8:9], v14, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[8:9], v8, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[6:7], v9, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[6:7], v17, v14
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v15, v18
+; NOMAD-NEXT:    v_add_i32_e64 v16, s[6:7], v16, v19
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v12
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[6:7], v8, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[6:7], v9, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[6:7], v15, v12
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v14
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[6:7], v10, v12
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v13, v11
+; NOMAD-NEXT:    v_addc_u32_e32 v6, vcc, v6, v10, vcc
+; NOMAD-NEXT:    v_addc_u32_e64 v7, vcc, v7, v11, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v6, vcc, 0, v6, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v1, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v1, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v2, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v0, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v1, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v0, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v6, v1, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v2, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v16, v3, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v2, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v7, v3, v7
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v13, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v16, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v13, v10
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v15, v9
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v16, v11
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v10, s10, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v12, 0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, s10, v5
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v11, s10, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v13, 0, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v4, s10, v4
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v7, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v6, s10, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v7, s10, v7
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v12, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v13, v7
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v7, v4
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v10
+; NOMAD-NEXT:    v_subb_u32_e64 v6, s[4:5], v1, v5, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v5
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[4:5], s10, v0
+; NOMAD-NEXT:    v_cndmask_b32_e64 v5, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v2, s[4:5], v2, v11
+; NOMAD-NEXT:    v_subb_u32_e64 v7, s[6:7], v3, v4, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v3, s[6:7], v3, v4
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[6:7], s10, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[6:7]
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[6:7], 0, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[6:7]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, -1, vcc
+; NOMAD-NEXT:    v_subbrev_u32_e64 v3, vcc, 0, v3, s[4:5]
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v6
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v8, vcc, s10, v0
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s10, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v7
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v9, v4, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v9, vcc, s10, v2
+; NOMAD-NEXT:    v_subbrev_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, s10, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v13, vcc, s10, v8
+; NOMAD-NEXT:    v_subbrev_u32_e32 v14, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, -1, vcc
+; NOMAD-NEXT:    v_subrev_i32_e32 v16, vcc, s10, v9
+; NOMAD-NEXT:    v_subbrev_u32_e32 v17, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v10, v12, v10, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v11, v15, v11, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v10
+; NOMAD-NEXT:    v_cndmask_b32_e32 v8, v8, v13, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, v9, v16, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v14, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v3, v3, v17, s[4:5]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, v2, v9, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v3, v7, v3, s[4:5]
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %result = urem <2 x i64> %num, <i64 1235195, i64 1235195>
   ret <2 x i64> %result
 }
 
 define i64 @v_urem_i64_pow2_shl_denom(i64 %x, i64 %y) {
 ; CHECK-LABEL: v_urem_i64_pow2_shl_denom:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

 ; CHECK-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
 ; CHECK-NEXT:    v_cndmask_b32_e64 v2, v3, v0, s[4:5]
 ; CHECK-NEXT:    v_mov_b32_e32 v3, 0
 ; CHECK-NEXT:  BB7_4:
 ; CHECK-NEXT:    s_or_b64 exec, exec, s[8:9]
 ; CHECK-NEXT:    v_mov_b32_e32 v0, v2
 ; CHECK-NEXT:    v_mov_b32_e32 v1, v3
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_urem_i64_pow2_shl_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_movk_i32 s4, 0x1000
+; NOMAD-NEXT:    s_mov_b32 s5, 0
+; NOMAD-NEXT:    v_mov_b32_e32 v6, 0
+; NOMAD-NEXT:    v_lshl_b64 v[4:5], s[4:5], v2
+; NOMAD-NEXT:    v_or_b32_e32 v7, v1, v5
+; NOMAD-NEXT:    v_cmp_eq_u64_e32 vcc, 0, v[6:7]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, 1
+; NOMAD-NEXT:    s_xor_b64 s[4:5], vcc, s[4:5]
+; NOMAD-NEXT:    ; implicit-def: $vgpr2_vgpr3
+; NOMAD-NEXT:    s_and_saveexec_b64 s[6:7], s[4:5]
+; NOMAD-NEXT:    s_xor_b64 s[6:7], exec, s[6:7]
+; NOMAD-NEXT:    s_cbranch_execz BB7_2
+; NOMAD-NEXT:  ; %bb.1:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v2, v4
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v3, v5
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, 0, v4
+; NOMAD-NEXT:    v_subb_u32_e32 v7, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_mul_f32_e32 v3, 0x4f800000, v3
+; NOMAD-NEXT:    v_add_f32_e32 v2, v3, v2
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
+; NOMAD-NEXT:    v_trunc_f32_e32 v3, v3
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0xcf800000, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v3, v3
+; NOMAD-NEXT:    v_add_f32_e32 v2, v8, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v3
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v6, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v7, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v6, v2
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v3, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v2, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v3, v9
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v2, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v3, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v2, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v3, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v13, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v13, v12
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v9
+; NOMAD-NEXT:    v_addc_u32_e64 v9, s[4:5], v3, v8, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v3, s[4:5], v3, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v7, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v6, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v6, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v9, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v2, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v9, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v7, v6
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v2, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v9, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v2, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v6, v9, v6
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v11, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v10, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v7, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v8, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v9, v7
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v10, v11
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[4:5], v8, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v9, v8
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, v3, v6, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v7
+; NOMAD-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v1, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v7, v0, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v1, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v0, v3
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v1, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v0, v3
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v1, v3
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v9, v2
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v8, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v9, v7
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v2, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v7, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v4, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v5, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v4, v2
+; NOMAD-NEXT:    v_add_i32_e32 v3, vcc, v3, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v3, v4, v3
+; NOMAD-NEXT:    v_add_i32_e32 v3, vcc, v8, v3
+; NOMAD-NEXT:    v_add_i32_e32 v2, vcc, v3, v2
+; NOMAD-NEXT:    v_sub_i32_e32 v3, vcc, v0, v7
+; NOMAD-NEXT:    v_subb_u32_e64 v6, s[4:5], v1, v2, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v2
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v3, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v6, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v1, vcc, v1, v5, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v6, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v7, v2, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, v3, v4
+; NOMAD-NEXT:    v_subbrev_u32_e64 v8, s[4:5], 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v7, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_subb_u32_e32 v1, vcc, v1, v5, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v8, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v11, vcc, v7, v4
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v8, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v10, v9, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v7, v11, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v3, v5, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v6, v1, vcc
+; NOMAD-NEXT:  BB7_2: ; %Flow
+; NOMAD-NEXT:    s_or_saveexec_b64 s[8:9], s[6:7]
+; NOMAD-NEXT:    s_xor_b64 exec, exec, s[8:9]
+; NOMAD-NEXT:    s_cbranch_execz BB7_4
+; NOMAD-NEXT:  ; %bb.3:
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v1, v4
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v2, v1, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v1, v4
+; NOMAD-NEXT:    v_sub_i32_e32 v5, vcc, 0, v2
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v2, v1
+; NOMAD-NEXT:    v_add_i32_e64 v3, s[4:5], v1, v2
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v2
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v1, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v1, v1, v4
+; NOMAD-NEXT:    v_sub_i32_e32 v2, vcc, v0, v1
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v2, v4
+; NOMAD-NEXT:    v_add_i32_e64 v3, s[4:5], v2, v4
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v1
+; NOMAD-NEXT:    v_sub_i32_e64 v0, s[6:7], v2, v4
+; NOMAD-NEXT:    s_and_b64 vcc, vcc, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, v3, v0, s[4:5]
+; NOMAD-NEXT:    v_mov_b32_e32 v3, 0
+; NOMAD-NEXT:  BB7_4:
+; NOMAD-NEXT:    s_or_b64 exec, exec, s[8:9]
+; NOMAD-NEXT:    v_mov_b32_e32 v0, v2
+; NOMAD-NEXT:    v_mov_b32_e32 v1, v3
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %shl.y = shl i64 4096, %y
   %r = urem i64 %x, %shl.y
   ret i64 %r
 }
 
 define <2 x i64> @v_urem_v2i64_pow2_shl_denom(<2 x i64> %x, <2 x i64> %y) {
 ; GISEL-LABEL: v_urem_v2i64_pow2_shl_denom:
 ; GISEL:       ; %bb.0:

 ; CGP-NEXT:    v_cndmask_b32_e32 v2, v4, v2, vcc
 ; CGP-NEXT:    v_cndmask_b32_e64 v4, v5, v2, s[4:5]
 ; CGP-NEXT:    v_mov_b32_e32 v5, 0
 ; CGP-NEXT:  BB8_8:
 ; CGP-NEXT:    s_or_b64 exec, exec, s[8:9]
 ; CGP-NEXT:    v_mov_b32_e32 v2, v4
 ; CGP-NEXT:    v_mov_b32_e32 v3, v5
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_urem_v2i64_pow2_shl_denom:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_movk_i32 s4, 0x1000
+; NOMAD-NEXT:    s_mov_b32 s5, 0
+; NOMAD-NEXT:    v_lshl_b64 v[4:5], s[4:5], v4
+; NOMAD-NEXT:    v_lshl_b64 v[6:7], s[4:5], v6
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v8, v4
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v9, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v9, 0x4f800000, v9
+; NOMAD-NEXT:    v_add_f32_e32 v8, v9, v8
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v8, v8
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0x5f7ffffc, v8
+; NOMAD-NEXT:    v_mul_f32_e32 v9, 0x2f800000, v8
+; NOMAD-NEXT:    v_trunc_f32_e32 v9, v9
+; NOMAD-NEXT:    v_mul_f32_e32 v10, 0xcf800000, v9
+; NOMAD-NEXT:    v_add_f32_e32 v8, v10, v8
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v8, v8
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v9, v9
+; NOMAD-NEXT:    v_sub_i32_e32 v10, vcc, 0, v4
+; NOMAD-NEXT:    v_subb_u32_e32 v11, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v11, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v10, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v10, v8
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v14
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v15
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v9, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v8, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v8, v12
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v15, v14
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v9, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v9, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v8, v13
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v15, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v15, v16
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v15, v14
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v9, v13
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v14
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_addc_u32_e64 v12, s[4:5], v9, v13, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v9, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v10, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v11, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v10, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v10, v8
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v14
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v11, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v12, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v8, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v8, v13
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v14, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v12, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v12, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v8, v10
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v14, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v13, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[4:5], v14, v15
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v13, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[4:5], v14, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v12, v10
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[4:5], v10, v13
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v8, v11
+; NOMAD-NEXT:    v_addc_u32_e32 v9, vcc, v9, v10, vcc
+; NOMAD-NEXT:    v_addc_u32_e64 v9, vcc, 0, v9, s[4:5]
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v1, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v0, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v0, v8
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v1, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v1, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v0, v9
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v11, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v12
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v1, v9
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v4, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v4, v9
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v4, v8
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v11, v9
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v10
+; NOMAD-NEXT:    v_subb_u32_e64 v9, s[4:5], v1, v8, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[4:5], v1, v8
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_eq_u32_e64 s[4:5], v9, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, v8, v10, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v10, s[4:5], v0, v4
+; NOMAD-NEXT:    v_subb_u32_e32 v1, vcc, v1, v5, vcc
+; NOMAD-NEXT:    v_subbrev_u32_e64 v11, vcc, 0, v1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v11, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v10, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v11, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v12, v12, v13, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v4, vcc, v10, v4
+; NOMAD-NEXT:    v_subb_u32_e64 v1, s[4:5], v1, v5, s[4:5]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; NOMAD-NEXT:    v_cndmask_b32_e32 v4, v10, v4, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v11, v1, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v8
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v4, v6
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v5, v7
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x4f800000, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v5, v4
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v4, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v4, 0x5f7ffffc, v4
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x2f800000, v4
+; NOMAD-NEXT:    v_trunc_f32_e32 v5, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v8, 0xcf800000, v5
+; NOMAD-NEXT:    v_add_f32_e32 v4, v8, v4
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v4, v4
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_sub_i32_e32 v8, vcc, 0, v6
+; NOMAD-NEXT:    v_subb_u32_e32 v9, vcc, 0, v7, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v8, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v8, v4
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v12
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v5, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v4, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v10
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v5, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v5, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v4, v11
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v13, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v14
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v13, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v5, v11
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v11, v12
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v10
+; NOMAD-NEXT:    v_addc_u32_e64 v10, s[4:5], v5, v11, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v5, s[4:5], v5, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v8, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v9, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v8, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v8, v4
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v9, v12
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v9, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v10, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v4, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v4, v11
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v9, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v9, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v12, v9
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v10, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v10, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v4, v8
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v12, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v11, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[4:5], v12, v13
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[4:5], v11, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[4:5], v12, v11
+; NOMAD-NEXT:    v_mul_hi_u32 v8, v10, v8
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[4:5], v8, v11
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v4, v9
+; NOMAD-NEXT:    v_addc_u32_e32 v5, vcc, v5, v8, vcc
+; NOMAD-NEXT:    v_addc_u32_e64 v5, vcc, 0, v5, s[4:5]
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v3, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v2, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v4
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v3, v4
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v2, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v9, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v9, v10
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v4, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v5
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v6, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v7, v4
+; NOMAD-NEXT:    v_mul_lo_u32 v5, v6, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v6, v4
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
+; NOMAD-NEXT:    v_add_i32_e32 v4, vcc, v5, v4
+; NOMAD-NEXT:    v_sub_i32_e32 v2, vcc, v2, v8
+; NOMAD-NEXT:    v_subb_u32_e64 v5, s[4:5], v3, v4, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v3, s[4:5], v3, v4
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v5, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v2, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_cmp_eq_u32_e64 s[4:5], v5, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, v4, v8, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v8, s[4:5], v2, v6
+; NOMAD-NEXT:    v_subb_u32_e32 v3, vcc, v3, v7, vcc
+; NOMAD-NEXT:    v_subbrev_u32_e64 v9, vcc, 0, v3, s[4:5]
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v9, v7
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v8, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, v9, v7
+; NOMAD-NEXT:    v_cndmask_b32_e32 v10, v10, v11, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v6, vcc, v8, v6
+; NOMAD-NEXT:    v_subb_u32_e64 v3, s[4:5], v3, v7, s[4:5]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v3, vcc, 0, v3, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v10
+; NOMAD-NEXT:    v_cndmask_b32_e32 v6, v8, v6, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v9, v3, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v5, v3, vcc
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %shl.y = shl <2 x i64> <i64 4096, i64 4096>, %y
   %r = urem <2 x i64> %x, %shl.y
   ret <2 x i64> %r
 }
 
 define i64 @v_urem_i64_24bit(i64 %num, i64 %den) {
 ; GISEL-LABEL: v_urem_i64_24bit:
 ; GISEL:       ; %bb.0:

 ; CGP-NEXT:    s_mov_b32 s4, 0xffffff
 ; CGP-NEXT:    v_and_b32_e32 v0, s4, v0
 ; CGP-NEXT:    v_and_b32_e32 v1, s4, v2
 ; CGP-NEXT:    v_cvt_f32_u32_e32 v2, v0
 ; CGP-NEXT:    v_cvt_f32_u32_e32 v3, v1
 ; CGP-NEXT:    v_rcp_f32_e32 v4, v3
 ; CGP-NEXT:    v_mul_f32_e32 v4, v2, v4
 ; CGP-NEXT:    v_trunc_f32_e32 v4, v4
-; CGP-NEXT:    v_fma_f32 v2, -v4, v3, v2
+; CGP-NEXT:    v_mad_f32 v2, -v4, v3, v2
 ; CGP-NEXT:    v_cvt_u32_f32_e32 v4, v4
 ; CGP-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v2|, v3
 ; CGP-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s[4:5]
 ; CGP-NEXT:    v_add_i32_e32 v2, vcc, v4, v2
 ; CGP-NEXT:    v_mul_lo_u32 v1, v2, v1
 ; CGP-NEXT:    v_sub_i32_e32 v0, vcc, v0, v1
 ; CGP-NEXT:    v_and_b32_e32 v0, 0xffffff, v0
 ; CGP-NEXT:    v_mov_b32_e32 v1, 0
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_urem_i64_24bit:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_mov_b32 s4, 0xffffff
+; NOMAD-NEXT:    v_and_b32_e32 v0, s4, v0
+; NOMAD-NEXT:    v_and_b32_e32 v1, s4, v2
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v2, v1
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_f32_e32 v2, 0x4f800000, v2
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v2, v2
+; NOMAD-NEXT:    v_mul_lo_u32 v3, v2, v1
+; NOMAD-NEXT:    v_mul_hi_u32 v4, v2, v1
+; NOMAD-NEXT:    v_sub_i32_e32 v5, vcc, 0, v3
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v4
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v3, v3, v2
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v2, v3
+; NOMAD-NEXT:    v_sub_i32_e64 v2, s[4:5], v2, v3
+; NOMAD-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc
+; NOMAD-NEXT:    v_mul_hi_u32 v2, v2, v0
+; NOMAD-NEXT:    v_mul_lo_u32 v2, v2, v1
+; NOMAD-NEXT:    v_sub_i32_e32 v3, vcc, v0, v2
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v3, v1
+; NOMAD-NEXT:    v_add_i32_e64 v4, s[4:5], v3, v1
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v2
+; NOMAD-NEXT:    v_sub_i32_e64 v0, s[6:7], v3, v1
+; NOMAD-NEXT:    s_and_b64 vcc, vcc, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v0, v4, v0, s[4:5]
+; NOMAD-NEXT:    v_mov_b32_e32 v1, 0
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %num.mask = and i64 %num, 16777215
   %den.mask = and i64 %den, 16777215
   %result = urem i64 %num.mask, %den.mask
   ret i64 %result
 }
 
 define <2 x i64> @v_urem_v2i64_24bit(<2 x i64> %num, <2 x i64> %den) {
 ; GISEL-LABEL: v_urem_v2i64_24bit:

 ; CGP-NEXT:    v_cvt_f32_u32_e32 v7, v2
 ; CGP-NEXT:    v_cvt_f32_u32_e32 v8, v4
 ; CGP-NEXT:    v_rcp_f32_e32 v9, v6
 ; CGP-NEXT:    v_rcp_f32_e32 v10, v8
 ; CGP-NEXT:    v_mul_f32_e32 v9, v5, v9
 ; CGP-NEXT:    v_mul_f32_e32 v10, v7, v10
 ; CGP-NEXT:    v_trunc_f32_e32 v9, v9
 ; CGP-NEXT:    v_trunc_f32_e32 v10, v10
-; CGP-NEXT:    v_fma_f32 v5, -v9, v6, v5
+; CGP-NEXT:    v_mad_f32 v5, -v9, v6, v5
 ; CGP-NEXT:    v_cvt_u32_f32_e32 v9, v9
-; CGP-NEXT:    v_fma_f32 v7, -v10, v8, v7
+; CGP-NEXT:    v_mad_f32 v7, -v10, v8, v7
 ; CGP-NEXT:    v_cvt_u32_f32_e32 v10, v10
 ; CGP-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v5|, v6
 ; CGP-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s[4:5]
 ; CGP-NEXT:    v_cmp_ge_f32_e64 s[4:5], |v7|, v8
 ; CGP-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s[4:5]
 ; CGP-NEXT:    v_add_i32_e32 v5, vcc, v9, v5
 ; CGP-NEXT:    v_add_i32_e32 v6, vcc, v10, v6
 ; CGP-NEXT:    v_mul_lo_u32 v3, v5, v3
 ; CGP-NEXT:    v_mul_lo_u32 v4, v6, v4
 ; CGP-NEXT:    v_sub_i32_e32 v0, vcc, v0, v3
 ; CGP-NEXT:    v_sub_i32_e32 v2, vcc, v2, v4
 ; CGP-NEXT:    v_and_b32_e32 v0, s6, v0
 ; CGP-NEXT:    v_and_b32_e32 v2, s6, v2
 ; CGP-NEXT:    v_mov_b32_e32 v3, v1
 ; CGP-NEXT:    s_setpc_b64 s[30:31]
+;
+; NOMAD-LABEL: v_urem_v2i64_24bit:
+; NOMAD:       ; %bb.0:
+; NOMAD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; NOMAD-NEXT:    s_mov_b32 s6, 0xffffff
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v1, 0
+; NOMAD-NEXT:    v_and_b32_e32 v3, s6, v4
+; NOMAD-NEXT:    v_and_b32_e32 v4, s6, v6
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x4f800000, v1
+; NOMAD-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v6, v3
+; NOMAD-NEXT:    v_sub_i32_e32 v7, vcc, 0, v3
+; NOMAD-NEXT:    v_subb_u32_e64 v8, s[4:5], 0, 0, vcc
+; NOMAD-NEXT:    v_cvt_f32_u32_e32 v9, v4
+; NOMAD-NEXT:    v_sub_i32_e32 v10, vcc, 0, v4
+; NOMAD-NEXT:    v_subb_u32_e64 v11, s[4:5], 0, 0, vcc
+; NOMAD-NEXT:    v_add_f32_e32 v5, v5, v6
+; NOMAD-NEXT:    v_add_f32_e32 v1, v1, v9
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v5, v5
+; NOMAD-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_f32_e32 v5, 0x5f7ffffc, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v1, 0x5f7ffffc, v1
+; NOMAD-NEXT:    v_mul_f32_e32 v6, 0x2f800000, v5
+; NOMAD-NEXT:    v_mul_f32_e32 v9, 0x2f800000, v1
+; NOMAD-NEXT:    v_trunc_f32_e32 v6, v6
+; NOMAD-NEXT:    v_trunc_f32_e32 v9, v9
+; NOMAD-NEXT:    v_mul_f32_e32 v12, 0xcf800000, v6
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v6, v6
+; NOMAD-NEXT:    v_mul_f32_e32 v13, 0xcf800000, v9
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v9, v9
+; NOMAD-NEXT:    v_add_f32_e32 v5, v12, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v7, v6
+; NOMAD-NEXT:    v_add_f32_e32 v1, v13, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v10, v9
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v5, v5
+; NOMAD-NEXT:    v_cvt_u32_f32_e32 v1, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v7, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v8, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v7, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v17, v10, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v18, v11, v1
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v10, v1
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v15, v12
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v18, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v9, v17
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v1, v17
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v19
+; NOMAD-NEXT:    v_mul_lo_u32 v19, v1, v13
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v15, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v15, v18
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v6, v14
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v5, v14
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v6, v14
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v9, v17
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[4:5], v12, v16
+; NOMAD-NEXT:    v_mul_lo_u32 v16, v5, v12
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[4:5], v15, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[4:5], v15, v18
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v6, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v16, s[4:5], v16, v18
+; NOMAD-NEXT:    v_mul_hi_u32 v18, v5, v12
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[4:5], v15, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v14, s[4:5], v14, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[4:5], v15, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v18, vcc, v19, v18
+; NOMAD-NEXT:    v_mul_lo_u32 v19, v9, v13
+; NOMAD-NEXT:    v_add_i32_e32 v17, vcc, v19, v17
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v1, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v20, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v17, vcc, v17, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v19, vcc, v20, v19
+; NOMAD-NEXT:    v_and_b32_e32 v0, s6, v0
+; NOMAD-NEXT:    v_and_b32_e32 v2, s6, v2
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v6, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v9, v13
+; NOMAD-NEXT:    v_add_i32_e32 v14, vcc, v14, v16
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v17, vcc, v17, v18
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v15, vcc, v15, v16
+; NOMAD-NEXT:    v_add_i32_e32 v16, vcc, v19, v18
+; NOMAD-NEXT:    v_add_i32_e32 v12, vcc, v12, v15
+; NOMAD-NEXT:    v_add_i32_e32 v13, vcc, v13, v16
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v14
+; NOMAD-NEXT:    v_addc_u32_e64 v14, s[4:5], v6, v12, vcc
+; NOMAD-NEXT:    v_add_i32_e64 v6, s[4:5], v6, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v7, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v8, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v15, v7, v5
+; NOMAD-NEXT:    v_add_i32_e64 v1, s[4:5], v1, v17
+; NOMAD-NEXT:    v_addc_u32_e64 v16, s[6:7], v9, v13, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e64 v9, s[6:7], v9, v13
+; NOMAD-NEXT:    v_mul_lo_u32 v13, v10, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v11, v1
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v10, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v7, v14
+; NOMAD-NEXT:    v_mul_lo_u32 v18, v14, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v19, v5, v12
+; NOMAD-NEXT:    v_mul_hi_u32 v12, v14, v12
+; NOMAD-NEXT:    v_mul_lo_u32 v10, v10, v16
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[6:7], v8, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v16, v13
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[6:7], v11, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v1, v13
+; NOMAD-NEXT:    v_mul_hi_u32 v13, v16, v13
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[6:7], v7, v15
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[6:7], v10, v17
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v5, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v17, v1, v10
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[6:7], v8, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v17, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[6:7], v8, v11
+; NOMAD-NEXT:    v_mul_lo_u32 v8, v14, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v5, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v7, v14, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v14, v16, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v16, v16, v10
+; NOMAD-NEXT:    v_mul_hi_u32 v10, v1, v10
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[8:9], v18, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v18, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[8:9], v8, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[8:9], v14, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[8:9], v15, v19
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[8:9], v8, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, s[8:9]
+; NOMAD-NEXT:    v_cndmask_b32_e64 v19, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[6:7], v13, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v15, s[6:7], v18, v15
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v12, v11
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[6:7], v17, v19
+; NOMAD-NEXT:    v_add_i32_e64 v13, s[6:7], v14, v13
+; NOMAD-NEXT:    v_add_i32_e64 v8, s[6:7], v8, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v10, s[6:7], v10, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, s[6:7]
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v11, v14
+; NOMAD-NEXT:    v_add_i32_e64 v12, s[6:7], v13, v12
+; NOMAD-NEXT:    v_add_i32_e64 v7, s[6:7], v7, v11
+; NOMAD-NEXT:    v_add_i32_e64 v11, s[6:7], v16, v12
+; NOMAD-NEXT:    v_addc_u32_e32 v6, vcc, v6, v7, vcc
+; NOMAD-NEXT:    v_addc_u32_e64 v7, vcc, v9, v11, s[4:5]
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_addc_u32_e32 v6, vcc, 0, v6, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v8, 0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v9, v0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, 0, v5
+; NOMAD-NEXT:    v_add_i32_e32 v1, vcc, v1, v10
+; NOMAD-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; NOMAD-NEXT:    v_mul_lo_u32 v10, 0, v1
+; NOMAD-NEXT:    v_mul_hi_u32 v11, v2, v1
+; NOMAD-NEXT:    v_mul_hi_u32 v1, 0, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v12, v0, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v13, 0, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v14, v0, v6
+; NOMAD-NEXT:    v_mul_hi_u32 v6, 0, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v15, v2, v7
+; NOMAD-NEXT:    v_mul_lo_u32 v16, 0, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v17, v2, v7
+; NOMAD-NEXT:    v_mul_hi_u32 v7, 0, v7
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v12
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v13, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v15
+; NOMAD-NEXT:    v_cndmask_b32_e64 v15, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v1, vcc, v16, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v8, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v14
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v10, v11
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v1, vcc, v1, v17
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v12, v8
+; NOMAD-NEXT:    v_add_i32_e32 v9, vcc, v13, v9
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v15, v10
+; NOMAD-NEXT:    v_add_i32_e32 v11, vcc, v16, v11
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v5, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v1, vcc, v1, v10
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
+; NOMAD-NEXT:    v_add_i32_e32 v8, vcc, v9, v8
+; NOMAD-NEXT:    v_mul_lo_u32 v9, v3, v5
+; NOMAD-NEXT:    v_mul_lo_u32 v12, 0, v5
+; NOMAD-NEXT:    v_mul_hi_u32 v5, v3, v5
+; NOMAD-NEXT:    v_add_i32_e32 v10, vcc, v11, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v11, v4, v1
+; NOMAD-NEXT:    v_mul_lo_u32 v13, 0, v1
+; NOMAD-NEXT:    v_mul_hi_u32 v1, v4, v1
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v6, v8
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v7, v10
+; NOMAD-NEXT:    v_mul_lo_u32 v6, v3, v6
+; NOMAD-NEXT:    v_mul_lo_u32 v7, v4, v7
+; NOMAD-NEXT:    v_add_i32_e32 v6, vcc, v12, v6
+; NOMAD-NEXT:    v_add_i32_e32 v7, vcc, v13, v7
+; NOMAD-NEXT:    v_add_i32_e32 v5, vcc, v6, v5
+; NOMAD-NEXT:    v_add_i32_e32 v1, vcc, v7, v1
+; NOMAD-NEXT:    v_sub_i32_e32 v0, vcc, v0, v9
+; NOMAD-NEXT:    v_subb_u32_e64 v6, s[4:5], 0, v5, vcc
+; NOMAD-NEXT:    v_sub_i32_e64 v5, s[4:5], 0, v5
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[4:5], v0, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v2, s[4:5], v2, v11
+; NOMAD-NEXT:    v_subb_u32_e64 v8, s[6:7], 0, v1, s[4:5]
+; NOMAD-NEXT:    v_sub_i32_e64 v1, s[6:7], 0, v1
+; NOMAD-NEXT:    v_cmp_ge_u32_e64 s[6:7], v2, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v9, 0, -1, s[6:7]
+; NOMAD-NEXT:    v_cmp_le_u32_e64 s[6:7], 0, v6
+; NOMAD-NEXT:    v_cndmask_b32_e64 v10, 0, -1, s[6:7]
+; NOMAD-NEXT:    v_subbrev_u32_e32 v5, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v8
+; NOMAD-NEXT:    v_cndmask_b32_e64 v11, 0, -1, vcc
+; NOMAD-NEXT:    v_subbrev_u32_e64 v1, vcc, 0, v1, s[4:5]
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v6
+; NOMAD-NEXT:    v_cndmask_b32_e32 v7, v10, v7, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v10, vcc, v0, v3
+; NOMAD-NEXT:    v_subbrev_u32_e32 v5, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v10, v3
+; NOMAD-NEXT:    v_cndmask_b32_e64 v12, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v8
+; NOMAD-NEXT:    v_cndmask_b32_e32 v9, v11, v9, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v11, vcc, v2, v4
+; NOMAD-NEXT:    v_subbrev_u32_e32 v1, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_ge_u32_e32 vcc, v11, v4
+; NOMAD-NEXT:    v_cndmask_b32_e64 v13, 0, -1, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e64 v14, 0, -1, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v3, vcc, v10, v3
+; NOMAD-NEXT:    v_subbrev_u32_e32 v15, vcc, 0, v5, vcc
+; NOMAD-NEXT:    v_cmp_le_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e64 v16, 0, -1, vcc
+; NOMAD-NEXT:    v_sub_i32_e32 v4, vcc, v11, v4
+; NOMAD-NEXT:    v_subbrev_u32_e32 v17, vcc, 0, v1, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v5
+; NOMAD-NEXT:    v_cndmask_b32_e32 v12, v14, v12, vcc
+; NOMAD-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; NOMAD-NEXT:    v_cndmask_b32_e32 v13, v16, v13, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
+; NOMAD-NEXT:    v_cndmask_b32_e32 v3, v10, v3, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v13
+; NOMAD-NEXT:    v_cndmask_b32_e64 v4, v11, v4, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v5, v5, v15, vcc
+; NOMAD-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v7
+; NOMAD-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v3, v1, v17, s[4:5]
+; NOMAD-NEXT:    v_cmp_ne_u32_e64 s[4:5], 0, v9
+; NOMAD-NEXT:    v_cndmask_b32_e64 v2, v2, v4, s[4:5]
+; NOMAD-NEXT:    v_cndmask_b32_e32 v1, v6, v5, vcc
+; NOMAD-NEXT:    v_cndmask_b32_e64 v3, v8, v3, s[4:5]
+; NOMAD-NEXT:    s_setpc_b64 s[30:31]
   %num.mask = and <2 x i64> %num, <i64 16777215, i64 16777215>
   %den.mask = and <2 x i64> %den, <i64 16777215, i64 16777215>
   %result = urem <2 x i64> %num.mask, %den.mask
   ret <2 x i64> %result
 }