PeepholeOpt: Allow introducing subregister uses on reg_sequence (#127052)

This reverts d246cc6. We now handle
composing subregister extracts through reg_sequence.

Author: arsenm

llvm/lib/CodeGen/PeepholeOptimizer.cpp

@@ -421,12 +421,6 @@ class RegSequenceRewriter : public Rewriter {
   }
 
   bool RewriteCurrentSource(Register NewReg, unsigned NewSubReg) override {
-    // Do not introduce new subregister uses in a reg_sequence. Until composing
-    // subregister indices is supported while folding, we're just blocking
-    // folding of subregister copies later in the function.
-    if (NewSubReg)
-      return false;
-
     MachineOperand &MO = CopyLike.getOperand(CurrentSrcIdx);
     MO.setReg(NewReg);
     MO.setSubReg(NewSubReg);

llvm/test/CodeGen/AMDGPU/GlobalISel/sdivrem.ll

@@ -1635,7 +1635,6 @@ define amdgpu_kernel void @sdivrem_v2i64(ptr addrspace(1) %out0, ptr addrspace(1
 ; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, v3, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v4, vcc, v4, v1, vcc
 ; GFX9-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], s16, v3, 0
-; GFX9-NEXT:    v_mov_b32_e32 v7, s11
 ; GFX9-NEXT:    v_mad_u64_u32 v[1:2], s[0:1], s16, v4, v[1:2]
 ; GFX9-NEXT:    v_mul_hi_u32 v6, v3, v0
 ; GFX9-NEXT:    v_mad_u64_u32 v[1:2], s[0:1], s17, v3, v[1:2]
@@ -1683,149 +1682,150 @@ define amdgpu_kernel void @sdivrem_v2i64(ptr addrspace(1) %out0, ptr addrspace(1
 ; GFX9-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], s8, v5, 0
 ; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
 ; GFX9-NEXT:    v_add_u32_e32 v3, v4, v3
-; GFX9-NEXT:    v_add3_u32 v6, v3, v2, v6
-; GFX9-NEXT:    v_mad_u64_u32 v[1:2], s[0:1], s8, v6, v[1:2]
-; GFX9-NEXT:    v_sub_co_u32_e32 v0, vcc, s10, v0
-; GFX9-NEXT:    v_mad_u64_u32 v[2:3], s[0:1], s9, v5, v[1:2]
+; GFX9-NEXT:    v_add3_u32 v3, v3, v2, v6
+; GFX9-NEXT:    v_mad_u64_u32 v[1:2], s[0:1], s8, v3, v[1:2]
+; GFX9-NEXT:    v_mov_b32_e32 v6, s11
+; GFX9-NEXT:    v_sub_co_u32_e32 v7, vcc, s10, v0
+; GFX9-NEXT:    v_mad_u64_u32 v[1:2], s[0:1], s9, v5, v[1:2]
 ; GFX9-NEXT:    v_mov_b32_e32 v4, s9
 ; GFX9-NEXT:    s_ashr_i32 s10, s3, 31
-; GFX9-NEXT:    v_subb_co_u32_e64 v1, s[0:1], v7, v2, vcc
-; GFX9-NEXT:    v_sub_u32_e32 v2, s11, v2
-; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v1
-; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, -1, s[0:1]
-; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v0
-; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v4, vcc
-; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, -1, s[0:1]
-; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v1
-; GFX9-NEXT:    v_subrev_co_u32_e32 v8, vcc, s8, v0
-; GFX9-NEXT:    v_cndmask_b32_e64 v7, v3, v7, s[0:1]
-; GFX9-NEXT:    v_subbrev_co_u32_e64 v9, s[0:1], 0, v2, vcc
-; GFX9-NEXT:    v_add_co_u32_e64 v10, s[0:1], 1, v5
-; GFX9-NEXT:    v_addc_co_u32_e64 v11, s[0:1], 0, v6, s[0:1]
-; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v9
-; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, -1, s[0:1]
-; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v8
+; GFX9-NEXT:    v_subb_co_u32_e64 v6, s[0:1], v6, v1, vcc
+; GFX9-NEXT:    v_sub_u32_e32 v0, s11, v1
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v6
+; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v7
+; GFX9-NEXT:    v_subb_co_u32_e32 v0, vcc, v0, v4, vcc
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v6
+; GFX9-NEXT:    v_subrev_co_u32_e32 v9, vcc, s8, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, v1, v2, s[0:1]
+; GFX9-NEXT:    v_subbrev_co_u32_e64 v10, s[0:1], 0, v0, vcc
+; GFX9-NEXT:    v_add_co_u32_e64 v2, s[0:1], 1, v5
+; GFX9-NEXT:    v_addc_co_u32_e64 v11, s[0:1], 0, v3, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s9, v10
+; GFX9-NEXT:    v_cndmask_b32_e64 v1, 0, -1, s[0:1]
+; GFX9-NEXT:    v_cmp_le_u32_e64 s[0:1], s8, v9
 ; GFX9-NEXT:    v_cndmask_b32_e64 v12, 0, -1, s[0:1]
-; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v9
-; GFX9-NEXT:    v_cndmask_b32_e64 v12, v3, v12, s[0:1]
-; GFX9-NEXT:    v_add_co_u32_e64 v13, s[0:1], 1, v10
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[0:1], s9, v10
+; GFX9-NEXT:    v_cndmask_b32_e64 v12, v1, v12, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e64 v13, s[0:1], 1, v2
 ; GFX9-NEXT:    v_addc_co_u32_e64 v14, s[0:1], 0, v11, s[0:1]
 ; GFX9-NEXT:    s_add_u32 s0, s18, s6
 ; GFX9-NEXT:    s_addc_u32 s1, s19, s6
 ; GFX9-NEXT:    s_add_u32 s2, s2, s10
 ; GFX9-NEXT:    s_mov_b32 s11, s10
 ; GFX9-NEXT:    s_addc_u32 s3, s3, s10
 ; GFX9-NEXT:    s_xor_b64 s[2:3], s[2:3], s[10:11]
-; GFX9-NEXT:    v_cvt_f32_u32_e32 v3, s3
+; GFX9-NEXT:    v_cvt_f32_u32_e32 v1, s3
 ; GFX9-NEXT:    v_cvt_f32_u32_e32 v15, s2
-; GFX9-NEXT:    v_subb_co_u32_e32 v2, vcc, v2, v4, vcc
-; GFX9-NEXT:    v_mul_f32_e32 v3, 0x4f800000, v3
-; GFX9-NEXT:    v_add_f32_e32 v3, v3, v15
-; GFX9-NEXT:    v_rcp_iflag_f32_e32 v3, v3
-; GFX9-NEXT:    v_subrev_co_u32_e32 v15, vcc, s8, v8
-; GFX9-NEXT:    v_subbrev_co_u32_e32 v16, vcc, 0, v2, vcc
-; GFX9-NEXT:    v_mul_f32_e32 v2, 0x5f7ffffc, v3
-; GFX9-NEXT:    v_mul_f32_e32 v3, 0x2f800000, v2
-; GFX9-NEXT:    v_trunc_f32_e32 v4, v3
-; GFX9-NEXT:    v_mul_f32_e32 v3, 0xcf800000, v4
-; GFX9-NEXT:    v_add_f32_e32 v2, v3, v2
-; GFX9-NEXT:    v_cvt_u32_f32_e32 v17, v2
+; GFX9-NEXT:    v_subb_co_u32_e32 v0, vcc, v0, v4, vcc
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x4f800000, v1
+; GFX9-NEXT:    v_add_f32_e32 v1, v1, v15
+; GFX9-NEXT:    v_rcp_iflag_f32_e32 v1, v1
+; GFX9-NEXT:    v_subrev_co_u32_e32 v4, vcc, s8, v9
+; GFX9-NEXT:    v_subbrev_co_u32_e32 v15, vcc, 0, v0, vcc
+; GFX9-NEXT:    v_mul_f32_e32 v0, 0x5f7ffffc, v1
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0x2f800000, v0
+; GFX9-NEXT:    v_trunc_f32_e32 v16, v1
+; GFX9-NEXT:    v_mul_f32_e32 v1, 0xcf800000, v16
+; GFX9-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v17, v0
 ; GFX9-NEXT:    s_xor_b64 s[8:9], s[0:1], s[6:7]
 ; GFX9-NEXT:    s_sub_u32 s5, 0, s2
 ; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v12
-; GFX9-NEXT:    v_mad_u64_u32 v[2:3], s[0:1], s5, v17, 0
-; GFX9-NEXT:    v_cvt_u32_f32_e32 v12, v4
+; GFX9-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], s5, v17, 0
+; GFX9-NEXT:    v_cndmask_b32_e32 v12, v2, v13, vcc
+; GFX9-NEXT:    v_cvt_u32_f32_e32 v13, v16
 ; GFX9-NEXT:    s_subb_u32 s20, 0, s3
 ; GFX9-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc
-; GFX9-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc
-; GFX9-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], s5, v12, v[3:4]
-; GFX9-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v7
-; GFX9-NEXT:    v_mul_lo_u32 v7, v12, v2
-; GFX9-NEXT:    v_mad_u64_u32 v[3:4], s[18:19], s20, v17, v[3:4]
-; GFX9-NEXT:    v_cndmask_b32_e64 v4, v6, v11, s[0:1]
-; GFX9-NEXT:    v_cndmask_b32_e32 v6, v8, v15, vcc
-; GFX9-NEXT:    v_mul_lo_u32 v8, v17, v3
-; GFX9-NEXT:    v_cndmask_b32_e64 v5, v5, v10, s[0:1]
-; GFX9-NEXT:    v_mul_hi_u32 v10, v17, v2
-; GFX9-NEXT:    v_cndmask_b32_e32 v9, v9, v16, vcc
-; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v7, v8
-; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
-; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v7, v10
-; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
-; GFX9-NEXT:    v_mul_lo_u32 v10, v12, v3
-; GFX9-NEXT:    v_mul_hi_u32 v2, v12, v2
-; GFX9-NEXT:    v_add_u32_e32 v7, v8, v7
-; GFX9-NEXT:    v_mul_hi_u32 v8, v17, v3
-; GFX9-NEXT:    v_mul_hi_u32 v3, v12, v3
-; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v10, v2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v9, v4, vcc
+; GFX9-NEXT:    v_mad_u64_u32 v[1:2], s[0:1], s5, v13, v[1:2]
+; GFX9-NEXT:    v_cmp_ne_u32_e64 s[0:1], 0, v8
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, v3, v11, s[0:1]
+; GFX9-NEXT:    v_mad_u64_u32 v[1:2], s[18:19], s20, v17, v[1:2]
+; GFX9-NEXT:    v_mul_lo_u32 v2, v13, v0
+; GFX9-NEXT:    v_cndmask_b32_e32 v9, v10, v15, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v3, v17, v1
+; GFX9-NEXT:    v_mul_hi_u32 v10, v17, v0
+; GFX9-NEXT:    v_mul_hi_u32 v0, v13, v0
+; GFX9-NEXT:    v_cndmask_b32_e64 v5, v5, v12, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v10
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v10, v13, v1
+; GFX9-NEXT:    v_add_u32_e32 v2, v3, v2
+; GFX9-NEXT:    v_mul_hi_u32 v3, v17, v1
+; GFX9-NEXT:    v_mul_hi_u32 v1, v13, v1
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v10, v0
 ; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
-; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v8
-; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
-; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v7
-; GFX9-NEXT:    v_add_u32_e32 v8, v10, v8
-; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
-; GFX9-NEXT:    v_add3_u32 v3, v8, v7, v3
-; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, v17, v2
-; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, v12, v3, vcc
-; GFX9-NEXT:    v_mad_u64_u32 v[2:3], s[18:19], s5, v7, 0
-; GFX9-NEXT:    v_cndmask_b32_e64 v6, v0, v6, s[0:1]
-; GFX9-NEXT:    v_cndmask_b32_e64 v9, v1, v9, s[0:1]
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v3, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
+; GFX9-NEXT:    v_add_u32_e32 v3, v10, v3
+; GFX9-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, v17, v0
+; GFX9-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX9-NEXT:    v_mad_u64_u32 v[2:3], s[18:19], s5, v10, 0
+; GFX9-NEXT:    v_addc_co_u32_e32 v11, vcc, v13, v1, vcc
 ; GFX9-NEXT:    v_mov_b32_e32 v0, v3
-; GFX9-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], s5, v8, v[0:1]
-; GFX9-NEXT:    v_xor_b32_e32 v10, s17, v4
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, v7, v4, s[0:1]
+; GFX9-NEXT:    v_cndmask_b32_e64 v6, v6, v9, s[0:1]
+; GFX9-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], s5, v11, v[0:1]
 ; GFX9-NEXT:    v_xor_b32_e32 v5, s16, v5
-; GFX9-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], s20, v7, v[0:1]
-; GFX9-NEXT:    v_mov_b32_e32 v11, s17
+; GFX9-NEXT:    v_xor_b32_e32 v8, s17, v8
+; GFX9-NEXT:    v_mad_u64_u32 v[3:4], s[0:1], s20, v10, v[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v9, s17
 ; GFX9-NEXT:    v_subrev_co_u32_e32 v0, vcc, s16, v5
-; GFX9-NEXT:    v_xor_b32_e32 v4, s4, v6
-; GFX9-NEXT:    v_mul_lo_u32 v5, v8, v2
-; GFX9-NEXT:    v_mul_lo_u32 v6, v7, v3
-; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v10, v11, vcc
-; GFX9-NEXT:    v_mul_hi_u32 v10, v7, v2
-; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, v5, v6
-; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
-; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, v5, v10
+; GFX9-NEXT:    v_xor_b32_e32 v4, s4, v7
+; GFX9-NEXT:    v_mul_lo_u32 v5, v11, v2
+; GFX9-NEXT:    v_mul_lo_u32 v7, v10, v3
+; GFX9-NEXT:    v_subb_co_u32_e32 v1, vcc, v8, v9, vcc
+; GFX9-NEXT:    v_mul_hi_u32 v8, v10, v2
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, v5, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, v5, v8
 ; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
-; GFX9-NEXT:    v_mul_lo_u32 v10, v8, v3
-; GFX9-NEXT:    v_mul_hi_u32 v2, v8, v2
-; GFX9-NEXT:    v_add_u32_e32 v5, v6, v5
-; GFX9-NEXT:    v_mul_hi_u32 v6, v7, v3
-; GFX9-NEXT:    v_mul_hi_u32 v3, v8, v3
-; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v10, v2
-; GFX9-NEXT:    v_cndmask_b32_e64 v10, 0, 1, vcc
-; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v6
-; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v8, v11, v3
+; GFX9-NEXT:    v_mul_hi_u32 v2, v11, v2
+; GFX9-NEXT:    v_add_u32_e32 v5, v7, v5
+; GFX9-NEXT:    v_mul_hi_u32 v7, v10, v3
+; GFX9-NEXT:    v_mul_hi_u32 v3, v11, v3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v8, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
 ; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v5
-; GFX9-NEXT:    v_add_u32_e32 v6, v10, v6
+; GFX9-NEXT:    v_add_u32_e32 v7, v8, v7
 ; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
-; GFX9-NEXT:    v_add3_u32 v3, v6, v5, v3
-; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v7, v2
-; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, v8, v3, vcc
+; GFX9-NEXT:    v_add3_u32 v3, v7, v5, v3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v10, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, v11, v3, vcc
 ; GFX9-NEXT:    v_mul_lo_u32 v5, s9, v2
-; GFX9-NEXT:    v_mul_lo_u32 v6, s8, v3
-; GFX9-NEXT:    v_mul_hi_u32 v8, s8, v2
+; GFX9-NEXT:    v_mul_lo_u32 v7, s8, v3
+; GFX9-NEXT:    v_mul_hi_u32 v9, s8, v2
 ; GFX9-NEXT:    v_mul_hi_u32 v2, s9, v2
 ; GFX9-NEXT:    v_mul_hi_u32 v12, s9, v3
-; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, v5, v6
-; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
-; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, v5, v8
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, v5, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, v5, v9
 ; GFX9-NEXT:    v_cndmask_b32_e64 v5, 0, 1, vcc
-; GFX9-NEXT:    v_mul_lo_u32 v8, s9, v3
-; GFX9-NEXT:    v_add_u32_e32 v5, v6, v5
-; GFX9-NEXT:    v_mul_hi_u32 v6, s8, v3
-; GFX9-NEXT:    v_xor_b32_e32 v9, s4, v9
-; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v8, v2
-; GFX9-NEXT:    v_cndmask_b32_e64 v8, 0, 1, vcc
-; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v6
-; GFX9-NEXT:    v_cndmask_b32_e64 v6, 0, 1, vcc
+; GFX9-NEXT:    v_mul_lo_u32 v9, s9, v3
+; GFX9-NEXT:    v_add_u32_e32 v5, v7, v5
+; GFX9-NEXT:    v_mul_hi_u32 v7, s8, v3
+; GFX9-NEXT:    v_xor_b32_e32 v6, s4, v6
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v9, v2
+; GFX9-NEXT:    v_cndmask_b32_e64 v9, 0, 1, vcc
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v7
+; GFX9-NEXT:    v_cndmask_b32_e64 v7, 0, 1, vcc
 ; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, v2, v5
 ; GFX9-NEXT:    v_mad_u64_u32 v[2:3], s[0:1], s2, v10, 0
+; GFX9-NEXT:    v_mov_b32_e32 v8, s4
 ; GFX9-NEXT:    v_cndmask_b32_e64 v11, 0, 1, vcc
-; GFX9-NEXT:    v_add_u32_e32 v6, v8, v6
-; GFX9-NEXT:    v_mov_b32_e32 v7, s4
 ; GFX9-NEXT:    v_subrev_co_u32_e32 v4, vcc, s4, v4
+; GFX9-NEXT:    v_subb_co_u32_e32 v5, vcc, v6, v8, vcc
+; GFX9-NEXT:    v_add_u32_e32 v6, v9, v7
 ; GFX9-NEXT:    v_add3_u32 v8, v6, v11, v12
-; GFX9-NEXT:    v_subb_co_u32_e32 v5, vcc, v9, v7, vcc
 ; GFX9-NEXT:    v_mad_u64_u32 v[6:7], s[0:1], s2, v8, v[3:4]
 ; GFX9-NEXT:    v_mov_b32_e32 v9, s9
 ; GFX9-NEXT:    v_sub_co_u32_e32 v2, vcc, s8, v2

llvm/test/CodeGen/AMDGPU/atomic_optimizations_global_pointer.ll

@@ -2218,31 +2218,31 @@ define amdgpu_kernel void @add_i64_uniform(ptr addrspace(1) %out, ptr addrspace(
 ; GFX1264-NEXT:    s_clause 0x1
 ; GFX1264-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 ; GFX1264-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX1264-NEXT:    s_mov_b64 s[6:7], exec
-; GFX1264-NEXT:    s_mov_b32 s11, 0
-; GFX1264-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
 ; GFX1264-NEXT:    s_mov_b64 s[8:9], exec
+; GFX1264-NEXT:    s_mov_b32 s11, 0
+; GFX1264-NEXT:    v_mbcnt_lo_u32_b32 v0, s8, 0
+; GFX1264-NEXT:    s_mov_b64 s[6:7], exec
 ; GFX1264-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1264-NEXT:    v_mbcnt_hi_u32_b32 v2, s7, v0
+; GFX1264-NEXT:    v_mbcnt_hi_u32_b32 v2, s9, v0
 ; GFX1264-NEXT:    ; implicit-def: $vgpr0_vgpr1
 ; GFX1264-NEXT:    v_cmpx_eq_u32_e32 0, v2
 ; GFX1264-NEXT:    s_cbranch_execz .LBB4_2
 ; GFX1264-NEXT:  ; %bb.1:
-; GFX1264-NEXT:    s_bcnt1_i32_b64 s10, s[6:7]
-; GFX1264-NEXT:    s_mov_b32 s15, 0x31016000
+; GFX1264-NEXT:    s_bcnt1_i32_b64 s10, s[8:9]
 ; GFX1264-NEXT:    s_wait_kmcnt 0x0
-; GFX1264-NEXT:    s_mul_u64 s[6:7], s[4:5], s[10:11]
-; GFX1264-NEXT:    s_mov_b32 s14, -1
+; GFX1264-NEXT:    s_mul_u64 s[8:9], s[4:5], s[10:11]
+; GFX1264-NEXT:    s_mov_b32 s11, 0x31016000
 ; GFX1264-NEXT:    s_wait_alu 0xfffe
-; GFX1264-NEXT:    v_mov_b32_e32 v0, s6
-; GFX1264-NEXT:    v_mov_b32_e32 v1, s7
-; GFX1264-NEXT:    s_mov_b32 s12, s2
-; GFX1264-NEXT:    s_mov_b32 s13, s3
-; GFX1264-NEXT:    buffer_atomic_add_u64 v[0:1], off, s[12:15], null th:TH_ATOMIC_RETURN scope:SCOPE_DEV
+; GFX1264-NEXT:    v_mov_b32_e32 v0, s8
+; GFX1264-NEXT:    v_mov_b32_e32 v1, s9
+; GFX1264-NEXT:    s_mov_b32 s10, -1
+; GFX1264-NEXT:    s_mov_b32 s8, s2
+; GFX1264-NEXT:    s_mov_b32 s9, s3
+; GFX1264-NEXT:    buffer_atomic_add_u64 v[0:1], off, s[8:11], null th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX1264-NEXT:    s_wait_loadcnt 0x0
 ; GFX1264-NEXT:    global_inv scope:SCOPE_DEV
 ; GFX1264-NEXT:  .LBB4_2:
-; GFX1264-NEXT:    s_or_b64 exec, exec, s[8:9]
+; GFX1264-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GFX1264-NEXT:    s_wait_kmcnt 0x0
 ; GFX1264-NEXT:    v_readfirstlane_b32 s3, v1
 ; GFX1264-NEXT:    v_readfirstlane_b32 s2, v0
@@ -5800,31 +5800,31 @@ define amdgpu_kernel void @sub_i64_uniform(ptr addrspace(1) %out, ptr addrspace(
 ; GFX1264-NEXT:    s_clause 0x1
 ; GFX1264-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 ; GFX1264-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX1264-NEXT:    s_mov_b64 s[6:7], exec
-; GFX1264-NEXT:    s_mov_b32 s11, 0
-; GFX1264-NEXT:    v_mbcnt_lo_u32_b32 v0, s6, 0
 ; GFX1264-NEXT:    s_mov_b64 s[8:9], exec
+; GFX1264-NEXT:    s_mov_b32 s11, 0
+; GFX1264-NEXT:    v_mbcnt_lo_u32_b32 v0, s8, 0
+; GFX1264-NEXT:    s_mov_b64 s[6:7], exec
 ; GFX1264-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1264-NEXT:    v_mbcnt_hi_u32_b32 v2, s7, v0
+; GFX1264-NEXT:    v_mbcnt_hi_u32_b32 v2, s9, v0
 ; GFX1264-NEXT:    ; implicit-def: $vgpr0_vgpr1
 ; GFX1264-NEXT:    v_cmpx_eq_u32_e32 0, v2
 ; GFX1264-NEXT:    s_cbranch_execz .LBB10_2
 ; GFX1264-NEXT:  ; %bb.1:
-; GFX1264-NEXT:    s_bcnt1_i32_b64 s10, s[6:7]
-; GFX1264-NEXT:    s_mov_b32 s15, 0x31016000
+; GFX1264-NEXT:    s_bcnt1_i32_b64 s10, s[8:9]
 ; GFX1264-NEXT:    s_wait_kmcnt 0x0
-; GFX1264-NEXT:    s_mul_u64 s[6:7], s[4:5], s[10:11]
-; GFX1264-NEXT:    s_mov_b32 s14, -1
+; GFX1264-NEXT:    s_mul_u64 s[8:9], s[4:5], s[10:11]
+; GFX1264-NEXT:    s_mov_b32 s11, 0x31016000
 ; GFX1264-NEXT:    s_wait_alu 0xfffe
-; GFX1264-NEXT:    v_mov_b32_e32 v0, s6
-; GFX1264-NEXT:    v_mov_b32_e32 v1, s7
-; GFX1264-NEXT:    s_mov_b32 s12, s2
-; GFX1264-NEXT:    s_mov_b32 s13, s3
-; GFX1264-NEXT:    buffer_atomic_sub_u64 v[0:1], off, s[12:15], null th:TH_ATOMIC_RETURN scope:SCOPE_DEV
+; GFX1264-NEXT:    v_mov_b32_e32 v0, s8
+; GFX1264-NEXT:    v_mov_b32_e32 v1, s9
+; GFX1264-NEXT:    s_mov_b32 s10, -1
+; GFX1264-NEXT:    s_mov_b32 s8, s2
+; GFX1264-NEXT:    s_mov_b32 s9, s3
+; GFX1264-NEXT:    buffer_atomic_sub_u64 v[0:1], off, s[8:11], null th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX1264-NEXT:    s_wait_loadcnt 0x0
 ; GFX1264-NEXT:    global_inv scope:SCOPE_DEV
 ; GFX1264-NEXT:  .LBB10_2:
-; GFX1264-NEXT:    s_or_b64 exec, exec, s[8:9]
+; GFX1264-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GFX1264-NEXT:    s_wait_kmcnt 0x0
 ; GFX1264-NEXT:    v_mad_co_u64_u32 v[3:4], null, s4, v2, 0
 ; GFX1264-NEXT:    v_readfirstlane_b32 s2, v0