[AMDGPU] illegal VGPR to SGPR copy

[BaoshanPang]

llc fail to compile this function:

define amdgpu_ps <4 x i32> @v_test(<4 x i32> %src0, <4 x i32> %bit.index) {
  %set.bit.at.index = shl <4 x i32> <i32 1, i32 1, i32 1, i32 1>, %bit.index
  %or = or <4 x i32> %src0, %set.bit.at.index
  ret <4 x i32> %or
}

https://godbolt.org/z/fhKTa76zn

llc works fine with option "--global-isel".

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Baoshan (BaoshanPang)

llc fail to compile this function:

define amdgpu_ps &lt;4 x i32&gt; @<!-- -->v_test(&lt;4 x i32&gt; %src0, &lt;4 x i32&gt; %bit.index) {
  %set.bit.at.index = shl &lt;4 x i32&gt; &lt;i32 1, i32 1, i32 1, i32 1&gt;, %bit.index
  %or = or &lt;4 x i32&gt; %src0, %set.bit.at.index
  ret &lt;4 x i32&gt; %or
}

https://godbolt.org/z/fhKTa76zn

llc works fine with option "--global-isel".

[arsenm]

Should need #129159 or similar

[BaoshanPang]

@arsenm with "-global-isel", the genereated asm code is like this:

6 v_test:                                 ; @v_test
   7 ; %bb.0:
   8 ›   v_lshl_or_b32 v0, 1, v4, v0
   9 ›   v_lshl_or_b32 v1, 1, v5, v1
  10 ›   v_lshl_or_b32 v2, 1, v6, v2
  11 ›   v_lshl_or_b32 v3, 1, v7, v3
  12 ›   v_readfirstlane_b32 s0, v0
  13 ›   v_readfirstlane_b32 s1, v1
  14 ›   v_readfirstlane_b32 s2, v2
  15 ›   v_readfirstlane_b32 s3, v3

Why we need to use v_readfirstlane_b32 to copy the value from vgpr to sgpr? Is there any document that I can check with?

[arsenm]

@arsenm with "-global-isel", the genereated asm code is like this:

6 v_test:                                 ; @v_test
   7 ; %bb.0:
   8 ›   v_lshl_or_b32 v0, 1, v4, v0
   9 ›   v_lshl_or_b32 v1, 1, v5, v1
  10 ›   v_lshl_or_b32 v2, 1, v6, v2
  11 ›   v_lshl_or_b32 v3, 1, v7, v3
  12 ›   v_readfirstlane_b32 s0, v0
  13 ›   v_readfirstlane_b32 s1, v1
  14 ›   v_readfirstlane_b32 s2, v2
  15 ›   v_readfirstlane_b32 s3, v3

Why we need to use v_readfirstlane_b32 to copy the value from vgpr to sgpr? Is there any document that I can check with?

Because it's not really a copy, it's a reduction. You have 64 values and need to select 1 to return the SGPR result