[llvm] AMDGPU/GlobalISel: Regbanklegalize rules for G_FSQRT (PR #179817)

[Matt Arsenault via llvm-commits]

================
@@ -1,40 +1,57 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -mtriple=amdgcn -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck %s
-# RUN: llc -mtriple=amdgcn -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass='amdgpu-regbankselect,amdgpu-regbanklegalize' %s -o - | FileCheck -check-prefixes=GCN,GFX9 %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -run-pass='amdgpu-regbankselect,amdgpu-regbanklegalize' %s -o - | FileCheck -check-prefixes=GCN,GFX12 %s
 
 ---
-name: fsqrt_s
+name: fsqrt_s16_uniform
 legalized: true
 
 body: |
   bb.0:
-    liveins: $sgpr0_sgpr1
-    ; CHECK-LABEL: name: fsqrt_s
-    ; CHECK: liveins: $sgpr0_sgpr1
-    ; CHECK-NEXT: {{  $}}
-    ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
-    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
-    ; CHECK-NEXT: [[FSQRT:%[0-9]+]]:vgpr(s32) = G_FSQRT [[COPY1]]
-    ; CHECK-NEXT: $vgpr0 = COPY [[FSQRT]](s32)
+    liveins: $sgpr0
+    ; GFX9-LABEL: name: fsqrt_s16_uniform
+    ; GFX9: liveins: $sgpr0
+    ; GFX9-NEXT: {{  $}}
+    ; GFX9-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
+    ; GFX9-NEXT: [[TRUNC:%[0-9]+]]:sgpr(s16) = G_TRUNC [[COPY]](s32)
+    ; GFX9-NEXT: [[COPY1:%[0-9]+]]:vgpr(s16) = COPY [[TRUNC]](s16)
+    ; GFX9-NEXT: [[FSQRT:%[0-9]+]]:vgpr(s16) = G_FSQRT [[COPY1]]
+    ; GFX9-NEXT: [[ANYEXT:%[0-9]+]]:vgpr(s32) = G_ANYEXT [[FSQRT]](s16)
+    ; GFX9-NEXT: $vgpr0 = COPY [[ANYEXT]](s32)
+    ;
+    ; GFX12-LABEL: name: fsqrt_s16_uniform
+    ; GFX12: liveins: $sgpr0
+    ; GFX12-NEXT: {{  $}}
+    ; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
+    ; GFX12-NEXT: [[TRUNC:%[0-9]+]]:sgpr(s16) = G_TRUNC [[COPY]](s32)
+    ; GFX12-NEXT: [[FSQRT:%[0-9]+]]:sgpr(s16) = G_FSQRT [[TRUNC]]
+    ; GFX12-NEXT: [[ANYEXT:%[0-9]+]]:sgpr(s32) = G_ANYEXT [[FSQRT]](s16)
+    ; GFX12-NEXT: $vgpr0 = COPY [[ANYEXT]](s32)
     %0:_(s32) = COPY $sgpr0
-    %1:_(s32) = G_FSQRT %0
-    $vgpr0 = COPY %1
+    %1:_(s16) = G_TRUNC %0
+    %2:_(s16) = G_FSQRT %1
+    %3:_(s32) = G_ANYEXT %2
+    $vgpr0 = COPY %3
 ...
 
 ---
-name: fsqrt_v
+name: fsqrt_s16_divergent
----------------
arsenm wrote:

v / VGPR and s / SGPR are more accurate naming convention. The uniformity is a proxy for the register bank 

https://github.com/llvm/llvm-project/pull/179817