[llvm] [AMDGPU] SIPeepholeSDWA: Handle V_CNDMASK_B32_e64 (PR #137930)

[Matt Arsenault via llvm-commits]

================
@@ -0,0 +1,201 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc %s -mtriple=amdgcn -mcpu=gfx1030 -run-pass=si-peephole-sdwa -o - | FileCheck %s
+
+---
+name: change-compare-dest-to-vcc # carry-in def is a comparison that can write to VCC
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: change-compare-dest-to-vcc
+    ; CHECK: liveins: $vgpr0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
+    ; CHECK-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 undef [[DEF]], 1, implicit $exec
+    ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, undef [[DEF1]], implicit $exec
+    ; CHECK-NEXT: $vcc_lo = COPY killed [[V_CMP_EQ_U32_e64_]]
+    ; CHECK-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    ; CHECK-NEXT: [[V_CNDMASK_B32_sdwa:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_sdwa 0, [[V_MOV_B32_e32_]], 0, undef [[DEF1]], 0, 6, 0, 6, 5, implicit $vcc_lo, implicit $exec
+    ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_sdwa]]
+    ; CHECK-NEXT: SI_RETURN implicit $vgpr0
+    %0:sreg_32_xm0_xexec = IMPLICIT_DEF
+    %1:vgpr_32 = IMPLICIT_DEF
+    %2:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 undef %0, 1, implicit $exec
+    %3:vgpr_32 = V_LSHRREV_B32_e64 16, undef %1, implicit $exec
+    %4:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, %3, killed %2, implicit $exec
+    $vgpr0 = COPY %4
+    SI_RETURN implicit $vgpr0
+...
+
+---
+name: change-compare-class-dest-to-vcc # check that non-compare instr V_CMP_CLASS is also handled
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: change-compare-class-dest-to-vcc
+    ; CHECK: liveins: $vgpr0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
+    ; CHECK-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; CHECK-NEXT: [[V_CMP_CLASS_F32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_CLASS_F32_e64 2, undef [[DEF]], 1, implicit $exec
+    ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, undef [[DEF1]], implicit $exec
+    ; CHECK-NEXT: $vcc_lo = COPY killed [[V_CMP_CLASS_F32_e64_]]
+    ; CHECK-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    ; CHECK-NEXT: [[V_CNDMASK_B32_sdwa:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_sdwa 0, [[V_MOV_B32_e32_]], 0, undef [[DEF1]], 0, 6, 0, 6, 5, implicit $vcc_lo, implicit $exec
+    ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_sdwa]]
+    ; CHECK-NEXT: SI_RETURN implicit $vgpr0
+    %0:sreg_32_xm0_xexec = IMPLICIT_DEF
+    %1:vgpr_32 = IMPLICIT_DEF
+    %2:sreg_32_xm0_xexec = V_CMP_CLASS_F32_e64 2, undef %0, 1, implicit $exec
+    %3:vgpr_32 = V_LSHRREV_B32_e64 16, undef %1, implicit $exec
+    %4:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, %3, killed %2, implicit $exec
+    $vgpr0 = COPY %4
+    SI_RETURN implicit $vgpr0
+...
+
+...
+---
+name: carry-copy-non-compare # copy of carry-in necessary because def. instr. cannot be changed to write to VCC
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: carry-copy-non-compare
+    ; CHECK: liveins: $vgpr0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
+    ; CHECK-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; CHECK-NEXT: [[V_PK_MAX_F16_:%[0-9]+]]:vgpr_32 = V_PK_MAX_F16 8, undef [[DEF1]], 8, undef [[DEF1]], 0, 0, 0, 0, 0, implicit $mode, implicit $exec
+    ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, [[V_PK_MAX_F16_]], implicit $exec
+    ; CHECK-NEXT: $vcc_lo = COPY killed undef [[DEF]]
+    ; CHECK-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    ; CHECK-NEXT: [[V_CNDMASK_B32_sdwa:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_sdwa 0, [[V_MOV_B32_e32_]], 0, [[V_PK_MAX_F16_]], 0, 6, 0, 6, 5, implicit $vcc_lo, implicit $exec
+    ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_sdwa]]
+    ; CHECK-NEXT: SI_RETURN implicit $vgpr0
+    %0:sreg_32_xm0_xexec = IMPLICIT_DEF
+    %1:vgpr_32 = IMPLICIT_DEF
+    %2:vgpr_32 = V_PK_MAX_F16 8, undef %1, 8, undef %1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
----------------
arsenm wrote:

```suggestion
    %2:vgpr_32 = V_PK_MAX_F16 8, %1, 8, %1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
```

https://github.com/llvm/llvm-project/pull/137930