[AMDGPU] Fix SIFixSGPRCopies handling of STRICT_WWM and friends (#142122)

SIFixSGPRCopies handled STRICT_WWM (and similar WWM/WQM pseudos) like a
COPY. In particular, if the source was a VGPR and the result was an
SGPR, lowerVGPR2SGPRCopies would replace it with a readfirstlane,
erasing the original pseudo and hence sabotaging the WWM region marking
which is supposed to be performed by SIWholeQuadMode.

Fix this by handling it more like INSERT_SUBREG, PHI and REG_SEQUENCE:
if the source is a VGPR then move the result to a VGPR, and keep the
pseudo.

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

@@ -634,11 +634,7 @@ bool SIFixSGPRCopies::run(MachineFunction &MF) {
       switch (MI.getOpcode()) {
       default:
         continue;
-      case AMDGPU::COPY:
+      case AMDGPU::COPY: {
-      case AMDGPU::WQM:
-      case AMDGPU::STRICT_WQM:
-      case AMDGPU::SOFT_WQM:
-      case AMDGPU::STRICT_WWM: {
         const TargetRegisterClass *SrcRC, *DstRC;
         std::tie(SrcRC, DstRC) = getCopyRegClasses(MI, *TRI, *MRI);
 
@@ -662,6 +658,10 @@ bool SIFixSGPRCopies::run(MachineFunction &MF) {
 
         break;
       }
+      case AMDGPU::WQM:
+      case AMDGPU::STRICT_WQM:
+      case AMDGPU::SOFT_WQM:
+      case AMDGPU::STRICT_WWM:
       case AMDGPU::INSERT_SUBREG:
       case AMDGPU::PHI:
       case AMDGPU::REG_SEQUENCE: {

llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-wwm.ll

@@ -0,0 +1,40 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 < %s | FileCheck %s
+
+define amdgpu_gs i32 @main() {
+; CHECK-LABEL: main:
+; CHECK:       ; %bb.0: ; %bb
+; CHECK-NEXT:    s_bitcmp1_b32 0, 0
+; CHECK-NEXT:    s_mov_b32 s0, 0
+; CHECK-NEXT:    s_cselect_b32 s1, -1, 0
+; CHECK-NEXT:    s_or_saveexec_b32 s2, -1
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s1
+; CHECK-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; CHECK-NEXT:    v_readfirstlane_b32 s1, v0
+; CHECK-NEXT:    s_mov_b32 exec_lo, s2
+; CHECK-NEXT:    s_or_b32 s0, s0, s1
+; CHECK-NEXT:    s_wait_alu 0xfffe
+; CHECK-NEXT:    s_bitcmp1_b32 s0, 0
+; CHECK-NEXT:    s_cselect_b32 s0, -1, 0
+; CHECK-NEXT:    s_wait_alu 0xfffe
+; CHECK-NEXT:    s_xor_b32 s0, s0, -1
+; CHECK-NEXT:    s_wait_alu 0xfffe
+; CHECK-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; CHECK-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; CHECK-NEXT:    v_readfirstlane_b32 s0, v1
+; CHECK-NEXT:    s_wait_alu 0xf1ff
+; CHECK-NEXT:    ; return to shader part epilog
+bb:
+  %i = call i1 @llvm.amdgcn.readfirstlane.i1(i1 false)
+  br label %bb1
+
+bb1:
+  %i2 = zext i1 %i to i32
+  %i3 = call i32 @llvm.amdgcn.wwm.i32(i32 0)
+  %i4 = call i32 @llvm.amdgcn.wwm.i32(i32 %i2)
+  %i5 = trunc i32 %i4 to i1
+  %i6 = trunc i32 %i3 to i1
+  %i7 = or i1 %i6, %i5
+  %i8 = select i1 %i7, i32 0, i32 1
+  ret i32 %i8
+}

llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-wwm.mir

@@ -0,0 +1,27 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -run-pass=si-fix-sgpr-copies %s -o - | FileCheck %s
+
+---
+name: main
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: main
+    ; CHECK: [[DEF:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
+    ; CHECK-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; CHECK-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[DEF1]], implicit $exec
+    ; CHECK-NEXT: early-clobber %2:sreg_32 = STRICT_WWM killed undef [[V_READFIRSTLANE_B32_]], implicit $exec
+    ; CHECK-NEXT: [[S_OR_B32_:%[0-9]+]]:sreg_32 = S_OR_B32 killed undef [[DEF]], killed undef %2, implicit-def dead $scc
+    ; CHECK-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32 = S_AND_B32 1, killed undef [[S_OR_B32_]], implicit-def dead $scc
+    ; CHECK-NEXT: S_CMP_EQ_U32 killed undef [[S_AND_B32_]], 1, implicit-def $scc
+    ; CHECK-NEXT: [[S_XOR_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_XOR_B32 killed undef [[S_AND_B32_]], killed undef [[S_AND_B32_]], implicit-def dead $scc
+    ; CHECK-NEXT: SI_RETURN_TO_EPILOG undef $sgpr0
+    %0:sreg_32 = IMPLICIT_DEF
+    %1:vgpr_32 = IMPLICIT_DEF
+    early-clobber %2:sreg_32 = STRICT_WWM killed undef %1, implicit $exec
+    %3:sreg_32 = S_OR_B32 killed undef %0, killed undef %2, implicit-def dead $scc
+    %4:sreg_32 = S_AND_B32 1, killed undef %3, implicit-def dead $scc
+    S_CMP_EQ_U32 killed undef %4, 1, implicit-def $scc
+    %5:sreg_32_xm0_xexec = S_XOR_B32 killed undef %4, killed undef %4, implicit-def dead $scc
+    SI_RETURN_TO_EPILOG undef $sgpr0
+...