AMDGPU: Fix not using s33 for scratch wave offset in kernels

Fixes missing piece from r363990.

llvm-svn: 364099

lib/Target/AMDGPU/SIISelLowering.cpp

@@ -1829,11 +1829,8 @@ static void reservePrivateMemoryRegs(const TargetMachine &TM,
     Info.setScratchRSrcReg(ReservedBufferReg);
   }
 
-  // This should be accurate for kernels even before the frame is finalized.
-  const bool HasFP = ST.getFrameLowering()->hasFP(MF);
-  if (HasFP) {
-    unsigned ReservedOffsetReg =
-        TRI.reservedPrivateSegmentWaveByteOffsetReg(MF);
+  // hasFP should be accurate for kernels even before the frame is finalized.
+  if (ST.getFrameLowering()->hasFP(MF)) {
     MachineRegisterInfo &MRI = MF.getRegInfo();
 
     // Try to use s32 as the SP, but move it if it would interfere with input
@@ -1860,8 +1857,15 @@ static void reservePrivateMemoryRegs(const TargetMachine &TM,
         report_fatal_error("failed to find register for SP");
     }
 
-    Info.setScratchWaveOffsetReg(ReservedOffsetReg);
-    Info.setFrameOffsetReg(ReservedOffsetReg);
+    if (MFI.hasCalls()) {
+      Info.setScratchWaveOffsetReg(AMDGPU::SGPR33);
+      Info.setFrameOffsetReg(AMDGPU::SGPR33);
+    } else {
+      unsigned ReservedOffsetReg =
+        TRI.reservedPrivateSegmentWaveByteOffsetReg(MF);
+      Info.setScratchWaveOffsetReg(ReservedOffsetReg);
+      Info.setFrameOffsetReg(ReservedOffsetReg);
+    }
   } else if (RequiresStackAccess) {
     assert(!MFI.hasCalls());
     // We know there are accesses and they will be done relative to SP, so just

test/CodeGen/AMDGPU/call-preserved-registers.ll

@@ -104,9 +104,9 @@ define amdgpu_kernel void @test_call_void_func_void_clobber_vcc(i32 addrspace(1)
 }
 
 ; GCN-LABEL: {{^}}test_call_void_func_void_mayclobber_s31:
-; GCN: s_mov_b32 s33, s31
+; GCN: s_mov_b32 s34, s31
 ; GCN-NEXT: s_swappc_b64
-; GCN-NEXT: s_mov_b32 s31, s33
+; GCN-NEXT: s_mov_b32 s31, s34
 define amdgpu_kernel void @test_call_void_func_void_mayclobber_s31(i32 addrspace(1)* %out) #0 {
   %s31 = call i32 asm sideeffect "; def $0", "={s31}"()
   call void @external_void_func_void()
@@ -128,15 +128,14 @@ define amdgpu_kernel void @test_call_void_func_void_mayclobber_v31(i32 addrspace
 ; FIXME: What is the expected behavior for reserved registers here?
 
 ; GCN-LABEL: {{^}}test_call_void_func_void_preserves_s33:
-; GCN: s_mov_b32 s34, s9
-; GCN-NOT: s33
+; GCN: s_mov_b32 s33, s9
+; GCN: s_mov_b32 s32, s33
 ; GCN: #ASMSTART
 ; GCN-NEXT: ; def s33
 ; GCN-NEXT: #ASMEND
 ; GCN: s_getpc_b64 s[4:5]
 ; GCN-NEXT: s_add_u32 s4, s4, external_void_func_void@rel32@lo+4
 ; GCN-NEXT: s_addc_u32 s5, s5, external_void_func_void@rel32@hi+4
-; GCN: s_mov_b32 s32, s34
 ; GCN: s_swappc_b64 s[30:31], s[4:5]
 ; GCN: ;;#ASMSTART
 ; GCN-NEXT: ; use s33

test/CodeGen/AMDGPU/ipra.ll

@@ -30,7 +30,7 @@ define hidden void @func() #1 {
 ; GCN-NOT: writelane
 ; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, v8
 
-; GCN: ; NumSgprs: 37
+; GCN: ; NumSgprs: 38
 ; GCN: ; NumVgprs: 9
 define amdgpu_kernel void @kernel_call() #0 {
   %vgpr = load volatile i32, i32 addrspace(1)* undef

test/CodeGen/AMDGPU/stack-realign.ll

@@ -91,7 +91,8 @@ define void @force_realign4(i32 %idx) #1 {
 }
 
 ; GCN-LABEL: {{^}}kernel_call_align16_from_8:
-; GCN: s_add_u32 s32, s8, 0x400{{$}}
+; GCN: s_mov_b32 s33, s7{{$}}
+; GCN-NEXT: s_add_u32 s32, s33, 0x400{{$}}
 ; GCN-NOT: s32
 ; GCN: s_swappc_b64
 define amdgpu_kernel void @kernel_call_align16_from_8() #0 {
@@ -103,7 +104,8 @@ define amdgpu_kernel void @kernel_call_align16_from_8() #0 {
 
 ; The call sequence should keep the stack on call aligned to 4
 ; GCN-LABEL: {{^}}kernel_call_align16_from_5:
-; GCN: s_add_u32 s32, s8, 0x400
+; GCN: s_mov_b32 s33, s7{{$}}
+; GCN-NEXT: s_add_u32 s32, s33, 0x400
 ; GCN: s_swappc_b64
 define amdgpu_kernel void @kernel_call_align16_from_5() {
   %alloca0 = alloca i8, align 1, addrspace(5)
@@ -114,7 +116,8 @@ define amdgpu_kernel void @kernel_call_align16_from_5() {
 }
 
 ; GCN-LABEL: {{^}}kernel_call_align4_from_5:
-; GCN: s_add_u32 s32, s8, 0x400
+; GCN: s_mov_b32 s33, s7{{$}}
+; GCN: s_add_u32 s32, s33, 0x400
 ; GCN: s_swappc_b64
 define amdgpu_kernel void @kernel_call_align4_from_5() {
   %alloca0 = alloca i8, align 1, addrspace(5)