AMDGPU: Don't align callable functions to 256

[arsenm]

Add new test

[arsenm]

ping

[nhaehnle]

There might be a benefit to choosing a larger alignment, like 64 bytes, due to cache line alignment.

Apart from that, LGTM.

[arsenm]

r300720

lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp

void AMDGPUAsmPrinter::EmitGlobalVariable(const GlobalVariable *GV) {
   // Group segment variables aren't emitted in HSA.
   if (AMDGPU::isGroupSegment(GV, AMDGPUASI))
     return;
 
   AsmPrinter::EmitGlobalVariable(GV);
 }
 
 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
+  const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
 
   // The starting address of all shader programs must be 256 bytes aligned.
-  MF.setAlignment(8);
+  // Regular functions just need the basic required instruction alignment.
+  MF.setAlignment(MFI->isEntryFunction() ? 8 : 2);
 
   SetupMachineFunction(MF);
 
   const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
   MCContext &Context = getObjFileLowering().getContext();
   if (!STM.isAmdHsaOS()) {
     MCSectionELF *ConfigSection =
         Context.getELFSection(".AMDGPU.config", ELF::SHT_PROGBITS, 0);

test/CodeGen/AMDGPU/hsa-func-align.ll

+; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri < %s | FileCheck -check-prefix=HSA %s
+; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -filetype=obj < %s | llvm-readobj -symbols -s -sd | FileCheck -check-prefix=ELF %s
+
+; ELF: Section {
+; ELF: Name: .text
+; ELF: SHF_ALLOC (0x2)
+; ELF: SHF_EXECINSTR (0x4)
+; ELF: AddressAlignment: 32
+; ELF: }
+
+; HSA: .globl simple_align16
+; HSA: .p2align 5
+define void @simple_align16(i32 addrspace(1)* addrspace(2)* %ptr.out) align 32 {
+entry:
+  %out = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(2)* %ptr.out
+  store i32 0, i32 addrspace(1)* %out
+  ret void
+}

test/CodeGen/AMDGPU/hsa-func.ll

 ; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=kaveri | FileCheck --check-prefix=HSA %s
 ; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=kaveri -mattr=-flat-for-global | FileCheck --check-prefix=HSA-CI %s
 ; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=carrizo  | FileCheck --check-prefix=HSA %s
 ; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=carrizo -mattr=-flat-for-global | FileCheck --check-prefix=HSA-VI %s
 ; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=kaveri -filetype=obj | llvm-readobj -symbols -s -sd | FileCheck --check-prefix=ELF %s
 ; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=kaveri | llvm-mc -filetype=obj -triple amdgcn--amdhsa -mcpu=kaveri | llvm-readobj -symbols -s -sd | FileCheck %s --check-prefix=ELF
 
 ; The SHT_NOTE section contains the output from the .hsa_code_object_*
 ; directives.
 
 ; ELF: Section {
 ; ELF: Name: .text
 ; ELF: Type: SHT_PROGBITS (0x1)
 ; ELF: Flags [ (0x6)
 ; ELF: SHF_ALLOC (0x2)
 ; ELF: SHF_EXECINSTR (0x4)
+; ELF: AddressAlignment: 4
 ; ELF: }
 
 ; ELF: SHT_NOTE
 ; ELF: 0000: 04000000 08000000 01000000 414D4400
 ; ELF: 0010: 02000000 01000000 04000000 1B000000
 
 ; ELF: 0020: 03000000 414D4400 04000700 07000000
 ; ELF: 0030: 00000000 00000000 414D4400 414D4447
 ; ELF: 0040: 50550000
 
 ; ELF: Symbol {
 ; ELF: Name: simple
 ; ELF: Size: 304
 ; ELF: Type: Function (0x2)
 ; ELF: }
 
 ; HSA: .text
 ; HSA: .hsa_code_object_version 2,1
 ; HSA-CI: .hsa_code_object_isa 7,0,0,"AMD","AMDGPU"
 ; HSA-VI: .hsa_code_object_isa 8,0,1,"AMD","AMDGPU"
 
 ; HSA-NOT: .amdgpu_hsa_kernel simple
+; HSA: .globl simple
+; HSA: .p2align 2
 ; HSA: {{^}}simple:
 ; HSA: .amd_kernel_code_t
 ; HSA: enable_sgpr_private_segment_buffer = 0
 ; HSA: enable_sgpr_kernarg_segment_ptr = 0
 ; HSA: .end_amd_kernel_code_t
 ; HSA-NOT: s_load_dwordx2 s[{{[0-9]+:[0-9]+}}], s[4:5], 0x0
 
 ; Make sure we are setting the ATC bit:
 ; HSA-CI: s_mov_b32 s[[HI:[0-9]+]], 0x100f000
 ; On VI+ we also need to set MTYPE = 2
 ; HSA-VI: s_mov_b32 s[[HI:[0-9]+]], 0x1100f000
 ; Make sure we generate flat store for HSA
 ; HSA: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}
 
 ; HSA: .Lfunc_end0:
 ; HSA: .size   simple, .Lfunc_end0-simple
 
 define void @simple(i32 addrspace(1)* addrspace(2)* %ptr.out) {
 entry:
   %out = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(2)* %ptr.out
   store i32 0, i32 addrspace(1)* %out
   ret void
 }
+
+; Ignore explicit alignment that is too low.
+; HSA: .globl simple_align2
+; HSA: .p2align 2
+define void @simple_align2(i32 addrspace(1)* addrspace(2)* %ptr.out) align 2 {
+entry:
+  %out = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(2)* %ptr.out
+  store i32 0, i32 addrspace(1)* %out
+  ret void
+}