AMDGPU: Fix sext_inreg for i1 in i16

[arsenm]

Summary

This produces worse code when i16 is legal, mostly
due to combines getting confused by conversions inserted
for uniform 16-bit operations.

[arsenm]

ping

[arsenm]

r291717

lib/Target/AMDGPU/SIInstructions.td

 
 // Handle sext_inreg in i64
 def : Pat <
   (i64 (sext_inreg i64:$src, i1)),
   (S_BFE_I64 i64:$src, (i32 0x10000)) // 0 | 1 << 16
 >;
 
 def : Pat <
+  (i16 (sext_inreg i16:$src, i1)),
+  (S_BFE_I32 $src, (i32 0x00010000)) // 0 | 1 << 16
+>;
+
+def : Pat <
   (i16 (sext_inreg i16:$src, i8)),
   (S_BFE_I32 $src, (i32 0x80000)) // 0 | 8 << 16
 >;
 
 def : Pat <
   (i64 (sext_inreg i64:$src, i8)),
   (S_BFE_I64 i64:$src, (i32 0x80000)) // 0 | 8 << 16
 >;

test/CodeGen/AMDGPU/sext-in-reg.ll

 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI -check-prefix=FUNC %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI -check-prefix=FUNC %s
 ; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
 
+; FIXME: i16 promotion pass ruins the scalar cases when legal.
+
 ; FUNC-LABEL: {{^}}sext_in_reg_i1_i32:
 ; GCN: s_load_dword [[ARG:s[0-9]+]],
 ; GCN: s_bfe_i32 [[SEXTRACT:s[0-9]+]], [[ARG]], 0x10000
 ; GCN: v_mov_b32_e32 [[EXTRACT:v[0-9]+]], [[SEXTRACT]]
 ; GCN: buffer_store_dword [[EXTRACT]],
 
 ; EG: MEM_{{.*}} STORE_{{.*}} [[RES:T[0-9]+\.[XYZW]]], [[ADDR:T[0-9]+.[XYZW]]]
 ; EG: LSHR * [[ADDR]]
define void @v_sext_in_reg_i32_to_i64_move_use(i64 addrspace(1)* %out, i64 addrspace(1)* %aptr, i64 addrspace(1)* %bptr, i64 %s.val) #0 {
   %c = shl i64 %a, %b
   %shl = shl i64 %c, 32
   %ashr = ashr i64 %shl, 32
   %and = and i64 %ashr, %s.val
   store i64 %and, i64 addrspace(1)* %out.gep, align 8
   ret void
 }
 
+; FUNC-LABEL: {{^}}s_sext_in_reg_i1_i16:
+; GCN: s_load_dword [[VAL:s[0-9]+]]
+
+; SI: s_bfe_i32 [[BFE:s[0-9]+]], [[VAL]], 0x10000
+; SI: v_mov_b32_e32 [[VBFE:v[0-9]+]], [[BFE]]
+; SI: buffer_store_short [[VBFE]]
+
+; VI: s_lshl_b32 s{{[0-9]+}}, s{{[0-9]+}}, 15
+; VI: s_sext_i32_i16 s{{[0-9]+}}, s{{[0-9]+}}
+; VI: s_lshr_b32 s{{[0-9]+}}, s{{[0-9]+}}, 15
+define void @s_sext_in_reg_i1_i16(i16 addrspace(1)* %out, i32 addrspace(2)* %ptr) #0 {
+  %ld = load i32, i32 addrspace(2)* %ptr
+  %in = trunc i32 %ld to i16
+  %shl = shl i16 %in, 15
+  %sext = ashr i16 %shl, 15
+  store i16 %sext, i16 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}s_sext_in_reg_i2_i16:
+; GCN: s_load_dword [[VAL:s[0-9]+]]
+
+; SI: s_bfe_i32 [[BFE:s[0-9]+]], [[VAL]], 0x20000
+; SI: v_mov_b32_e32 [[VBFE:v[0-9]+]], [[BFE]]
+; SI: buffer_store_short [[VBFE]]
+
+; VI: s_lshl_b32 s{{[0-9]+}}, s{{[0-9]+}}, 14
+; VI: s_sext_i32_i16 s{{[0-9]+}}, s{{[0-9]+}}
+; VI: s_lshr_b32 s{{[0-9]+}}, s{{[0-9]+}}, 14
+define void @s_sext_in_reg_i2_i16(i16 addrspace(1)* %out, i32 addrspace(2)* %ptr) #0 {
+  %ld = load i32, i32 addrspace(2)* %ptr
+  %in = trunc i32 %ld to i16
+  %shl = shl i16 %in, 14
+  %sext = ashr i16 %shl, 14
+  store i16 %sext, i16 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}v_sext_in_reg_i1_i16:
+; GCN: {{buffer|flat}}_load_ushort [[VAL:v[0-9]+]]
+; GCN: v_bfe_i32 [[BFE:v[0-9]+]], [[VAL]], 0, 1{{$}}
+
+; GCN: ds_write_b16 v{{[0-9]+}}, [[BFE]]
+define void @v_sext_in_reg_i1_i16(i16 addrspace(3)* %out, i16 addrspace(1)* %ptr) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x()
+  %gep = getelementptr i16, i16 addrspace(1)* %ptr, i32 %tid
+  %out.gep = getelementptr i16, i16 addrspace(3)* %out, i32 %tid
+
+  %in = load i16, i16 addrspace(1)* %gep
+  %shl = shl i16 %in, 15
+  %sext = ashr i16 %shl, 15
+  store i16 %sext, i16 addrspace(3)* %out.gep
+  ret void
+}
+
+; FUNC-LABEL: {{^}}v_sext_in_reg_i1_i16_nonload:
+; GCN: {{buffer|flat}}_load_ushort [[VAL0:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_ushort [[VAL1:v[0-9]+]]
+
+; SI: v_lshlrev_b32_e32 [[REG:v[0-9]+]], [[VAL1]], [[VAL0]]
+; VI: v_lshlrev_b16_e32 [[REG:v[0-9]+]], [[VAL1]], [[VAL0]]
+
+; GCN: v_bfe_i32 [[BFE:v[0-9]+]], [[REG]], 0, 1{{$}}
+; GCN: ds_write_b16 v{{[0-9]+}}, [[BFE]]
+define void @v_sext_in_reg_i1_i16_nonload(i16 addrspace(3)* %out, i16 addrspace(1)* %aptr, i16 addrspace(1)* %bptr, i16 %s.val) nounwind {
+  %tid = call i32 @llvm.r600.read.tidig.x()
+  %a.gep = getelementptr i16, i16 addrspace(1)* %aptr, i32 %tid
+  %b.gep = getelementptr i16, i16 addrspace(1)* %bptr, i32 %tid
+  %out.gep = getelementptr i16, i16 addrspace(3)* %out, i32 %tid
+  %a = load volatile i16, i16 addrspace(1)* %a.gep, align 2
+  %b = load volatile i16, i16 addrspace(1)* %b.gep, align 2
+
+  %c = shl i16 %a, %b
+  %shl = shl i16 %c, 15
+  %ashr = ashr i16 %shl, 15
+
+  store i16 %ashr, i16 addrspace(3)* %out.gep, align 2
+  ret void
+}
+
+; FUNC-LABEL: {{^}}s_sext_in_reg_i2_i16_arg:
+; GCN: s_load_dword [[VAL:s[0-9]+]]
+
+; SI: s_bfe_i32 [[BFE:s[0-9]+]], [[VAL]], 0x20000
+; SI: v_mov_b32_e32 [[VBFE:v[0-9]+]], [[BFE]]
+; SI: buffer_store_short [[VBFE]]
+
+; VI: s_lshl_b32 s{{[0-9]+}}, s{{[0-9]+}}, 14{{$}}
+; VI: s_sext_i32_i16 s{{[0-9]+}}, s{{[0-9]+}}
+; VI: s_lshr_b32 s{{[0-9]+}}, s{{[0-9]+}}, 14{{$}}
+define void @s_sext_in_reg_i2_i16_arg(i16 addrspace(1)* %out, i16 %in) #0 {
+  %shl = shl i16 %in, 14
+  %sext = ashr i16 %shl, 14
+  store i16 %sext, i16 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}s_sext_in_reg_i8_i16_arg:
+; GCN: s_load_dword [[VAL:s[0-9]+]]
+
+; SI: s_sext_i32_i8 [[SSEXT:s[0-9]+]], [[VAL]]
+; SI: v_mov_b32_e32 [[VSEXT:v[0-9]+]], [[SSEXT]]
+; SI: buffer_store_short [[VBFE]]
+
+; VI: s_lshl_b32 s{{[0-9]+}}, s{{[0-9]+}}, 8{{$}}
+; VI: s_sext_i32_i16 s{{[0-9]+}}, s{{[0-9]+}}
+; VI: s_lshr_b32 s{{[0-9]+}}, s{{[0-9]+}}, 8{{$}}
+define void @s_sext_in_reg_i8_i16_arg(i16 addrspace(1)* %out, i16 %in) #0 {
+  %shl = shl i16 %in, 8
+  %sext = ashr i16 %shl, 8
+  store i16 %sext, i16 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}s_sext_in_reg_i15_i16_arg:
+; GCN: s_load_dword [[VAL:s[0-9]+]]
+
+; SI: s_bfe_i32 [[BFE:s[0-9]+]], [[VAL]], 0xf0000
+; SI: v_mov_b32_e32 [[VBFE:v[0-9]+]], [[BFE]]
+; SI: buffer_store_short [[VBFE]]
+
+; VI: s_lshl_b32 s{{[0-9]+}}, s{{[0-9]+}}, 1{{$}}
+; VI: s_sext_i32_i16 s{{[0-9]+}}, s{{[0-9]+}}
+; VI: s_lshr_b32 s{{[0-9]+}}, s{{[0-9]+}}, 1{{$}}
+define void @s_sext_in_reg_i15_i16_arg(i16 addrspace(1)* %out, i16 %in) #0 {
+  %shl = shl i16 %in, 1
+  %sext = ashr i16 %shl, 1
+  store i16 %sext, i16 addrspace(1)* %out
+  ret void
+}
+
 declare i32 @llvm.r600.read.tidig.x() #1
 
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone }