aboutsummaryrefslogtreecommitdiffstats
path: root/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
diff options
context:
space:
mode:
Diffstat (limited to 'test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll')
-rw-r--r--test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll430
1 files changed, 226 insertions, 204 deletions
diff --git a/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll b/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
index 1a62253..0794ac4 100644
--- a/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
+++ b/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
@@ -3,8 +3,8 @@
declare i32 @llvm.AMDGPU.bfe.u32(i32, i32, i32) nounwind readnone
-; FUNC-LABEL: @bfe_u32_arg_arg_arg
-; SI: V_BFE_U32
+; FUNC-LABEL: {{^}}bfe_u32_arg_arg_arg:
+; SI: v_bfe_u32
; EG: BFE_UINT
define void @bfe_u32_arg_arg_arg(i32 addrspace(1)* %out, i32 %src0, i32 %src1, i32 %src2) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 %src0, i32 %src1, i32 %src1) nounwind readnone
@@ -12,8 +12,8 @@ define void @bfe_u32_arg_arg_arg(i32 addrspace(1)* %out, i32 %src0, i32 %src1, i
ret void
}
-; FUNC-LABEL: @bfe_u32_arg_arg_imm
-; SI: V_BFE_U32
+; FUNC-LABEL: {{^}}bfe_u32_arg_arg_imm:
+; SI: v_bfe_u32
; EG: BFE_UINT
define void @bfe_u32_arg_arg_imm(i32 addrspace(1)* %out, i32 %src0, i32 %src1) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 %src0, i32 %src1, i32 123) nounwind readnone
@@ -21,8 +21,8 @@ define void @bfe_u32_arg_arg_imm(i32 addrspace(1)* %out, i32 %src0, i32 %src1) n
ret void
}
-; FUNC-LABEL: @bfe_u32_arg_imm_arg
-; SI: V_BFE_U32
+; FUNC-LABEL: {{^}}bfe_u32_arg_imm_arg:
+; SI: v_bfe_u32
; EG: BFE_UINT
define void @bfe_u32_arg_imm_arg(i32 addrspace(1)* %out, i32 %src0, i32 %src2) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 %src0, i32 123, i32 %src2) nounwind readnone
@@ -30,8 +30,8 @@ define void @bfe_u32_arg_imm_arg(i32 addrspace(1)* %out, i32 %src0, i32 %src2) n
ret void
}
-; FUNC-LABEL: @bfe_u32_imm_arg_arg
-; SI: V_BFE_U32
+; FUNC-LABEL: {{^}}bfe_u32_imm_arg_arg:
+; SI: v_bfe_u32
; EG: BFE_UINT
define void @bfe_u32_imm_arg_arg(i32 addrspace(1)* %out, i32 %src1, i32 %src2) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 123, i32 %src1, i32 %src2) nounwind readnone
@@ -39,9 +39,9 @@ define void @bfe_u32_imm_arg_arg(i32 addrspace(1)* %out, i32 %src1, i32 %src2) n
ret void
}
-; FUNC-LABEL: @bfe_u32_arg_0_width_reg_offset
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_arg_0_width_reg_offset:
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_arg_0_width_reg_offset(i32 addrspace(1)* %out, i32 %src0, i32 %src1) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 %src0, i32 %src1, i32 0) nounwind readnone
@@ -49,9 +49,9 @@ define void @bfe_u32_arg_0_width_reg_offset(i32 addrspace(1)* %out, i32 %src0, i
ret void
}
-; FUNC-LABEL: @bfe_u32_arg_0_width_imm_offset
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_arg_0_width_imm_offset:
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_arg_0_width_imm_offset(i32 addrspace(1)* %out, i32 %src0, i32 %src1) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 %src0, i32 8, i32 0) nounwind readnone
@@ -59,10 +59,10 @@ define void @bfe_u32_arg_0_width_imm_offset(i32 addrspace(1)* %out, i32 %src0, i
ret void
}
-; FUNC-LABEL: @bfe_u32_zextload_i8
-; SI: BUFFER_LOAD_UBYTE
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_zextload_i8:
+; SI: buffer_load_ubyte
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_zextload_i8(i32 addrspace(1)* %out, i8 addrspace(1)* %in) nounwind {
%load = load i8 addrspace(1)* %in
%ext = zext i8 %load to i32
@@ -71,12 +71,12 @@ define void @bfe_u32_zextload_i8(i32 addrspace(1)* %out, i8 addrspace(1)* %in) n
ret void
}
-; FUNC-LABEL: @bfe_u32_zext_in_reg_i8
-; SI: BUFFER_LOAD_DWORD
-; SI: V_ADD_I32
-; SI-NEXT: V_AND_B32_e32
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_zext_in_reg_i8:
+; SI: buffer_load_dword
+; SI: v_add_i32
+; SI-NEXT: v_and_b32_e32
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_zext_in_reg_i8(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%load = load i32 addrspace(1)* %in, align 4
%add = add i32 %load, 1
@@ -86,12 +86,12 @@ define void @bfe_u32_zext_in_reg_i8(i32 addrspace(1)* %out, i32 addrspace(1)* %i
ret void
}
-; FUNC-LABEL: @bfe_u32_zext_in_reg_i16
-; SI: BUFFER_LOAD_DWORD
-; SI: V_ADD_I32
-; SI-NEXT: V_AND_B32_e32
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_zext_in_reg_i16:
+; SI: buffer_load_dword
+; SI: v_add_i32
+; SI-NEXT: v_and_b32_e32
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_zext_in_reg_i16(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%load = load i32 addrspace(1)* %in, align 4
%add = add i32 %load, 1
@@ -101,11 +101,11 @@ define void @bfe_u32_zext_in_reg_i16(i32 addrspace(1)* %out, i32 addrspace(1)* %
ret void
}
-; FUNC-LABEL: @bfe_u32_zext_in_reg_i8_offset_1
-; SI: BUFFER_LOAD_DWORD
-; SI: V_ADD_I32
-; SI: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_zext_in_reg_i8_offset_1:
+; SI: buffer_load_dword
+; SI: v_add_i32
+; SI: bfe
+; SI: s_endpgm
define void @bfe_u32_zext_in_reg_i8_offset_1(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%load = load i32 addrspace(1)* %in, align 4
%add = add i32 %load, 1
@@ -115,12 +115,12 @@ define void @bfe_u32_zext_in_reg_i8_offset_1(i32 addrspace(1)* %out, i32 addrspa
ret void
}
-; FUNC-LABEL: @bfe_u32_zext_in_reg_i8_offset_3
-; SI: BUFFER_LOAD_DWORD
-; SI: V_ADD_I32
-; SI-NEXT: V_AND_B32_e32 {{v[0-9]+}}, 0xf8
-; SI-NEXT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_zext_in_reg_i8_offset_3:
+; SI: buffer_load_dword
+; SI: v_add_i32
+; SI-NEXT: v_and_b32_e32 {{v[0-9]+}}, 0xf8
+; SI-NEXT: bfe
+; SI: s_endpgm
define void @bfe_u32_zext_in_reg_i8_offset_3(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%load = load i32 addrspace(1)* %in, align 4
%add = add i32 %load, 1
@@ -130,12 +130,12 @@ define void @bfe_u32_zext_in_reg_i8_offset_3(i32 addrspace(1)* %out, i32 addrspa
ret void
}
-; FUNC-LABEL: @bfe_u32_zext_in_reg_i8_offset_7
-; SI: BUFFER_LOAD_DWORD
-; SI: V_ADD_I32
-; SI-NEXT: V_AND_B32_e32 {{v[0-9]+}}, 0x80
-; SI-NEXT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_zext_in_reg_i8_offset_7:
+; SI: buffer_load_dword
+; SI: v_add_i32
+; SI-NEXT: v_and_b32_e32 {{v[0-9]+}}, 0x80
+; SI-NEXT: bfe
+; SI: s_endpgm
define void @bfe_u32_zext_in_reg_i8_offset_7(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%load = load i32 addrspace(1)* %in, align 4
%add = add i32 %load, 1
@@ -145,11 +145,11 @@ define void @bfe_u32_zext_in_reg_i8_offset_7(i32 addrspace(1)* %out, i32 addrspa
ret void
}
-; FUNC-LABEL: @bfe_u32_zext_in_reg_i16_offset_8
-; SI: BUFFER_LOAD_DWORD
-; SI: V_ADD_I32
-; SI-NEXT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_zext_in_reg_i16_offset_8:
+; SI: buffer_load_dword
+; SI: v_add_i32
+; SI-NEXT: bfe
+; SI: s_endpgm
define void @bfe_u32_zext_in_reg_i16_offset_8(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%load = load i32 addrspace(1)* %in, align 4
%add = add i32 %load, 1
@@ -159,10 +159,10 @@ define void @bfe_u32_zext_in_reg_i16_offset_8(i32 addrspace(1)* %out, i32 addrsp
ret void
}
-; FUNC-LABEL: @bfe_u32_test_1
-; SI: BUFFER_LOAD_DWORD
-; SI: V_AND_B32_e32 {{v[0-9]+}}, 1, {{v[0-9]+}}
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_1:
+; SI: buffer_load_dword
+; SI: v_and_b32_e32 {{v[0-9]+}}, 1, {{v[0-9]+}}
+; SI: s_endpgm
; EG: AND_INT T{{[0-9]\.[XYZW]}}, T{{[0-9]\.[XYZW]}}, 1,
define void @bfe_u32_test_1(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
@@ -187,13 +187,13 @@ define void @bfe_u32_test_3(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounw
ret void
}
-; FUNC-LABEL: @bfe_u32_test_4
-; SI-NOT: LSHL
-; SI-NOT: SHR
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_4:
+; SI-NOT: lshl
+; SI-NOT: shr
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
define void @bfe_u32_test_4(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%shl = shl i32 %x, 31
@@ -203,12 +203,12 @@ define void @bfe_u32_test_4(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounw
ret void
}
-; FUNC-LABEL: @bfe_u32_test_5
-; SI: BUFFER_LOAD_DWORD
-; SI-NOT: LSHL
-; SI-NOT: SHR
-; SI: V_BFE_I32 {{v[0-9]+}}, {{v[0-9]+}}, 0, 1
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_5:
+; SI: buffer_load_dword
+; SI-NOT: lshl
+; SI-NOT: shr
+; SI: v_bfe_i32 {{v[0-9]+}}, {{v[0-9]+}}, 0, 1
+; SI: s_endpgm
define void @bfe_u32_test_5(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%shl = shl i32 %x, 31
@@ -218,10 +218,10 @@ define void @bfe_u32_test_5(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounw
ret void
}
-; FUNC-LABEL: @bfe_u32_test_6
-; SI: V_LSHLREV_B32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
-; SI: V_LSHRREV_B32_e32 v{{[0-9]+}}, 1, v{{[0-9]+}}
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_6:
+; SI: v_lshlrev_b32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
+; SI: v_lshrrev_b32_e32 v{{[0-9]+}}, 1, v{{[0-9]+}}
+; SI: s_endpgm
define void @bfe_u32_test_6(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%shl = shl i32 %x, 31
@@ -230,10 +230,10 @@ define void @bfe_u32_test_6(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounw
ret void
}
-; FUNC-LABEL: @bfe_u32_test_7
-; SI: V_LSHLREV_B32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_7:
+; SI: v_lshlrev_b32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_test_7(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%shl = shl i32 %x, 31
@@ -242,11 +242,11 @@ define void @bfe_u32_test_7(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounw
ret void
}
-; FUNC-LABEL: @bfe_u32_test_8
-; SI-NOT: BFE
-; SI: V_AND_B32_e32 {{v[0-9]+}}, 1, {{v[0-9]+}}
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_8:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_and_b32_e32 {{v[0-9]+}}, 1, {{v[0-9]+}}
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_test_8(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%shl = shl i32 %x, 31
@@ -255,11 +255,11 @@ define void @bfe_u32_test_8(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounw
ret void
}
-; FUNC-LABEL: @bfe_u32_test_9
-; SI-NOT: BFE
-; SI: V_LSHRREV_B32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_9:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_lshrrev_b32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_test_9(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %x, i32 31, i32 1)
@@ -267,11 +267,11 @@ define void @bfe_u32_test_9(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounw
ret void
}
-; FUNC-LABEL: @bfe_u32_test_10
-; SI-NOT: BFE
-; SI: V_LSHRREV_B32_e32 v{{[0-9]+}}, 1, v{{[0-9]+}}
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_10:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_lshrrev_b32_e32 v{{[0-9]+}}, 1, v{{[0-9]+}}
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_test_10(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %x, i32 1, i32 31)
@@ -279,11 +279,11 @@ define void @bfe_u32_test_10(i32 addrspace(1)* %out, i32 addrspace(1)* %in) noun
ret void
}
-; FUNC-LABEL: @bfe_u32_test_11
-; SI-NOT: BFE
-; SI: V_LSHRREV_B32_e32 v{{[0-9]+}}, 8, v{{[0-9]+}}
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_11:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_lshrrev_b32_e32 v{{[0-9]+}}, 8, v{{[0-9]+}}
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_test_11(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %x, i32 8, i32 24)
@@ -291,11 +291,11 @@ define void @bfe_u32_test_11(i32 addrspace(1)* %out, i32 addrspace(1)* %in) noun
ret void
}
-; FUNC-LABEL: @bfe_u32_test_12
-; SI-NOT: BFE
-; SI: V_LSHRREV_B32_e32 v{{[0-9]+}}, 24, v{{[0-9]+}}
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_12:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_lshrrev_b32_e32 v{{[0-9]+}}, 24, v{{[0-9]+}}
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_test_12(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %x, i32 24, i32 8)
@@ -303,10 +303,10 @@ define void @bfe_u32_test_12(i32 addrspace(1)* %out, i32 addrspace(1)* %in) noun
ret void
}
-; FUNC-LABEL: @bfe_u32_test_13
+; FUNC-LABEL: {{^}}bfe_u32_test_13:
; V_ASHRREV_U32_e32 {{v[0-9]+}}, 31, {{v[0-9]+}}
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_test_13(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%shl = ashr i32 %x, 31
@@ -314,10 +314,10 @@ define void @bfe_u32_test_13(i32 addrspace(1)* %out, i32 addrspace(1)* %in) noun
store i32 %bfe, i32 addrspace(1)* %out, align 4 ret void
}
-; FUNC-LABEL: @bfe_u32_test_14
-; SI-NOT: LSHR
-; SI-NOT: BFE
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_test_14:
+; SI-NOT: lshr
+; SI-NOT: {{[^@]}}bfe
+; SI: s_endpgm
define void @bfe_u32_test_14(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%x = load i32 addrspace(1)* %in, align 4
%shl = lshr i32 %x, 31
@@ -325,11 +325,11 @@ define void @bfe_u32_test_14(i32 addrspace(1)* %out, i32 addrspace(1)* %in) noun
store i32 %bfe, i32 addrspace(1)* %out, align 4 ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_0
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_0:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_0(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 0, i32 0, i32 0) nounwind readnone
@@ -337,11 +337,11 @@ define void @bfe_u32_constant_fold_test_0(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_1
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_1:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_1(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 12334, i32 0, i32 0) nounwind readnone
@@ -349,11 +349,11 @@ define void @bfe_u32_constant_fold_test_1(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_2
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_2:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_2(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 0, i32 0, i32 1) nounwind readnone
@@ -361,11 +361,11 @@ define void @bfe_u32_constant_fold_test_2(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_3
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 1
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_3:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 1
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_3(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 1, i32 0, i32 1) nounwind readnone
@@ -373,11 +373,11 @@ define void @bfe_u32_constant_fold_test_3(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_4
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], -1
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_4:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], -1
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_4(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 4294967295, i32 0, i32 1) nounwind readnone
@@ -385,11 +385,11 @@ define void @bfe_u32_constant_fold_test_4(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_5
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 1
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_5:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 1
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_5(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 128, i32 7, i32 1) nounwind readnone
@@ -397,11 +397,11 @@ define void @bfe_u32_constant_fold_test_5(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_6
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0x80
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_6:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0x80
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_6(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 128, i32 0, i32 8) nounwind readnone
@@ -409,11 +409,11 @@ define void @bfe_u32_constant_fold_test_6(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_7
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0x7f
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_7:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0x7f
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_7(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 127, i32 0, i32 8) nounwind readnone
@@ -421,11 +421,11 @@ define void @bfe_u32_constant_fold_test_7(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_8
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 1
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_8:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 1
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_8(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 127, i32 6, i32 8) nounwind readnone
@@ -433,11 +433,11 @@ define void @bfe_u32_constant_fold_test_8(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_9
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 1
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_9:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 1
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFEfppppppppppppp
define void @bfe_u32_constant_fold_test_9(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 65536, i32 16, i32 8) nounwind readnone
@@ -445,11 +445,11 @@ define void @bfe_u32_constant_fold_test_9(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_10
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_10:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_10(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 65535, i32 16, i32 16) nounwind readnone
@@ -457,11 +457,11 @@ define void @bfe_u32_constant_fold_test_10(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_11
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 10
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_11:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 10
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_11(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 160, i32 4, i32 4) nounwind readnone
@@ -469,11 +469,11 @@ define void @bfe_u32_constant_fold_test_11(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_12
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_12:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_12(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 160, i32 31, i32 1) nounwind readnone
@@ -481,11 +481,11 @@ define void @bfe_u32_constant_fold_test_12(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_13
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 1
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_13:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 1
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_13(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 131070, i32 16, i32 16) nounwind readnone
@@ -493,11 +493,11 @@ define void @bfe_u32_constant_fold_test_13(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_14
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 40
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_14:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 40
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_14(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 160, i32 2, i32 30) nounwind readnone
@@ -505,11 +505,11 @@ define void @bfe_u32_constant_fold_test_14(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_15
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 10
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_15:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 10
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_15(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 160, i32 4, i32 28) nounwind readnone
@@ -517,11 +517,11 @@ define void @bfe_u32_constant_fold_test_15(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_16
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0x7f
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_16:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0x7f
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_16(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 4294967295, i32 1, i32 7) nounwind readnone
@@ -529,11 +529,11 @@ define void @bfe_u32_constant_fold_test_16(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_17
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0x7f
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_17:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0x7f
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_17(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 255, i32 1, i32 31) nounwind readnone
@@ -541,14 +541,36 @@ define void @bfe_u32_constant_fold_test_17(i32 addrspace(1)* %out) nounwind {
ret void
}
-; FUNC-LABEL: @bfe_u32_constant_fold_test_18
-; SI-NOT: BFE
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
-; SI: BUFFER_STORE_DWORD [[VREG]],
-; SI: S_ENDPGM
+; FUNC-LABEL: {{^}}bfe_u32_constant_fold_test_18:
+; SI-NOT: {{[^@]}}bfe
+; SI: v_mov_b32_e32 [[VREG:v[0-9]+]], 0
+; SI: buffer_store_dword [[VREG]],
+; SI: s_endpgm
; EG-NOT: BFE
define void @bfe_u32_constant_fold_test_18(i32 addrspace(1)* %out) nounwind {
%bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 255, i32 31, i32 1) nounwind readnone
store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
ret void
}
+
+; Make sure that SimplifyDemandedBits doesn't cause the and to be
+; reduced to the bits demanded by the bfe.
+
+; XXX: The operand to v_bfe_u32 could also just directly be the load register.
+; FUNC-LABEL: {{^}}simplify_bfe_u32_multi_use_arg:
+; SI: buffer_load_dword [[ARG:v[0-9]+]]
+; SI: v_and_b32_e32 [[AND:v[0-9]+]], 63, [[ARG]]
+; SI: v_bfe_u32 [[BFE:v[0-9]+]], [[AND]], 2, 2
+; SI-DAG: buffer_store_dword [[AND]]
+; SI-DAG: buffer_store_dword [[BFE]]
+; SI: s_endpgm
+define void @simplify_bfe_u32_multi_use_arg(i32 addrspace(1)* %out0,
+ i32 addrspace(1)* %out1,
+ i32 addrspace(1)* %in) nounwind {
+ %src = load i32 addrspace(1)* %in, align 4
+ %and = and i32 %src, 63
+ %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 %and, i32 2, i32 2) nounwind readnone
+ store i32 %bfe_u32, i32 addrspace(1)* %out0, align 4
+ store i32 %and, i32 addrspace(1)* %out1, align 4
+ ret void
+}
generated by cgit v1.1 at 2024-05-19 02:11:12 +0000