Skip to content

Commit 264b1b5

Browse files
ViacheslavRbigcbot
ViacheslavRb
authored and
igcbot
committed
Fixing ConstantCoalescing optimiztion
Fixing ConstantCoalescing optimiztion. The fix removes an unreasonable limitation on base address calculations for OCL shaders.
1 parent a9e1ef2 commit 264b1b5

File tree

3 files changed

+35
-41
lines changed

3 files changed

+35
-41
lines changed

‎IGC/Compiler/CISACodeGen/ConstantCoalescing.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1596,7 +1596,7 @@ bool ConstantCoalescing::DecomposePtrExp(
15961596
int64_t Offset = 0;
15971597
auto *Ptr = getPointerBaseWithConstantOffset(ptr_val, Offset, *dataLayout);
15981598

1599-
if (Ptr == ptr_val || Offset < 0)
1599+
if ((m_ctx->type != ShaderType::OPENCL_SHADER && Ptr == ptr_val) || Offset < 0)
16001600
return false;
16011601

16021602
buf_idxv = Ptr;

‎IGC/Compiler/tests/ConstantCoalescing/get-base-const-offset-typed-pointers.ll‎

Lines changed: 14 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -17,10 +17,10 @@
1717
; or a getelementptr instruction.
1818

1919
; Merge
20-
define void @test_merge(float addrspace(2)* %src) {
20+
define void @test_merge(i32 addrspace(2)* %src) {
2121
; CHECK-LABEL: define void @test_merge(
22-
; CHECK-SAME: float addrspace(2)* [[SRC:%.*]]) {
23-
; CHECK-NEXT: [[TMP1:%.*]] = ptrtoint float addrspace(2)* [[SRC]] to i64
22+
; CHECK-SAME: i32 addrspace(2)* [[SRC:%.*]]) {
23+
; CHECK-NEXT: [[TMP1:%.*]] = ptrtoint i32 addrspace(2)* [[SRC]] to i64
2424
; CHECK-NEXT: [[CHUNKPTR:%.*]] = inttoptr i64 [[TMP1]] to <2 x i32> addrspace(2)*
2525
; CHECK-NEXT: [[TMP2:%.*]] = load <2 x i32>, <2 x i32> addrspace(2)* [[CHUNKPTR]], align 4
2626
; CHECK-NEXT: [[TMP3:%.*]] = extractelement <2 x i32> [[TMP2]], i32 0
@@ -29,12 +29,11 @@ define void @test_merge(float addrspace(2)* %src) {
2929
; CHECK-NEXT: call void @use.i32(i32 [[TMP4]])
3030
; CHECK-NEXT: ret void
3131
;
32-
%1 = bitcast float addrspace(2)* %src to i32 addrspace(2)*
33-
%2 = getelementptr i32, i32 addrspace(2)* %1, i32 1
32+
%1 = getelementptr i32, i32 addrspace(2)* %src, i32 1
33+
%2 = load i32, i32 addrspace(2)* %src
3434
%3 = load i32, i32 addrspace(2)* %1
35-
%4 = load i32, i32 addrspace(2)* %2
35+
call void @use.i32(i32 %2)
3636
call void @use.i32(i32 %3)
37-
call void @use.i32(i32 %4)
3837
ret void
3938
}
4039

@@ -66,14 +65,13 @@ define void @test_nonconst_gep(i32 addrspace(2)* %src, i32 %off) {
6665
; CHECK-LABEL: define void @test_nonconst_gep(
6766
; CHECK-SAME: i32 addrspace(2)* [[SRC:%.*]], i32 [[OFF:%.*]]) {
6867
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr i32, i32 addrspace(2)* [[SRC]], i32 [[OFF]]
69-
; CHECK-NEXT: [[TMP2:%.*]] = load i32, i32 addrspace(2)* [[TMP1]], align 4
70-
; CHECK-NEXT: [[TMP3:%.*]] = ptrtoint i32 addrspace(2)* [[TMP1]] to i64
71-
; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[TMP3]], 4
72-
; CHECK-NEXT: [[CHUNKPTR:%.*]] = inttoptr i64 [[TMP4]] to <1 x i32> addrspace(2)*
73-
; CHECK-NEXT: [[TMP5:%.*]] = load <1 x i32>, <1 x i32> addrspace(2)* [[CHUNKPTR]], align 4
74-
; CHECK-NEXT: [[TMP6:%.*]] = extractelement <1 x i32> [[TMP5]], i32 0
75-
; CHECK-NEXT: call void @use.i32(i32 [[TMP2]])
76-
; CHECK-NEXT: call void @use.i32(i32 [[TMP6]])
68+
; CHECK-NEXT: [[TMP2:%.*]] = ptrtoint i32 addrspace(2)* [[TMP1]] to i64
69+
; CHECK-NEXT: [[CHUNKPTR:%.*]] = inttoptr i64 [[TMP2]] to <2 x i32> addrspace(2)*
70+
; CHECK-NEXT: [[TMP3:%.*]] = load <2 x i32>, <2 x i32> addrspace(2)* [[CHUNKPTR]], align 4
71+
; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x i32> [[TMP3]], i32 0
72+
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x i32> [[TMP3]], i32 1
73+
; CHECK-NEXT: call void @use.i32(i32 [[TMP4]])
74+
; CHECK-NEXT: call void @use.i32(i32 [[TMP5]])
7775
; CHECK-NEXT: ret void
7876
;
7977
%1 = getelementptr i32, i32 addrspace(2)* %src, i32 %off
@@ -89,7 +87,7 @@ declare void @use.i32(i32)
8987

9088
!igc.functions = !{!0, !4, !5}
9189

92-
!0 = !{void (float addrspace(2)*)* @test_merge, !1}
90+
!0 = !{void (i32 addrspace(2)*)* @test_merge, !1}
9391
!1 = !{!2, !3}
9492
!2 = !{!"function_type", i32 0}
9593
!3 = !{!"implicit_arg_desc"}

‎IGC/Compiler/tests/ConstantCoalescing/get-base-const-offset.ll‎

Lines changed: 20 additions & 24 deletions
Original file line numberDiff line numberDiff line change
@@ -16,19 +16,16 @@
1616
; which is part of DecomposePtrExp, used to calculate base from bitcast
1717
; or a getelementptr instruction.
1818

19-
; TODO: check, with opaque pointers on, pass doesn't optimize test_merge case
20-
2119
define void @test_merge(ptr addrspace(2) %src) {
2220
; CHECK-LABEL: define void @test_merge(
2321
; CHECK-SAME: ptr addrspace(2) [[SRC:%.*]]) {
24-
; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr addrspace(2) [[SRC]], align 4
25-
; CHECK-NEXT: [[TMP2:%.*]] = ptrtoint ptr addrspace(2) [[SRC]] to i64
26-
; CHECK-NEXT: [[TMP3:%.*]] = add i64 [[TMP2]], 4
27-
; CHECK-NEXT: [[CHUNKPTR:%.*]] = inttoptr i64 [[TMP3]] to ptr addrspace(2)
28-
; CHECK-NEXT: [[TMP4:%.*]] = load <1 x i32>, ptr addrspace(2) [[CHUNKPTR]], align 4
29-
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <1 x i32> [[TMP4]], i32 0
30-
; CHECK-NEXT: call void @use.i32(i32 [[TMP1]])
31-
; CHECK-NEXT: call void @use.i32(i32 [[TMP5]])
22+
; CHECK-NEXT: [[TMP1:%.*]] = ptrtoint ptr addrspace(2) [[SRC]] to i64
23+
; CHECK-NEXT: [[CHUNKPTR:%.*]] = inttoptr i64 [[TMP1]] to ptr addrspace(2)
24+
; CHECK-NEXT: [[TMP2:%.*]] = load <2 x i32>, ptr addrspace(2) [[CHUNKPTR]], align 4
25+
; CHECK-NEXT: [[TMP3:%.*]] = extractelement <2 x i32> [[TMP2]], i32 0
26+
; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x i32> [[TMP2]], i32 1
27+
; CHECK-NEXT: call void @use.i32(i32 [[TMP3]])
28+
; CHECK-NEXT: call void @use.i32(i32 [[TMP4]])
3229
; CHECK-NEXT: ret void
3330
;
3431
%1 = getelementptr i32, ptr addrspace(2) %src, i32 1
@@ -42,13 +39,13 @@ define void @test_merge(ptr addrspace(2) %src) {
4239
define void @test_vectorize(ptr addrspace(2) %src) {
4340
; CHECK-LABEL: define void @test_vectorize(
4441
; CHECK-SAME: ptr addrspace(2) [[SRC:%.*]]) {
45-
; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr addrspace(2) [[SRC]], align 4
42+
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr i32, ptr addrspace(2) [[SRC]], i32 2
4643
; CHECK-NEXT: [[TMP2:%.*]] = ptrtoint ptr addrspace(2) [[SRC]] to i64
47-
; CHECK-NEXT: [[TMP3:%.*]] = add i64 [[TMP2]], 8
48-
; CHECK-NEXT: [[CHUNKPTR:%.*]] = inttoptr i64 [[TMP3]] to ptr addrspace(2)
49-
; CHECK-NEXT: [[TMP4:%.*]] = load <1 x i32>, ptr addrspace(2) [[CHUNKPTR]], align 4
50-
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <1 x i32> [[TMP4]], i32 0
51-
; CHECK-NEXT: call void @use.i32(i32 [[TMP1]])
44+
; CHECK-NEXT: [[CHUNKPTR:%.*]] = inttoptr i64 [[TMP2]] to ptr addrspace(2)
45+
; CHECK-NEXT: [[TMP3:%.*]] = load <1 x i32>, ptr addrspace(2) [[CHUNKPTR]], align 4
46+
; CHECK-NEXT: [[TMP4:%.*]] = extractelement <1 x i32> [[TMP3]], i32 0
47+
; CHECK-NEXT: [[TMP5:%.*]] = load i32, ptr addrspace(2) [[TMP1]], align 4
48+
; CHECK-NEXT: call void @use.i32(i32 [[TMP4]])
5249
; CHECK-NEXT: call void @use.i32(i32 [[TMP5]])
5350
; CHECK-NEXT: ret void
5451
;
@@ -64,14 +61,13 @@ define void @test_nonconst_gep(ptr addrspace(2) %src, i32 %off) {
6461
; CHECK-LABEL: define void @test_nonconst_gep(
6562
; CHECK-SAME: ptr addrspace(2) [[SRC:%.*]], i32 [[OFF:%.*]]) {
6663
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr i32, ptr addrspace(2) [[SRC]], i32 [[OFF]]
67-
; CHECK-NEXT: [[TMP2:%.*]] = load i32, ptr addrspace(2) [[TMP1]], align 4
68-
; CHECK-NEXT: [[TMP3:%.*]] = ptrtoint ptr addrspace(2) [[TMP1]] to i64
69-
; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[TMP3]], 4
70-
; CHECK-NEXT: [[CHUNKPTR:%.*]] = inttoptr i64 [[TMP4]] to ptr addrspace(2)
71-
; CHECK-NEXT: [[TMP5:%.*]] = load <1 x i32>, ptr addrspace(2) [[CHUNKPTR]], align 4
72-
; CHECK-NEXT: [[TMP6:%.*]] = extractelement <1 x i32> [[TMP5]], i32 0
73-
; CHECK-NEXT: call void @use.i32(i32 [[TMP2]])
74-
; CHECK-NEXT: call void @use.i32(i32 [[TMP6]])
64+
; CHECK-NEXT: [[TMP2:%.*]] = ptrtoint ptr addrspace(2) [[TMP1]] to i64
65+
; CHECK-NEXT: [[CHUNKPTR:%.*]] = inttoptr i64 [[TMP2]] to ptr addrspace(2)
66+
; CHECK-NEXT: [[TMP3:%.*]] = load <2 x i32>, ptr addrspace(2) [[CHUNKPTR]], align 4
67+
; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x i32> [[TMP3]], i32 0
68+
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x i32> [[TMP3]], i32 1
69+
; CHECK-NEXT: call void @use.i32(i32 [[TMP4]])
70+
; CHECK-NEXT: call void @use.i32(i32 [[TMP5]])
7571
; CHECK-NEXT: ret void
7672
;
7773
%1 = getelementptr i32, ptr addrspace(2) %src, i32 %off

0 commit comments

Comments
 (0)