4cd836181f
Added support for lowering the scalar S_ABSDIFF_I32 instruction to equivalent VALU operations.
143 lines
5.1 KiB
LLVM
143 lines
5.1 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
|
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck %s
|
|
|
|
|
|
define amdgpu_gs float @absdiff_valu_input_regression() {
|
|
; CHECK-LABEL: absdiff_valu_input_regression:
|
|
; CHECK: ; %bb.0: ; %bb
|
|
; CHECK-NEXT: s_mov_b32 s0, 0
|
|
; CHECK-NEXT: .LBB0_1: ; %bb1
|
|
; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
|
|
; CHECK-NEXT: s_mov_b32 s1, s0
|
|
; CHECK-NEXT: s_or_b32 s0, s0, 1
|
|
; CHECK-NEXT: s_cmp_gt_i32 s1, 0
|
|
; CHECK-NEXT: s_cbranch_scc1 .LBB0_1
|
|
; CHECK-NEXT: ; %bb.2: ; %bb11
|
|
; CHECK-NEXT: v_med3_i32 v0, s1, 0, 1
|
|
; CHECK-NEXT: v_sub_u32_e32 v0, 0, v0
|
|
; CHECK-NEXT: v_sub_u32_e32 v1, 0, v0
|
|
; CHECK-NEXT: v_max_i32_e32 v0, v0, v1
|
|
; CHECK-NEXT: ; return to shader part epilog
|
|
bb:
|
|
br label %bb1
|
|
|
|
bb1: ; preds = %bb1, %bb
|
|
%i = phi i32 [ 0, %bb ], [ %i9, %bb1 ]
|
|
%i2 = phi i32 [ 0, %bb ], [ %i5, %bb1 ]
|
|
%i3 = or i32 %i2, 1
|
|
%i4 = or i32 %i3, 0
|
|
%i5 = call i32 @llvm.smax.i32(i32 %i, i32 0)
|
|
%i6 = call i32 @llvm.umin.i32(i32 %i5, i32 1)
|
|
%i7 = sub i32 0, %i6
|
|
%i8 = call i32 @llvm.abs.i32(i32 %i7, i1 false)
|
|
%i9 = or i32 %i, 1
|
|
%i10 = icmp sgt i32 %i, 0
|
|
br i1 %i10, label %bb1, label %bb11
|
|
|
|
bb11: ; preds = %bb1
|
|
%i12 = bitcast i32 %i8 to float
|
|
ret float %i12
|
|
}
|
|
|
|
define amdgpu_ps i16 @absdiff_i16_false(i16 inreg %arg0, i16 inreg %arg1) {
|
|
; CHECK-LABEL: absdiff_i16_false:
|
|
; CHECK: ; %bb.0:
|
|
; CHECK-NEXT: s_sub_i32 s0, s0, s1
|
|
; CHECK-NEXT: s_sext_i32_i16 s1, s0
|
|
; CHECK-NEXT: s_sub_i32 s0, 0, s0
|
|
; CHECK-NEXT: s_sext_i32_i16 s0, s0
|
|
; CHECK-NEXT: s_max_i32 s0, s1, s0
|
|
; CHECK-NEXT: ; return to shader part epilog
|
|
%diff = sub i16 %arg0, %arg1
|
|
%res = call i16 @llvm.abs.i16(i16 %diff, i1 false) ; INT_MIN input returns INT_MIN
|
|
ret i16 %res
|
|
}
|
|
|
|
define amdgpu_ps i16 @absdiff_i16_true(i16 inreg %arg0, i16 inreg %arg1) {
|
|
; CHECK-LABEL: absdiff_i16_true:
|
|
; CHECK: ; %bb.0:
|
|
; CHECK-NEXT: s_sub_i32 s0, s0, s1
|
|
; CHECK-NEXT: s_sext_i32_i16 s1, s0
|
|
; CHECK-NEXT: s_sub_i32 s0, 0, s0
|
|
; CHECK-NEXT: s_sext_i32_i16 s0, s0
|
|
; CHECK-NEXT: s_max_i32 s0, s1, s0
|
|
; CHECK-NEXT: ; return to shader part epilog
|
|
%diff = sub i16 %arg0, %arg1
|
|
%res = call i16 @llvm.abs.i16(i16 %diff, i1 true) ; INT_MIN input returns poison
|
|
ret i16 %res
|
|
}
|
|
|
|
define amdgpu_ps i32 @absdiff_i32_false(i32 inreg %arg0, i32 inreg %arg1) {
|
|
; CHECK-LABEL: absdiff_i32_false:
|
|
; CHECK: ; %bb.0:
|
|
; CHECK-NEXT: s_absdiff_i32 s0, s0, s1
|
|
; CHECK-NEXT: ; return to shader part epilog
|
|
%diff = sub i32 %arg0, %arg1
|
|
%res = call i32 @llvm.abs.i32(i32 %diff, i1 false) ; INT_MIN input returns INT_MIN
|
|
ret i32 %res
|
|
}
|
|
|
|
define amdgpu_ps i32 @absdiff_i32_true(i32 inreg %arg0, i32 inreg %arg1) {
|
|
; CHECK-LABEL: absdiff_i32_true:
|
|
; CHECK: ; %bb.0:
|
|
; CHECK-NEXT: s_absdiff_i32 s0, s0, s1
|
|
; CHECK-NEXT: ; return to shader part epilog
|
|
%diff = sub i32 %arg0, %arg1
|
|
%res = call i32 @llvm.abs.i32(i32 %diff, i1 true) ; INT_MIN input returns poison
|
|
ret i32 %res
|
|
}
|
|
|
|
; Multiple uses of %diff. No benefit for using s_absdiff_i32.
|
|
define amdgpu_ps i32 @absdiff_i32_false_multi_use(i32 inreg %arg0, i32 inreg %arg1) {
|
|
; CHECK-LABEL: absdiff_i32_false_multi_use:
|
|
; CHECK: ; %bb.0:
|
|
; CHECK-NEXT: s_sub_i32 s1, s0, s1
|
|
; CHECK-NEXT: s_abs_i32 s0, s1
|
|
; CHECK-NEXT: ;;#ASMSTART
|
|
; CHECK-NEXT: ; use s1
|
|
; CHECK-NEXT: ;;#ASMEND
|
|
; CHECK-NEXT: ; return to shader part epilog
|
|
%diff = sub i32 %arg0, %arg1
|
|
%res = call i32 @llvm.abs.i32(i32 %diff, i1 false) ; INT_MIN input returns INT_MIN
|
|
call void asm "; use $0", "s"(i32 %diff)
|
|
ret i32 %res
|
|
}
|
|
|
|
define <2 x i32> @absdiff_2xi32_false(<2 x i32> %arg0, <2 x i32> %arg1) {
|
|
; CHECK-LABEL: absdiff_2xi32_false:
|
|
; CHECK: ; %bb.0:
|
|
; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; CHECK-NEXT: v_sub_u32_e32 v0, v0, v2
|
|
; CHECK-NEXT: v_sub_u32_e32 v1, v1, v3
|
|
; CHECK-NEXT: v_sub_u32_e32 v2, 0, v0
|
|
; CHECK-NEXT: v_max_i32_e32 v0, v2, v0
|
|
; CHECK-NEXT: v_sub_u32_e32 v2, 0, v1
|
|
; CHECK-NEXT: v_max_i32_e32 v1, v2, v1
|
|
; CHECK-NEXT: s_setpc_b64 s[30:31]
|
|
%diff = sub <2 x i32> %arg0, %arg1
|
|
%res = call <2 x i32> @llvm.abs.v2i32(<2 x i32> %diff, i1 false) ; INT_MIN input returns INT_MIN
|
|
ret <2 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @absdiff_4xi32_false(<4 x i32> %arg0, <4 x i32> %arg1) {
|
|
; CHECK-LABEL: absdiff_4xi32_false:
|
|
; CHECK: ; %bb.0:
|
|
; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; CHECK-NEXT: v_sub_u32_e32 v0, v0, v4
|
|
; CHECK-NEXT: v_sub_u32_e32 v1, v1, v5
|
|
; CHECK-NEXT: v_sub_u32_e32 v4, 0, v0
|
|
; CHECK-NEXT: v_sub_u32_e32 v2, v2, v6
|
|
; CHECK-NEXT: v_max_i32_e32 v0, v4, v0
|
|
; CHECK-NEXT: v_sub_u32_e32 v4, 0, v1
|
|
; CHECK-NEXT: v_sub_u32_e32 v3, v3, v7
|
|
; CHECK-NEXT: v_max_i32_e32 v1, v4, v1
|
|
; CHECK-NEXT: v_sub_u32_e32 v4, 0, v2
|
|
; CHECK-NEXT: v_max_i32_e32 v2, v4, v2
|
|
; CHECK-NEXT: v_sub_u32_e32 v4, 0, v3
|
|
; CHECK-NEXT: v_max_i32_e32 v3, v4, v3
|
|
; CHECK-NEXT: s_setpc_b64 s[30:31]
|
|
%diff = sub <4 x i32> %arg0, %arg1
|
|
%res = call <4 x i32> @llvm.abs.v4i32(<4 x i32> %diff, i1 false) ; INT_MIN input returns INT_MIN
|
|
ret <4 x i32> %res
|
|
}
|