1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-24 03:33:20 +01:00
llvm-mirror/test/CodeGen/NVPTX/shfl.ll
Justin Lebar 63fe7a59a1 [NVPTX] Remove NVPTXFavorNonGenericAddrSpaces pass.
Summary:
This has been replaced by the NVPTXInferAddressSpaces pass.  We've had
the new one as the default with the old one accessible via a flag for
some months now, and we've had no problems.

Reviewers: tra

Subscribers: llvm-commits, jholewinski, jingyue, mgorny

Differential Revision: https://reviews.llvm.org/D26165

llvm-svn: 285642
2016-10-31 21:51:42 +00:00

91 lines
3.5 KiB
LLVM

; RUN: llc < %s -march=nvptx64 -mcpu=sm_30 | FileCheck %s
declare i32 @llvm.nvvm.shfl.down.i32(i32, i32, i32)
declare float @llvm.nvvm.shfl.down.f32(float, i32, i32)
declare i32 @llvm.nvvm.shfl.up.i32(i32, i32, i32)
declare float @llvm.nvvm.shfl.up.f32(float, i32, i32)
declare i32 @llvm.nvvm.shfl.bfly.i32(i32, i32, i32)
declare float @llvm.nvvm.shfl.bfly.f32(float, i32, i32)
declare i32 @llvm.nvvm.shfl.idx.i32(i32, i32, i32)
declare float @llvm.nvvm.shfl.idx.f32(float, i32, i32)
; Try all four permutations of register and immediate parameters with
; shfl.down.
; CHECK-LABEL: .func{{.*}}shfl.down1
define i32 @shfl.down1(i32 %in) {
; CHECK: ld.param.u32 [[IN:%r[0-9]+]]
; CHECK: shfl.down.b32 [[OUT:%r[0-9]+]], [[IN]], 1, 2;
; CHECK: st.param.{{.}}32 {{.*}}, [[OUT]]
%val = call i32 @llvm.nvvm.shfl.down.i32(i32 %in, i32 1, i32 2)
ret i32 %val
}
; CHECK-LABEL: .func{{.*}}shfl.down2
define i32 @shfl.down2(i32 %in, i32 %width) {
; CHECK: ld.param.u32 [[IN1:%r[0-9]+]]
; CHECK: ld.param.u32 [[IN2:%r[0-9]+]]
; CHECK: shfl.down.{{.}}32 %r{{[0-9]+}}, [[IN1]], [[IN2]], 3;
%val = call i32 @llvm.nvvm.shfl.down.i32(i32 %in, i32 %width, i32 3)
ret i32 %val
}
; CHECK-LABEL: .func{{.*}}shfl.down3
define i32 @shfl.down3(i32 %in, i32 %mask) {
; CHECK: ld.param.u32 [[IN1:%r[0-9]+]]
; CHECK: ld.param.u32 [[IN2:%r[0-9]+]]
; CHECK: shfl.down.{{.}}32 %r{{[0-9]+}}, [[IN1]], 4, [[IN2]];
%val = call i32 @llvm.nvvm.shfl.down.i32(i32 %in, i32 4, i32 %mask)
ret i32 %val
}
; CHECK-LABEL: .func{{.*}}shfl.down4
define i32 @shfl.down4(i32 %in, i32 %width, i32 %mask) {
; CHECK: ld.param.u32 [[IN1:%r[0-9]+]]
; CHECK: ld.param.u32 [[IN2:%r[0-9]+]]
; CHECK: ld.param.u32 [[IN3:%r[0-9]+]]
; CHECK: shfl.down.{{.}}32 %r{{[0-9]+}}, [[IN1]], [[IN2]], [[IN3]];
%val = call i32 @llvm.nvvm.shfl.down.i32(i32 %in, i32 %width, i32 %mask)
ret i32 %val
}
; Try shfl.down with floating-point params.
; CHECK-LABEL: .func{{.*}}shfl.down.float
define float @shfl.down.float(float %in) {
; CHECK: ld.param.f32 [[IN:%f[0-9]+]]
; CHECK: shfl.down.b32 [[OUT:%f[0-9]+]], [[IN]], 5, 6;
; CHECK: st.param.{{.}}32 {{.*}}, [[OUT]]
%out = call float @llvm.nvvm.shfl.down.f32(float %in, i32 5, i32 6)
ret float %out
}
; Try the rest of the shfl modes. Hopefully they're declared in such a way
; that if shfl.down works correctly, they also work correctly.
define void @shfl.rest(i32 %in_i32, float %in_float, i32* %out_i32, float* %out_float) {
; CHECK: shfl.up.b32 %r{{[0-9]+}}, %r{{[0-9]+}}, 1, 2;
%up_i32 = call i32 @llvm.nvvm.shfl.up.i32(i32 %in_i32, i32 1, i32 2)
store i32 %up_i32, i32* %out_i32
; CHECK: shfl.up.b32 %f{{[0-9]+}}, %f{{[0-9]+}}, 3, 4;
%up_float = call float @llvm.nvvm.shfl.up.f32(float %in_float, i32 3, i32 4)
store float %up_float, float* %out_float
; CHECK: shfl.bfly.b32 %r{{[0-9]+}}, %r{{[0-9]+}}, 5, 6;
%bfly_i32 = call i32 @llvm.nvvm.shfl.bfly.i32(i32 %in_i32, i32 5, i32 6)
store i32 %bfly_i32, i32* %out_i32
; CHECK: shfl.bfly.b32 %f{{[0-9]+}}, %f{{[0-9]+}}, 7, 8;
%bfly_float = call float @llvm.nvvm.shfl.bfly.f32(float %in_float, i32 7, i32 8)
store float %bfly_float, float* %out_float
; CHECK: shfl.idx.b32 %r{{[0-9]+}}, %r{{[0-9]+}}, 9, 10;
%idx_i32 = call i32 @llvm.nvvm.shfl.idx.i32(i32 %in_i32, i32 9, i32 10)
store i32 %idx_i32, i32* %out_i32
; CHECK: shfl.idx.b32 %f{{[0-9]+}}, %f{{[0-9]+}}, 11, 12;
%idx_float = call float @llvm.nvvm.shfl.idx.f32(float %in_float, i32 11, i32 12)
store float %idx_float, float* %out_float
ret void
}