diff options
author | Daniel Sanders <daniel.sanders@imgtec.com> | 2014-04-30 12:09:32 +0000 |
---|---|---|
committer | Daniel Sanders <daniel.sanders@imgtec.com> | 2014-04-30 12:09:32 +0000 |
commit | 1c8add99789a3066e94e97c56a6ce11a5c8e8740 (patch) | |
tree | 501cfd87ae4e95427c1482e0108e5d5ff1d1aeb6 /test/CodeGen | |
parent | 4e29e3a10ad3ef6ad9e14295eb24c3444aa107cb (diff) | |
download | llvm-1c8add99789a3066e94e97c56a6ce11a5c8e8740.tar.gz llvm-1c8add99789a3066e94e97c56a6ce11a5c8e8740.tar.bz2 llvm-1c8add99789a3066e94e97c56a6ce11a5c8e8740.tar.xz |
[mips][msa] Fix vector insertions where the index is variable
Summary:
This isn't supported directly so we rotate the vector by the desired number of
elements, insert to element zero, then rotate back.
The i64 case generates rather poor code on MIPS32. There is an obvious
optimisation to be made in future (do both insert.w's inside a shared
rotate/unrotate sequence) but for now it's sufficient to select valid code
instead of aborting.
Depends on D3536
Reviewers: matheusalmeida
Reviewed By: matheusalmeida
Differential Revision: http://reviews.llvm.org/D3537
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@207640 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'test/CodeGen')
-rw-r--r-- | test/CodeGen/Mips/msa/basic_operations.ll | 123 | ||||
-rw-r--r-- | test/CodeGen/Mips/msa/basic_operations_float.ll | 52 |
2 files changed, 175 insertions, 0 deletions
diff --git a/test/CodeGen/Mips/msa/basic_operations.ll b/test/CodeGen/Mips/msa/basic_operations.ll index cfe43b94ff..dbdf42be49 100644 --- a/test/CodeGen/Mips/msa/basic_operations.ll +++ b/test/CodeGen/Mips/msa/basic_operations.ll @@ -673,6 +673,129 @@ define void @insert_v2i64(i64 %a) nounwind { ; MIPS32-AE: .size insert_v2i64 } +define void @insert_v16i8_vidx(i32 %a) nounwind { + ; MIPS32-AE: insert_v16i8_vidx: + + %1 = load <16 x i8>* @v16i8 + ; MIPS32-AE-DAG: ld.b [[R1:\$w[0-9]+]], + + %2 = load i32* @i32 + ; MIPS32-AE-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)( + ; MIPS32-AE-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]]) + + %a2 = trunc i32 %a to i8 + %a3 = sext i8 %a2 to i32 + %a4 = trunc i32 %a3 to i8 + ; MIPS32-AE-NOT: andi + ; MIPS32-AE-NOT: sra + + %3 = insertelement <16 x i8> %1, i8 %a4, i32 %2 + ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[IDX]]] + ; MIPS32-AE-DAG: insert.b [[R1]][0], $4 + ; MIPS32-AE-DAG: neg [[NIDX:\$[0-9]+]], [[IDX]] + ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]] + + store <16 x i8> %3, <16 x i8>* @v16i8 + ; MIPS32-AE-DAG: st.b [[R1]] + + ret void + ; MIPS32-AE: .size insert_v16i8_vidx +} + +define void @insert_v8i16_vidx(i32 %a) nounwind { + ; MIPS32-AE: insert_v8i16_vidx: + + %1 = load <8 x i16>* @v8i16 + ; MIPS32-AE-DAG: ld.h [[R1:\$w[0-9]+]], + + %2 = load i32* @i32 + ; MIPS32-AE-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)( + ; MIPS32-AE-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]]) + + %a2 = trunc i32 %a to i16 + %a3 = sext i16 %a2 to i32 + %a4 = trunc i32 %a3 to i16 + ; MIPS32-AE-NOT: andi + ; MIPS32-AE-NOT: sra + + %3 = insertelement <8 x i16> %1, i16 %a4, i32 %2 + ; MIPS32-AE-DAG: sll [[BIDX:\$[0-9]+]], [[IDX]], 1 + ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]] + ; MIPS32-AE-DAG: insert.h [[R1]][0], $4 + ; MIPS32-AE-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]] + ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]] + + store <8 x i16> %3, <8 x i16>* @v8i16 + ; MIPS32-AE-DAG: st.h [[R1]] + + ret void + ; MIPS32-AE: .size insert_v8i16_vidx +} + +define void @insert_v4i32_vidx(i32 %a) nounwind { + ; MIPS32-AE: insert_v4i32_vidx: + + %1 = load <4 x i32>* @v4i32 + ; MIPS32-AE-DAG: ld.w [[R1:\$w[0-9]+]], + + %2 = load i32* @i32 + ; MIPS32-AE-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)( + ; MIPS32-AE-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]]) + + ; MIPS32-AE-NOT: andi + ; MIPS32-AE-NOT: sra + + %3 = insertelement <4 x i32> %1, i32 %a, i32 %2 + ; MIPS32-AE-DAG: sll [[BIDX:\$[0-9]+]], [[IDX]], 2 + ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]] + ; MIPS32-AE-DAG: insert.w [[R1]][0], $4 + ; MIPS32-AE-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]] + ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]] + + store <4 x i32> %3, <4 x i32>* @v4i32 + ; MIPS32-AE-DAG: st.w [[R1]] + + ret void + ; MIPS32-AE: .size insert_v4i32_vidx +} + +define void @insert_v2i64_vidx(i64 %a) nounwind { + ; MIPS32-AE: insert_v2i64_vidx: + + %1 = load <2 x i64>* @v2i64 + ; MIPS32-AE-DAG: ld.w [[R1:\$w[0-9]+]], + + %2 = load i32* @i32 + ; MIPS32-AE-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)( + ; MIPS32-AE-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]]) + + ; MIPS32-AE-NOT: andi + ; MIPS32-AE-NOT: sra + + %3 = insertelement <2 x i64> %1, i64 %a, i32 %2 + ; TODO: This code could be a lot better but it works. The legalizer splits + ; 64-bit inserts into two 32-bit inserts because there is no i64 type on + ; MIPS32. The obvious optimisation is to perform both insert.w's at once while + ; the vector is rotated. + ; MIPS32-AE-DAG: sll [[BIDX:\$[0-9]+]], [[IDX]], 2 + ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]] + ; MIPS32-AE-DAG: insert.w [[R1]][0], $4 + ; MIPS32-AE-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]] + ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]] + ; MIPS32-AE-DAG: addiu [[IDX2:\$[0-9]+]], [[IDX]], 1 + ; MIPS32-AE-DAG: sll [[BIDX:\$[0-9]+]], [[IDX2]], 2 + ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]] + ; MIPS32-AE-DAG: insert.w [[R1]][0], $5 + ; MIPS32-AE-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]] + ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]] + + store <2 x i64> %3, <2 x i64>* @v2i64 + ; MIPS32-AE-DAG: st.w [[R1]] + + ret void + ; MIPS32-AE: .size insert_v2i64_vidx +} + define void @truncstore() nounwind { ; MIPS32-AE-LABEL: truncstore: diff --git a/test/CodeGen/Mips/msa/basic_operations_float.ll b/test/CodeGen/Mips/msa/basic_operations_float.ll index ceefa9114b..a0c9d29e23 100644 --- a/test/CodeGen/Mips/msa/basic_operations_float.ll +++ b/test/CodeGen/Mips/msa/basic_operations_float.ll @@ -275,3 +275,55 @@ define void @insert_v2f64(double %a) nounwind { ret void ; MIPS32: .size insert_v2f64 } + +define void @insert_v4f32_vidx(float %a) nounwind { + ; MIPS32-LABEL: insert_v4f32_vidx: + + %1 = load <4 x float>* @v4f32 + ; MIPS32-DAG: lw [[PTR_V:\$[0-9]+]], %got(v4f32)( + ; MIPS32-DAG: ld.w [[R1:\$w[0-9]+]], 0([[PTR_V]]) + + %2 = load i32* @i32 + ; MIPS32-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)( + ; MIPS32-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]]) + + %3 = insertelement <4 x float> %1, float %a, i32 %2 + ; float argument passed in $f12 + ; MIPS32-DAG: sll [[BIDX:\$[0-9]+]], [[IDX]], 2 + ; MIPS32-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]] + ; MIPS32-DAG: insve.w [[R1]][0], $w12[0] + ; MIPS32-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]] + ; MIPS32-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]] + + store <4 x float> %3, <4 x float>* @v4f32 + ; MIPS32-DAG: st.w [[R1]] + + ret void + ; MIPS32: .size insert_v4f32_vidx +} + +define void @insert_v2f64_vidx(double %a) nounwind { + ; MIPS32-LABEL: insert_v2f64_vidx: + + %1 = load <2 x double>* @v2f64 + ; MIPS32-DAG: lw [[PTR_V:\$[0-9]+]], %got(v2f64)( + ; MIPS32-DAG: ld.d [[R1:\$w[0-9]+]], 0([[PTR_V]]) + + %2 = load i32* @i32 + ; MIPS32-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)( + ; MIPS32-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]]) + + %3 = insertelement <2 x double> %1, double %a, i32 %2 + ; double argument passed in $f12 + ; MIPS32-DAG: sll [[BIDX:\$[0-9]+]], [[IDX]], 3 + ; MIPS32-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]] + ; MIPS32-DAG: insve.d [[R1]][0], $w12[0] + ; MIPS32-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]] + ; MIPS32-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]] + + store <2 x double> %3, <2 x double>* @v2f64 + ; MIPS32-DAG: st.d [[R1]] + + ret void + ; MIPS32: .size insert_v2f64_vidx +} |