From 6d2316158f5100e25a8c063c661f2b96d57742ad Mon Sep 17 00:00:00 2001
From: Petar Avramovic <Petar.Avramovic@rt-rk.com>
Date: Wed, 20 Feb 2019 13:42:44 +0000
Subject: [PATCH] [MIPS MSA] Avoid some DAG combines for vector shifts

DAG combiner combines two shifts into shift + and with bitmask.
Avoid such combines for vectors since leaving two vector shifts
as they are produces better end results.

Differential Revision: https://reviews.llvm.org/D58225


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@354461 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/Mips/MipsISelLowering.cpp          |   7 +
 lib/Target/Mips/MipsISelLowering.h            |   2 +
 ...ines.ll => avoid_vector_shift_combines.ll} | 133 ++++++++----------
 3 files changed, 65 insertions(+), 77 deletions(-)
 rename test/CodeGen/Mips/msa/{vector_shift_combines.ll => avoid_vector_shift_combines.ll} (55%)

diff --git a/lib/Target/Mips/MipsISelLowering.cpp b/lib/Target/Mips/MipsISelLowering.cpp
index 103d1fcb44b..d69effb6224 100644
--- a/lib/Target/Mips/MipsISelLowering.cpp
+++ b/lib/Target/Mips/MipsISelLowering.cpp
@@ -1190,6 +1190,13 @@ bool MipsTargetLowering::isCheapToSpeculateCtlz() const {
   return Subtarget.hasMips32();
 }
 
+bool MipsTargetLowering::shouldFoldShiftPairToMask(const SDNode *N,
+                                                   CombineLevel Level) const {
+  if (N->getOperand(0).getValueType().isVector())
+    return false;
+  return true;
+}
+
 void
 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
                                           SmallVectorImpl<SDValue> &Results,
diff --git a/lib/Target/Mips/MipsISelLowering.h b/lib/Target/Mips/MipsISelLowering.h
index e25810c8741..3a9e61b97b4 100644
--- a/lib/Target/Mips/MipsISelLowering.h
+++ b/lib/Target/Mips/MipsISelLowering.h
@@ -284,6 +284,8 @@ class TargetRegisterClass;
 
     bool isCheapToSpeculateCttz() const override;
     bool isCheapToSpeculateCtlz() const override;
+    bool shouldFoldShiftPairToMask(const SDNode *N,
+                                   CombineLevel Level) const override;
 
     /// Return the register type for a given MVT, ensuring vectors are treated
     /// as a series of gpr sized integers.
diff --git a/test/CodeGen/Mips/msa/vector_shift_combines.ll b/test/CodeGen/Mips/msa/avoid_vector_shift_combines.ll
similarity index 55%
rename from test/CodeGen/Mips/msa/vector_shift_combines.ll
rename to test/CodeGen/Mips/msa/avoid_vector_shift_combines.ll
index 2cd97d2bcac..04633cb575c 100644
--- a/test/CodeGen/Mips/msa/vector_shift_combines.ll
+++ b/test/CodeGen/Mips/msa/avoid_vector_shift_combines.ll
@@ -8,21 +8,18 @@ declare <2 x i64> @llvm.mips.srli.d(<2 x i64>, i32)
 declare <4 x i32> @llvm.mips.slli.w(<4 x i32>, i32)
 declare <4 x i32> @llvm.mips.srli.w(<4 x i32>, i32)
 
-; fold (shl (srl x, c1), c2) -> (and (srl x, (sub c1, c2), MASK) if C1 < C2
+; do not fold (shl (srl x, c1), c2) -> (and (srl x, (sub c1, c2), MASK) if C1 < C2
 ; MASK_TYPE1 = C2-C1 0s | 1s | ends with C1 0s
-define void @combine_shifts_to_shift_plus_and_mask_type1_i64(<2 x i64>* %a, <2 x i64>* %b) {
-; MIPSEL64R6-LABEL: combine_shifts_to_shift_plus_and_mask_type1_i64:
+define void @avoid_to_combine_shifts_to_shift_plus_and_mask_type1_i64(<2 x i64>* %a, <2 x i64>* %b) {
+; MIPSEL64R6-LABEL: avoid_to_combine_shifts_to_shift_plus_and_mask_type1_i64:
 ; MIPSEL64R6:       # %bb.0: # %entry
 ; MIPSEL64R6-NEXT:    ld.d $w0, 0($4)
-; MIPSEL64R6-NEXT:    srli.d $w0, $w0, 1
-; MIPSEL64R6-NEXT:    lui $1, 32760
-; MIPSEL64R6-NEXT:    dsll32 $1, $1, 0
-; MIPSEL64R6-NEXT:    fill.d $w1, $1
-; MIPSEL64R6-NEXT:    and.v $w0, $w0, $w1
+; MIPSEL64R6-NEXT:    srli.d $w0, $w0, 52
+; MIPSEL64R6-NEXT:    slli.d $w0, $w0, 51
 ; MIPSEL64R6-NEXT:    jr $ra
 ; MIPSEL64R6-NEXT:    st.d $w0, 0($5)
 ;
-; MIPSEL32R5-LABEL: combine_shifts_to_shift_plus_and_mask_type1_i64:
+; MIPSEL32R5-LABEL: avoid_to_combine_shifts_to_shift_plus_and_mask_type1_i64:
 ; MIPSEL32R5:       # %bb.0: # %entry
 ; MIPSEL32R5-NEXT:    ld.d $w0, 0($4)
 ; MIPSEL32R5-NEXT:    srli.d $w0, $w0, 52
@@ -37,23 +34,17 @@ entry:
   ret void
 }
 
-; fold (shl (srl x, c1), c2) -> (and (srl x, (sub c1, c2), MASK) if C1 < C2
-define void @combine_shifts_to_shift_plus_and_mask_type1_i64_long(<2 x i64>* %a, <2 x i64>* %b) {
-; MIPSEL64R6-LABEL: combine_shifts_to_shift_plus_and_mask_type1_i64_long:
+; do not fold (shl (srl x, c1), c2) -> (and (srl x, (sub c1, c2), MASK) if C1 < C2
+define void @avoid_to_combine_shifts_to_shift_plus_and_mask_type1_i64_long(<2 x i64>* %a, <2 x i64>* %b) {
+; MIPSEL64R6-LABEL: avoid_to_combine_shifts_to_shift_plus_and_mask_type1_i64_long:
 ; MIPSEL64R6:       # %bb.0: # %entry
-; MIPSEL64R6-NEXT:    lui $1, 65535
-; MIPSEL64R6-NEXT:    ori $1, $1, 65520
-; MIPSEL64R6-NEXT:    lui $2, 16383
-; MIPSEL64R6-NEXT:    ori $2, $2, 65535
-; MIPSEL64R6-NEXT:    dinsu $1, $2, 32, 32
 ; MIPSEL64R6-NEXT:    ld.d $w0, 0($4)
-; MIPSEL64R6-NEXT:    srli.d $w0, $w0, 2
-; MIPSEL64R6-NEXT:    fill.d $w1, $1
-; MIPSEL64R6-NEXT:    and.v $w0, $w0, $w1
+; MIPSEL64R6-NEXT:    srli.d $w0, $w0, 6
+; MIPSEL64R6-NEXT:    slli.d $w0, $w0, 4
 ; MIPSEL64R6-NEXT:    jr $ra
 ; MIPSEL64R6-NEXT:    st.d $w0, 0($5)
 ;
-; MIPSEL32R5-LABEL: combine_shifts_to_shift_plus_and_mask_type1_i64_long:
+; MIPSEL32R5-LABEL: avoid_to_combine_shifts_to_shift_plus_and_mask_type1_i64_long:
 ; MIPSEL32R5:       # %bb.0: # %entry
 ; MIPSEL32R5-NEXT:    ld.d $w0, 0($4)
 ; MIPSEL32R5-NEXT:    srli.d $w0, $w0, 6
@@ -68,19 +59,18 @@ entry:
   ret void
 }
 
-; fold (shl (srl x, c1), c2) -> (and (shl x, (sub c1, c2), MASK) if C1 >= C2
+; do not fold (shl (srl x, c1), c2) -> (and (shl x, (sub c1, c2), MASK) if C1 >= C2
 ; MASK_TYPE2 = 1s | C1 zeros
-define void @combine_shifts_to_shift_plus_and_mask_type2_i32(<2 x i64>* %a, <2 x i64>* %b) {
-; MIPSEL64R6-LABEL: combine_shifts_to_shift_plus_and_mask_type2_i32:
+define void @avoid_to_combine_shifts_to_shift_plus_and_mask_type2_i32(<2 x i64>* %a, <2 x i64>* %b) {
+; MIPSEL64R6-LABEL: avoid_to_combine_shifts_to_shift_plus_and_mask_type2_i32:
 ; MIPSEL64R6:       # %bb.0: # %entry
 ; MIPSEL64R6-NEXT:    ld.d $w0, 0($4)
-; MIPSEL64R6-NEXT:    slli.d $w0, $w0, 2
-; MIPSEL64R6-NEXT:    ldi.d $w1, -64
-; MIPSEL64R6-NEXT:    and.v $w0, $w0, $w1
+; MIPSEL64R6-NEXT:    srli.d $w0, $w0, 4
+; MIPSEL64R6-NEXT:    slli.d $w0, $w0, 6
 ; MIPSEL64R6-NEXT:    jr $ra
 ; MIPSEL64R6-NEXT:    st.d $w0, 0($5)
 ;
-; MIPSEL32R5-LABEL: combine_shifts_to_shift_plus_and_mask_type2_i32:
+; MIPSEL32R5-LABEL: avoid_to_combine_shifts_to_shift_plus_and_mask_type2_i32:
 ; MIPSEL32R5:       # %bb.0: # %entry
 ; MIPSEL32R5-NEXT:    ld.d $w0, 0($4)
 ; MIPSEL32R5-NEXT:    srli.d $w0, $w0, 4
@@ -95,27 +85,21 @@ entry:
   ret void
 }
 
-; fold (shl (srl x, c1), c2) -> (and (srl x, (sub c1, c2), MASK) if C1 < C2
-define void @combine_shifts_to_shift_plus_and_mask_type1_i32_long(<4 x i32>* %a, <4 x i32>* %b) {
-; MIPSEL64R6-LABEL: combine_shifts_to_shift_plus_and_mask_type1_i32_long:
+; do not fold (shl (srl x, c1), c2) -> (and (srl x, (sub c1, c2), MASK) if C1 < C2
+define void @avoid_to_combine_shifts_to_shift_plus_and_mask_type1_i32_long(<4 x i32>* %a, <4 x i32>* %b) {
+; MIPSEL64R6-LABEL: avoid_to_combine_shifts_to_shift_plus_and_mask_type1_i32_long:
 ; MIPSEL64R6:       # %bb.0: # %entry
 ; MIPSEL64R6-NEXT:    ld.w $w0, 0($4)
-; MIPSEL64R6-NEXT:    srli.w $w0, $w0, 4
-; MIPSEL64R6-NEXT:    lui $1, 4095
-; MIPSEL64R6-NEXT:    ori $1, $1, 65528
-; MIPSEL64R6-NEXT:    fill.w $w1, $1
-; MIPSEL64R6-NEXT:    and.v $w0, $w0, $w1
+; MIPSEL64R6-NEXT:    srli.w $w0, $w0, 7
+; MIPSEL64R6-NEXT:    slli.w $w0, $w0, 3
 ; MIPSEL64R6-NEXT:    jr $ra
 ; MIPSEL64R6-NEXT:    st.w $w0, 0($5)
 ;
-; MIPSEL32R5-LABEL: combine_shifts_to_shift_plus_and_mask_type1_i32_long:
+; MIPSEL32R5-LABEL: avoid_to_combine_shifts_to_shift_plus_and_mask_type1_i32_long:
 ; MIPSEL32R5:       # %bb.0: # %entry
 ; MIPSEL32R5-NEXT:    ld.w $w0, 0($4)
-; MIPSEL32R5-NEXT:    srli.w $w0, $w0, 4
-; MIPSEL32R5-NEXT:    lui $1, 4095
-; MIPSEL32R5-NEXT:    ori $1, $1, 65528
-; MIPSEL32R5-NEXT:    fill.w $w1, $1
-; MIPSEL32R5-NEXT:    and.v $w0, $w0, $w1
+; MIPSEL32R5-NEXT:    srli.w $w0, $w0, 7
+; MIPSEL32R5-NEXT:    slli.w $w0, $w0, 3
 ; MIPSEL32R5-NEXT:    jr $ra
 ; MIPSEL32R5-NEXT:    st.w $w0, 0($5)
 entry:
@@ -126,20 +110,17 @@ entry:
   ret void
 }
 
-; fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
-define void @combine_shifts_to_and_mask_type2_i64_long(<2 x i64>* %a, <2 x i64>* %b) {
-; MIPSEL64R6-LABEL: combine_shifts_to_and_mask_type2_i64_long:
+; do not fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
+define void @avoid_to_combine_shifts_to_and_mask_type2_i64_long(<2 x i64>* %a, <2 x i64>* %b) {
+; MIPSEL64R6-LABEL: avoid_to_combine_shifts_to_and_mask_type2_i64_long:
 ; MIPSEL64R6:       # %bb.0: # %entry
-; MIPSEL64R6-NEXT:    lui $1, 65535
-; MIPSEL64R6-NEXT:    ori $1, $1, 65472
-; MIPSEL64R6-NEXT:    dsll32 $1, $1, 0
 ; MIPSEL64R6-NEXT:    ld.d $w0, 0($4)
-; MIPSEL64R6-NEXT:    fill.d $w1, $1
-; MIPSEL64R6-NEXT:    and.v $w0, $w0, $w1
+; MIPSEL64R6-NEXT:    srli.d $w0, $w0, 38
+; MIPSEL64R6-NEXT:    slli.d $w0, $w0, 38
 ; MIPSEL64R6-NEXT:    jr $ra
 ; MIPSEL64R6-NEXT:    st.d $w0, 0($5)
 ;
-; MIPSEL32R5-LABEL: combine_shifts_to_and_mask_type2_i64_long:
+; MIPSEL32R5-LABEL: avoid_to_combine_shifts_to_and_mask_type2_i64_long:
 ; MIPSEL32R5:       # %bb.0: # %entry
 ; MIPSEL32R5-NEXT:    ld.d $w0, 0($4)
 ; MIPSEL32R5-NEXT:    srli.d $w0, $w0, 38
@@ -154,17 +135,17 @@ entry:
   ret void
 }
 
-; fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
-define void @combine_shifts_to_and_mask_type2_i64(<2 x i64>* %a, <2 x i64>* %b) {
-; MIPSEL64R6-LABEL: combine_shifts_to_and_mask_type2_i64:
+; do not fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
+define void @avoid_to_combine_shifts_to_and_mask_type2_i64(<2 x i64>* %a, <2 x i64>* %b) {
+; MIPSEL64R6-LABEL: avoid_to_combine_shifts_to_and_mask_type2_i64:
 ; MIPSEL64R6:       # %bb.0: # %entry
 ; MIPSEL64R6-NEXT:    ld.d $w0, 0($4)
-; MIPSEL64R6-NEXT:    ldi.d $w1, -8
-; MIPSEL64R6-NEXT:    and.v $w0, $w0, $w1
+; MIPSEL64R6-NEXT:    srli.d $w0, $w0, 3
+; MIPSEL64R6-NEXT:    slli.d $w0, $w0, 3
 ; MIPSEL64R6-NEXT:    jr $ra
 ; MIPSEL64R6-NEXT:    st.d $w0, 0($5)
 ;
-; MIPSEL32R5-LABEL: combine_shifts_to_and_mask_type2_i64:
+; MIPSEL32R5-LABEL: avoid_to_combine_shifts_to_and_mask_type2_i64:
 ; MIPSEL32R5:       # %bb.0: # %entry
 ; MIPSEL32R5-NEXT:    ld.d $w0, 0($4)
 ; MIPSEL32R5-NEXT:    srli.d $w0, $w0, 3
@@ -179,21 +160,21 @@ entry:
   ret void
 }
 
-; fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
-define void @combine_shifts_to_and_mask_type1_long_i32_a(<4 x i32>* %a, <4 x i32>* %b) {
-; MIPSEL64R6-LABEL: combine_shifts_to_and_mask_type1_long_i32_a:
+; do not fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
+define void @avoid_to_combine_shifts_to_and_mask_type1_long_i32_a(<4 x i32>* %a, <4 x i32>* %b) {
+; MIPSEL64R6-LABEL: avoid_to_combine_shifts_to_and_mask_type1_long_i32_a:
 ; MIPSEL64R6:       # %bb.0: # %entry
 ; MIPSEL64R6-NEXT:    ld.w $w0, 0($4)
-; MIPSEL64R6-NEXT:    ldi.w $w1, -32
-; MIPSEL64R6-NEXT:    and.v $w0, $w0, $w1
+; MIPSEL64R6-NEXT:    srli.w $w0, $w0, 5
+; MIPSEL64R6-NEXT:    slli.w $w0, $w0, 5
 ; MIPSEL64R6-NEXT:    jr $ra
 ; MIPSEL64R6-NEXT:    st.w $w0, 0($5)
 ;
-; MIPSEL32R5-LABEL: combine_shifts_to_and_mask_type1_long_i32_a:
+; MIPSEL32R5-LABEL: avoid_to_combine_shifts_to_and_mask_type1_long_i32_a:
 ; MIPSEL32R5:       # %bb.0: # %entry
 ; MIPSEL32R5-NEXT:    ld.w $w0, 0($4)
-; MIPSEL32R5-NEXT:    ldi.w $w1, -32
-; MIPSEL32R5-NEXT:    and.v $w0, $w0, $w1
+; MIPSEL32R5-NEXT:    srli.w $w0, $w0, 5
+; MIPSEL32R5-NEXT:    slli.w $w0, $w0, 5
 ; MIPSEL32R5-NEXT:    jr $ra
 ; MIPSEL32R5-NEXT:    st.w $w0, 0($5)
 entry:
@@ -204,23 +185,21 @@ entry:
   ret void
 }
 
-; fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
-define void @combine_shifts_to_and_mask_type1_long_i32_b(<4 x i32>* %a, <4 x i32>* %b) {
-; MIPSEL64R6-LABEL: combine_shifts_to_and_mask_type1_long_i32_b:
+; do not fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
+define void @avoid_to_combine_shifts_to_and_mask_type1_long_i32_b(<4 x i32>* %a, <4 x i32>* %b) {
+; MIPSEL64R6-LABEL: avoid_to_combine_shifts_to_and_mask_type1_long_i32_b:
 ; MIPSEL64R6:       # %bb.0: # %entry
-; MIPSEL64R6-NEXT:    lui $1, 49152
-; MIPSEL64R6-NEXT:    fill.w $w0, $1
-; MIPSEL64R6-NEXT:    ld.w $w1, 0($4)
-; MIPSEL64R6-NEXT:    and.v $w0, $w1, $w0
+; MIPSEL64R6-NEXT:    ld.w $w0, 0($4)
+; MIPSEL64R6-NEXT:    srli.w $w0, $w0, 30
+; MIPSEL64R6-NEXT:    slli.w $w0, $w0, 30
 ; MIPSEL64R6-NEXT:    jr $ra
 ; MIPSEL64R6-NEXT:    st.w $w0, 0($5)
 ;
-; MIPSEL32R5-LABEL: combine_shifts_to_and_mask_type1_long_i32_b:
+; MIPSEL32R5-LABEL: avoid_to_combine_shifts_to_and_mask_type1_long_i32_b:
 ; MIPSEL32R5:       # %bb.0: # %entry
-; MIPSEL32R5-NEXT:    lui $1, 49152
-; MIPSEL32R5-NEXT:    fill.w $w0, $1
-; MIPSEL32R5-NEXT:    ld.w $w1, 0($4)
-; MIPSEL32R5-NEXT:    and.v $w0, $w1, $w0
+; MIPSEL32R5-NEXT:    ld.w $w0, 0($4)
+; MIPSEL32R5-NEXT:    srli.w $w0, $w0, 30
+; MIPSEL32R5-NEXT:    slli.w $w0, $w0, 30
 ; MIPSEL32R5-NEXT:    jr $ra
 ; MIPSEL32R5-NEXT:    st.w $w0, 0($5)
 entry:
-- 
2.40.0