[MIPS GlobalISel] Fold load/store + G_GEP + G_CONSTANT

Fold load/store + G_GEP + G_CONSTANT when immediate in G_CONSTANT fits into 16 bit signed integer. Differential Revision: https://reviews.llvm.org/D65507 llvm-svn: 367535
llvm · Aug 1, 2019 · 8a40ced · 8a40ced
1 parent 0cfa72e
commit 8a40ced
Show file tree

Hide file tree

Showing 8 changed files with 357 additions and 43 deletions.
diff --git a/llvm/lib/Target/Mips/MipsInstructionSelector.cpp b/llvm/lib/Target/Mips/MipsInstructionSelector.cpp
@@ -318,10 +318,31 @@ bool MipsInstructionSelector::select(MachineInstr &I,
     if (NewOpc == I.getOpcode())
       return false;
 
+    MachineOperand BaseAddr = I.getOperand(1);
+    int64_t SignedOffset = 0;
+    // Try to fold load/store + G_GEP + G_CONSTANT
+    // %SignedOffset:(s32) = G_CONSTANT i32 16_bit_signed_immediate
+    // %Addr:(p0) = G_GEP %BaseAddr, %SignedOffset
+    // %LoadResult/%StoreSrc = load/store %Addr(p0)
+    // into:
+    // %LoadResult/%StoreSrc = NewOpc %BaseAddr(p0), 16_bit_signed_immediate
+
+    MachineInstr *Addr = MRI.getVRegDef(I.getOperand(1).getReg());
+    if (Addr->getOpcode() == G_GEP) {
+      MachineInstr *Offset = MRI.getVRegDef(Addr->getOperand(2).getReg());
+      if (Offset->getOpcode() == G_CONSTANT) {
+        APInt OffsetValue = Offset->getOperand(1).getCImm()->getValue();
+        if (OffsetValue.isSignedIntN(16)) {
+          BaseAddr = Addr->getOperand(1);
+          SignedOffset = OffsetValue.getSExtValue();
+        }
+      }
+    }
+
     MI = BuildMI(MBB, I, I.getDebugLoc(), TII.get(NewOpc))
              .add(I.getOperand(0))
-             .add(I.getOperand(1))
-             .addImm(0)
+             .add(BaseAddr)
+             .addImm(SignedOffset)
              .addMemOperand(*I.memoperands_begin());
     break;
   }

diff --git a/llvm/test/CodeGen/Mips/GlobalISel/instruction-select/load_store_fold.mir b/llvm/test/CodeGen/Mips/GlobalISel/instruction-select/load_store_fold.mir
@@ -0,0 +1,214 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -O0 -mtriple=mipsel-linux-gnu -run-pass=instruction-select -verify-machineinstrs %s -o - | FileCheck %s -check-prefixes=MIPS32
+--- |
+
+  define void @_16_bit_positive_offset() {entry: ret void}
+  define void @_16_bit_negative_offset() {entry: ret void}
+  define void @_large_positive_offset() {entry: ret void}
+  define void @_large_negative_offset() {entry: ret void}
+  define void @fold_f32_load() {entry: ret void}
+  define void @fold_f64_store() {entry: ret void}
+  define void @fold_i16_load() {entry: ret void}
+  define void @fold_i32_store() {entry: ret void}
+
+...
+---
+name:            _16_bit_positive_offset
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $a0
+
+    ; MIPS32-LABEL: name: _16_bit_positive_offset
+    ; MIPS32: liveins: $a0
+    ; MIPS32: [[COPY:%[0-9]+]]:gpr32 = COPY $a0
+    ; MIPS32: [[LBu:%[0-9]+]]:gpr32 = LBu [[COPY]], 32767 :: (load 1)
+    ; MIPS32: $v0 = COPY [[LBu]]
+    ; MIPS32: RetRA implicit $v0
+    %0:gprb(p0) = COPY $a0
+    %1:gprb(s32) = G_CONSTANT i32 32767
+    %2:gprb(p0) = G_GEP %0, %1(s32)
+    %4:gprb(s32) = G_ZEXTLOAD %2(p0) :: (load 1)
+    $v0 = COPY %4(s32)
+    RetRA implicit $v0
+
+...
+---
+name:            _16_bit_negative_offset
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $a0, $a1
+
+    ; MIPS32-LABEL: name: _16_bit_negative_offset
+    ; MIPS32: liveins: $a0, $a1
+    ; MIPS32: [[COPY:%[0-9]+]]:gpr32 = COPY $a0
+    ; MIPS32: [[COPY1:%[0-9]+]]:gpr32 = COPY $a1
+    ; MIPS32: SB [[COPY]], [[COPY1]], -32768 :: (store 1)
+    ; MIPS32: RetRA
+    %2:gprb(s32) = COPY $a0
+    %1:gprb(p0) = COPY $a1
+    %3:gprb(s32) = G_CONSTANT i32 -32768
+    %4:gprb(p0) = G_GEP %1, %3(s32)
+    %5:gprb(s32) = COPY %2(s32)
+    G_STORE %5(s32), %4(p0) :: (store 1)
+    RetRA
+
+...
+---
+name:            _large_positive_offset
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $a0, $a1
+
+    ; MIPS32-LABEL: name: _large_positive_offset
+    ; MIPS32: liveins: $a0, $a1
+    ; MIPS32: [[COPY:%[0-9]+]]:gpr32 = COPY $a0
+    ; MIPS32: [[COPY1:%[0-9]+]]:gpr32 = COPY $a1
+    ; MIPS32: [[ORi:%[0-9]+]]:gpr32 = ORi $zero, 32768
+    ; MIPS32: [[ADDu:%[0-9]+]]:gpr32 = ADDu [[COPY1]], [[ORi]]
+    ; MIPS32: SB [[COPY]], [[ADDu]], 0 :: (store 1)
+    ; MIPS32: RetRA
+    %2:gprb(s32) = COPY $a0
+    %1:gprb(p0) = COPY $a1
+    %3:gprb(s32) = G_CONSTANT i32 32768
+    %4:gprb(p0) = G_GEP %1, %3(s32)
+    %5:gprb(s32) = COPY %2(s32)
+    G_STORE %5(s32), %4(p0) :: (store 1)
+    RetRA
+
+...
+---
+name:            _large_negative_offset
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $a0
+
+    ; MIPS32-LABEL: name: _large_negative_offset
+    ; MIPS32: liveins: $a0
+    ; MIPS32: [[COPY:%[0-9]+]]:gpr32 = COPY $a0
+    ; MIPS32: [[LUi:%[0-9]+]]:gpr32 = LUi 65535
+    ; MIPS32: [[ORi:%[0-9]+]]:gpr32 = ORi [[LUi]], 32767
+    ; MIPS32: [[ADDu:%[0-9]+]]:gpr32 = ADDu [[COPY]], [[ORi]]
+    ; MIPS32: [[LB:%[0-9]+]]:gpr32 = LB [[ADDu]], 0 :: (load 1)
+    ; MIPS32: $v0 = COPY [[LB]]
+    ; MIPS32: RetRA implicit $v0
+    %0:gprb(p0) = COPY $a0
+    %1:gprb(s32) = G_CONSTANT i32 -32769
+    %2:gprb(p0) = G_GEP %0, %1(s32)
+    %4:gprb(s32) = G_SEXTLOAD %2(p0) :: (load 1)
+    $v0 = COPY %4(s32)
+    RetRA implicit $v0
+
+...
+---
+name:            fold_f32_load
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $a0
+
+    ; MIPS32-LABEL: name: fold_f32_load
+    ; MIPS32: liveins: $a0
+    ; MIPS32: [[COPY:%[0-9]+]]:gpr32 = COPY $a0
+    ; MIPS32: [[LWC1_:%[0-9]+]]:fgr32 = LWC1 [[COPY]], 40 :: (load 4)
+    ; MIPS32: $f0 = COPY [[LWC1_]]
+    ; MIPS32: RetRA implicit $f0
+    %0:gprb(p0) = COPY $a0
+    %1:gprb(s32) = G_CONSTANT i32 40
+    %2:gprb(p0) = G_GEP %0, %1(s32)
+    %3:fprb(s32) = G_LOAD %2(p0) :: (load 4)
+    $f0 = COPY %3(s32)
+    RetRA implicit $f0
+
+...
+---
+name:            fold_f64_store
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $a2, $d6
+
+    ; MIPS32-LABEL: name: fold_f64_store
+    ; MIPS32: liveins: $a2, $d6
+    ; MIPS32: [[COPY:%[0-9]+]]:afgr64 = COPY $d6
+    ; MIPS32: [[COPY1:%[0-9]+]]:gpr32 = COPY $a2
+    ; MIPS32: SDC1 [[COPY]], [[COPY1]], -80 :: (store 8)
+    ; MIPS32: RetRA
+    %0:fprb(s64) = COPY $d6
+    %1:gprb(p0) = COPY $a2
+    %2:gprb(s32) = G_CONSTANT i32 -80
+    %3:gprb(p0) = G_GEP %1, %2(s32)
+    G_STORE %0(s64), %3(p0) :: (store 8)
+    RetRA
+
+...
+---
+name:            fold_i16_load
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $a0
+
+    ; MIPS32-LABEL: name: fold_i16_load
+    ; MIPS32: liveins: $a0
+    ; MIPS32: [[COPY:%[0-9]+]]:gpr32 = COPY $a0
+    ; MIPS32: [[LHu:%[0-9]+]]:gpr32 = LHu [[COPY]], -20 :: (load 2)
+    ; MIPS32: $v0 = COPY [[LHu]]
+    ; MIPS32: RetRA implicit $v0
+    %0:gprb(p0) = COPY $a0
+    %1:gprb(s32) = G_CONSTANT i32 -20
+    %2:gprb(p0) = G_GEP %0, %1(s32)
+    %4:gprb(s32) = G_LOAD %2(p0) :: (load 2)
+    $v0 = COPY %4(s32)
+    RetRA implicit $v0
+
+...
+---
+name:            fold_i32_store
+alignment:       2
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $a0, $a1
+
+    ; MIPS32-LABEL: name: fold_i32_store
+    ; MIPS32: liveins: $a0, $a1
+    ; MIPS32: [[COPY:%[0-9]+]]:gpr32 = COPY $a0
+    ; MIPS32: [[COPY1:%[0-9]+]]:gpr32 = COPY $a1
+    ; MIPS32: SW [[COPY]], [[COPY1]], 40 :: (store 4)
+    ; MIPS32: RetRA
+    %0:gprb(s32) = COPY $a0
+    %1:gprb(p0) = COPY $a1
+    %2:gprb(s32) = G_CONSTANT i32 40
+    %3:gprb(p0) = G_GEP %1, %2(s32)
+    G_STORE %0(s32), %3(p0) :: (store 4)
+    RetRA
+
+...
+
diff --git a/llvm/test/CodeGen/Mips/GlobalISel/instruction-select/stack_args.mir b/llvm/test/CodeGen/Mips/GlobalISel/instruction-select/stack_args.mir
@@ -32,9 +32,7 @@ body:             |
     ; MIPS32: $a2 = COPY [[COPY2]]
     ; MIPS32: $a3 = COPY [[COPY3]]
     ; MIPS32: [[COPY4:%[0-9]+]]:gpr32 = COPY $sp
-    ; MIPS32: [[ORi:%[0-9]+]]:gpr32 = ORi $zero, 1
-    ; MIPS32: [[ADDu:%[0-9]+]]:gpr32 = ADDu [[COPY4]], [[ORi]]
-    ; MIPS32: SW [[LW]], [[ADDu]], 0 :: (store 4 into stack + 16)
+    ; MIPS32: SW [[LW]], [[COPY4]], 16 :: (store 4 into stack + 16)
     ; MIPS32: JAL @f, csr_o32, implicit-def $ra, implicit-def $sp, implicit $a0, implicit $a1, implicit $a2, implicit $a3, implicit-def $v0
     ; MIPS32: [[COPY5:%[0-9]+]]:gpr32 = COPY $v0
     ; MIPS32: ADJCALLSTACKUP 24, 0, implicit-def $sp, implicit $sp

diff --git a/llvm/test/CodeGen/Mips/GlobalISel/llvm-ir/load.ll b/llvm/test/CodeGen/Mips/GlobalISel/llvm-ir/load.ll
@@ -16,9 +16,7 @@ define i64 @load_i64(i64* %ptr) {
 ; MIPS32-LABEL: load_i64:
 ; MIPS32:       # %bb.0: # %entry
 ; MIPS32-NEXT:    lw $2, 0($4)
-; MIPS32-NEXT:    ori $1, $zero, 4
-; MIPS32-NEXT:    addu $1, $4, $1
-; MIPS32-NEXT:    lw $3, 0($1)
+; MIPS32-NEXT:    lw $3, 4($4)
 ; MIPS32-NEXT:    jr $ra
 ; MIPS32-NEXT:    nop
 entry:

diff --git a/llvm/test/CodeGen/Mips/GlobalISel/llvm-ir/load_store_fold.ll b/llvm/test/CodeGen/Mips/GlobalISel/llvm-ir/load_store_fold.ll
@@ -0,0 +1,105 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc  -O0 -mtriple=mipsel-linux-gnu -global-isel  -verify-machineinstrs %s -o -| FileCheck %s -check-prefixes=MIPS32
+
+define zeroext i8 @_16_bit_positive_offset(i8* %a) {
+; MIPS32-LABEL: _16_bit_positive_offset:
+; MIPS32:       # %bb.0: # %entry
+; MIPS32-NEXT:    lbu $2, 32767($4)
+; MIPS32-NEXT:    jr $ra
+; MIPS32-NEXT:    nop
+entry:
+  %arrayidx = getelementptr inbounds i8, i8* %a, i32 32767
+  %0 = load i8, i8* %arrayidx
+  ret i8 %0
+}
+
+define void @_16_bit_negative_offset(i8 %val, i8* %a) {
+; MIPS32-LABEL: _16_bit_negative_offset:
+; MIPS32:       # %bb.0: # %entry
+; MIPS32-NEXT:    sb $4, -32768($5)
+; MIPS32-NEXT:    jr $ra
+; MIPS32-NEXT:    nop
+entry:
+  %arrayidx = getelementptr inbounds i8, i8* %a, i32 -32768
+  store i8 %val, i8* %arrayidx
+  ret void
+}
+
+define void @_large_positive_offset(i8 %val, i8* %a) {
+; MIPS32-LABEL: _large_positive_offset:
+; MIPS32:       # %bb.0: # %entry
+; MIPS32-NEXT:    ori $1, $zero, 32768
+; MIPS32-NEXT:    addu $1, $5, $1
+; MIPS32-NEXT:    sb $4, 0($1)
+; MIPS32-NEXT:    jr $ra
+; MIPS32-NEXT:    nop
+entry:
+  %arrayidx = getelementptr inbounds i8, i8* %a, i32 32768
+  store i8 %val, i8* %arrayidx
+  ret void
+}
+
+define signext i8 @_large_negative_offset(i8* %a) {
+; MIPS32-LABEL: _large_negative_offset:
+; MIPS32:       # %bb.0: # %entry
+; MIPS32-NEXT:    lui $1, 65535
+; MIPS32-NEXT:    ori $1, $1, 32767
+; MIPS32-NEXT:    addu $1, $4, $1
+; MIPS32-NEXT:    lb $2, 0($1)
+; MIPS32-NEXT:    jr $ra
+; MIPS32-NEXT:    nop
+entry:
+  %arrayidx = getelementptr inbounds i8, i8* %a, i32 -32769
+  %0 = load i8, i8* %arrayidx
+  ret i8 %0
+}
+
+define float @fold_f32_load(float* %a) {
+; MIPS32-LABEL: fold_f32_load:
+; MIPS32:       # %bb.0: # %entry
+; MIPS32-NEXT:    lwc1 $f0, 40($4)
+; MIPS32-NEXT:    jr $ra
+; MIPS32-NEXT:    nop
+entry:
+  %arrayidx = getelementptr inbounds float, float* %a, i32 10
+  %0 = load float, float* %arrayidx
+  ret float %0
+}
+
+define void @fold_f64_store(double %val, double* %a) {
+; MIPS32-LABEL: fold_f64_store:
+; MIPS32:       # %bb.0: # %entry
+; MIPS32-NEXT:    sdc1 $f12, -80($6)
+; MIPS32-NEXT:    jr $ra
+; MIPS32-NEXT:    nop
+entry:
+  %arrayidx = getelementptr inbounds double, double* %a, i32 -10
+  store double %val, double* %arrayidx
+  ret void
+}
+
+define i16 @fold_i16_load(i16* %a) {
+; MIPS32-LABEL: fold_i16_load:
+; MIPS32:       # %bb.0: # %entry
+; MIPS32-NEXT:    lhu $2, -20($4)
+; MIPS32-NEXT:    jr $ra
+; MIPS32-NEXT:    nop
+entry:
+  %arrayidx = getelementptr inbounds i16, i16* %a, i32 -10
+  %0 = load i16, i16* %arrayidx
+  ret i16 %0
+}
+
+define void @fold_i32_store(i32 %val, i32* %a) {
+; MIPS32-LABEL: fold_i32_store:
+; MIPS32:       # %bb.0: # %entry
+; MIPS32-NEXT:    sw $4, 40($5)
+; MIPS32-NEXT:    jr $ra
+; MIPS32-NEXT:    nop
+entry:
+  %arrayidx = getelementptr inbounds i32, i32* %a, i32 10
+  store i32 %val, i32* %arrayidx
+  ret void
+}
+
+