[CSKY] Optimize subtraction with SUBI32/SUBI16

[benshi001]

llvm/test/CodeGen/CSKY/base-i.ll

46

4096 is the upper boundary of ADDI32.

291

4096 is upper boundary of SUBI32.

[benshi001]

llvm/test/CodeGen/CSKY/base-i.ll

44

256 is the upper boundary of ADDI16.

292

256 is the upper boundary of SUBI16.

349

4097 is not optimized to SUBI32 due to out of range.

[benshi001]

ping ...

[zixuan-wu]

llvm/lib/Target/CSKY/CSKYInstrInfo.td

105

I think there is no oimm specific meaning, just use imm_neg or neg_imm would be fine.

llvm/test/CodeGen/CSKY/base-i.ll

43

Is this case related to this patch change?

[benshi001]

llvm/test/CodeGen/CSKY/base-i.ll

43

Yes. I would like to show the upper boundary of ADDI16.

[zixuan-wu]

llvm/lib/Target/CSKY/CSKYInstrInfo.td

537

I think the similar Pat for SUBI16 should also be added when there is no SUBI32

[benshi001]

llvm/lib/Target/CSKY/CSKYInstrInfo.td

537

I have added in the file CSKYInstrInfo16Instr.td.

[zixuan-wu]

BTW, as the check-all test is not enough, have you run llvm-test-suite sometimes although it's not needed every patch? I think should run llvm-test-suite in qemu from time to time.

[benshi001]

In D153326#4462433, @zixuan-wu wrote:

BTW, as the check-all test is not enough, have you run llvm-test-suite sometimes although it's not needed every patch? I think should run llvm-test-suite in qemu from time to time.

Thanks. I will try. Further questions,

1. Is there any QEMU pre-build package I can use?
2. Also csky-binutils pre-build package, where I can download?
3. Any other tools are needed to run llvm-test-suite ? Is there a wiki page for that ?

[zixuan-wu]

In D153326#4462460, @benshi001 wrote:

In D153326#4462433, @zixuan-wu wrote:

BTW, as the check-all test is not enough, have you run llvm-test-suite sometimes although it's not needed every patch? I think should run llvm-test-suite in qemu from time to time.

Thanks. I will try. Further questions,

1. Is there any QEMU pre-build package I can use?
2. Also csky-binutils pre-build package, where I can download?
3. Any other tools are needed to run llvm-test-suite ? Is there a wiki page for that ?

1. You can download at https://xuantie.t-head.cn/community/download?id=4168444414324183040
2. clang on csky depends to gcc runtime so it needs gcc toolchain, you can download at https://xuantie.t-head.cn/community/download?id=3854599713706024960
3. Sorry about no document to describe the run process of llvm-test-suite in csky. But there is llvm build bot you can refer to. https://lab.llvm.org/staging/#/builders/215/builds/2795

[benshi001]

In D153326#4462511, @zixuan-wu wrote:

In D153326#4462460, @benshi001 wrote:

In D153326#4462433, @zixuan-wu wrote:

BTW, as the check-all test is not enough, have you run llvm-test-suite sometimes although it's not needed every patch? I think should run llvm-test-suite in qemu from time to time.

Thanks. I will try. Further questions,

1. Is there any QEMU pre-build package I can use?
2. Also csky-binutils pre-build package, where I can download?
3. Any other tools are needed to run llvm-test-suite ? Is there a wiki page for that ?

1. You can download at https://xuantie.t-head.cn/community/download?id=4168444414324183040
2. clang on csky depends to gcc runtime so it needs gcc toolchain, you can download at https://xuantie.t-head.cn/community/download?id=3854599713706024960
3. Sorry about no document to describe the run process of llvm-test-suite in csky. But there is llvm build bot you can refer to. https://lab.llvm.org/staging/#/builders/215/builds/2795

Thanks. I will try!

[benshi001]

In D153326#4462511, @zixuan-wu wrote:

In D153326#4462460, @benshi001 wrote:

In D153326#4462433, @zixuan-wu wrote:

BTW, as the check-all test is not enough, have you run llvm-test-suite sometimes although it's not needed every patch? I think should run llvm-test-suite in qemu from time to time.

Thanks. I will try. Further questions,

1. Is there any QEMU pre-build package I can use?
2. Also csky-binutils pre-build package, where I can download?
3. Any other tools are needed to run llvm-test-suite ? Is there a wiki page for that ?

1. You can download at https://xuantie.t-head.cn/community/download?id=4168444414324183040
2. clang on csky depends to gcc runtime so it needs gcc toolchain, you can download at https://xuantie.t-head.cn/community/download?id=3854599713706024960
3. Sorry about no document to describe the run process of llvm-test-suite in csky. But there is llvm build bot you can refer to. https://lab.llvm.org/staging/#/builders/215/builds/2795

I have run the SingleSource and MultiSource of the llvm-test-suite with this patch applied, there are several failures, and several link errors.

Failures of SingleSource

********************
********************
Failed Tests (1):
  test-suite :: SingleSource/Benchmarks/SmallPT/smallpt.test


Testing Time: 11971.65s
  Passed: 324
  Failed:   1

Failures of MultiSource

********************
********************
Failed Tests (4):
  test-suite :: MultiSource/Applications/ClamAV/clamscan.test
  test-suite :: MultiSource/Benchmarks/SciMark2-C/scimark2.test
  test-suite :: MultiSource/Benchmarks/TSVC/NodeSplitting-dbl/NodeSplitting-dbl.test
  test-suite :: MultiSource/Benchmarks/mafft/pairlocalalign.test


Testing Time: 11903.17s
  Passed: 185
  Failed:   4

And the cases disabled for link errors

MultiSource/Benchmarks/Bullet
MultiSource/Benchmarks/tramp3d-v4
MultiSource/Benchmarks/7zip
MultiSource/Benchmarks/PAQ8p
SingleSource/Benchmarks/Misc-C++-EH
SingleSource/Regression/C++/2003-06-08-BaseType.cpp
SingleSource/UnitTests/AtomicOps.c

Is that result OK?

[zixuan-wu]

In D153326#4466875, @benshi001 wrote:

In D153326#4462511, @zixuan-wu wrote:

In D153326#4462460, @benshi001 wrote:

In D153326#4462433, @zixuan-wu wrote:

BTW, as the check-all test is not enough, have you run llvm-test-suite sometimes although it's not needed every patch? I think should run llvm-test-suite in qemu from time to time.

Thanks. I will try. Further questions,

1. Is there any QEMU pre-build package I can use?
2. Also csky-binutils pre-build package, where I can download?
3. Any other tools are needed to run llvm-test-suite ? Is there a wiki page for that ?

1. You can download at https://xuantie.t-head.cn/community/download?id=4168444414324183040
2. clang on csky depends to gcc runtime so it needs gcc toolchain, you can download at https://xuantie.t-head.cn/community/download?id=3854599713706024960
3. Sorry about no document to describe the run process of llvm-test-suite in csky. But there is llvm build bot you can refer to. https://lab.llvm.org/staging/#/builders/215/builds/2795

I have run the SingleSource and MultiSource of the llvm-test-suite with this patch applied, there are several failures, and several link errors.

Failures of SingleSource

********************
********************
Failed Tests (1):
  test-suite :: SingleSource/Benchmarks/SmallPT/smallpt.test


Testing Time: 11971.65s
  Passed: 324
  Failed:   1

Failures of MultiSource

********************
********************
Failed Tests (4):
  test-suite :: MultiSource/Applications/ClamAV/clamscan.test
  test-suite :: MultiSource/Benchmarks/SciMark2-C/scimark2.test
  test-suite :: MultiSource/Benchmarks/TSVC/NodeSplitting-dbl/NodeSplitting-dbl.test
  test-suite :: MultiSource/Benchmarks/mafft/pairlocalalign.test


Testing Time: 11903.17s
  Passed: 185
  Failed:   4

And the cases disabled for link errors

MultiSource/Benchmarks/Bullet
MultiSource/Benchmarks/tramp3d-v4
MultiSource/Benchmarks/7zip
MultiSource/Benchmarks/PAQ8p
SingleSource/Benchmarks/Misc-C++-EH
SingleSource/Regression/C++/2003-06-08-BaseType.cpp
SingleSource/UnitTests/AtomicOps.c

Is that result OK?

There is only one fail case clamscan.test, all of others should be passed.link error is about to add extra compiling flag of "-latomic".

As the two buildbot are success(exception for the only clamscan.test fails), https://lab.llvm.org/staging/#/builders/216/builds/1806 and https://lab.llvm.org/staging/#/builders/215/builds/2819, I think maybe you got wrong run command of qemu. And could you please provide your run command of qemu?

[benshi001]

In D153326#4467374, @zixuan-wu wrote:

In D153326#4466875, @benshi001 wrote:

In D153326#4462511, @zixuan-wu wrote:

In D153326#4462460, @benshi001 wrote:

In D153326#4462433, @zixuan-wu wrote:

BTW, as the check-all test is not enough, have you run llvm-test-suite sometimes although it's not needed every patch? I think should run llvm-test-suite in qemu from time to time.

Thanks. I will try. Further questions,

1. Is there any QEMU pre-build package I can use?
2. Also csky-binutils pre-build package, where I can download?
3. Any other tools are needed to run llvm-test-suite ? Is there a wiki page for that ?

1. You can download at https://xuantie.t-head.cn/community/download?id=4168444414324183040
2. clang on csky depends to gcc runtime so it needs gcc toolchain, you can download at https://xuantie.t-head.cn/community/download?id=3854599713706024960
3. Sorry about no document to describe the run process of llvm-test-suite in csky. But there is llvm build bot you can refer to. https://lab.llvm.org/staging/#/builders/215/builds/2795

I have run the SingleSource and MultiSource of the llvm-test-suite with this patch applied, there are several failures, and several link errors.

Failures of SingleSource

********************
********************
Failed Tests (1):
  test-suite :: SingleSource/Benchmarks/SmallPT/smallpt.test


Testing Time: 11971.65s
  Passed: 324
  Failed:   1

Failures of MultiSource

********************
********************
Failed Tests (4):
  test-suite :: MultiSource/Applications/ClamAV/clamscan.test
  test-suite :: MultiSource/Benchmarks/SciMark2-C/scimark2.test
  test-suite :: MultiSource/Benchmarks/TSVC/NodeSplitting-dbl/NodeSplitting-dbl.test
  test-suite :: MultiSource/Benchmarks/mafft/pairlocalalign.test


Testing Time: 11903.17s
  Passed: 185
  Failed:   4

And the cases disabled for link errors

MultiSource/Benchmarks/Bullet
MultiSource/Benchmarks/tramp3d-v4
MultiSource/Benchmarks/7zip
MultiSource/Benchmarks/PAQ8p
SingleSource/Benchmarks/Misc-C++-EH
SingleSource/Regression/C++/2003-06-08-BaseType.cpp
SingleSource/UnitTests/AtomicOps.c

Is that result OK?

There is only one fail case clamscan.test, all of others should be passed.link error is about to add extra compiling flag of "-latomic".

As the two buildbot are success(exception for the only clamscan.test fails), https://lab.llvm.org/staging/#/builders/216/builds/1806 and https://lab.llvm.org/staging/#/builders/215/builds/2819, I think maybe you got wrong run command of qemu. And could you please provide your run command of qemu?

I run Qemu by defining -DTEST_SUITE_RUN_UNDER="/usr/bin/qemu-cskyv2" when doing cmake, without specifying other flags. And that is fine, I see that my patch does not cause any harm, this is enough. Thanks!

[zixuan-wu]

In D153326#4467462, @benshi001 wrote:

In D153326#4467374, @zixuan-wu wrote:

In D153326#4466875, @benshi001 wrote:

In D153326#4462511, @zixuan-wu wrote:

In D153326#4462460, @benshi001 wrote:

In D153326#4462433, @zixuan-wu wrote:

BTW, as the check-all test is not enough, have you run llvm-test-suite sometimes although it's not needed every patch? I think should run llvm-test-suite in qemu from time to time.

Thanks. I will try. Further questions,

1. Is there any QEMU pre-build package I can use?
2. Also csky-binutils pre-build package, where I can download?
3. Any other tools are needed to run llvm-test-suite ? Is there a wiki page for that ?

1. You can download at https://xuantie.t-head.cn/community/download?id=4168444414324183040
2. clang on csky depends to gcc runtime so it needs gcc toolchain, you can download at https://xuantie.t-head.cn/community/download?id=3854599713706024960
3. Sorry about no document to describe the run process of llvm-test-suite in csky. But there is llvm build bot you can refer to. https://lab.llvm.org/staging/#/builders/215/builds/2795

I have run the SingleSource and MultiSource of the llvm-test-suite with this patch applied, there are several failures, and several link errors.

Failures of SingleSource

********************
********************
Failed Tests (1):
  test-suite :: SingleSource/Benchmarks/SmallPT/smallpt.test


Testing Time: 11971.65s
  Passed: 324
  Failed:   1

Failures of MultiSource

********************
********************
Failed Tests (4):
  test-suite :: MultiSource/Applications/ClamAV/clamscan.test
  test-suite :: MultiSource/Benchmarks/SciMark2-C/scimark2.test
  test-suite :: MultiSource/Benchmarks/TSVC/NodeSplitting-dbl/NodeSplitting-dbl.test
  test-suite :: MultiSource/Benchmarks/mafft/pairlocalalign.test


Testing Time: 11903.17s
  Passed: 185
  Failed:   4

And the cases disabled for link errors

MultiSource/Benchmarks/Bullet
MultiSource/Benchmarks/tramp3d-v4
MultiSource/Benchmarks/7zip
MultiSource/Benchmarks/PAQ8p
SingleSource/Benchmarks/Misc-C++-EH
SingleSource/Regression/C++/2003-06-08-BaseType.cpp
SingleSource/UnitTests/AtomicOps.c

Is that result OK?

There is only one fail case clamscan.test, all of others should be passed.link error is about to add extra compiling flag of "-latomic".

As the two buildbot are success(exception for the only clamscan.test fails), https://lab.llvm.org/staging/#/builders/216/builds/1806 and https://lab.llvm.org/staging/#/builders/215/builds/2819, I think maybe you got wrong run command of qemu. And could you please provide your run command of qemu?

I run Qemu by defining -DTEST_SUITE_RUN_UNDER="/usr/bin/qemu-cskyv2" when doing cmake, without specifying other flags. And that is fine, I see that my patch does not cause any harm, this is enough. Thanks!

Also need specify library path and other options. Please see https://github.com/llvm/llvm-zorg/blob/8006da069b5776b01fc5204d267e65e48f86c80e/buildbot/osuosl/master/config/builders.py#L2615.

llvm/lib/Target/CSKY/CSKYInstrInfo.td

 
 class oimm<int num> : Operand<i32>,
   ImmLeaf<i32, "return isUInt<"#num#">(Imm - 1);"> {
   let EncoderMethod = "getOImmOpValue";
   let ParserMatchClass = OImmAsmOperand<num>;
   let DecoderMethod = "decodeOImmOperand<"#num#">";
 }
 
+def oimm_neg_XFORM : SDNodeXForm<imm, [{
+  return CurDAG->getTargetConstant(-N->getSExtValue(), SDLoc(N), MVT::i32);
+}]>;
+class oimm_neg<int num> : Operand<i32>,
+  ImmLeaf<i32, "return isUInt<"#num#">(-Imm - 1);"> {
+}
+
 class uimm<int num, int shift = 0> : Operand<i32>,
   ImmLeaf<i32, "return isShiftedUInt<"#num#", "#shift#">(Imm);"> {
   let EncoderMethod = "getImmOpValue<"#shift#">";
   let ParserMatchClass =
     !if(!ne(shift, 0),
         UImmAsmOperand<num, "Shift"#shift>,
         UImmAsmOperand<num>);
   let DecoderMethod = "decodeUImmOperand<"#num#", "#shift#">";
def oimm16 : oimm<16> {
   let MCOperandPredicate = [{
     int64_t Imm;
     if (MCOp.evaluateAsConstantImm(Imm))
       return isUInt<16>(Imm - 1);
     return MCOp.isBareSymbolRef();
   }];
 }
 
+def oimm12_neg : oimm_neg<12> {
+  let MCOperandPredicate = [{
+    int64_t Imm;
+    if (MCOp.evaluateAsConstantImm(Imm))
+      return isUInt<12>(-Imm - 1);
+    return MCOp.isBareSymbolRef();
+  }];
+}
+
 def nimm12 : nimm<12>;
 
 def uimm1 : uimm<1>;
 def uimm2 : uimm<2>;
 
 
 def uimm2_jmpix : Operand<i32>,
   ImmLeaf<i32, "return Imm == 16 || Imm == 24 || Imm == 32 || Imm == 40;"> {
def NOT32 : R_XXZ<0b001001, 0b00100, (outs GPR:$rz), (ins GPR:$rx),
     "not32", [(set GPR:$rz, (not GPR:$rx))]>;
 
   let Size = 8 in
   def NEG32 : CSKYPseudo<(outs GPR:$rd), (ins GPR:$rx), "neg32 $rd, $rx", []>;
 
   let Size = 8 in
   def RSUBI32 : CSKYPseudo<(outs GPR:$rd), (ins GPR:$rx, uimm12:$imm12), "rsubi32 $rd, $rx, $imm12", []>;
 
+  def : Pat<(add GPR:$rs1, (oimm12_neg:$im)),
+            (SUBI32 GPR:$rs1, (oimm_neg_XFORM oimm12_neg:$im))>;
+
   def LSL32 : R_YXZ_SP_F1<0x10, 0x1,
     BinOpFrag<(shl node:$LHS, node:$RHS)>, "lsl32">;
   def LSR32 : R_YXZ_SP_F1<0x10, 0x2,
     BinOpFrag<(srl node:$LHS, node:$RHS)>, "lsr32">;
   def ASR32 : R_YXZ_SP_F1<0x10, 0x4,
     BinOpFrag<(sra node:$LHS, node:$RHS)>, "asr32">;
   def ROTL32 : R_YXZ_SP_F1<0x10, 0x8,
     BinOpFrag<(rotl node:$LHS, (and node:$RHS, 0x1f))>, "rotl32">;

llvm/test/CodeGen/CSKY/base-i.ll

 ; GENERIC-NEXT:    addi16 a0, 10
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 entry:
   %add = add nsw i32 %x, 10
   ret i32 %add
 }
 
+define i32 @addRI_256(i32 %x) {
+; CHECK-LABEL: addRI_256:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    addi16 a0, 256
+; CHECK-NEXT:    rts16
+;
+; GENERIC-LABEL: addRI_256:
+; GENERIC:       # %bb.0: # %entry
+; GENERIC-NEXT:    .cfi_def_cfa_offset 0
+; GENERIC-NEXT:    subi16 sp, sp, 4
+; GENERIC-NEXT:    .cfi_def_cfa_offset 4
+; GENERIC-NEXT:    addi16 a0, 256
+; GENERIC-NEXT:    addi16 sp, sp, 4
+; GENERIC-NEXT:    rts16
+entry:
+  %add = add nsw i32 %x, 256
+  ret i32 %add
+}
+
+define i32 @addRI_4096(i32 %x) {
+; CHECK-LABEL: addRI_4096:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    addi32 a0, a0, 4096
+; CHECK-NEXT:    rts16
+;
+; GENERIC-LABEL: addRI_4096:
+; GENERIC:       # %bb.0: # %entry
+; GENERIC-NEXT:    .cfi_def_cfa_offset 0
+; GENERIC-NEXT:    subi16 sp, sp, 4
+; GENERIC-NEXT:    .cfi_def_cfa_offset 4
+; GENERIC-NEXT:    movi16 a1, 0
+; GENERIC-NEXT:    lsli16 a2, a1, 24
+; GENERIC-NEXT:    lsli16 a3, a1, 16
+; GENERIC-NEXT:    or16 a3, a2
+; GENERIC-NEXT:    movi16 a2, 16
+; GENERIC-NEXT:    lsli16 a2, a2, 8
+; GENERIC-NEXT:    or16 a2, a3
+; GENERIC-NEXT:    or16 a2, a1
+; GENERIC-NEXT:    addu16 a0, a0, a2
+; GENERIC-NEXT:    addi16 sp, sp, 4
+; GENERIC-NEXT:    rts16
+entry:
+  %add = add nsw i32 %x, 4096
+  ret i32 %add
+}
+
 define i32 @addRI_X(i32 %x) {
 ; CHECK-LABEL: addRI_X:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    movi32 a1, 4097
 ; CHECK-NEXT:    addu16 a0, a1
 ; CHECK-NEXT:    rts16
 ;
 ; GENERIC-LABEL: addRI_X:

 entry:
   %sub = sub nsw i32 %y, %x
   ret i32 %sub
 }
 
 define i32 @subRI(i32 %x) {
 ; CHECK-LABEL: subRI:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    movih32 a1, 65535
-; CHECK-NEXT:    ori32 a1, a1, 65526
-; CHECK-NEXT:    addu16 a0, a1
+; CHECK-NEXT:    subi16 a0, 10
 ; CHECK-NEXT:    rts16
 ;
 ; GENERIC-LABEL: subRI:
 ; GENERIC:       # %bb.0: # %entry
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 0
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 4
 ; GENERIC-NEXT:    movi16 a1, 255
 ; GENERIC-NEXT:    lsli16 a2, a1, 24
 ; GENERIC-NEXT:    lsli16 a3, a1, 16
 ; GENERIC-NEXT:    or16 a3, a2
 ; GENERIC-NEXT:    lsli16 a1, a1, 8
 ; GENERIC-NEXT:    or16 a1, a3
 ; GENERIC-NEXT:    movi16 a2, 246
 ; GENERIC-NEXT:    or16 a2, a1
 ; GENERIC-NEXT:    addu16 a0, a0, a2
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 entry:
   %sub = sub nsw i32 %x, 10
   ret i32 %sub
 }
 
+define i32 @subRI_256(i32 %x) {
+; CHECK-LABEL: subRI_256:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    subi16 a0, 256
+; CHECK-NEXT:    rts16
+;
+; GENERIC-LABEL: subRI_256:
+; GENERIC:       # %bb.0: # %entry
+; GENERIC-NEXT:    .cfi_def_cfa_offset 0
+; GENERIC-NEXT:    subi16 sp, sp, 4
+; GENERIC-NEXT:    .cfi_def_cfa_offset 4
+; GENERIC-NEXT:    movi16 a1, 255
+; GENERIC-NEXT:    lsli16 a2, a1, 24
+; GENERIC-NEXT:    lsli16 a3, a1, 16
+; GENERIC-NEXT:    or16 a3, a2
+; GENERIC-NEXT:    lsli16 a1, a1, 8
+; GENERIC-NEXT:    or16 a1, a3
+; GENERIC-NEXT:    movi16 a2, 0
+; GENERIC-NEXT:    or16 a2, a1
+; GENERIC-NEXT:    addu16 a0, a0, a2
+; GENERIC-NEXT:    addi16 sp, sp, 4
+; GENERIC-NEXT:    rts16
+entry:
+  %sub = sub nsw i32 %x, 256
+  ret i32 %sub
+}
+
+define i32 @subRI_4096(i32 %x) {
+; CHECK-LABEL: subRI_4096:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    subi32 a0, a0, 4096
+; CHECK-NEXT:    rts16
+;
+; GENERIC-LABEL: subRI_4096:
+; GENERIC:       # %bb.0: # %entry
+; GENERIC-NEXT:    .cfi_def_cfa_offset 0
+; GENERIC-NEXT:    subi16 sp, sp, 4
+; GENERIC-NEXT:    .cfi_def_cfa_offset 4
+; GENERIC-NEXT:    movi16 a1, 255
+; GENERIC-NEXT:    lsli16 a2, a1, 24
+; GENERIC-NEXT:    lsli16 a1, a1, 16
+; GENERIC-NEXT:    or16 a1, a2
+; GENERIC-NEXT:    movi16 a2, 240
+; GENERIC-NEXT:    lsli16 a2, a2, 8
+; GENERIC-NEXT:    or16 a2, a1
+; GENERIC-NEXT:    movi16 a1, 0
+; GENERIC-NEXT:    or16 a1, a2
+; GENERIC-NEXT:    addu16 a0, a0, a1
+; GENERIC-NEXT:    addi16 sp, sp, 4
+; GENERIC-NEXT:    rts16
+entry:
+  %sub = sub nsw i32 %x, 4096
+  ret i32 %sub
+}
+
 define i32 @subRI_X(i32 %x) {
 ; CHECK-LABEL: subRI_X:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    movih32 a1, 65535
 ; CHECK-NEXT:    ori32 a1, a1, 61439
 ; CHECK-NEXT:    addu16 a0, a1
 ; CHECK-NEXT:    rts16
 ;
 ; GENERIC-LABEL: subRI_X:
 ; GENERIC:       # %bb.0: # %entry
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 0
 ; GENERIC-NEXT:    subi16 sp, sp, 4

 entry:
   %sub = sub nsw i16 %y, %x
   ret i16 %sub
 }
 
 define i16 @SUB_SHORT_I(i16 %x) {
 ; CHECK-LABEL: SUB_SHORT_I:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    movih32 a1, 65535
-; CHECK-NEXT:    ori32 a1, a1, 65535
-; CHECK-NEXT:    addu16 a0, a1
+; CHECK-NEXT:    subi16 a0, a0, 1
 ; CHECK-NEXT:    rts16
 ;
 ; GENERIC-LABEL: SUB_SHORT_I:
 ; GENERIC:       # %bb.0: # %entry
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 0
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 4
 ; GENERIC-NEXT:    movi16 a1, 255

 entry:
   %sub = sub nsw i8 %y, %x
   ret i8 %sub
 }
 
 define i8 @SUB_CHAR_I(i8 %x) {
 ; CHECK-LABEL: SUB_CHAR_I:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    movih32 a1, 65535
-; CHECK-NEXT:    ori32 a1, a1, 65535
-; CHECK-NEXT:    addu16 a0, a1
+; CHECK-NEXT:    subi16 a0, a0, 1
 ; CHECK-NEXT:    rts16
 ;
 ; GENERIC-LABEL: SUB_CHAR_I:
 ; GENERIC:       # %bb.0: # %entry
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 0
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 4
 ; GENERIC-NEXT:    movi16 a1, 255

 ; GENERIC:       # %bb.0: # %entry
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 4
 ; GENERIC-NEXT:    st32.w lr, (sp, 0) # 4-byte Folded Spill
 ; GENERIC-NEXT:    .cfi_offset lr, -4
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 8
 ; GENERIC-NEXT:    mov16 a2, a0
-; GENERIC-NEXT:    lrw32 a3, [.LCPI25_0]
+; GENERIC-NEXT:    lrw32 a3, [.LCPI29_0]
 ; GENERIC-NEXT:    mov16 a0, a1
 ; GENERIC-NEXT:    mov16 a1, a2
 ; GENERIC-NEXT:    jsr16 a3
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    ld32.w lr, (sp, 0) # 4-byte Folded Reload
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 ; GENERIC-NEXT:    .p2align 1
 ; GENERIC-NEXT:  # %bb.1:
 ; GENERIC-NEXT:    .p2align 2, 0x0
-; GENERIC-NEXT:  .LCPI25_0:
+; GENERIC-NEXT:  .LCPI29_0:
 ; GENERIC-NEXT:    .long __udivsi3
 entry:
   %udiv = udiv  i32 %y, %x
   ret i32 %udiv
 }
 
 define i32 @udivRI(i32 %x) {
 ; CHECK-LABEL: udivRI:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    movi16 a1, 10
 ; CHECK-NEXT:    divu32 a0, a0, a1
 ; CHECK-NEXT:    rts16
 ;
 ; GENERIC-LABEL: udivRI:
 ; GENERIC:       # %bb.0: # %entry
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 4
 ; GENERIC-NEXT:    st32.w lr, (sp, 0) # 4-byte Folded Spill
 ; GENERIC-NEXT:    .cfi_offset lr, -4
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 8
-; GENERIC-NEXT:    lrw32 a2, [.LCPI26_0]
+; GENERIC-NEXT:    lrw32 a2, [.LCPI30_0]
 ; GENERIC-NEXT:    movi16 a1, 10
 ; GENERIC-NEXT:    jsr16 a2
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    ld32.w lr, (sp, 0) # 4-byte Folded Reload
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 ; GENERIC-NEXT:    .p2align 1
 ; GENERIC-NEXT:  # %bb.1:
 ; GENERIC-NEXT:    .p2align 2, 0x0
-; GENERIC-NEXT:  .LCPI26_0:
+; GENERIC-NEXT:  .LCPI30_0:
 ; GENERIC-NEXT:    .long __udivsi3
 entry:
   %udiv = udiv  i32 %x, 10
   ret i32 %udiv
 }
 
 define i32 @udivRI_X(i32 %x) {
 ; CHECK-LABEL: udivRI_X:

 ; GENERIC-NEXT:    lsli16 a2, a1, 24
 ; GENERIC-NEXT:    lsli16 a1, a1, 16
 ; GENERIC-NEXT:    or16 a1, a2
 ; GENERIC-NEXT:    movi16 a2, 16
 ; GENERIC-NEXT:    lsli16 a2, a2, 8
 ; GENERIC-NEXT:    or16 a2, a1
 ; GENERIC-NEXT:    movi16 a1, 1
 ; GENERIC-NEXT:    or16 a1, a2
-; GENERIC-NEXT:    lrw32 a2, [.LCPI27_0]
+; GENERIC-NEXT:    lrw32 a2, [.LCPI31_0]
 ; GENERIC-NEXT:    jsr16 a2
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    ld32.w lr, (sp, 0) # 4-byte Folded Reload
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 ; GENERIC-NEXT:    .p2align 1
 ; GENERIC-NEXT:  # %bb.1:
 ; GENERIC-NEXT:    .p2align 2, 0x0
-; GENERIC-NEXT:  .LCPI27_0:
+; GENERIC-NEXT:  .LCPI31_0:
 ; GENERIC-NEXT:    .long __udivsi3
 entry:
   %udiv = udiv  i32 %x, 4097
   ret i32 %udiv
 }
 
 define i16 @UDIV_SHORT(i16 %x, i16 %y) {
 ; CHECK-LABEL: UDIV_SHORT:

 ; GENERIC-NEXT:    lsli16 l0, a1, 16
 ; GENERIC-NEXT:    or16 l0, a3
 ; GENERIC-NEXT:    movi16 a3, 255
 ; GENERIC-NEXT:    lsli16 a1, a3, 8
 ; GENERIC-NEXT:    or16 a1, l0
 ; GENERIC-NEXT:    or16 a1, a3
 ; GENERIC-NEXT:    and16 a2, a1
 ; GENERIC-NEXT:    and16 a1, a0
-; GENERIC-NEXT:    lrw32 a3, [.LCPI28_0]
+; GENERIC-NEXT:    lrw32 a3, [.LCPI32_0]
 ; GENERIC-NEXT:    mov16 a0, a2
 ; GENERIC-NEXT:    jsr16 a3
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    ld32.w lr, (sp, 0) # 4-byte Folded Reload
 ; GENERIC-NEXT:    ld16.w l0, (sp, 4) # 4-byte Folded Reload
 ; GENERIC-NEXT:    addi16 sp, sp, 8
 ; GENERIC-NEXT:    rts16
 ; GENERIC-NEXT:    .p2align 1
 ; GENERIC-NEXT:  # %bb.1:
 ; GENERIC-NEXT:    .p2align 2, 0x0
-; GENERIC-NEXT:  .LCPI28_0:
+; GENERIC-NEXT:  .LCPI32_0:
 ; GENERIC-NEXT:    .long __udivsi3
 entry:
   %udiv = udiv  i16 %y, %x
   ret i16 %udiv
 }
 
 define i16 @UDIV_SHORT_I(i16 %x) {
 ; CHECK-LABEL: UDIV_SHORT_I:

 ; GENERIC-NEXT:    st32.w lr, (sp, 0) # 4-byte Folded Spill
 ; GENERIC-NEXT:    .cfi_offset lr, -4
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 8
 ; GENERIC-NEXT:    mov16 a2, a1
 ; GENERIC-NEXT:    movi16 a1, 255
 ; GENERIC-NEXT:    and16 a2, a1
 ; GENERIC-NEXT:    and16 a1, a0
-; GENERIC-NEXT:    lrw32 a3, [.LCPI30_0]
+; GENERIC-NEXT:    lrw32 a3, [.LCPI34_0]
 ; GENERIC-NEXT:    mov16 a0, a2
 ; GENERIC-NEXT:    jsr16 a3
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    ld32.w lr, (sp, 0) # 4-byte Folded Reload
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 ; GENERIC-NEXT:    .p2align 1
 ; GENERIC-NEXT:  # %bb.1:
 ; GENERIC-NEXT:    .p2align 2, 0x0
-; GENERIC-NEXT:  .LCPI30_0:
+; GENERIC-NEXT:  .LCPI34_0:
 ; GENERIC-NEXT:    .long __udivsi3
 entry:
   %udiv = udiv  i8 %y, %x
   ret i8 %udiv
 }
 
 define i8 @UDIV_CHAR_I(i8 %x) {
 ; CHECK-LABEL: UDIV_CHAR_I:

 ; GENERIC:       # %bb.0: # %entry
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 4
 ; GENERIC-NEXT:    st32.w lr, (sp, 0) # 4-byte Folded Spill
 ; GENERIC-NEXT:    .cfi_offset lr, -4
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 8
 ; GENERIC-NEXT:    mov16 a2, a0
-; GENERIC-NEXT:    lrw32 a3, [.LCPI32_0]
+; GENERIC-NEXT:    lrw32 a3, [.LCPI36_0]
 ; GENERIC-NEXT:    mov16 a0, a1
 ; GENERIC-NEXT:    mov16 a1, a2
 ; GENERIC-NEXT:    jsr16 a3
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    ld32.w lr, (sp, 0) # 4-byte Folded Reload
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 ; GENERIC-NEXT:    .p2align 1
 ; GENERIC-NEXT:  # %bb.1:
 ; GENERIC-NEXT:    .p2align 2, 0x0
-; GENERIC-NEXT:  .LCPI32_0:
+; GENERIC-NEXT:  .LCPI36_0:
 ; GENERIC-NEXT:    .long __divsi3
 entry:
   %sdiv = sdiv  i32 %y, %x
   ret i32 %sdiv
 }
 
 define i32 @sdivRI(i32 %x) {
 ; CHECK-LABEL: sdivRI:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    movi16 a1, 10
 ; CHECK-NEXT:    divs32 a0, a0, a1
 ; CHECK-NEXT:    rts16
 ;
 ; GENERIC-LABEL: sdivRI:
 ; GENERIC:       # %bb.0: # %entry
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 4
 ; GENERIC-NEXT:    st32.w lr, (sp, 0) # 4-byte Folded Spill
 ; GENERIC-NEXT:    .cfi_offset lr, -4
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 8
-; GENERIC-NEXT:    lrw32 a2, [.LCPI33_0]
+; GENERIC-NEXT:    lrw32 a2, [.LCPI37_0]
 ; GENERIC-NEXT:    movi16 a1, 10
 ; GENERIC-NEXT:    jsr16 a2
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    ld32.w lr, (sp, 0) # 4-byte Folded Reload
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 ; GENERIC-NEXT:    .p2align 1
 ; GENERIC-NEXT:  # %bb.1:
 ; GENERIC-NEXT:    .p2align 2, 0x0
-; GENERIC-NEXT:  .LCPI33_0:
+; GENERIC-NEXT:  .LCPI37_0:
 ; GENERIC-NEXT:    .long __divsi3
 entry:
   %sdiv = sdiv  i32 %x, 10
   ret i32 %sdiv
 }
 
 define i32 @sdivRI_X(i32 %x) {
 ; CHECK-LABEL: sdivRI_X:

 ; GENERIC-NEXT:    lsli16 a2, a1, 24
 ; GENERIC-NEXT:    lsli16 a1, a1, 16
 ; GENERIC-NEXT:    or16 a1, a2
 ; GENERIC-NEXT:    movi16 a2, 16
 ; GENERIC-NEXT:    lsli16 a2, a2, 8
 ; GENERIC-NEXT:    or16 a2, a1
 ; GENERIC-NEXT:    movi16 a1, 1
 ; GENERIC-NEXT:    or16 a1, a2
-; GENERIC-NEXT:    lrw32 a2, [.LCPI34_0]
+; GENERIC-NEXT:    lrw32 a2, [.LCPI38_0]
 ; GENERIC-NEXT:    jsr16 a2
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    ld32.w lr, (sp, 0) # 4-byte Folded Reload
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 ; GENERIC-NEXT:    .p2align 1
 ; GENERIC-NEXT:  # %bb.1:
 ; GENERIC-NEXT:    .p2align 2, 0x0
-; GENERIC-NEXT:  .LCPI34_0:
+; GENERIC-NEXT:  .LCPI38_0:
 ; GENERIC-NEXT:    .long __divsi3
 entry:
   %sdiv = sdiv  i32 %x, 4097
   ret i32 %sdiv
 }
 
 define i16 @SDIV_SHORT(i16 %x, i16 %y) {
 ; CHECK-LABEL: SDIV_SHORT:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    sexth16 a0, a0
 ; CHECK-NEXT:    sexth16 a1, a1
 ; CHECK-NEXT:    divs32 a0, a1, a0
 ; CHECK-NEXT:    rts16
 ;
 ; GENERIC-LABEL: SDIV_SHORT:
 ; GENERIC:       # %bb.0: # %entry
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 4
 ; GENERIC-NEXT:    st32.w lr, (sp, 0) # 4-byte Folded Spill
 ; GENERIC-NEXT:    .cfi_offset lr, -4
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 8
 ; GENERIC-NEXT:    sexth16 a2, a1
 ; GENERIC-NEXT:    sexth16 a1, a0
-; GENERIC-NEXT:    lrw32 a3, [.LCPI35_0]
+; GENERIC-NEXT:    lrw32 a3, [.LCPI39_0]
 ; GENERIC-NEXT:    mov16 a0, a2
 ; GENERIC-NEXT:    jsr16 a3
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    ld32.w lr, (sp, 0) # 4-byte Folded Reload
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 ; GENERIC-NEXT:    .p2align 1
 ; GENERIC-NEXT:  # %bb.1:
 ; GENERIC-NEXT:    .p2align 2, 0x0
-; GENERIC-NEXT:  .LCPI35_0:
+; GENERIC-NEXT:  .LCPI39_0:
 ; GENERIC-NEXT:    .long __divsi3
 entry:
   %sdiv = sdiv  i16 %y, %x
   ret i16 %sdiv
 }
 
 define i16 @SDIV_SHORT_I(i16 %x) {
 ; CHECK-LABEL: SDIV_SHORT_I:

 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 4
 ; GENERIC-NEXT:    st32.w lr, (sp, 0) # 4-byte Folded Spill
 ; GENERIC-NEXT:    .cfi_offset lr, -4
 ; GENERIC-NEXT:    subi16 sp, sp, 4
 ; GENERIC-NEXT:    .cfi_def_cfa_offset 8
 ; GENERIC-NEXT:    sextb16 a2, a1
 ; GENERIC-NEXT:    sextb16 a1, a0
-; GENERIC-NEXT:    lrw32 a3, [.LCPI37_0]
+; GENERIC-NEXT:    lrw32 a3, [.LCPI41_0]
 ; GENERIC-NEXT:    mov16 a0, a2
 ; GENERIC-NEXT:    jsr16 a3
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    ld32.w lr, (sp, 0) # 4-byte Folded Reload
 ; GENERIC-NEXT:    addi16 sp, sp, 4
 ; GENERIC-NEXT:    rts16
 ; GENERIC-NEXT:    .p2align 1
 ; GENERIC-NEXT:  # %bb.1:
 ; GENERIC-NEXT:    .p2align 2, 0x0
-; GENERIC-NEXT:  .LCPI37_0:
+; GENERIC-NEXT:  .LCPI41_0:
 ; GENERIC-NEXT:    .long __divsi3
 entry:
   %sdiv = sdiv  i8 %y, %x
   ret i8 %sdiv
 }
 
 define i8 @SDIV_CHAR_I(i8 %x) {
 ; CHECK-LABEL: SDIV_CHAR_I: