From: Saleem Abdulrasool <compnerd@compnerd.org>
Date: Mon, 31 Oct 2016 22:12:37 +0000 (+0000)
Subject: CodeGen: further loosen -O0 CG for WoA division
X-Git-Tag: llvmorg-4.0.0-rc1~5796
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=e1aa782bd073925dae4c3069e396bdd77e2199a1;p=platform%2Fupstream%2Fllvm.git

CodeGen: further loosen -O0 CG for WoA division

Generate the slowest possible codepath for noopt CodeGen.  Even trying to be
clever with the negated jump can cause out-of-range jumps.  Use a wide branch
instead. Although the code is modelled simplistically, the later optimizations
would recombine the branching into `cbz` if possible.  This re-enables the
previous optimization as well as hopefully gives us working code in all cases.

Addresses PR30356!

llvm-svn: 285649
---

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 12f3247..a87cf50 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -8593,12 +8593,20 @@ ARMTargetLowering::EmitLowered__dbzchk(MachineInstr &MI,
   ContBB->splice(ContBB->begin(), MBB,
                  std::next(MachineBasicBlock::iterator(MI)), MBB->end());
   ContBB->transferSuccessorsAndUpdatePHIs(MBB);
-
-  BuildMI(*MBB, MI, DL, TII->get(ARM::tCBNZ))
-      .addReg(MI.getOperand(0).getReg())
-      .addMBB(ContBB);
   MBB->addSuccessor(ContBB);
-  BuildMI(*MBB, MI, DL, TII->get(ARM::t__brkdiv0));
+
+  MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
+  BuildMI(TrapBB, DL, TII->get(ARM::t__brkdiv0));
+  MF->push_back(TrapBB);
+  MBB->addSuccessor(TrapBB);
+
+  AddDefaultPred(BuildMI(*MBB, MI, DL, TII->get(ARM::tCMPi8))
+                     .addReg(MI.getOperand(0).getReg())
+                     .addImm(0));
+  BuildMI(*MBB, MI, DL, TII->get(ARM::t2Bcc))
+      .addMBB(TrapBB)
+      .addImm(ARMCC::EQ)
+      .addReg(ARM::CPSR);
 
   MI.eraseFromParent();
   return ContBB;
diff --git a/llvm/test/CodeGen/ARM/Windows/dbzchk.ll b/llvm/test/CodeGen/ARM/Windows/dbzchk.ll
index af23cb2..aea3799 100644
--- a/llvm/test/CodeGen/ARM/Windows/dbzchk.ll
+++ b/llvm/test/CodeGen/ARM/Windows/dbzchk.ll
@@ -75,7 +75,7 @@ return:
 ; CHECK-MOD-DAG: BB#2
 
 ; RUN: llc -mtriple thumbv7--windows-itanium -print-machineinstrs=expand-isel-pseudos -verify-machineinstrs -filetype asm -o /dev/null %s 2>&1 | FileCheck %s -check-prefix CHECK-CFG
-; RUN: llc -mtriple thumbv7--windows-itanium -print-machineinstrs=expand-isel-pseudos -verify-machineinstrs -filetype asm -o - %s | FileCheck %s -check-prefix CHECK-CFG-ASM
+; RUN: llc -mtriple thumbv7--windows-itanium -verify-machineinstrs -filetype asm -o - %s | FileCheck %s -check-prefix CHECK-CFG-ASM
 
 ; unsigned c;
 ; extern unsigned long g(void);
@@ -119,21 +119,24 @@ attributes #0 = { optsize }
 ; CHECK-CFG-DAG: t2B <BB#3>
 
 ; CHECK-CFG-DAG: BB#2
-; CHECK-CFG-DAG: tCBNZ %vreg{{[0-9]}}, <BB#4>
-; CHECK-CFG-DAG: t__brkdiv0
+; CHECK-CFG-DAG: tCMPi8 %vreg{{[0-9]}}, 0
+; CHECK-CFG-DAG: t2Bcc <BB#5>
 
 ; CHECK-CFG-DAG: BB#4
 
 ; CHECK-CFG-DAG: BB#3
 ; CHECK-CFG-DAG: tBX_RET
 
+; CHECK-CFG-DAG: BB#5
+; CHECK-CFG-DAG: t__brkdiv0
+
 ; CHECK-CFG-ASM-LABEL: h:
-; CHECK-CFG-ASM: cbnz r{{[0-9]}}, .LBB2_2
-; CHECK-CFG-ASM: __brkdiv0
-; CHECK-CFG-ASM-LABEL: .LBB2_2:
+; CHECK-CFG-ASM: cbz r{{[0-9]}}, .LBB2_4
 ; CHECK-CFG-ASM: bl __rt_udiv
+; CHECK-CFG-ASM-LABEL: .LBB2_4:
+; CHECK-CFG-ASM: __brkdiv0
 
-; RUN: llc -O0 -mtriple thumbv7--windows-itanium -verify-machineinstrs -filetype asm -o - %s | FileCheck %s -check-prefix CHECK-WIN__DBZCHK
+; RUN: llc -O1 -mtriple thumbv7--windows-itanium -verify-machineinstrs -filetype asm -o - %s | FileCheck %s -check-prefix CHECK-WIN__DBZCHK
 
 ; long k(void);
 ; int l(void);
@@ -172,11 +175,11 @@ return:
 }
 
 ; CHECK-WIN__DBZCHK-LABEL: j:
-; CHECK-WIN__DBZCHK: cbnz r{{[0-7]}}, .LBB
-; CHECK-WIN__DBZCHK-NOT: cbnz r8, .LBB
-; CHECK-WIN__DBZCHK-NOT: cbnz r9, .LBB
-; CHECK-WIN__DBZCHK-NOT: cbnz r10, .LBB
-; CHECK-WIN__DBZCHK-NOT: cbnz r11, .LBB
-; CHECK-WIN__DBZCHK-NOT: cbnz ip, .LBB
-; CHECK-WIN__DBZCHK-NOT: cbnz lr, .LBB
+; CHECK-WIN__DBZCHK: cbz r{{[0-7]}}, .LBB
+; CHECK-WIN__DBZCHK-NOT: cbz r8, .LBB
+; CHECK-WIN__DBZCHK-NOT: cbz r9, .LBB
+; CHECK-WIN__DBZCHK-NOT: cbz r10, .LBB
+; CHECK-WIN__DBZCHK-NOT: cbz r11, .LBB
+; CHECK-WIN__DBZCHK-NOT: cbz ip, .LBB
+; CHECK-WIN__DBZCHK-NOT: cbz lr, .LBB
 
diff --git a/llvm/test/CodeGen/ARM/Windows/division-range.ll b/llvm/test/CodeGen/ARM/Windows/division-range.ll
index 9b1920c..b14339e 100644
--- a/llvm/test/CodeGen/ARM/Windows/division-range.ll
+++ b/llvm/test/CodeGen/ARM/Windows/division-range.ll
@@ -9,7 +9,7 @@ entry:
   ret i32 %div
 }
 
-; CHECK: cbnz r1, #0
-; CHECK: __brkdiv0
+; CHECK: cmp r1, #0
+; CHECK: beq #
 ; CHECK: bl
 
diff --git a/llvm/test/CodeGen/ARM/Windows/division.ll b/llvm/test/CodeGen/ARM/Windows/division.ll
index b149cb0..10f86b0 100644
--- a/llvm/test/CodeGen/ARM/Windows/division.ll
+++ b/llvm/test/CodeGen/ARM/Windows/division.ll
@@ -8,9 +8,9 @@ entry:
 }
 
 ; CHECK-LABEL: sdiv32:
-; CHECK: cbnz r0
-; CHECK: __brkdiv0
+; CHECK: cbz r0
 ; CHECK: bl __rt_sdiv
+; CHECK: __brkdiv0
 
 define arm_aapcs_vfpcc i32 @udiv32(i32 %divisor, i32 %divident) {
 entry:
@@ -19,9 +19,9 @@ entry:
 }
 
 ; CHECK-LABEL: udiv32:
-; CHECK: cbnz r0
-; CHECK: __brkdiv0
+; CHECK: cbz r0
 ; CHECK: bl __rt_udiv
+; CHECK: __brkdiv0
 
 define arm_aapcs_vfpcc i64 @sdiv64(i64 %divisor, i64 %divident) {
 entry:
@@ -30,10 +30,10 @@ entry:
 }
 
 ; CHECK-LABEL: sdiv64:
-; CHECK: orr.w r4, r0, r1
-; CHECK-NEXT: cbnz r4
-; CHECK: __brkdiv0
+; CHECK: orrs.w r4, r0, r1
+; CHECK-NEXT: beq
 ; CHECK: bl __rt_sdiv64
+; CHECK: __brkdiv0
 
 define arm_aapcs_vfpcc i64 @udiv64(i64 %divisor, i64 %divident) {
 entry:
@@ -42,8 +42,8 @@ entry:
 }
 
 ; CHECK-LABEL: udiv64:
-; CHECK: orr.w r4, r0, r1
-; CHECK-NEXT: cbnz r4
-; CHECK: __brkdiv0
+; CHECK: orrs.w r4, r0, r1
+; CHECK-NEXT: beq
 ; CHECK: bl __rt_udiv64
+; CHECK: __brkdiv0