Fix dynamic probing scheme

If we probe *after* each static stack allocation, we need to probe *before* each
dynamic stack allocation. Provide a scheme to describe the possible scenario.

Thanks a lot to @jonpa for motivating this fix.

Differential Revision: https://reviews.llvm.org/D81067

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index fe06bb8..7e84dba 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -31684,17 +31684,28 @@ X86TargetLowering::EmitLoweredProbedAlloca(MachineInstr &MI,
   testMBB->addSuccessor(blockMBB);
   testMBB->addSuccessor(tailMBB);
 
-  // allocate a block and touch it
+  // Touch the block then extend it. This is done on the opposite side of
+  // static probe where we allocate then touch, to avoid the need of probing the
+  // tail of the static alloca. Possible scenarios are:
+  //
+  //       + ---- <- ------------ <- ------------- <- ------------ +
+  //       |                                                       |
+  // [free probe] -> [page alloc] -> [alloc probe] -> [tail alloc] + -> [dyn probe] -> [page alloc] -> [dyn probe] -> [tail alloc] +
+  //                                                               |                                                               |
+  //                                                               + <- ----------- <- ------------ <- ----------- <- ------------ +
+  //
+  // The property we want to enforce is to never have more than [page alloc] between two probes.
+
+  const unsigned MovMIOpc =
+      TFI.Uses64BitFramePtr ? X86::MOV64mi32 : X86::MOV32mi;
+  addRegOffset(BuildMI(blockMBB, DL, TII->get(MovMIOpc)), physSPReg, false, 0)
+      .addImm(0);
 
   BuildMI(blockMBB, DL,
           TII->get(getSUBriOpcode(TFI.Uses64BitFramePtr, ProbeSize)), physSPReg)
       .addReg(physSPReg)
       .addImm(ProbeSize);
 
-  const unsigned MovMIOpc =
-      TFI.Uses64BitFramePtr ? X86::MOV64mi32 : X86::MOV32mi;
-  addRegOffset(BuildMI(blockMBB, DL, TII->get(MovMIOpc)), physSPReg, false, 0)
-      .addImm(0);
 
   BuildMI(blockMBB, DL, TII->get(X86::JMP_1)).addMBB(testMBB);
   blockMBB->addSuccessor(testMBB);

diff --git a/llvm/test/CodeGen/X86/stack-clash-dynamic-alloca.ll b/llvm/test/CodeGen/X86/stack-clash-dynamic-alloca.ll
index 140da42..928b7a9 100644 (file)
--- a/llvm/test/CodeGen/X86/stack-clash-dynamic-alloca.ll
+++ b/llvm/test/CodeGen/X86/stack-clash-dynamic-alloca.ll
@@ -26,8 +26,8 @@ attributes #0 =  {"probe-stack"="inline-asm"}
 ; CHECK-X86-64-NEXT:   cmpq    %rax, %rsp
 ; CHECK-X86-64-NEXT:   jl      .LBB0_3
 ; CHECK-X86-64-NEXT:  .LBB0_2: # =>This Inner Loop Header: Depth=1
-; CHECK-X86-64-NEXT:   subq    $4096, %rsp # imm = 0x1000
 ; CHECK-X86-64-NEXT:   movq    $0, (%rsp)
+; CHECK-X86-64-NEXT:   subq    $4096, %rsp # imm = 0x1000
 ; CHECK-X86-64-NEXT:   cmpq    %rax, %rsp
 ; CHECK-X86-64-NEXT:   jge     .LBB0_2
 ; CHECK-X86-64-NEXT:  .LBB0_3:
@@ -56,8 +56,8 @@ attributes #0 =  {"probe-stack"="inline-asm"}
 ; CHECK-X86-32-NEXT:    cmpl    %eax, %esp
 ; CHECK-X86-32-NEXT:    jl  .LBB0_3
 ; CHECK-X86-32-NEXT:  .LBB0_2: # =>This Inner Loop Header: Depth=1
-; CHECK-X86-32-NEXT:    subl    $4096, %esp # imm = 0x1000
 ; CHECK-X86-32-NEXT:    movl    $0, (%esp)
+; CHECK-X86-32-NEXT:    subl    $4096, %esp # imm = 0x1000
 ; CHECK-X86-32-NEXT:    cmpl    %eax, %esp
 ; CHECK-X86-32-NEXT:    jge .LBB0_2
 ; CHECK-X86-32-NEXT:  .LBB0_3: