Check for stack space more intelligently.

libgcc sets the stack limit field in TCB to 256 bytes above the actual
allocated stack limit.  This means if the function's stack frame needs
less than 256 bytes, we can just compare the stack pointer with the
stack limit.  This should result in lesser calls to __morestack.

llvm-svn: 145766

2 files changed, 43 insertions, 13 deletions

diff --git a/llvm/lib/Target/X86/X86FrameLowering.cpp b/llvm/lib/Target/X86/X86FrameLowering.cpp
index 182af84..27b762a 100644
--- a/llvm/lib/Target/X86/X86FrameLowering.cpp
+++ b/llvm/lib/Target/X86/X86FrameLowering.cpp
@@ -1306,6 +1306,10 @@ GetScratchRegister(bool Is64Bit, const MachineFunction &MF) {
   }
 }
 
+// The stack limit in the TCB is set to this many bytes above the actual stack
+// limit.
+static const uint64_t kSplitStackAvailable = 256;
+
 void
 X86FrameLowering::adjustForSegmentedStacks(MachineFunction &MF) const {
   MachineBasicBlock &prologueMBB = MF.front();
@@ -1360,16 +1364,24 @@ X86FrameLowering::adjustForSegmentedStacks(MachineFunction &MF) const {
     TlsReg = X86::FS;
     TlsOffset = 0x70;
 
-    BuildMI(checkMBB, DL, TII.get(X86::LEA64r), ScratchReg).addReg(X86::RSP)
-      .addImm(0).addReg(0).addImm(-StackSize).addReg(0);
+    if (StackSize < kSplitStackAvailable)
+      ScratchReg = X86::RSP;
+    else
+      BuildMI(checkMBB, DL, TII.get(X86::LEA64r), ScratchReg).addReg(X86::RSP)
+        .addImm(0).addReg(0).addImm(-StackSize).addReg(0);
+
     BuildMI(checkMBB, DL, TII.get(X86::CMP64rm)).addReg(ScratchReg)
       .addReg(0).addImm(0).addReg(0).addImm(TlsOffset).addReg(TlsReg);
   } else {
     TlsReg = X86::GS;
     TlsOffset = 0x30;
 
-    BuildMI(checkMBB, DL, TII.get(X86::LEA32r), ScratchReg).addReg(X86::ESP)
-      .addImm(0).addReg(0).addImm(-StackSize).addReg(0);
+    if (StackSize < kSplitStackAvailable)
+      ScratchReg = X86::ESP;
+    else
+      BuildMI(checkMBB, DL, TII.get(X86::LEA32r), ScratchReg).addReg(X86::ESP)
+        .addImm(0).addReg(0).addImm(-StackSize).addReg(0);
+
     BuildMI(checkMBB, DL, TII.get(X86::CMP32rm)).addReg(ScratchReg)
       .addReg(0).addImm(0).addReg(0).addImm(TlsOffset).addReg(TlsReg);
   }
diff --git a/llvm/test/CodeGen/X86/segmented-stacks.ll b/llvm/test/CodeGen/X86/segmented-stacks.ll
index c28c319..4f529c1 100644
--- a/llvm/test/CodeGen/X86/segmented-stacks.ll
+++ b/llvm/test/CodeGen/X86/segmented-stacks.ll
@@ -20,8 +20,7 @@ false:
 
 ; X32:      test_basic:
 
-; X32:      leal -12(%esp), %ecx
-; X32-NEXT: cmpl %gs:48, %ecx
+; X32:      cmpl %gs:48, %esp
 
 ; X32:      pushl $4
 ; X32-NEXT: pushl $12
@@ -41,8 +40,7 @@ false:
 
 ; X64:      test_basic:
 
-; X64:      leaq -24(%rsp), %r11
-; X64-NEXT: cmpq %fs:112, %r11
+; X64:      cmpq %fs:112, %rsp
 
 ; X64:      movabsq $24, %r10
 ; X64-NEXT: movabsq $0, %r11
@@ -66,17 +64,14 @@ define i32 @test_nested(i32 * nest %closure, i32 %other) {
        %result = add i32 %other, %addend
        ret i32 %result
 
-; X32:      leal (%esp), %edx
-; X32-NEXT: cmpl %gs:48, %edx
-
+; X32:      cmpl %gs:48, %esp
 
 ; X32:      pushl $4
 ; X32-NEXT: pushl $0
 ; X32-NEXT: calll __morestack
 ; X32-NEXT: ret
 
-; X64:      leaq (%rsp), %r11
-; X64-NEXT: cmpq %fs:112, %r11
+; X64:      cmpq %fs:112, %rsp
 
 ; X64:      movq %r10, %rax
 ; X64-NEXT: movabsq $0, %r10
@@ -86,3 +81,26 @@ define i32 @test_nested(i32 * nest %closure, i32 %other) {
 ; X64-NEXT: movq %rax, %r10
 
 }
+
+define void @test_large() {
+        %mem = alloca i32, i32 10000
+        call void @dummy_use (i32* %mem, i32 0)
+        ret void
+
+; X32:      leal -40012(%esp), %ecx
+; X32-NEXT: cmpl %gs:48, %ecx
+
+; X32:      pushl $0
+; X32-NEXT: pushl $40012
+; X32-NEXT: calll __morestack
+; X32-NEXT: ret
+
+; X64:      leaq -40008(%rsp), %r11
+; X64-NEXT: cmpq %fs:112, %r11
+
+; X64:      movabsq $40008, %r10
+; X64-NEXT: movabsq $0, %r11
+; X64-NEXT: callq __morestack
+; X64-NEXT: ret
+
+}