gentoo-overlay/sys-devel/llvm/files/llvm-3.4-fix_varargs.patch

148 lines
6.3 KiB
Diff

https://bugs.gentoo.org/show_bug.cgi?id=497298
http://llvm.org/bugs/show_bug.cgi?id=18346
SVN revisions:
http://llvm.org/viewvc/llvm-project?view=revision&revision=197503
http://llvm.org/viewvc/llvm-project?view=revision&revision=197505
http://llvm.org/viewvc/llvm-project?view=revision&revision=197520
diff -Naur llvm-3.4.orig/lib/Target/X86/X86ISelLowering.cpp llvm-3.4/lib/Target/X86/X86ISelLowering.cpp
--- llvm-3.4.orig/lib/Target/X86/X86ISelLowering.cpp 2014-01-08 11:22:23.373060109 +0100
+++ llvm-3.4/lib/Target/X86/X86ISelLowering.cpp 2014-01-08 11:23:58.517125907 +0100
@@ -15226,9 +15226,15 @@
MBB->addSuccessor(EndMBB);
}
+ // Make sure the last operand is EFLAGS, which gets clobbered by the branch
+ // that was just emitted, but clearly shouldn't be "saved".
+ assert((MI->getNumOperands() <= 3 ||
+ !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
+ MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
+ && "Expected last argument to be EFLAGS");
unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
// In the XMM save block, save all the XMM argument registers.
- for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
+ for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
MachineMemOperand *MMO =
F->getMachineMemOperand(
diff -Naur llvm-3.4.orig/lib/Target/X86/X86InstrCompiler.td llvm-3.4/lib/Target/X86/X86InstrCompiler.td
--- llvm-3.4.orig/lib/Target/X86/X86InstrCompiler.td 2014-01-08 11:22:23.374060109 +0100
+++ llvm-3.4/lib/Target/X86/X86InstrCompiler.td 2014-01-08 11:23:39.654112857 +0100
@@ -72,7 +72,7 @@
// x86-64 va_start lowering magic.
-let usesCustomInserter = 1 in {
+let usesCustomInserter = 1, Defs = [EFLAGS] in {
def VASTART_SAVE_XMM_REGS : I<0, Pseudo,
(outs),
(ins GR8:$al,
@@ -81,7 +81,8 @@
"#VASTART_SAVE_XMM_REGS $al, $regsavefi, $offset",
[(X86vastart_save_xmm_regs GR8:$al,
imm:$regsavefi,
- imm:$offset)]>;
+ imm:$offset),
+ (implicit EFLAGS)]>;
// The VAARG_64 pseudo-instruction takes the address of the va_list,
// and places the address of the next argument into a register.
diff -Naur llvm-3.4.orig/test/CodeGen/X86/vaargs.ll llvm-3.4/test/CodeGen/X86/vaargs.ll
--- llvm-3.4.orig/test/CodeGen/X86/vaargs.ll 1970-01-01 01:00:00.000000000 +0100
+++ llvm-3.4/test/CodeGen/X86/vaargs.ll 2014-01-08 11:23:53.451122402 +0100
@@ -0,0 +1,67 @@
+; RUN: llc -mcpu=corei7-avx %s -o - | FileCheck %s --check-prefix=CHECK --check-prefix=NO-FLAGS
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.9.0"
+
+%struct.__va_list_tag = type { i32, i32, i8*, i8* }
+
+; Check that vastart gets the right thing.
+define i32 @sum(i32 %count, ...) nounwind optsize ssp uwtable {
+; CHECK: testb %al, %al
+; CHECK-NEXT: je
+; CHECK-NEXT: ## BB#{{[0-9]+}}:
+; CHECK-NEXT: vmovaps %xmm0, 48(%rsp)
+; CHECK-NEXT: vmovaps %xmm1, 64(%rsp)
+; CHECK-NEXT: vmovaps %xmm2, 80(%rsp)
+; CHECK-NEXT: vmovaps %xmm3, 96(%rsp)
+; CHECK-NEXT: vmovaps %xmm4, 112(%rsp)
+; CHECK-NEXT: vmovaps %xmm5, 128(%rsp)
+; CHECK-NEXT: vmovaps %xmm6, 144(%rsp)
+; CHECK-NEXT: vmovaps %xmm7, 160(%rsp)
+
+; Check that [EFLAGS] hasn't been pulled in.
+; NO-FLAGS-NOT: %flags
+
+ %ap = alloca [1 x %struct.__va_list_tag], align 16
+ %1 = bitcast [1 x %struct.__va_list_tag]* %ap to i8*
+ call void @llvm.va_start(i8* %1)
+ %2 = icmp sgt i32 %count, 0
+ br i1 %2, label %.lr.ph, label %._crit_edge
+
+.lr.ph: ; preds = %0
+ %3 = getelementptr inbounds [1 x %struct.__va_list_tag]* %ap, i64 0, i64 0, i32 0
+ %4 = getelementptr inbounds [1 x %struct.__va_list_tag]* %ap, i64 0, i64 0, i32 2
+ %.pre = load i32* %3, align 16
+ br label %5
+
+; <label>:5 ; preds = %.lr.ph, %13
+ %6 = phi i32 [ %.pre, %.lr.ph ], [ %14, %13 ]
+ %.01 = phi i32 [ %count, %.lr.ph ], [ %15, %13 ]
+ %7 = icmp ult i32 %6, 41
+ br i1 %7, label %8, label %10
+
+; <label>:8 ; preds = %5
+ %9 = add i32 %6, 8
+ store i32 %9, i32* %3, align 16
+ br label %13
+
+; <label>:10 ; preds = %5
+ %11 = load i8** %4, align 8
+ %12 = getelementptr i8* %11, i64 8
+ store i8* %12, i8** %4, align 8
+ br label %13
+
+; <label>:13 ; preds = %10, %8
+ %14 = phi i32 [ %6, %10 ], [ %9, %8 ]
+ %15 = add nsw i32 %.01, 1
+ %16 = icmp sgt i32 %15, 0
+ br i1 %16, label %5, label %._crit_edge
+
+._crit_edge: ; preds = %13, %0
+ %.0.lcssa = phi i32 [ %count, %0 ], [ %15, %13 ]
+ call void @llvm.va_end(i8* %1)
+ ret i32 %.0.lcssa
+}
+
+declare void @llvm.va_start(i8*) nounwind
+
+declare void @llvm.va_end(i8*) nounwind
diff -Naur llvm-3.4.orig/test/CodeGen/X86/vastart-defs-eflags.ll llvm-3.4/test/CodeGen/X86/vastart-defs-eflags.ll
--- llvm-3.4.orig/test/CodeGen/X86/vastart-defs-eflags.ll 1970-01-01 01:00:00.000000000 +0100
+++ llvm-3.4/test/CodeGen/X86/vastart-defs-eflags.ll 2014-01-08 11:23:39.654112857 +0100
@@ -0,0 +1,23 @@
+; RUN: llc %s -o - | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.10.0"
+
+; Check that vastart handling doesn't get between testb and je for the branch.
+define i32 @check_flag(i32 %flags, ...) nounwind {
+entry:
+; CHECK: {{^}} testb $2, %bh
+; CHECK-NOT: test
+; CHECK: {{^}} je
+ %and = and i32 %flags, 512
+ %tobool = icmp eq i32 %and, 0
+ br i1 %tobool, label %if.end, label %if.then
+
+if.then: ; preds = %entry
+ br label %if.end
+
+if.end: ; preds = %entry, %if.then
+ %hasflag = phi i32 [ 1, %if.then ], [ 0, %entry ]
+ ret i32 %hasflag
+}
+