[clang] a9fba2b - [X86][ABI] Do not return float/double from x87 registers when x87 is disabled
Phoebe Wang via cfe-commits
cfe-commits at lists.llvm.org
Wed Nov 17 20:31:17 PST 2021
Author: Phoebe Wang
Date: 2021-11-18T12:31:08+08:00
New Revision: a9fba2be35db674971382e38b99a31403444d9bf
URL: https://github.com/llvm/llvm-project/commit/a9fba2be35db674971382e38b99a31403444d9bf
DIFF: https://github.com/llvm/llvm-project/commit/a9fba2be35db674971382e38b99a31403444d9bf.diff
LOG: [X86][ABI] Do not return float/double from x87 registers when x87 is disabled
This is aligned with GCC's behavior.
Also, alias `-mno-fp-ret-in-387` to `-mno-x87`, by which we can fix pr51498.
Reviewed By: nickdesaulniers
Differential Revision: https://reviews.llvm.org/D112143
Added:
llvm/test/CodeGen/X86/no-ret-in-x87-reg.ll
Modified:
clang/include/clang/Driver/Options.td
clang/test/Driver/x86-target-features.c
llvm/lib/Target/X86/X86CallingConv.td
llvm/lib/Target/X86/X86ISelLowering.cpp
Removed:
################################################################################
diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td
index 7730c0b982bb..9bde64cf49fd 100644
--- a/clang/include/clang/Driver/Options.td
+++ b/clang/include/clang/Driver/Options.td
@@ -4184,6 +4184,7 @@ def mx87 : Flag<["-"], "mx87">, Group<m_x86_Features_Group>;
def mno_x87 : Flag<["-"], "mno-x87">, Group<m_x86_Features_Group>;
def m80387 : Flag<["-"], "m80387">, Alias<mx87>;
def mno_80387 : Flag<["-"], "mno-80387">, Alias<mno_x87>;
+def mno_fp_ret_in_387 : Flag<["-"], "mno-fp-ret-in-387">, Alias<mno_x87>;
def mmmx : Flag<["-"], "mmmx">, Group<m_x86_Features_Group>;
def mno_mmx : Flag<["-"], "mno-mmx">, Group<m_x86_Features_Group>;
def m3dnow : Flag<["-"], "m3dnow">, Group<m_x86_Features_Group>;
diff --git a/clang/test/Driver/x86-target-features.c b/clang/test/Driver/x86-target-features.c
index ce9afcf51dc4..42385dd5f0a4 100644
--- a/clang/test/Driver/x86-target-features.c
+++ b/clang/test/Driver/x86-target-features.c
@@ -2,6 +2,7 @@
// RUN: %clang -target i386-unknown-linux-gnu -march=i386 -mno-x87 %s -### -o %t.o 2>&1 | FileCheck -check-prefix=NO-X87 %s
// RUN: %clang -target i386-unknown-linux-gnu -march=i386 -m80387 %s -### -o %t.o 2>&1 | FileCheck -check-prefix=X87 %s
// RUN: %clang -target i386-unknown-linux-gnu -march=i386 -mno-80387 %s -### -o %t.o 2>&1 | FileCheck -check-prefix=NO-X87 %s
+// RUN: %clang -target i386-unknown-linux-gnu -march=i386 -mno-fp-ret-in-387 %s -### -o %t.o 2>&1 | FileCheck -check-prefix=NO-X87 %s
// X87: "-target-feature" "+x87"
// NO-X87: "-target-feature" "-x87"
diff --git a/llvm/lib/Target/X86/X86CallingConv.td b/llvm/lib/Target/X86/X86CallingConv.td
index 9bf621355d0a..4dd8a6cdd898 100644
--- a/llvm/lib/Target/X86/X86CallingConv.td
+++ b/llvm/lib/Target/X86/X86CallingConv.td
@@ -273,7 +273,10 @@ def RetCC_X86_32_C : CallingConv<[
// conv.
CCIfInReg<CCIfSubtarget<"hasSSE2()",
CCIfType<[f32, f64], CCAssignToReg<[XMM0,XMM1,XMM2]>>>>,
- CCIfType<[f32,f64], CCAssignToReg<[FP0, FP1]>>,
+ CCIfSubtarget<"hasX87()",
+ CCIfType<[f32, f64], CCAssignToReg<[FP0, FP1]>>>,
+ CCIfNotSubtarget<"hasX87()",
+ CCIfType<[f32], CCAssignToReg<[EAX, EDX, ECX]>>>,
CCIfType<[f16], CCAssignToReg<[XMM0,XMM1,XMM2]>>,
CCDelegateTo<RetCC_X86Common>
]>;
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 09e51977e095..bb5bc77b9e37 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -2371,6 +2371,11 @@ MVT X86TargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
if (VT == MVT::v3f16 && Subtarget.hasFP16())
return MVT::v8f16;
+ // We will use more GPRs for f64 and f80 on 32 bits when x87 is disabled.
+ if ((VT == MVT::f64 || VT == MVT::f80) && !Subtarget.is64Bit() &&
+ !Subtarget.hasX87())
+ return MVT::i32;
+
return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);
}
@@ -2394,6 +2399,15 @@ unsigned X86TargetLowering::getNumRegistersForCallingConv(LLVMContext &Context,
if (VT == MVT::v3f16 && Subtarget.hasFP16())
return 1;
+ // We have to split f64 to 2 registers and f80 to 3 registers on 32 bits if
+ // x87 is disabled.
+ if (!Subtarget.is64Bit() && !Subtarget.hasX87()) {
+ if (VT == MVT::f64)
+ return 2;
+ if (VT == MVT::f80)
+ return 3;
+ }
+
return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT);
}
diff --git a/llvm/test/CodeGen/X86/no-ret-in-x87-reg.ll b/llvm/test/CodeGen/X86/no-ret-in-x87-reg.ll
new file mode 100644
index 000000000000..e1ccede445d8
--- /dev/null
+++ b/llvm/test/CodeGen/X86/no-ret-in-x87-reg.ll
@@ -0,0 +1,193 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-- | FileCheck %s -check-prefix=X87
+; RUN: llc < %s -mtriple=i686-- -mattr=-x87 | FileCheck %s -check-prefixes=NOX87,NOSSE-NOX87
+; RUN: llc < %s -mtriple=i686-- -mattr=-x87,-sse2 | FileCheck %s -check-prefixes=NOX87,NOSSE-NOX87
+; RUN: llc < %s -mtriple=i686-- -mattr=-x87,+sse2 | FileCheck %s -check-prefixes=NOX87,SSE-NOX87
+
+define float @f1(float %a, float %b) nounwind {
+; X87-LABEL: f1:
+; X87: # %bb.0: # %entry
+; X87-NEXT: flds {{[0-9]+}}(%esp)
+; X87-NEXT: retl
+;
+; NOSSE-NOX87-LABEL: f1:
+; NOSSE-NOX87: # %bb.0: # %entry
+; NOSSE-NOX87-NEXT: movl {{[0-9]+}}(%esp), %eax
+; NOSSE-NOX87-NEXT: retl
+;
+; SSE-NOX87-LABEL: f1:
+; SSE-NOX87: # %bb.0: # %entry
+; SSE-NOX87-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NOX87-NEXT: movd %xmm0, %eax
+; SSE-NOX87-NEXT: retl
+entry:
+ ret float %b
+}
+
+define double @f2(double %a, double %b) nounwind {
+; X87-LABEL: f2:
+; X87: # %bb.0: # %entry
+; X87-NEXT: fldl {{[0-9]+}}(%esp)
+; X87-NEXT: retl
+;
+; NOX87-LABEL: f2:
+; NOX87: # %bb.0: # %entry
+; NOX87-NEXT: movl {{[0-9]+}}(%esp), %eax
+; NOX87-NEXT: movl {{[0-9]+}}(%esp), %edx
+; NOX87-NEXT: retl
+entry:
+ ret double %b
+}
+
+define x86_fp80 @f3(x86_fp80 %a, x86_fp80 %b) nounwind {
+; X87-LABEL: f3:
+; X87: # %bb.0: # %entry
+; X87-NEXT: fldt {{[0-9]+}}(%esp)
+; X87-NEXT: retl
+;
+; NOX87-LABEL: f3:
+; NOX87: # %bb.0: # %entry
+; NOX87-NEXT: movl {{[0-9]+}}(%esp), %eax
+; NOX87-NEXT: movl {{[0-9]+}}(%esp), %edx
+; NOX87-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; NOX87-NEXT: retl
+entry:
+ ret x86_fp80 %b
+}
+
+define float @f4(float %a, float %b) nounwind {
+; X87-LABEL: f4:
+; X87: # %bb.0: # %entry
+; X87-NEXT: flds {{[0-9]+}}(%esp)
+; X87-NEXT: fadds {{[0-9]+}}(%esp)
+; X87-NEXT: retl
+;
+; NOSSE-NOX87-LABEL: f4:
+; NOSSE-NOX87: # %bb.0: # %entry
+; NOSSE-NOX87-NEXT: pushl {{[0-9]+}}(%esp)
+; NOSSE-NOX87-NEXT: pushl {{[0-9]+}}(%esp)
+; NOSSE-NOX87-NEXT: calll __addsf3
+; NOSSE-NOX87-NEXT: addl $8, %esp
+; NOSSE-NOX87-NEXT: retl
+;
+; SSE-NOX87-LABEL: f4:
+; SSE-NOX87: # %bb.0: # %entry
+; SSE-NOX87-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NOX87-NEXT: addss {{[0-9]+}}(%esp), %xmm0
+; SSE-NOX87-NEXT: movd %xmm0, %eax
+; SSE-NOX87-NEXT: retl
+entry:
+ %0 = fadd float %a, %b
+ ret float %0
+}
+
+define double @f5(double %a, double %b) nounwind {
+; X87-LABEL: f5:
+; X87: # %bb.0: # %entry
+; X87-NEXT: fldl {{[0-9]+}}(%esp)
+; X87-NEXT: faddl {{[0-9]+}}(%esp)
+; X87-NEXT: retl
+;
+; NOSSE-NOX87-LABEL: f5:
+; NOSSE-NOX87: # %bb.0: # %entry
+; NOSSE-NOX87-NEXT: pushl {{[0-9]+}}(%esp)
+; NOSSE-NOX87-NEXT: pushl {{[0-9]+}}(%esp)
+; NOSSE-NOX87-NEXT: pushl {{[0-9]+}}(%esp)
+; NOSSE-NOX87-NEXT: pushl {{[0-9]+}}(%esp)
+; NOSSE-NOX87-NEXT: calll __adddf3
+; NOSSE-NOX87-NEXT: addl $16, %esp
+; NOSSE-NOX87-NEXT: retl
+;
+; SSE-NOX87-LABEL: f5:
+; SSE-NOX87: # %bb.0: # %entry
+; SSE-NOX87-NEXT: pushl %ebp
+; SSE-NOX87-NEXT: movl %esp, %ebp
+; SSE-NOX87-NEXT: andl $-8, %esp
+; SSE-NOX87-NEXT: subl $8, %esp
+; SSE-NOX87-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
+; SSE-NOX87-NEXT: addsd 16(%ebp), %xmm0
+; SSE-NOX87-NEXT: movsd %xmm0, (%esp)
+; SSE-NOX87-NEXT: movl (%esp), %eax
+; SSE-NOX87-NEXT: movl {{[0-9]+}}(%esp), %edx
+; SSE-NOX87-NEXT: movl %ebp, %esp
+; SSE-NOX87-NEXT: popl %ebp
+; SSE-NOX87-NEXT: retl
+entry:
+ %0 = fadd double %a, %b
+ ret double %0
+}
+
+; FIXME: We should not generate x87 instructions when x87 is disabled.
+define x86_fp80 @f6(x86_fp80 %a, x86_fp80 %b) nounwind {
+; X87-LABEL: f6:
+; X87: # %bb.0: # %entry
+; X87-NEXT: fldt {{[0-9]+}}(%esp)
+; X87-NEXT: fldt {{[0-9]+}}(%esp)
+; X87-NEXT: faddp %st, %st(1)
+; X87-NEXT: retl
+;
+; NOX87-LABEL: f6:
+; NOX87: # %bb.0: # %entry
+; NOX87-NEXT: pushl %ebp
+; NOX87-NEXT: movl %esp, %ebp
+; NOX87-NEXT: andl $-8, %esp
+; NOX87-NEXT: subl $48, %esp
+; NOX87-NEXT: movl 20(%ebp), %eax
+; NOX87-NEXT: movl 24(%ebp), %ecx
+; NOX87-NEXT: movl %ecx, {{[0-9]+}}(%esp)
+; NOX87-NEXT: movl %eax, {{[0-9]+}}(%esp)
+; NOX87-NEXT: movl 28(%ebp), %eax
+; NOX87-NEXT: movw %ax, {{[0-9]+}}(%esp)
+; NOX87-NEXT: movl 8(%ebp), %eax
+; NOX87-NEXT: movl 12(%ebp), %ecx
+; NOX87-NEXT: movl %ecx, {{[0-9]+}}(%esp)
+; NOX87-NEXT: movl %eax, {{[0-9]+}}(%esp)
+; NOX87-NEXT: movl 16(%ebp), %eax
+; NOX87-NEXT: movw %ax, {{[0-9]+}}(%esp)
+; NOX87-NEXT: fldt {{[0-9]+}}(%esp)
+; NOX87-NEXT: fldt {{[0-9]+}}(%esp)
+; NOX87-NEXT: faddp %st, %st(1)
+; NOX87-NEXT: fstpt (%esp)
+; NOX87-NEXT: movl (%esp), %eax
+; NOX87-NEXT: movl {{[0-9]+}}(%esp), %edx
+; NOX87-NEXT: movzwl {{[0-9]+}}(%esp), %ecx
+; NOX87-NEXT: movl %ebp, %esp
+; NOX87-NEXT: popl %ebp
+; NOX87-NEXT: retl
+entry:
+ %0 = fadd x86_fp80 %a, %b
+ ret x86_fp80 %0
+}
+
+define {float, float, float} @f7(float %a, float %b) nounwind {
+; X87-LABEL: f7:
+; X87: # %bb.0: # %entry
+; X87-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X87-NEXT: flds {{[0-9]+}}(%esp)
+; X87-NEXT: flds {{[0-9]+}}(%esp)
+; X87-NEXT: fstps 8(%eax)
+; X87-NEXT: fsts 4(%eax)
+; X87-NEXT: fstps (%eax)
+; X87-NEXT: retl $4
+;
+; NOSSE-NOX87-LABEL: f7:
+; NOSSE-NOX87: # %bb.0: # %entry
+; NOSSE-NOX87-NEXT: movl {{[0-9]+}}(%esp), %eax
+; NOSSE-NOX87-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; NOSSE-NOX87-NEXT: movl %eax, %edx
+; NOSSE-NOX87-NEXT: retl
+;
+; SSE-NOX87-LABEL: f7:
+; SSE-NOX87: # %bb.0: # %entry
+; SSE-NOX87-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NOX87-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; SSE-NOX87-NEXT: movd %xmm0, %eax
+; SSE-NOX87-NEXT: movd %xmm0, %edx
+; SSE-NOX87-NEXT: movd %xmm1, %ecx
+; SSE-NOX87-NEXT: retl
+entry:
+ %0 = insertvalue {float, float, float} undef, float %a, 0
+ %1 = insertvalue {float, float, float} %0, float %a, 1
+ %2 = insertvalue {float, float, float} %1, float %b, 2
+ ret {float, float, float} %2
+}
More information about the cfe-commits
mailing list