19860ce3c9
Fixes #146428 _fltused reference generated for vector floating point operations Currently references to _fltused are incorrectly emitted due to the presence of vector floating point operations. This causes spurious references to _fltused in vector floating point system code where the CRT is not used. This issue is limited to the X86 MSVC environment. As described in the bug: * _fltused should only be emitted for floating point operations as the reference is used to initialize some parts of FP CRT support. * Vector floating point operations on their own don't require that CRT support. Have confirmed intended behavior with MSVC team. Fix alters usesMSVCFloatingPoint() to look for floating point instructions/operands rather than floating point or vector floating point.
30 lines
963 B
LLVM
30 lines
963 B
LLVM
; The purpose of this test to verify that the fltused symbol is
|
|
; not emitted when purely vector floating point operations are used on Windows.
|
|
|
|
; RUN: llc < %s -mtriple i686-pc-win32 | FileCheck %s
|
|
; RUN: llc < %s -mtriple x86_64-pc-win32 | FileCheck %s
|
|
|
|
@foo = external dso_local global [4 x float], align 16
|
|
|
|
; Function Attrs: noinline nounwind optnone sspstrong uwtable
|
|
define dso_local <4 x float> @func() #0 {
|
|
entry:
|
|
%__p.addr.i = alloca ptr, align 8
|
|
%vector1 = alloca <4 x float>, align 16
|
|
store ptr @foo, ptr %__p.addr.i, align 8
|
|
%0 = load ptr, ptr %__p.addr.i, align 8
|
|
%1 = load <4 x float>, ptr %0, align 16
|
|
store <4 x float> %1, ptr %vector1, align 16
|
|
%2 = load <4 x float>, ptr %vector1, align 16
|
|
ret <4 x float> %2
|
|
}
|
|
|
|
define <4 x float> @mul_vectors(<4 x float> %a, <4 x float> %b) {
|
|
entry:
|
|
%result = fmul <4 x float> %a, %b
|
|
ret <4 x float> %result
|
|
}
|
|
|
|
; _fltused is determined at a module level
|
|
; CHECK-NOT: .globl {{_?}}_fltused
|