maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/SPARC/32abi.ll

258 lines
7.9 KiB
LLVM
Raw Normal View History

[Sparc] Add Soft Float support This change adds support for software floating point operations for Sparc targets. This is the first in a set of patches to enable software floating point on Sparc. The next patch will enable the option to be used with Clang. Differential Revision: http://reviews.llvm.org/D19265 llvm-svn: 269892
2016-05-18 17:14:13 +08:00
; RUN: llc < %s -march=sparc -disable-sparc-delay-filler -disable-sparc-leaf-proc | FileCheck %s --check-prefix=CHECK --check-prefix=HARD --check-prefix=CHECK-BE
; RUN: llc < %s -march=sparcel -disable-sparc-delay-filler -disable-sparc-leaf-proc | FileCheck %s --check-prefix=CHECK --check-prefix=HARD --check-prefix=CHECK-LE
; RUN: llc < %s -march=sparc -disable-sparc-delay-filler -disable-sparc-leaf-proc -mattr=soft-float | FileCheck %s --check-prefix=CHECK --check-prefix=SOFT --check-prefix=CHECK-BE
[Sparc] Fix handling of double incoming arguments on sparc little-endian. On SparcV8, doubles get passed in two 32-bit integer registers. The call code was already handling endianness correctly, but the incoming argument code was not -- it got the two halves in opposite order. Also remove some dead code in LowerFormalArguments_32 to handle less-than-32bit values, which can't actually happen. Finally, add some test cases for the 32-bit calling convention, cribbed from the 64abi.ll test, and run for both big and little-endian. llvm-svn: 255668
2015-12-16 03:23:12 +08:00
; CHECK-LABEL: intarg:
; The save/restore frame is not strictly necessary here, but we would need to
; refer to %o registers instead.
; CHECK: save %sp, -96, %sp
; CHECK: ld [%fp+96], [[R2:%[gilo][0-7]]]
; CHECK: ld [%fp+92], [[R1:%[gilo][0-7]]]
; CHECK: stb %i0, [%i4]
; CHECK: stb %i1, [%i4]
; CHECK: sth %i2, [%i4]
; CHECK: st %i3, [%i4]
; CHECK: st %i4, [%i4]
; CHECK: st %i5, [%i4]
; CHECK: st [[R1]], [%i4]
; CHECK: st [[R2]], [%i4]
; CHECK: restore
define void @intarg(i8 %a0, ; %i0
i8 %a1, ; %i1
i16 %a2, ; %i2
i32 %a3, ; %i3
i8* %a4, ; %i4
i32 %a5, ; %i5
i32 signext %a6, ; [%fp+92]
i8* %a7) { ; [%fp+96]
Elide stores which are overwritten without being observed. Summary: In SelectionDAG, when a store is immediately chained to another store to the same address, elide the first store as it has no observable effects. This is causes small improvements dealing with intrinsics lowered to stores. Test notes: * Many testcases overwrite store addresses multiple times and needed minor changes, mainly making stores volatile to prevent the optimization from optimizing the test away. * Many X86 test cases optimized out instructions associated with associated with va_start. * Note that test_splat in CodeGen/AArch64/misched-stp.ll no longer has dependencies to check and can probably be removed and potentially replaced with another test. Reviewers: rnk, john.brawn Subscribers: aemerson, rengolin, qcolombet, jyknight, nemanjai, nhaehnle, javed.absar, llvm-commits Differential Revision: https://reviews.llvm.org/D33206 llvm-svn: 303198
2017-05-17 03:43:56 +08:00
store volatile i8 %a0, i8* %a4
store volatile i8 %a1, i8* %a4
[Sparc] Fix handling of double incoming arguments on sparc little-endian. On SparcV8, doubles get passed in two 32-bit integer registers. The call code was already handling endianness correctly, but the incoming argument code was not -- it got the two halves in opposite order. Also remove some dead code in LowerFormalArguments_32 to handle less-than-32bit values, which can't actually happen. Finally, add some test cases for the 32-bit calling convention, cribbed from the 64abi.ll test, and run for both big and little-endian. llvm-svn: 255668
2015-12-16 03:23:12 +08:00
%p16 = bitcast i8* %a4 to i16*
Elide stores which are overwritten without being observed. Summary: In SelectionDAG, when a store is immediately chained to another store to the same address, elide the first store as it has no observable effects. This is causes small improvements dealing with intrinsics lowered to stores. Test notes: * Many testcases overwrite store addresses multiple times and needed minor changes, mainly making stores volatile to prevent the optimization from optimizing the test away. * Many X86 test cases optimized out instructions associated with associated with va_start. * Note that test_splat in CodeGen/AArch64/misched-stp.ll no longer has dependencies to check and can probably be removed and potentially replaced with another test. Reviewers: rnk, john.brawn Subscribers: aemerson, rengolin, qcolombet, jyknight, nemanjai, nhaehnle, javed.absar, llvm-commits Differential Revision: https://reviews.llvm.org/D33206 llvm-svn: 303198
2017-05-17 03:43:56 +08:00
store volatile i16 %a2, i16* %p16
[Sparc] Fix handling of double incoming arguments on sparc little-endian. On SparcV8, doubles get passed in two 32-bit integer registers. The call code was already handling endianness correctly, but the incoming argument code was not -- it got the two halves in opposite order. Also remove some dead code in LowerFormalArguments_32 to handle less-than-32bit values, which can't actually happen. Finally, add some test cases for the 32-bit calling convention, cribbed from the 64abi.ll test, and run for both big and little-endian. llvm-svn: 255668
2015-12-16 03:23:12 +08:00
%p32 = bitcast i8* %a4 to i32*
Elide stores which are overwritten without being observed. Summary: In SelectionDAG, when a store is immediately chained to another store to the same address, elide the first store as it has no observable effects. This is causes small improvements dealing with intrinsics lowered to stores. Test notes: * Many testcases overwrite store addresses multiple times and needed minor changes, mainly making stores volatile to prevent the optimization from optimizing the test away. * Many X86 test cases optimized out instructions associated with associated with va_start. * Note that test_splat in CodeGen/AArch64/misched-stp.ll no longer has dependencies to check and can probably be removed and potentially replaced with another test. Reviewers: rnk, john.brawn Subscribers: aemerson, rengolin, qcolombet, jyknight, nemanjai, nhaehnle, javed.absar, llvm-commits Differential Revision: https://reviews.llvm.org/D33206 llvm-svn: 303198
2017-05-17 03:43:56 +08:00
store volatile i32 %a3, i32* %p32
[Sparc] Fix handling of double incoming arguments on sparc little-endian. On SparcV8, doubles get passed in two 32-bit integer registers. The call code was already handling endianness correctly, but the incoming argument code was not -- it got the two halves in opposite order. Also remove some dead code in LowerFormalArguments_32 to handle less-than-32bit values, which can't actually happen. Finally, add some test cases for the 32-bit calling convention, cribbed from the 64abi.ll test, and run for both big and little-endian. llvm-svn: 255668
2015-12-16 03:23:12 +08:00
%pp = bitcast i8* %a4 to i8**
Elide stores which are overwritten without being observed. Summary: In SelectionDAG, when a store is immediately chained to another store to the same address, elide the first store as it has no observable effects. This is causes small improvements dealing with intrinsics lowered to stores. Test notes: * Many testcases overwrite store addresses multiple times and needed minor changes, mainly making stores volatile to prevent the optimization from optimizing the test away. * Many X86 test cases optimized out instructions associated with associated with va_start. * Note that test_splat in CodeGen/AArch64/misched-stp.ll no longer has dependencies to check and can probably be removed and potentially replaced with another test. Reviewers: rnk, john.brawn Subscribers: aemerson, rengolin, qcolombet, jyknight, nemanjai, nhaehnle, javed.absar, llvm-commits Differential Revision: https://reviews.llvm.org/D33206 llvm-svn: 303198
2017-05-17 03:43:56 +08:00
store volatile i8* %a4, i8** %pp
store volatile i32 %a5, i32* %p32
store volatile i32 %a6, i32* %p32
store volatile i8* %a7, i8** %pp
[Sparc] Fix handling of double incoming arguments on sparc little-endian. On SparcV8, doubles get passed in two 32-bit integer registers. The call code was already handling endianness correctly, but the incoming argument code was not -- it got the two halves in opposite order. Also remove some dead code in LowerFormalArguments_32 to handle less-than-32bit values, which can't actually happen. Finally, add some test cases for the 32-bit calling convention, cribbed from the 64abi.ll test, and run for both big and little-endian. llvm-svn: 255668
2015-12-16 03:23:12 +08:00
ret void
}
; CHECK-LABEL: call_intarg:
; CHECK: save %sp, -104, %sp
; Use %o0-%o5 for outgoing arguments
; CHECK: mov 5, %o5
; CHECK: st %i0, [%sp+92]
; CHECK: call intarg
; CHECK-NOT: add %sp
; CHECK: restore
define void @call_intarg(i32 %i0, i8* %i1) {
call void @intarg(i8 0, i8 1, i16 2, i32 3, i8* undef, i32 5, i32 %i0, i8* %i1)
ret void
}
;; Verify doubles starting with an even reg, starting with an odd reg,
;; straddling the boundary of regs and mem, and floats in regs and mem.
;
; CHECK-LABEL: floatarg:
[Sparc] Add Soft Float support This change adds support for software floating point operations for Sparc targets. This is the first in a set of patches to enable software floating point on Sparc. The next patch will enable the option to be used with Clang. Differential Revision: http://reviews.llvm.org/D19265 llvm-svn: 269892
2016-05-18 17:14:13 +08:00
; HARD: save %sp, -120, %sp
; HARD: mov %i5, %g2
; HARD-NEXT: ld [%fp+92], %g3
; HARD-NEXT: mov %i4, %i5
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; HARD-NEXT: ! kill
[Sparc] Add Soft Float support This change adds support for software floating point operations for Sparc targets. This is the first in a set of patches to enable software floating point on Sparc. The next patch will enable the option to be used with Clang. Differential Revision: http://reviews.llvm.org/D19265 llvm-svn: 269892
2016-05-18 17:14:13 +08:00
; HARD-NEXT: std %g2, [%fp+-24]
; HARD-NEXT: mov %i3, %i4
; HARD-NEXT: std %i4, [%fp+-16]
VirtRegMap: Replace some identity copies with KILL instructions. An identity COPY like this: %AL = COPY %AL, %EAX<imp-def> has no semantic effect, but encodes liveness information: Further users of %EAX only depend on this instruction even though it does not define the full register. Replace the COPY with a KILL instruction in those cases to maintain this liveness information. (This reverts a small part of r238588 but this time adds a comment explaining why a KILL instruction is useful). llvm-svn: 274952
2016-07-09 08:19:07 +08:00
; HARD-NEXT: ! kill
[Sparc] Add Soft Float support This change adds support for software floating point operations for Sparc targets. This is the first in a set of patches to enable software floating point on Sparc. The next patch will enable the option to be used with Clang. Differential Revision: http://reviews.llvm.org/D19265 llvm-svn: 269892
2016-05-18 17:14:13 +08:00
; HARD-NEXT: std %i0, [%fp+-8]
; HARD-NEXT: st %i2, [%fp+-28]
; HARD-NEXT: ld [%fp+104], %f0
; HARD-NEXT: ldd [%fp+96], %f2
; HARD-NEXT: ld [%fp+-28], %f1
; HARD-NEXT: ldd [%fp+-8], %f4
; HARD-NEXT: ldd [%fp+-16], %f6
; HARD-NEXT: ldd [%fp+-24], %f8
; HARD-NEXT: fstod %f1, %f10
; HARD-NEXT: faddd %f4, %f10, %f4
; HARD-NEXT: faddd %f6, %f4, %f4
; HARD-NEXT: faddd %f8, %f4, %f4
; HARD-NEXT: faddd %f2, %f4, %f2
; HARD-NEXT: fstod %f0, %f0
; HARD-NEXT: faddd %f0, %f2, %f0
; SOFT: save %sp, -96, %sp
; SOFT: ld [%fp+104], %l0
; SOFT-NEXT: ld [%fp+96], %l1
; SOFT-NEXT: ld [%fp+100], %l2
; SOFT-NEXT: ld [%fp+92], %l3
; SOFT-NEXT: mov %i2, %o0
; SOFT-NEXT: call __extendsfdf2
; SOFT-NEXT: nop
; SOFT-NEXT: mov %o0, %i2
; SOFT-NEXT: mov %o1, %g2
; SOFT-NEXT: mov %i0, %o0
; SOFT-NEXT: mov %i1, %o1
; SOFT-NEXT: mov %i2, %o2
; SOFT-NEXT: mov %g2, %o3
; SOFT-NEXT: call __adddf3
; SOFT-NEXT: nop
; SOFT-NEXT: mov %o0, %i0
; SOFT-NEXT: mov %o1, %i1
; SOFT-NEXT: mov %i3, %o0
; SOFT-NEXT: mov %i4, %o1
; SOFT-NEXT: mov %i0, %o2
; SOFT-NEXT: mov %i1, %o3
; SOFT-NEXT: call __adddf3
; SOFT-NEXT: nop
; SOFT-NEXT: mov %o0, %i0
; SOFT-NEXT: mov %o1, %i1
; SOFT-NEXT: mov %i5, %o0
; SOFT-NEXT: mov %l3, %o1
; SOFT-NEXT: mov %i0, %o2
; SOFT-NEXT: mov %i1, %o3
; SOFT-NEXT: call __adddf3
; SOFT-NEXT: nop
; SOFT-NEXT: mov %o0, %i0
; SOFT-NEXT: mov %o1, %i1
; SOFT-NEXT: mov %l1, %o0
; SOFT-NEXT: mov %l2, %o1
; SOFT-NEXT: mov %i0, %o2
; SOFT-NEXT: mov %i1, %o3
; SOFT-NEXT: call __adddf3
; SOFT-NEXT: nop
; SOFT-NEXT: mov %o0, %i0
; SOFT-NEXT: mov %o1, %i1
; SOFT-NEXT: mov %l0, %o0
; SOFT-NEXT: call __extendsfdf2
; SOFT-NEXT: nop
; SOFT-NEXT: mov %i0, %o2
; SOFT-NEXT: mov %i1, %o3
; SOFT-NEXT: call __adddf3
; SOFT-NEXT: nop
; SOFT-NEXT: mov %o0, %i0
; SOFT-NEXT: mov %o1, %i1
; CHECK: restore
[Sparc] Fix handling of double incoming arguments on sparc little-endian. On SparcV8, doubles get passed in two 32-bit integer registers. The call code was already handling endianness correctly, but the incoming argument code was not -- it got the two halves in opposite order. Also remove some dead code in LowerFormalArguments_32 to handle less-than-32bit values, which can't actually happen. Finally, add some test cases for the 32-bit calling convention, cribbed from the 64abi.ll test, and run for both big and little-endian. llvm-svn: 255668
2015-12-16 03:23:12 +08:00
define double @floatarg(double %a0, ; %i0,%i1
float %a1, ; %i2
double %a2, ; %i3, %i4
double %a3, ; %i5, [%fp+92] (using 4 bytes)
double %a4, ; [%fp+96] (using 8 bytes)
float %a5) { ; [%fp+104] (using 4 bytes)
%d1 = fpext float %a1 to double
%s1 = fadd double %a0, %d1
%s2 = fadd double %a2, %s1
%s3 = fadd double %a3, %s2
%s4 = fadd double %a4, %s3
%d5 = fpext float %a5 to double
%s5 = fadd double %d5, %s4
ret double %s5
}
; CHECK-LABEL: call_floatarg:
[Sparc] Add Soft Float support This change adds support for software floating point operations for Sparc targets. This is the first in a set of patches to enable software floating point on Sparc. The next patch will enable the option to be used with Clang. Differential Revision: http://reviews.llvm.org/D19265 llvm-svn: 269892
2016-05-18 17:14:13 +08:00
; HARD: save %sp, -112, %sp
; HARD: mov %i2, %o1
; HARD-NEXT: mov %i1, %o0
; HARD-NEXT: st %i0, [%sp+104]
; HARD-NEXT: std %o0, [%sp+96]
; HARD-NEXT: st %o1, [%sp+92]
; HARD-NEXT: mov %i0, %o2
; HARD-NEXT: mov %o0, %o3
; HARD-NEXT: mov %o1, %o4
; HARD-NEXT: mov %o0, %o5
; HARD-NEXT: call floatarg
; HARD: std %f0, [%i4]
; SOFT: st %i0, [%sp+104]
; SOFT-NEXT: st %i2, [%sp+100]
; SOFT-NEXT: st %i1, [%sp+96]
; SOFT-NEXT: st %i2, [%sp+92]
; SOFT-NEXT: mov %i1, %o0
; SOFT-NEXT: mov %i2, %o1
; SOFT-NEXT: mov %i0, %o2
; SOFT-NEXT: mov %i1, %o3
; SOFT-NEXT: mov %i2, %o4
; SOFT-NEXT: mov %i1, %o5
; SOFT-NEXT: call floatarg
; SOFT: std %o0, [%i4]
[Sparc] Fix handling of double incoming arguments on sparc little-endian. On SparcV8, doubles get passed in two 32-bit integer registers. The call code was already handling endianness correctly, but the incoming argument code was not -- it got the two halves in opposite order. Also remove some dead code in LowerFormalArguments_32 to handle less-than-32bit values, which can't actually happen. Finally, add some test cases for the 32-bit calling convention, cribbed from the 64abi.ll test, and run for both big and little-endian. llvm-svn: 255668
2015-12-16 03:23:12 +08:00
; CHECK: restore
define void @call_floatarg(float %f1, double %d2, float %f5, double *%p) {
%r = call double @floatarg(double %d2, float %f1, double %d2, double %d2,
double %d2, float %f1)
store double %r, double* %p
ret void
}
;; i64 arguments should effectively work the same as double: split
;; into two locations. This is different for little-endian vs big
;; endian, since the 64-bit math needs to be split
; CHECK-LABEL: i64arg:
; CHECK: save %sp, -96, %sp
; CHECK-BE: ld [%fp+100], %g2
; CHECK-BE-NEXT: ld [%fp+96], %g3
; CHECK-BE-NEXT: ld [%fp+92], %g4
; CHECK-BE-NEXT: addcc %i1, %i2, %i1
; CHECK-BE-NEXT: addxcc %i0, 0, %i0
; CHECK-BE-NEXT: addcc %i4, %i1, %i1
; CHECK-BE-NEXT: addxcc %i3, %i0, %i0
; CHECK-BE-NEXT: addcc %g4, %i1, %i1
; CHECK-BE-NEXT: ld [%fp+104], %i2
; CHECK-BE-NEXT: addxcc %i5, %i0, %i0
; CHECK-BE-NEXT: addcc %g2, %i1, %i1
; CHECK-BE-NEXT: addxcc %g3, %i0, %i0
; CHECK-BE-NEXT: addcc %i2, %i1, %i1
; CHECK-BE-NEXT: addxcc %i0, 0, %i0
;
; CHECK-LE: ld [%fp+96], %g2
; CHECK-LE-NEXT: ld [%fp+100], %g3
; CHECK-LE-NEXT: ld [%fp+92], %g4
; CHECK-LE-NEXT: addcc %i0, %i2, %i0
; CHECK-LE-NEXT: addxcc %i1, 0, %i1
; CHECK-LE-NEXT: addcc %i3, %i0, %i0
; CHECK-LE-NEXT: addxcc %i4, %i1, %i1
; CHECK-LE-NEXT: addcc %i5, %i0, %i0
; CHECK-LE-NEXT: ld [%fp+104], %i2
; CHECK-LE-NEXT: addxcc %g4, %i1, %i1
; CHECK-LE-NEXT: addcc %g2, %i0, %i0
; CHECK-LE-NEXT: addxcc %g3, %i1, %i1
; CHECK-LE-NEXT: addcc %i2, %i0, %i0
; CHECK-LE-NEXT: addxcc %i1, 0, %i1
; CHECK-NEXT: restore
define i64 @i64arg(i64 %a0, ; %i0,%i1
i32 %a1, ; %i2
i64 %a2, ; %i3, %i4
i64 %a3, ; %i5, [%fp+92] (using 4 bytes)
i64 %a4, ; [%fp+96] (using 8 bytes)
i32 %a5) { ; [%fp+104] (using 4 bytes)
%a1L = zext i32 %a1 to i64
%s1 = add i64 %a0, %a1L
%s2 = add i64 %a2, %s1
%s3 = add i64 %a3, %s2
%s4 = add i64 %a4, %s3
%a5L = zext i32 %a5 to i64
%s5 = add i64 %a5L, %s4
ret i64 %s5
}
; CHECK-LABEL: call_i64arg:
; CHECK: save %sp, -112, %sp
; CHECK: st %i0, [%sp+104]
; CHECK-NEXT: st %i2, [%sp+100]
; CHECK-NEXT: st %i1, [%sp+96]
; CHECK-NEXT: st %i2, [%sp+92]
; CHECK-NEXT: mov %i1, %o0
; CHECK-NEXT: mov %i2, %o1
; CHECK-NEXT: mov %i0, %o2
; CHECK-NEXT: mov %i1, %o3
; CHECK-NEXT: mov %i2, %o4
; CHECK-NEXT: mov %i1, %o5
; CHECK-NEXT: call i64arg
; CHECK: std %o0, [%i3]
; CHECK-NEXT: restore
define void @call_i64arg(i32 %a0, i64 %a1, i64* %p) {
%r = call i64 @i64arg(i64 %a1, i32 %a0, i64 %a1, i64 %a1, i64 %a1, i32 %a0)
store i64 %r, i64* %p
ret void
}