[llvm-branch-commits] [llvm] 6a1cd33 - [X86][APX] Save R16-R31 in interrupt and no_caller_saved_registers functions (#237109)

Tobias Hieta via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Sun Nov 15 12:05:38 PST 2026


Author: Feng Zou
Date: 2026-11-15T21:04:04+01:00
New Revision: 6a1cd3390561366f6c3dd741bc1e9f4dd1afccce

URL: https://github.com/llvm/llvm-project/commit/6a1cd3390561366f6c3dd741bc1e9f4dd1afccce
DIFF: https://github.com/llvm/llvm-project/commit/6a1cd3390561366f6c3dd741bc1e9f4dd1afccce.diff

LOG: [X86][APX] Save R16-R31 in interrupt and no_caller_saved_registers functions (#237109)

x86_intrcc, "no_caller_saved_registers" and anyregcc functions must
preserve every register they modify. With EGPR their save lists
(CSR_64_AllRegs*) did not include R16-R31, so a handler that uses, or
calls something that clobbers, an extended GPR silently corrupts the
interrupted context. For example, a handler reading its interrupt frame
loads it into r16/r17 and returns with iretq without restoring them.

Add CSR_64_APX_AllRegs{,_NoSSE,_AVX,_AVX512}, the AllRegs lists plus
R16-R31, and use them as the save lists when the subtarget has EGPR. The
call-preserved masks are unchanged, so callers keep treating R16-R31 as
clobbered. This stays correct for callees built without EGPR or by other
or older compilers.

Assisted-by: Claude Opus 5.5 <noreply at anthropic.com>
(cherry picked from commit 6a30204d78bd45b56611bbd09cf8c53b78b7435d)

Added: 
    llvm/test/CodeGen/X86/apx/intr-egpr-csr-call.ll
    llvm/test/CodeGen/X86/apx/intr-egpr-csr.ll

Modified: 
    llvm/lib/Target/X86/X86CallingConv.td
    llvm/lib/Target/X86/X86RegisterInfo.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86CallingConv.td b/llvm/lib/Target/X86/X86CallingConv.td
index bde5f81eaa7b84..801b00f6d44531 100644
--- a/llvm/lib/Target/X86/X86CallingConv.td
+++ b/llvm/lib/Target/X86/X86CallingConv.td
@@ -1198,6 +1198,20 @@ def CSR_64_AllRegs_AVX512 : CalleeSavedRegs<(sub (add CSR_64_MostRegs, RAX,
                                                       (sequence "ZMM%u", 0, 31),
                                                       (sequence "K%u", 0, 7)),
                                                  (sequence "XMM%u", 0, 15))>;
+
+// CSR_64_AllRegs* plus R16-R31: the save lists of x86_intrcc,
+// "no_caller_saved_registers" and anyregcc functions when EGPR is available.
+// Call sites keep the CSR_64_AllRegs* masks and treat R16-R31 as clobbered,
+// since callees built without EGPR or by older compilers may not save them.
+def CSR_64_APX_AllRegs : CalleeSavedRegs<(add CSR_64_AllRegs,
+                                              (sequence "R%u", 16, 31))>;
+def CSR_64_APX_AllRegs_NoSSE : CalleeSavedRegs<(add CSR_64_AllRegs_NoSSE,
+                                                    (sequence "R%u", 16, 31))>;
+def CSR_64_APX_AllRegs_AVX : CalleeSavedRegs<(add CSR_64_AllRegs_AVX,
+                                                  (sequence "R%u", 16, 31))>;
+def CSR_64_APX_AllRegs_AVX512 : CalleeSavedRegs<(add CSR_64_AllRegs_AVX512,
+                                                     (sequence "R%u", 16, 31))>;
+
 def CSR_64_NoneRegs    : CalleeSavedRegs<(add RBP)>;
 
 // Standard C + YMM6-15

diff  --git a/llvm/lib/Target/X86/X86RegisterInfo.cpp b/llvm/lib/Target/X86/X86RegisterInfo.cpp
index ba137c7e7350dc..641fb990bd6dab 100644
--- a/llvm/lib/Target/X86/X86RegisterInfo.cpp
+++ b/llvm/lib/Target/X86/X86RegisterInfo.cpp
@@ -270,8 +270,9 @@ X86RegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
     return CSR_NoRegs_SaveList;
   case CallingConv::AnyReg:
     if (HasAVX)
-      return CSR_64_AllRegs_AVX_SaveList;
-    return CSR_64_AllRegs_SaveList;
+      return HasEGPR ? CSR_64_APX_AllRegs_AVX_SaveList
+                     : CSR_64_AllRegs_AVX_SaveList;
+    return HasEGPR ? CSR_64_APX_AllRegs_SaveList : CSR_64_AllRegs_SaveList;
   case CallingConv::PreserveMost:
     if (IsWin64)
       return HasEGPR ? CSR_Win64_APX_RT_MostRegs_SaveList
@@ -341,12 +342,15 @@ X86RegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
   case CallingConv::X86_INTR:
     if (Is64Bit) {
       if (HasAVX512)
-        return CSR_64_AllRegs_AVX512_SaveList;
+        return HasEGPR ? CSR_64_APX_AllRegs_AVX512_SaveList
+                       : CSR_64_AllRegs_AVX512_SaveList;
       if (HasAVX)
-        return CSR_64_AllRegs_AVX_SaveList;
+        return HasEGPR ? CSR_64_APX_AllRegs_AVX_SaveList
+                       : CSR_64_AllRegs_AVX_SaveList;
       if (HasSSE)
-        return CSR_64_AllRegs_SaveList;
-      return CSR_64_AllRegs_NoSSE_SaveList;
+        return HasEGPR ? CSR_64_APX_AllRegs_SaveList : CSR_64_AllRegs_SaveList;
+      return HasEGPR ? CSR_64_APX_AllRegs_NoSSE_SaveList
+                     : CSR_64_AllRegs_NoSSE_SaveList;
     }
     if (HasAVX512)
       return CSR_32_AllRegs_AVX512_SaveList;
@@ -410,6 +414,8 @@ X86RegisterInfo::getCallPreservedMask(const MachineFunction &MF,
   case CallingConv::HiPE:
     return CSR_NoRegs_RegMask;
   case CallingConv::AnyReg:
+    // Callers keep R16-R31 clobbered: callees built without EGPR or by older
+    // compilers may not save them (see CSR_64_APX_AllRegs).
     if (HasAVX)
       return CSR_64_AllRegs_AVX_RegMask;
     return CSR_64_AllRegs_RegMask;
@@ -480,6 +486,8 @@ X86RegisterInfo::getCallPreservedMask(const MachineFunction &MF,
   case CallingConv::X86_64_SysV:
     return CSR_64_RegMask;
   case CallingConv::X86_INTR:
+    // Callers keep R16-R31 clobbered: callees built without EGPR or by older
+    // compilers may not save them (see CSR_64_APX_AllRegs).
     if (Is64Bit) {
       if (HasAVX512)
         return CSR_64_AllRegs_AVX512_RegMask;

diff  --git a/llvm/test/CodeGen/X86/apx/intr-egpr-csr-call.ll b/llvm/test/CodeGen/X86/apx/intr-egpr-csr-call.ll
new file mode 100644
index 00000000000000..f253764a2bcb85
--- /dev/null
+++ b/llvm/test/CodeGen/X86/apx/intr-egpr-csr-call.ll
@@ -0,0 +1,127 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+egpr,-sse | FileCheck %s --check-prefix=EGPR
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+egpr,+push2pop2,+ppx,-sse | FileCheck %s --check-prefix=PPX
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=-sse | FileCheck %s --check-prefix=NOEGPR
+
+; An x86_intrcc handler that makes a call must save every register the callee
+; may clobber, R16-R31 included. -sse selects the NoSSE save list.
+
+%struct.frame = type { i64, i64, i64, i64, i64 }
+
+declare void @ext()
+
+define x86_intrcc void @intr_call(ptr byval(%struct.frame) %f) nounwind {
+; EGPR-LABEL: intr_call:
+; EGPR:       # %bb.0:
+; EGPR-NEXT:    pushq %r31
+; EGPR-NEXT:    pushq %r30
+; EGPR-NEXT:    pushq %r29
+; EGPR-NEXT:    pushq %r28
+; EGPR-NEXT:    pushq %r27
+; EGPR-NEXT:    pushq %r26
+; EGPR-NEXT:    pushq %r25
+; EGPR-NEXT:    pushq %r24
+; EGPR-NEXT:    pushq %r23
+; EGPR-NEXT:    pushq %r22
+; EGPR-NEXT:    pushq %r21
+; EGPR-NEXT:    pushq %r20
+; EGPR-NEXT:    pushq %r19
+; EGPR-NEXT:    pushq %r18
+; EGPR-NEXT:    pushq %r17
+; EGPR-NEXT:    pushq %r16
+; EGPR-NEXT:    pushq %r11
+; EGPR-NEXT:    pushq %r10
+; EGPR-NEXT:    pushq %r9
+; EGPR-NEXT:    pushq %r8
+; EGPR-NEXT:    pushq %rdi
+; EGPR-NEXT:    pushq %rsi
+; EGPR-NEXT:    pushq %rdx
+; EGPR-NEXT:    pushq %rcx
+; EGPR-NEXT:    pushq %rax
+; EGPR-NEXT:    cld
+; EGPR-NEXT:    callq ext at PLT
+; EGPR-NEXT:    popq %rax
+; EGPR-NEXT:    popq %rcx
+; EGPR-NEXT:    popq %rdx
+; EGPR-NEXT:    popq %rsi
+; EGPR-NEXT:    popq %rdi
+; EGPR-NEXT:    popq %r8
+; EGPR-NEXT:    popq %r9
+; EGPR-NEXT:    popq %r10
+; EGPR-NEXT:    popq %r11
+; EGPR-NEXT:    popq %r16
+; EGPR-NEXT:    popq %r17
+; EGPR-NEXT:    popq %r18
+; EGPR-NEXT:    popq %r19
+; EGPR-NEXT:    popq %r20
+; EGPR-NEXT:    popq %r21
+; EGPR-NEXT:    popq %r22
+; EGPR-NEXT:    popq %r23
+; EGPR-NEXT:    popq %r24
+; EGPR-NEXT:    popq %r25
+; EGPR-NEXT:    popq %r26
+; EGPR-NEXT:    popq %r27
+; EGPR-NEXT:    popq %r28
+; EGPR-NEXT:    popq %r29
+; EGPR-NEXT:    popq %r30
+; EGPR-NEXT:    popq %r31
+; EGPR-NEXT:    iretq
+;
+; PPX-LABEL: intr_call:
+; PPX:       # %bb.0:
+; PPX-NEXT:    pushp %r31
+; PPX-NEXT:    push2p %r29, %r30
+; PPX-NEXT:    push2p %r27, %r28
+; PPX-NEXT:    push2p %r25, %r26
+; PPX-NEXT:    push2p %r23, %r24
+; PPX-NEXT:    push2p %r21, %r22
+; PPX-NEXT:    push2p %r19, %r20
+; PPX-NEXT:    push2p %r17, %r18
+; PPX-NEXT:    push2p %r11, %r16
+; PPX-NEXT:    push2p %r9, %r10
+; PPX-NEXT:    push2p %rdi, %r8
+; PPX-NEXT:    push2p %rdx, %rsi
+; PPX-NEXT:    push2p %rax, %rcx
+; PPX-NEXT:    cld
+; PPX-NEXT:    callq ext at PLT
+; PPX-NEXT:    pop2p %rcx, %rax
+; PPX-NEXT:    pop2p %rsi, %rdx
+; PPX-NEXT:    pop2p %r8, %rdi
+; PPX-NEXT:    pop2p %r10, %r9
+; PPX-NEXT:    pop2p %r16, %r11
+; PPX-NEXT:    pop2p %r18, %r17
+; PPX-NEXT:    pop2p %r20, %r19
+; PPX-NEXT:    pop2p %r22, %r21
+; PPX-NEXT:    pop2p %r24, %r23
+; PPX-NEXT:    pop2p %r26, %r25
+; PPX-NEXT:    pop2p %r28, %r27
+; PPX-NEXT:    pop2p %r30, %r29
+; PPX-NEXT:    popp %r31
+; PPX-NEXT:    iretq
+;
+; NOEGPR-LABEL: intr_call:
+; NOEGPR:       # %bb.0:
+; NOEGPR-NEXT:    pushq %r11
+; NOEGPR-NEXT:    pushq %r10
+; NOEGPR-NEXT:    pushq %r9
+; NOEGPR-NEXT:    pushq %r8
+; NOEGPR-NEXT:    pushq %rdi
+; NOEGPR-NEXT:    pushq %rsi
+; NOEGPR-NEXT:    pushq %rdx
+; NOEGPR-NEXT:    pushq %rcx
+; NOEGPR-NEXT:    pushq %rax
+; NOEGPR-NEXT:    cld
+; NOEGPR-NEXT:    callq ext at PLT
+; NOEGPR-NEXT:    popq %rax
+; NOEGPR-NEXT:    popq %rcx
+; NOEGPR-NEXT:    popq %rdx
+; NOEGPR-NEXT:    popq %rsi
+; NOEGPR-NEXT:    popq %rdi
+; NOEGPR-NEXT:    popq %r8
+; NOEGPR-NEXT:    popq %r9
+; NOEGPR-NEXT:    popq %r10
+; NOEGPR-NEXT:    popq %r11
+; NOEGPR-NEXT:    iretq
+  call void @ext()
+  ret void
+}

diff  --git a/llvm/test/CodeGen/X86/apx/intr-egpr-csr.ll b/llvm/test/CodeGen/X86/apx/intr-egpr-csr.ll
new file mode 100644
index 00000000000000..8ba25d1d5e58cf
--- /dev/null
+++ b/llvm/test/CodeGen/X86/apx/intr-egpr-csr.ll
@@ -0,0 +1,227 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+egpr | FileCheck %s --check-prefixes=EGPR,SSE
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+egpr,+avx | FileCheck %s --check-prefixes=EGPR,AVX
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+egpr,+avx512f | FileCheck %s --check-prefixes=EGPR,AVX512
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+egpr,+push2pop2,+ppx | FileCheck %s --check-prefix=PPX
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu | FileCheck %s --check-prefix=NOEGPR
+
+; x86_intrcc, "no_caller_saved_registers" and anyregcc functions must preserve
+; every register they modify, including R16-R31 when EGPR is available.
+
+%struct.frame = type { i64, i64, i64, i64, i64 }
+
+define x86_intrcc void @intr_clobber(ptr byval(%struct.frame) %f) nounwind {
+; EGPR-LABEL: intr_clobber:
+; EGPR:       # %bb.0:
+; EGPR-NEXT:    pushq %r31
+; EGPR-NEXT:    pushq %r16
+; EGPR-NEXT:    pushq %rax
+; EGPR-NEXT:    cld
+; EGPR-NEXT:    #APP
+; EGPR-NEXT:    #NO_APP
+; EGPR-NEXT:    popq %rax
+; EGPR-NEXT:    popq %r16
+; EGPR-NEXT:    popq %r31
+; EGPR-NEXT:    iretq
+;
+; PPX-LABEL: intr_clobber:
+; PPX:       # %bb.0:
+; PPX-NEXT:    pushp %r31
+; PPX-NEXT:    push2p %rax, %r16
+; PPX-NEXT:    cld
+; PPX-NEXT:    #APP
+; PPX-NEXT:    #NO_APP
+; PPX-NEXT:    pop2p %r16, %rax
+; PPX-NEXT:    popp %r31
+; PPX-NEXT:    iretq
+;
+; NOEGPR-LABEL: intr_clobber:
+; NOEGPR:       # %bb.0:
+; NOEGPR-NEXT:    pushq %rax
+; NOEGPR-NEXT:    cld
+; NOEGPR-NEXT:    #APP
+; NOEGPR-NEXT:    #NO_APP
+; NOEGPR-NEXT:    popq %rax
+; NOEGPR-NEXT:    iretq
+  call void asm sideeffect "", "~{r16},~{r31},~{rax}"()
+  ret void
+}
+
+; The XMM clobber tells the SSE, AVX and AVX512 save lists apart.
+define x86_intrcc void @intr_clobber_vec(ptr byval(%struct.frame) %f) nounwind {
+; SSE-LABEL: intr_clobber_vec:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushq %r31
+; SSE-NEXT:    pushq %r16
+; SSE-NEXT:    pushq %rax
+; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; SSE-NEXT:    cld
+; SSE-NEXT:    #APP
+; SSE-NEXT:    #NO_APP
+; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; SSE-NEXT:    popq %rax
+; SSE-NEXT:    popq %r16
+; SSE-NEXT:    popq %r31
+; SSE-NEXT:    iretq
+;
+; AVX-LABEL: intr_clobber_vec:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %r31
+; AVX-NEXT:    pushq %r16
+; AVX-NEXT:    pushq %rax
+; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX-NEXT:    cld
+; AVX-NEXT:    #APP
+; AVX-NEXT:    #NO_APP
+; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX-NEXT:    popq %rax
+; AVX-NEXT:    popq %r16
+; AVX-NEXT:    popq %r31
+; AVX-NEXT:    iretq
+;
+; AVX512-LABEL: intr_clobber_vec:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    pushq %r31
+; AVX512-NEXT:    pushq %r16
+; AVX512-NEXT:    pushq %rax
+; AVX512-NEXT:    vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT:    cld
+; AVX512-NEXT:    #APP
+; AVX512-NEXT:    #NO_APP
+; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-NEXT:    popq %rax
+; AVX512-NEXT:    popq %r16
+; AVX512-NEXT:    popq %r31
+; AVX512-NEXT:    iretq
+;
+; PPX-LABEL: intr_clobber_vec:
+; PPX:       # %bb.0:
+; PPX-NEXT:    pushp %r31
+; PPX-NEXT:    push2p %rax, %r16
+; PPX-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; PPX-NEXT:    cld
+; PPX-NEXT:    #APP
+; PPX-NEXT:    #NO_APP
+; PPX-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; PPX-NEXT:    pop2p %r16, %rax
+; PPX-NEXT:    popp %r31
+; PPX-NEXT:    iretq
+;
+; NOEGPR-LABEL: intr_clobber_vec:
+; NOEGPR:       # %bb.0:
+; NOEGPR-NEXT:    pushq %rax
+; NOEGPR-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; NOEGPR-NEXT:    cld
+; NOEGPR-NEXT:    #APP
+; NOEGPR-NEXT:    #NO_APP
+; NOEGPR-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; NOEGPR-NEXT:    popq %rax
+; NOEGPR-NEXT:    iretq
+  call void asm sideeffect "", "~{r16},~{r31},~{rax},~{xmm1}"()
+  ret void
+}
+
+define void @ncsr_clobber() nounwind "no_caller_saved_registers" {
+; EGPR-LABEL: ncsr_clobber:
+; EGPR:       # %bb.0:
+; EGPR-NEXT:    pushq %r31
+; EGPR-NEXT:    pushq %r16
+; EGPR-NEXT:    pushq %rax
+; EGPR-NEXT:    #APP
+; EGPR-NEXT:    #NO_APP
+; EGPR-NEXT:    popq %rax
+; EGPR-NEXT:    popq %r16
+; EGPR-NEXT:    popq %r31
+; EGPR-NEXT:    retq
+;
+; PPX-LABEL: ncsr_clobber:
+; PPX:       # %bb.0:
+; PPX-NEXT:    pushp %r31
+; PPX-NEXT:    push2p %rax, %r16
+; PPX-NEXT:    #APP
+; PPX-NEXT:    #NO_APP
+; PPX-NEXT:    pop2p %r16, %rax
+; PPX-NEXT:    popp %r31
+; PPX-NEXT:    retq
+;
+; NOEGPR-LABEL: ncsr_clobber:
+; NOEGPR:       # %bb.0:
+; NOEGPR-NEXT:    pushq %rax
+; NOEGPR-NEXT:    #APP
+; NOEGPR-NEXT:    #NO_APP
+; NOEGPR-NEXT:    popq %rax
+; NOEGPR-NEXT:    retq
+  call void asm sideeffect "", "~{r16},~{r31},~{rax}"()
+  ret void
+}
+
+; anyregcc has no AVX512 save list, so the AVX512 run saves YMM, as AVX does.
+define anyregcc void @anyreg_clobber() nounwind {
+; SSE-LABEL: anyreg_clobber:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pushq %r31
+; SSE-NEXT:    pushq %r16
+; SSE-NEXT:    pushq %rax
+; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; SSE-NEXT:    #APP
+; SSE-NEXT:    #NO_APP
+; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; SSE-NEXT:    popq %rax
+; SSE-NEXT:    popq %r16
+; SSE-NEXT:    popq %r31
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: anyreg_clobber:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %r31
+; AVX-NEXT:    pushq %r16
+; AVX-NEXT:    pushq %rax
+; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX-NEXT:    #APP
+; AVX-NEXT:    #NO_APP
+; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX-NEXT:    popq %rax
+; AVX-NEXT:    popq %r16
+; AVX-NEXT:    popq %r31
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    retq
+;
+; AVX512-LABEL: anyreg_clobber:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    pushq %r31
+; AVX512-NEXT:    pushq %r16
+; AVX512-NEXT:    pushq %rax
+; AVX512-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT:    #APP
+; AVX512-NEXT:    #NO_APP
+; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512-NEXT:    popq %rax
+; AVX512-NEXT:    popq %r16
+; AVX512-NEXT:    popq %r31
+; AVX512-NEXT:    vzeroupper
+; AVX512-NEXT:    retq
+;
+; PPX-LABEL: anyreg_clobber:
+; PPX:       # %bb.0:
+; PPX-NEXT:    pushp %r31
+; PPX-NEXT:    push2p %rax, %r16
+; PPX-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; PPX-NEXT:    #APP
+; PPX-NEXT:    #NO_APP
+; PPX-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; PPX-NEXT:    pop2p %r16, %rax
+; PPX-NEXT:    popp %r31
+; PPX-NEXT:    retq
+;
+; NOEGPR-LABEL: anyreg_clobber:
+; NOEGPR:       # %bb.0:
+; NOEGPR-NEXT:    pushq %rax
+; NOEGPR-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; NOEGPR-NEXT:    #APP
+; NOEGPR-NEXT:    #NO_APP
+; NOEGPR-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; NOEGPR-NEXT:    popq %rax
+; NOEGPR-NEXT:    retq
+  call void asm sideeffect "", "~{r16},~{r31},~{rax},~{xmm1}"()
+  ret void
+}


        


More information about the llvm-branch-commits mailing list