replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[AArch64][FastISel] Always use an AND instruction when truncating to non-legal types. 

When truncating to non-legal types (such as i16, i8 and i1) always use an AND
instruction to mask out the upper bits. This was only done when the source type
was an i64, but not when the source type was an i32.

This commit fixes this and adds the missing i32 truncate tests.

This fixes rdar://problem/21990703.

llvm-svn: 243198
This commit is contained in:
Juergen Ributzka 2015-07-25 02:16:53 +00:00
parent 3afd5bfd7b
commit 6364985b58
3 changed files with 55 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

59
llvm/lib/Target/AArch64/AArch64FastISel.cpp
View File

@ -3795,40 +3795,33 @@ bool AArch64FastISel::selectTrunc(const Instruction *I) {
return false;
bool SrcIsKill = hasTrivialKill(Op);
// If we're truncating from i64 to a smaller non-legal type then generate an
// AND. Otherwise, we know the high bits are undefined and a truncate only
// generate a COPY. We cannot mark the source register also as result
// register, because this can incorrectly transfer the kill flag onto the
// source register.
unsigned ResultReg;
if (SrcVT == MVT::i64) {
uint64_t Mask = 0;
switch (DestVT.SimpleTy) {
default:
// Trunc i64 to i32 is handled by the target-independent fast-isel.
return false;
case MVT::i1:
Mask = 0x1;
break;
case MVT::i8:
Mask = 0xff;
break;
case MVT::i16:
Mask = 0xffff;
break;
}
// Issue an extract_subreg to get the lower 32-bits.
unsigned Reg32 = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
AArch64::sub_32);
// Create the AND instruction which performs the actual truncation.
ResultReg = emitAnd_ri(MVT::i32, Reg32, /*IsKill=*/true, Mask);
assert(ResultReg && "Unexpected AND instruction emission failure.");
} else {
ResultReg = createResultReg(&AArch64::GPR32RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY), ResultReg)
.addReg(SrcReg, getKillRegState(SrcIsKill));
// If we're truncating from i64/i32 to a smaller non-legal type then generate
// an AND.
uint64_t Mask = 0;
switch (DestVT.SimpleTy) {
default:
// Trunc i64 to i32 is handled by the target-independent fast-isel.
return false;
case MVT::i1:
Mask = 0x1;
break;
case MVT::i8:
Mask = 0xff;
break;
case MVT::i16:
Mask = 0xffff;
break;
}
if (SrcVT == MVT::i64) {
// Issue an extract_subreg to get the lower 32-bits.
SrcReg = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
AArch64::sub_32);
SrcIsKill = true;
}
// Create the AND instruction which performs the actual truncation.
unsigned ResultReg = emitAnd_ri(MVT::i32, SrcReg, SrcIsKill, Mask);
assert(ResultReg && "Unexpected AND instruction emission failure.");
updateValueMap(I, ResultReg);
return true;

27
llvm/test/CodeGen/AArch64/arm64-fast-isel-conversion.ll
View File

@ -398,6 +398,33 @@ entry:
ret i1 %conv
}
define zeroext i16 @i32_trunc_i16(i32 %a) nounwind ssp {
entry:
; CHECK-LABEL: i32_trunc_i16
; CHECK: and [[REG:w[0-9]+]], w0, #0xffff
; CHECK: uxth w0, [[REG]]
%conv = trunc i32 %a to i16
ret i16 %conv
}
define zeroext i8 @i32_trunc_i8(i32 %a) nounwind ssp {
entry:
; CHECK-LABEL: i32_trunc_i8
; CHECK: and [[REG:w[0-9]+]], w0, #0xff
; CHECK: uxtb w0, [[REG]]
%conv = trunc i32 %a to i8
ret i8 %conv
}
define zeroext i1 @i32_trunc_i1(i32 %a) nounwind ssp {
entry:
; CHECK-LABEL: i32_trunc_i1
; CHECK: and [[REG:w[0-9]+]], w0, #0x1
; CHECK: and w0, [[REG]], #0x1
%conv = trunc i32 %a to i1
ret i1 %conv
}
; rdar://15101939
define void @stack_trunc() nounwind {
; CHECK-LABEL: stack_trunc

4
llvm/test/CodeGen/AArch64/fast-isel-address-extends.ll
View File

@ -1,4 +1,4 @@
; RUN: llc %s -o - -O0 -verify-machineinstrs -fast-isel=true | FileCheck %s
; RUN: llc %s -o - -O2 -verify-machineinstrs -fast-isel=true | FileCheck %s
target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
target triple = "arm64-apple-ios8.0.0"
@ -7,7 +7,7 @@ target triple = "arm64-apple-ios8.0.0"
; This was incorrect as %.mux isn't available in the last bb.
; CHECK: sxtw [[REG:x[0-9]+]]
; CHECK: strh wzr, {{\[}}[[REG]], {{.*}}, lsl #1]
; CHECK: strh wzr, {{\[}}{{.*}}, [[REG]], lsl #1]
; Function Attrs: nounwind optsize ssp
define void @EdgeLoop(i32 %dir, i32 %edge, i32 %width, i16* %tmp89, i32 %tmp136, i16 %tmp144) #0 {