[Thumb] Improve THUMB code generation by moving movs around if it avoids emitting an extra instruction #88032
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-backend-arm Author: AtariDreams (AtariDreams) ChangesWhen the ARM backend decides it is best to copy a value from one register to another, it does not take THUMB limitations into account, which means we have to find a way to do a transfer AFTER it has been decided that a mov would work, sometimes using movs, and other times using two instructions, which was not what was intended by the compiler backend. Until we find a better solution, this is unfortunately the best I could do unless we can somehow find a more proper way that involves updating the cost model to take the possibility of pushing through stack or a high register into account. Full diff: https://github.com/llvm/llvm-project/pull/88032.diff 5 Files Affected:
diff --git a/llvm/lib/Target/ARM/Thumb1InstrInfo.cpp b/llvm/lib/Target/ARM/Thumb1InstrInfo.cpp
index 5b0b799880a35f..825e0b430d4d79 100644
--- a/llvm/lib/Target/ARM/Thumb1InstrInfo.cpp
+++ b/llvm/lib/Target/ARM/Thumb1InstrInfo.cpp
@@ -38,6 +38,48 @@ unsigned Thumb1InstrInfo::getUnindexedOpcode(unsigned Opc) const {
return 0;
}
+static bool tryToSinkCSPRDef(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator &I,
+ const BitVector &RegUnits, const DebugLoc &DL,
+ MCRegister DestReg, MCRegister SrcReg,
+ bool KillSrc, const TargetRegisterInfo *RegInfo) {
+
+ // Try to find the cspr instruction
+ LiveRegUnits UsedRegs(*RegInfo);
+
+ // Pick up where we left off with last RegUnits.
+ UsedRegs.addUnits(RegUnits);
+
+ // We are assuming at this point SrcReg and DestReg are both available
+ // Because we want to change where it is inserted.
+
+ auto InstUpToI = I;
+
+ auto begin = MBB.begin();
+ // Okay, so let's do this:
+ while (InstUpToI != begin && !UsedRegs.available(ARM::CPSR)) {
+
+ // Do not move any instruction across function call.
+ if (InstUpToI->isCall())
+ return false;
+
+ // If we end up going too far back so that either src or dst is not
+ // available, stop.
+ if (!UsedRegs.available(DestReg) || UsedRegs.available(SrcReg))
+ return false;
+
+ UsedRegs.stepBackward(*--InstUpToI);
+ }
+
+ // If we reached the beginning, then there is nothing we can do.
+ // FIXME: Can we keep going back if there is only one predecessor?
+ if (!UsedRegs.available(ARM::CPSR))
+ return false;
+
+ I = InstUpToI;
+ return true;
+}
+
void Thumb1InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
MachineBasicBlock::iterator I,
const DebugLoc &DL, MCRegister DestReg,
@@ -72,6 +114,34 @@ void Thumb1InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
return;
}
+ // Not ideal, but since the solution involves 2 instructions instead of 1,
+ // Which the scheduler did not account for, codegen is not ideal anyway, so
+ // lets see if we can manually sink this copy
+ // FIXME: Shouldn't this be done by the MachineSink pass?
+ // Though the sink pass won't see the two instructions as one copy but two.
+ // Here is the only change we could remedy that.
+
+ // TODO: What if the definition of the last is outside the basic block?
+ // FIXME: For now, we sink only to a successor which has a single
+ // predecessor
+ // so that we can directly sink COPY instructions to the successor without
+ // adding any new block or branch instruction.
+
+ // See if we can find the instruction where CSPR is defined.
+ // Bail if any reg dependencies will be violated
+
+ // InstUpToI is equal to I
+
+ if (tryToSinkCSPRDef(MBB, InstUpToI, UsedRegs.getBitVector(), DL, DestReg,
+ SrcReg, KillSrc, RegInfo)) {
+
+ // We found the place to insert the MOVS
+ BuildMI(MBB, InstUpToI, DL, get(ARM::tMOVSr), DestReg)
+ .addReg(SrcReg, getKillRegState(KillSrc))
+ ->addRegisterDead(ARM::CPSR, RegInfo);
+ return;
+ }
+
// Use high register to move source to destination
// if movs is not an option.
BitVector Allocatable = RegInfo->getAllocatableSet(
diff --git a/llvm/test/CodeGen/ARM/sadd_sat.ll b/llvm/test/CodeGen/ARM/sadd_sat.ll
index 0060b4458081bc..c94c4edd64ef85 100644
--- a/llvm/test/CodeGen/ARM/sadd_sat.ll
+++ b/llvm/test/CodeGen/ARM/sadd_sat.ll
@@ -129,9 +129,8 @@ define i64 @func2(i64 %x, i64 %y) nounwind {
; CHECK-T15TE-NEXT: eors r4, r3
; CHECK-T15TE-NEXT: bics r4, r1
; CHECK-T15TE-NEXT: asrs r1, r3, #31
+; CHECK-T15TE-NEXT: movs r0, r1
; CHECK-T15TE-NEXT: cmp r4, #0
-; CHECK-T15TE-NEXT: mov r12, r1
-; CHECK-T15TE-NEXT: mov r0, r12
; CHECK-T15TE-NEXT: bmi .LBB1_2
; CHECK-T15TE-NEXT: @ %bb.1:
; CHECK-T15TE-NEXT: movs r0, r2
@@ -151,28 +150,28 @@ define i64 @func2(i64 %x, i64 %y) nounwind {
}
define signext i16 @func16(i16 signext %x, i16 signext %y) nounwind {
-; CHECK-T1-LABEL: func16:
-; CHECK-T1: @ %bb.0:
-; CHECK-T1-NEXT: adds r0, r0, r1
-; CHECK-T1-NEXT: ldr r1, .LCPI2_0
-; CHECK-T1-NEXT: cmp r0, r1
-; CHECK-T1-NEXT: blt .LBB2_2
-; CHECK-T1-NEXT: @ %bb.1:
-; CHECK-T1-NEXT: {{movs|mov}} r0, r1
-; CHECK-T1-NEXT: .LBB2_2:
-; CHECK-T1-NEXT: ldr r1, .LCPI2_1
-; CHECK-T1-NEXT: cmp r0, r1
-; CHECK-T1-NEXT: bgt .LBB2_4
-; CHECK-T1-NEXT: @ %bb.3:
-; CHECK-T1-NEXT: {{movs|mov}} r0, r1
-; CHECK-T1-NEXT: .LBB2_4:
-; CHECK-T1-NEXT: bx lr
-; CHECK-T1-NEXT: .p2align 2
-; CHECK-T1-NEXT: @ %bb.5:
-; CHECK-T1-NEXT: .LCPI2_0:
-; CHECK-T1-NEXT: .long 32767 @ 0x7fff
-; CHECK-T1-NEXT: .LCPI2_1:
-; CHECK-T1-NEXT: .long 4294934528 @ 0xffff8000
+; CHECK-T16-LABEL: func16:
+; CHECK-T16: @ %bb.0:
+; CHECK-T16-NEXT: adds r0, r0, r1
+; CHECK-T16-NEXT: ldr r1, .LCPI2_0
+; CHECK-T16-NEXT: cmp r0, r1
+; CHECK-T16-NEXT: blt .LBB2_2
+; CHECK-T16-NEXT: @ %bb.1:
+; CHECK-T16-NEXT: mov r0, r1
+; CHECK-T16-NEXT: .LBB2_2:
+; CHECK-T16-NEXT: ldr r1, .LCPI2_1
+; CHECK-T16-NEXT: cmp r0, r1
+; CHECK-T16-NEXT: bgt .LBB2_4
+; CHECK-T16-NEXT: @ %bb.3:
+; CHECK-T16-NEXT: mov r0, r1
+; CHECK-T16-NEXT: .LBB2_4:
+; CHECK-T16-NEXT: bx lr
+; CHECK-T16-NEXT: .p2align 2
+; CHECK-T16-NEXT: @ %bb.5:
+; CHECK-T16-NEXT: .LCPI2_0:
+; CHECK-T16-NEXT: .long 32767 @ 0x7fff
+; CHECK-T16-NEXT: .LCPI2_1:
+; CHECK-T16-NEXT: .long 4294934528 @ 0xffff8000
;
; CHECK-T2NODSP-LABEL: func16:
; CHECK-T2NODSP: @ %bb.0:
@@ -210,6 +209,29 @@ define signext i16 @func16(i16 signext %x, i16 signext %y) nounwind {
; CHECK-ARMBASEDSP-NEXT: asr r0, r0, #16
; CHECK-ARMBASEDSP-NEXT: bx lr
;
+; CHECK-T15TE-LABEL: func16:
+; CHECK-T15TE: @ %bb.0:
+; CHECK-T15TE-NEXT: adds r0, r0, r1
+; CHECK-T15TE-NEXT: ldr r1, .LCPI2_0
+; CHECK-T15TE-NEXT: cmp r0, r1
+; CHECK-T15TE-NEXT: blt .LBB2_2
+; CHECK-T15TE-NEXT: @ %bb.1:
+; CHECK-T15TE-NEXT: movs r0, r1
+; CHECK-T15TE-NEXT: .LBB2_2:
+; CHECK-T15TE-NEXT: ldr r1, .LCPI2_1
+; CHECK-T15TE-NEXT: cmp r0, r1
+; CHECK-T15TE-NEXT: bgt .LBB2_4
+; CHECK-T15TE-NEXT: @ %bb.3:
+; CHECK-T15TE-NEXT: movs r0, r1
+; CHECK-T15TE-NEXT: .LBB2_4:
+; CHECK-T15TE-NEXT: bx lr
+; CHECK-T15TE-NEXT: .p2align 2
+; CHECK-T15TE-NEXT: @ %bb.5:
+; CHECK-T15TE-NEXT: .LCPI2_0:
+; CHECK-T15TE-NEXT: .long 32767 @ 0x7fff
+; CHECK-T15TE-NEXT: .LCPI2_1:
+; CHECK-T15TE-NEXT: .long 4294934528 @ 0xffff8000
+;
; CHECK-ARMDSP-LABEL: func16:
; CHECK-ARMDSP: @ %bb.0:
; CHECK-ARMDSP-NEXT: qadd16 r0, r0, r1
@@ -220,22 +242,22 @@ define signext i16 @func16(i16 signext %x, i16 signext %y) nounwind {
}
define signext i8 @func8(i8 signext %x, i8 signext %y) nounwind {
-; CHECK-T1-LABEL: func8:
-; CHECK-T1: @ %bb.0:
-; CHECK-T1-NEXT: adds r0, r0, r1
-; CHECK-T1-NEXT: movs r1, #127
-; CHECK-T1-NEXT: cmp r0, #127
-; CHECK-T1-NEXT: blt .LBB3_2
-; CHECK-T1-NEXT: @ %bb.1:
-; CHECK-T1-NEXT: {{movs|mov}} r0, r1
-; CHECK-T1-NEXT: .LBB3_2:
-; CHECK-T1-NEXT: mvns r1, r1
-; CHECK-T1-NEXT: cmp r0, r1
-; CHECK-T1-NEXT: bgt .LBB3_4
-; CHECK-T1-NEXT: @ %bb.3:
-; CHECK-T1-NEXT: {{movs|mov}} r0, r1
-; CHECK-T1-NEXT: .LBB3_4:
-; CHECK-T1-NEXT: bx lr
+; CHECK-T16-LABEL: func8:
+; CHECK-T16: @ %bb.0:
+; CHECK-T16-NEXT: adds r0, r0, r1
+; CHECK-T16-NEXT: movs r1, #127
+; CHECK-T16-NEXT: cmp r0, #127
+; CHECK-T16-NEXT: blt .LBB3_2
+; CHECK-T16-NEXT: @ %bb.1:
+; CHECK-T16-NEXT: mov r0, r1
+; CHECK-T16-NEXT: .LBB3_2:
+; CHECK-T16-NEXT: mvns r1, r1
+; CHECK-T16-NEXT: cmp r0, r1
+; CHECK-T16-NEXT: bgt .LBB3_4
+; CHECK-T16-NEXT: @ %bb.3:
+; CHECK-T16-NEXT: mov r0, r1
+; CHECK-T16-NEXT: .LBB3_4:
+; CHECK-T16-NEXT: bx lr
;
; CHECK-T2NODSP-LABEL: func8:
; CHECK-T2NODSP: @ %bb.0:
@@ -266,6 +288,23 @@ define signext i8 @func8(i8 signext %x, i8 signext %y) nounwind {
; CHECK-ARMBASEDSP-NEXT: asr r0, r0, #24
; CHECK-ARMBASEDSP-NEXT: bx lr
;
+; CHECK-T15TE-LABEL: func8:
+; CHECK-T15TE: @ %bb.0:
+; CHECK-T15TE-NEXT: adds r0, r0, r1
+; CHECK-T15TE-NEXT: movs r1, #127
+; CHECK-T15TE-NEXT: cmp r0, #127
+; CHECK-T15TE-NEXT: blt .LBB3_2
+; CHECK-T15TE-NEXT: @ %bb.1:
+; CHECK-T15TE-NEXT: movs r0, r1
+; CHECK-T15TE-NEXT: .LBB3_2:
+; CHECK-T15TE-NEXT: mvns r1, r1
+; CHECK-T15TE-NEXT: cmp r0, r1
+; CHECK-T15TE-NEXT: bgt .LBB3_4
+; CHECK-T15TE-NEXT: @ %bb.3:
+; CHECK-T15TE-NEXT: movs r0, r1
+; CHECK-T15TE-NEXT: .LBB3_4:
+; CHECK-T15TE-NEXT: bx lr
+;
; CHECK-ARMDSP-LABEL: func8:
; CHECK-ARMDSP: @ %bb.0:
; CHECK-ARMDSP-NEXT: qadd8 r0, r0, r1
@@ -276,22 +315,22 @@ define signext i8 @func8(i8 signext %x, i8 signext %y) nounwind {
}
define signext i4 @func3(i4 signext %x, i4 signext %y) nounwind {
-; CHECK-T1-LABEL: func3:
-; CHECK-T1: @ %bb.0:
-; CHECK-T1-NEXT: adds r0, r0, r1
-; CHECK-T1-NEXT: movs r1, #7
-; CHECK-T1-NEXT: cmp r0, #7
-; CHECK-T1-NEXT: blt .LBB4_2
-; CHECK-T1-NEXT: @ %bb.1:
-; CHECK-T1-NEXT: {{movs|mov}} r0, r1
-; CHECK-T1-NEXT: .LBB4_2:
-; CHECK-T1-NEXT: mvns r1, r1
-; CHECK-T1-NEXT: cmp r0, r1
-; CHECK-T1-NEXT: bgt .LBB4_4
-; CHECK-T1-NEXT: @ %bb.3:
-; CHECK-T1-NEXT: {{movs|mov}} r0, r1
-; CHECK-T1-NEXT: .LBB4_4:
-; CHECK-T1-NEXT: bx lr
+; CHECK-T16-LABEL: func3:
+; CHECK-T16: @ %bb.0:
+; CHECK-T16-NEXT: adds r0, r0, r1
+; CHECK-T16-NEXT: movs r1, #7
+; CHECK-T16-NEXT: cmp r0, #7
+; CHECK-T16-NEXT: blt .LBB4_2
+; CHECK-T16-NEXT: @ %bb.1:
+; CHECK-T16-NEXT: mov r0, r1
+; CHECK-T16-NEXT: .LBB4_2:
+; CHECK-T16-NEXT: mvns r1, r1
+; CHECK-T16-NEXT: cmp r0, r1
+; CHECK-T16-NEXT: bgt .LBB4_4
+; CHECK-T16-NEXT: @ %bb.3:
+; CHECK-T16-NEXT: mov r0, r1
+; CHECK-T16-NEXT: .LBB4_4:
+; CHECK-T16-NEXT: bx lr
;
; CHECK-T2NODSP-LABEL: func3:
; CHECK-T2NODSP: @ %bb.0:
@@ -324,6 +363,23 @@ define signext i4 @func3(i4 signext %x, i4 signext %y) nounwind {
; CHECK-ARMBASEDSP-NEXT: asr r0, r0, #28
; CHECK-ARMBASEDSP-NEXT: bx lr
;
+; CHECK-T15TE-LABEL: func3:
+; CHECK-T15TE: @ %bb.0:
+; CHECK-T15TE-NEXT: adds r0, r0, r1
+; CHECK-T15TE-NEXT: movs r1, #7
+; CHECK-T15TE-NEXT: cmp r0, #7
+; CHECK-T15TE-NEXT: blt .LBB4_2
+; CHECK-T15TE-NEXT: @ %bb.1:
+; CHECK-T15TE-NEXT: movs r0, r1
+; CHECK-T15TE-NEXT: .LBB4_2:
+; CHECK-T15TE-NEXT: mvns r1, r1
+; CHECK-T15TE-NEXT: cmp r0, r1
+; CHECK-T15TE-NEXT: bgt .LBB4_4
+; CHECK-T15TE-NEXT: @ %bb.3:
+; CHECK-T15TE-NEXT: movs r0, r1
+; CHECK-T15TE-NEXT: .LBB4_4:
+; CHECK-T15TE-NEXT: bx lr
+;
; CHECK-ARMDSP-LABEL: func3:
; CHECK-ARMDSP: @ %bb.0:
; CHECK-ARMDSP-NEXT: lsl r0, r0, #28
diff --git a/llvm/test/CodeGen/ARM/select_const.ll b/llvm/test/CodeGen/ARM/select_const.ll
index df95af313eac66..b412acb8e56ec8 100644
--- a/llvm/test/CodeGen/ARM/select_const.ll
+++ b/llvm/test/CodeGen/ARM/select_const.ll
@@ -664,9 +664,8 @@ define i64 @opaque_constant1(i1 %cond, i64 %x) {
; THUMB-NEXT: push {r4, r5, r6, r7, lr}
; THUMB-NEXT: movs r7, #1
; THUMB-NEXT: ands r0, r7
+; THUMB-NEXT: movs r4, r0
; THUMB-NEXT: subs r1, r0, #1
-; THUMB-NEXT: mov r12, r0
-; THUMB-NEXT: mov r4, r12
; THUMB-NEXT: sbcs r4, r1
; THUMB-NEXT: cmp r0, #0
; THUMB-NEXT: bne .LBB24_2
@@ -680,9 +679,8 @@ define i64 @opaque_constant1(i1 %cond, i64 %x) {
; THUMB-NEXT: ldr r0, .LCPI24_0
; THUMB-NEXT: ands r5, r0
; THUMB-NEXT: movs r6, #0
+; THUMB-NEXT: movs r1, r4
; THUMB-NEXT: subs r0, r5, #1
-; THUMB-NEXT: mov r12, r4
-; THUMB-NEXT: mov r1, r12
; THUMB-NEXT: sbcs r1, r6
; THUMB-NEXT: eors r3, r7
; THUMB-NEXT: ldr r6, .LCPI24_0
@@ -785,9 +783,8 @@ define i64 @func(i64 %arg) {
; THUMB-NEXT: .save {r4, lr}
; THUMB-NEXT: push {r4, lr}
; THUMB-NEXT: movs r2, #0
+; THUMB-NEXT: movs r3, r1
; THUMB-NEXT: adds r3, r0, #1
-; THUMB-NEXT: mov r12, r1
-; THUMB-NEXT: mov r3, r12
; THUMB-NEXT: adcs r3, r2
; THUMB-NEXT: mov r12, r2
; THUMB-NEXT: mov r3, r12
diff --git a/llvm/test/CodeGen/Thumb/pr35836.ll b/llvm/test/CodeGen/Thumb/pr35836.ll
index ba33a8184bcc71..9278e61a34d3db 100644
--- a/llvm/test/CodeGen/Thumb/pr35836.ll
+++ b/llvm/test/CodeGen/Thumb/pr35836.ll
@@ -1,3 +1,4 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
; RUN: llc < %s | FileCheck %s
target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
@@ -5,6 +6,48 @@ target triple = "thumbv5e-none-linux-gnueabi"
; Function Attrs: norecurse nounwind optsize
define void @f(i32,i32,i32,i32,ptr %x4p, ptr %x5p, ptr %x6p) {
+; CHECK-LABEL: f:
+; CHECK: @ %bb.0: @ %if.end
+; CHECK-NEXT: .save {r4, r5, r6, r7, lr}
+; CHECK-NEXT: push {r4, r5, r6, r7, lr}
+; CHECK-NEXT: .pad #24
+; CHECK-NEXT: sub sp, #24
+; CHECK-NEXT: add r4, sp, #8
+; CHECK-NEXT: stm r4!, {r0, r1, r2, r3} @ 16-byte Folded Spill
+; CHECK-NEXT: movs r5, #0
+; CHECK-NEXT: ldr r0, [sp, #52]
+; CHECK-NEXT: str r0, [sp, #4] @ 4-byte Spill
+; CHECK-NEXT: ldr r0, [sp, #48]
+; CHECK-NEXT: str r0, [sp] @ 4-byte Spill
+; CHECK-NEXT: ldr r7, [sp, #44]
+; CHECK-NEXT: .LBB0_1: @ %while.body
+; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: ldr r0, [sp, #20] @ 4-byte Reload
+; CHECK-NEXT: ldr r1, [sp, #16] @ 4-byte Reload
+; CHECK-NEXT: movs r1, r5
+; CHECK-NEXT: adds r3, r0, r1
+; CHECK-NEXT: adcs r1, r5
+; CHECK-NEXT: ldr r0, [sp, #12] @ 4-byte Reload
+; CHECK-NEXT: ldr r2, [sp, #8] @ 4-byte Reload
+; CHECK-NEXT: movs r4, r5
+; CHECK-NEXT: adds r2, r0, r2
+; CHECK-NEXT: adcs r4, r5
+; CHECK-NEXT: movs r0, r3
+; CHECK-NEXT: adds r0, r2, r5
+; CHECK-NEXT: adcs r0, r4
+; CHECK-NEXT: ldr r6, [sp, #4] @ 4-byte Reload
+; CHECK-NEXT: str r0, [r6]
+; CHECK-NEXT: ldr r0, [r7]
+; CHECK-NEXT: ldr r6, [sp] @ 4-byte Reload
+; CHECK-NEXT: ldr r6, [r6]
+; CHECK-NEXT: movs r6, r5
+; CHECK-NEXT: adds r0, r6, r0
+; CHECK-NEXT: adcs r6, r5
+; CHECK-NEXT: adds r2, r2, r5
+; CHECK-NEXT: adcs r4, r3
+; CHECK-NEXT: adcs r0, r1
+; CHECK-NEXT: adcs r6, r5
+; CHECK-NEXT: b .LBB0_1
if.end:
br label %while.body
diff --git a/llvm/test/CodeGen/Thumb/urem-seteq-illegal-types.ll b/llvm/test/CodeGen/Thumb/urem-seteq-illegal-types.ll
index 61a741445b81cf..c54e4547243eb3 100644
--- a/llvm/test/CodeGen/Thumb/urem-seteq-illegal-types.ll
+++ b/llvm/test/CodeGen/Thumb/urem-seteq-illegal-types.ll
@@ -121,9 +121,8 @@ define <3 x i1> @test_urem_vec(<3 x i11> %X) nounwind {
; CHECK-NEXT: lsrs r0, r3, #1
; CHECK-NEXT: movs r3, #1
; CHECK-NEXT: movs r4, #0
+; CHECK-NEXT: movs r0, r3
; CHECK-NEXT: cmp r0, #170
-; CHECK-NEXT: mov r12, r3
-; CHECK-NEXT: mov r0, r12
; CHECK-NEXT: bhi .LBB4_2
; CHECK-NEXT: @ %bb.1:
; CHECK-NEXT: movs r0, r4
@@ -133,9 +132,8 @@ define <3 x i1> @test_urem_vec(<3 x i11> %X) nounwind {
; CHECK-NEXT: muls r5, r1, r5
; CHECK-NEXT: movs r1, #73
; CHECK-NEXT: lsls r1, r1, #23
+; CHECK-NEXT: movs r1, r3
; CHECK-NEXT: cmp r5, r1
-; CHECK-NEXT: mov r12, r3
-; CHECK-NEXT: mov r1, r12
; CHECK-NEXT: bhi .LBB4_4
; CHECK-NEXT: @ %bb.3:
; CHECK-NEXT: movs r1, r4
|
AtariDreams
force-pushed
the
thumbs
branch
4 times, most recently
from
b856c4c
to
d24c1b2
Compare
…ids emitting an extra instruction When the ARM backend decides it is best to copy a value from one register to another, it does not take THUMB limitations into account, which means we have to find a way to do a transfer AFTER it has been decided that a mov would work, sometimes using movs, and other times using two instructions, which was not what was intended by the compiler backend. Until we find a better solution, this is unfortunately the best I could do unless we can somehow find a more proper way that involves updating the cost model to take the possibility of pushing through stack or a high register into account.
…ids emitting an extra instruction When the ARM backend decides it is best to copy a value from one register to another, it does not take THUMB limitations into account, which means we have to find a way to do a transfer AFTER it has been decided that a mov would work, sometimes using movs, and other times using two instructions, which was not what was inteded by the compiler backend. Until we find a better solution, this is unfortunately the best I could do unless we can somehow find a more proper way that involves updating the cost model to take the possibility of pushing through stack or a high register into account.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
When the ARM backend decides it is best to copy a value from one register to another, it does not take THUMB limitations into account, which means we have to find a way to do a transfer AFTER it has been decided that a mov would work, sometimes using movs, and other times using two instructions, which was not what was intended by the compiler backend.
Until we find a better solution, this is unfortunately the best I could do unless we can somehow find a more proper way that involves updating the cost model to take the possibility of pushing through stack or a high register into account.