RegAlloc: Use new approach to handling failed allocations #128469
+52
−91
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
This fixes an assert after allocation failure. Rather than collecting failed virtual registers and hacking on the uses after the fact, directly hack on the uses and rewrite the registers to the dummy assignment immediately. Previously we were bypassing LiveRegMatrix and directly assigning in the VirtRegMap. This resulted in inconsistencies where illegal overlapping assignments were missing. Rather than try to hack in some system to manage these in LiveRegMatrix (i.e. hacking around cases with invalid iterators), avoid this by directly using the physreg. This should also allow removal of special casing in virtregrewriter for failed allocations.
This was referenced Feb 24, 2025
|
Warning This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
This stack of pull requests is managed by Graphite. Learn more about stacking. |
|
@llvm/pr-subscribers-llvm-regalloc @llvm/pr-subscribers-backend-amdgpu Author: Matt Arsenault (arsenm) ChangesThis fixes an assert after allocation failure. Rather than collecting failed virtual registers and hacking Previously we were bypassing LiveRegMatrix and directly assigning Full diff: https://github.com/llvm/llvm-project/pull/128469.diff 10 Files Affected:
diff --git a/llvm/lib/CodeGen/RegAllocBase.cpp b/llvm/lib/CodeGen/RegAllocBase.cpp
index d66d396a15018..b9599e77fbf0a 100644
--- a/llvm/lib/CodeGen/RegAllocBase.cpp
+++ b/llvm/lib/CodeGen/RegAllocBase.cpp
@@ -128,8 +128,7 @@ void RegAllocBase::allocatePhysRegs() {
AvailablePhysReg = getErrorAssignment(*RC, MI);
// Keep going after reporting the error.
- VRM->assignVirt2Phys(VirtReg->reg(), AvailablePhysReg);
- FailedVRegs.insert(VirtReg->reg());
+ cleanupFailedVReg(VirtReg->reg(), AvailablePhysReg, SplitVRegs);
} else if (AvailablePhysReg)
Matrix->assign(*VirtReg, AvailablePhysReg);
@@ -163,58 +162,35 @@ void RegAllocBase::postOptimization() {
DeadRemats.clear();
}
-void RegAllocBase::cleanupFailedVRegs() {
- SmallSet<Register, 8> JunkRegs;
-
- for (Register FailedReg : FailedVRegs) {
- JunkRegs.insert(FailedReg);
-
- MCRegister PhysReg = VRM->getPhys(FailedReg);
- LiveInterval &FailedInterval = LIS->getInterval(FailedReg);
-
- // The liveness information for the failed register and anything interfering
- // with the physical register we arbitrarily chose is junk and needs to be
- // deleted.
- for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
- LiveIntervalUnion::Query &Q = Matrix->query(FailedInterval, *Units);
- for (const LiveInterval *InterferingReg : Q.interferingVRegs())
- JunkRegs.insert(InterferingReg->reg());
- LIS->removeRegUnit(*Units);
- }
+void RegAllocBase::cleanupFailedVReg(Register FailedReg, MCRegister PhysReg,
+ SmallVectorImpl<Register> &SplitRegs) {
+ // We still should produce valid IR. Kill all the uses and reduce the live
+ // ranges so that we don't think it's possible to introduce kill flags later
+ // which will fail the verifier.
+ for (MachineOperand &MO : MRI->reg_operands(FailedReg)) {
+ if (MO.readsReg())
+ MO.setIsUndef(true);
}
- for (Register JunkReg : JunkRegs) {
- MCRegister PhysReg = VRM->getPhys(JunkReg);
- // We still should produce valid IR. Kill all the uses and reduce the live
- // ranges so that we don't think it's possible to introduce kill flags
- // later which will fail the verifier.
- for (MachineOperand &MO : MRI->reg_operands(JunkReg)) {
- if (MO.readsReg())
- MO.setIsUndef(true);
- }
-
- // The liveness of the assigned physical register is also now unreliable.
+ if (!MRI->isReserved(PhysReg)) {
+ // Physical liveness for any aliasing registers is now unreliable, so delete
+ // the uses.
for (MCRegAliasIterator Aliases(PhysReg, TRI, true); Aliases.isValid();
++Aliases) {
for (MachineOperand &MO : MRI->reg_operands(*Aliases)) {
- if (MO.readsReg())
+ if (MO.readsReg()) {
MO.setIsUndef(true);
- }
- }
-
- LiveInterval &JunkLI = LIS->getInterval(JunkReg);
- if (LIS->shrinkToUses(&JunkLI)) {
- SmallVector<LiveInterval *, 8> SplitLIs;
- LIS->splitSeparateComponents(JunkLI, SplitLIs);
-
- VRM->grow();
- Register Original = VRM->getOriginal(JunkReg);
- for (LiveInterval *SplitLI : SplitLIs) {
- VRM->setIsSplitFromReg(SplitLI->reg(), Original);
- VRM->assignVirt2Phys(SplitLI->reg(), PhysReg);
+ LIS->removeAllRegUnitsForPhysReg(MO.getReg());
+ }
}
}
}
+
+ // Directly perform the rewrite, and do not leave it to VirtRegRewriter as
+ // usual. This avoids trying to manage illegal overlapping assignments in
+ // LiveRegMatrix.
+ MRI->replaceRegWith(FailedReg, PhysReg);
+ LIS->removeInterval(FailedReg);
}
void RegAllocBase::enqueue(const LiveInterval *LI) {
diff --git a/llvm/lib/CodeGen/RegAllocBase.h b/llvm/lib/CodeGen/RegAllocBase.h
index 1fdbab694bb0e..f1b5af8cd4d74 100644
--- a/llvm/lib/CodeGen/RegAllocBase.h
+++ b/llvm/lib/CodeGen/RegAllocBase.h
@@ -108,7 +108,8 @@ class RegAllocBase {
/// Perform cleanups on registers that failed to allocate. This hacks on the
/// liveness in order to avoid spurious verifier errors in later passes.
- void cleanupFailedVRegs();
+ void cleanupFailedVReg(Register FailedVReg, MCRegister PhysReg,
+ SmallVectorImpl<Register> &SplitRegs);
// Get a temporary reference to a Spiller instance.
virtual Spiller &spiller() = 0;
diff --git a/llvm/lib/CodeGen/RegAllocBasic.cpp b/llvm/lib/CodeGen/RegAllocBasic.cpp
index d240bf916ac05..51e047b2fa3f0 100644
--- a/llvm/lib/CodeGen/RegAllocBasic.cpp
+++ b/llvm/lib/CodeGen/RegAllocBasic.cpp
@@ -329,7 +329,6 @@ bool RABasic::runOnMachineFunction(MachineFunction &mf) {
allocatePhysRegs();
postOptimization();
- cleanupFailedVRegs();
// Diagnostic output before rewriting
LLVM_DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << *VRM << "\n");
diff --git a/llvm/lib/CodeGen/RegAllocGreedy.cpp b/llvm/lib/CodeGen/RegAllocGreedy.cpp
index 74a363b5bc4b5..f1e734dd2f933 100644
--- a/llvm/lib/CodeGen/RegAllocGreedy.cpp
+++ b/llvm/lib/CodeGen/RegAllocGreedy.cpp
@@ -2795,7 +2795,6 @@ bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
if (VerifyEnabled)
MF->verify(this, "Before post optimization", &errs());
postOptimization();
- cleanupFailedVRegs();
reportStats();
releaseMemory();
diff --git a/llvm/lib/CodeGen/VirtRegMap.cpp b/llvm/lib/CodeGen/VirtRegMap.cpp
index 4e4c89eba7aa9..81a82e898a901 100644
--- a/llvm/lib/CodeGen/VirtRegMap.cpp
+++ b/llvm/lib/CodeGen/VirtRegMap.cpp
@@ -620,8 +620,7 @@ void VirtRegRewriter::rewrite() {
assert(Register(PhysReg).isPhysical());
RewriteRegs.insert(PhysReg);
- assert((!MRI->isReserved(PhysReg) || !IsValidAlloc) &&
- "Reserved register assignment");
+ assert(!MRI->isReserved(PhysReg) && "Reserved register assignment");
// Preserve semantics of sub-register operands.
unsigned SubReg = MO.getSubReg();
diff --git a/llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers-assertion-after-ra-failure.xfail.ll b/llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers-assertion-after-ra-failure.ll
similarity index 75%
rename from llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers-assertion-after-ra-failure.xfail.ll
rename to llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers-assertion-after-ra-failure.ll
index 5f2e9af378f08..f3eb7a42cb823 100644
--- a/llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers-assertion-after-ra-failure.xfail.ll
+++ b/llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers-assertion-after-ra-failure.ll
@@ -1,8 +1,7 @@
-; REQUIRES: asserts
-; RUN: not --crash llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -filetype=null %s 2>&1 | FileCheck -check-prefix=CRASH %s
+; RUN: not llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -filetype=null %s 2>&1 | FileCheck -check-prefix=ERR %s
+; RUN: not llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -vgpr-regalloc=basic -filetype=null %s 2>&1 | FileCheck -check-prefix=ERR %s
-; CRASH: error: <unknown>:0:0: no registers from class available to allocate in function 'no_free_vgprs_at_agpr_to_agpr_copy'
-; CRASH: Cannot access invalid iterator
+; ERR: error: <unknown>:0:0: no registers from class available to allocate in function 'no_free_vgprs_at_agpr_to_agpr_copy'
define void @no_free_vgprs_at_agpr_to_agpr_copy(float %v0, float %v1) #0 {
%asm = call { <32 x i32>, <16 x float> } asm sideeffect "; def $0 $1", "=${v[0:31]},=${a[0:15]}"()
diff --git a/llvm/test/CodeGen/AMDGPU/illegal-eviction-assert.mir b/llvm/test/CodeGen/AMDGPU/illegal-eviction-assert.mir
index e583b168c15f7..e32fa90ead4f4 100644
--- a/llvm/test/CodeGen/AMDGPU/illegal-eviction-assert.mir
+++ b/llvm/test/CodeGen/AMDGPU/illegal-eviction-assert.mir
@@ -17,8 +17,8 @@
...
-# CHECK: S_NOP 0, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19, implicit-def $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27, implicit-def dead $vgpr0_vgpr1_vgpr2_vgpr3, implicit-def $vgpr28_vgpr29_vgpr30_vgpr31, implicit-def dead $vgpr0_vgpr1_vgpr2_vgpr3
-# CHECK: S_NOP 0, implicit killed undef $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19, implicit killed undef $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27, implicit undef $vgpr0_vgpr1_vgpr2_vgpr3, implicit killed undef $vgpr28_vgpr29_vgpr30_vgpr31, implicit undef $vgpr0_vgpr1_vgpr2_vgpr3
+# CHECK: S_NOP 0, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19, implicit-def $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit-def $vgpr28_vgpr29_vgpr30_vgpr31, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3
+# CHECK: S_NOP 0, implicit killed undef $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19, implicit killed undef $vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27, implicit undef $vgpr0_vgpr1_vgpr2_vgpr3, implicit killed undef $vgpr28_vgpr29_vgpr30_vgpr31, implicit killed undef $vgpr0_vgpr1_vgpr2_vgpr3
---
name: foo
diff --git a/llvm/test/CodeGen/AMDGPU/inflated-reg-class-snippet-copy-use-after-free.mir b/llvm/test/CodeGen/AMDGPU/inflated-reg-class-snippet-copy-use-after-free.mir
index 6d63a4a1cc0ab..c1e0d0716acae 100644
--- a/llvm/test/CodeGen/AMDGPU/inflated-reg-class-snippet-copy-use-after-free.mir
+++ b/llvm/test/CodeGen/AMDGPU/inflated-reg-class-snippet-copy-use-after-free.mir
@@ -27,10 +27,10 @@
# CHECK-LABEL: name: inflated_reg_class_copy_use_after_free
# CHECK: S_NOP 0, implicit-def [[ORIG_REG:%[0-9]+]].sub0_sub1_sub2_sub3
# CHECK-NEXT: SI_SPILL_AV512_SAVE [[ORIG_REG]], %stack.0, $sgpr32, 0, implicit $exec :: (store (s512) into %stack.0, align 4, addrspace 5)
-# CHECK-NEXT: dead [[RESTORE0:%[0-9]+]]:vreg_512_align2 = SI_SPILL_V512_RESTORE %stack.0, $sgpr32, 0, implicit $exec :: (load (s512) from %stack.0, align 4, addrspace 5)
-# CHECK-NEXT: dead early-clobber [[MFMA0:%[0-9]+]]:vreg_512_align2 = V_MFMA_F32_16X16X1F32_vgprcd_e64 undef %3:vgpr_32, undef %3:vgpr_32, undef [[RESTORE0]], 0, 0, 0, implicit $mode, implicit $exec, implicit $mode, implicit $exec
-# CHECK-NEXT: undef [[SPLIT0:%[0-9]+]].sub2_sub3:av_512_align2 = COPY undef [[MFMA0]].sub2_sub3 {
-# CHECK-NEXT: internal [[SPLIT0]].sub0:av_512_align2 = COPY undef [[MFMA0]].sub0
+# CHECK-NEXT: [[RESTORE0:%[0-9]+]]:vreg_512_align2 = SI_SPILL_V512_RESTORE %stack.0, $sgpr32, 0, implicit $exec :: (load (s512) from %stack.0, align 4, addrspace 5)
+# CHECK-NEXT: early-clobber $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = V_MFMA_F32_16X16X1F32_vgprcd_e64 undef %3:vgpr_32, undef %3:vgpr_32, [[RESTORE0]], 0, 0, 0, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+# CHECK-NEXT: undef [[SPLIT0:%[0-9]+]].sub2_sub3:av_512_align2 = COPY undef $vgpr2_vgpr3 {
+# CHECK-NEXT: internal [[SPLIT0]].sub0:av_512_align2 = COPY undef $vgpr0
# CHECK-NEXT: }
# CHECK-NEXT: undef [[SPLIT1:%[0-9]+]].sub2_sub3:av_512_align2 = COPY [[SPLIT0]].sub2_sub3 {
# CHECK-NEXT: internal [[SPLIT1]].sub0:av_512_align2 = COPY [[SPLIT0]].sub0
@@ -118,10 +118,10 @@ body: |
# CHECK-LABEL: name: inflated_reg_class_copy_use_after_free_lane_subset
# CHECK: S_NOP 0, implicit-def [[ORIG_REG:%[0-9]+]].sub0_sub1_sub2_sub3
# CHECK-NEXT: SI_SPILL_AV512_SAVE [[ORIG_REG]], %stack.0, $sgpr32, 0, implicit $exec :: (store (s512) into %stack.0, align 4, addrspace 5)
-# CHECK-NEXT: dead [[RESTORE_0:%[0-9]+]]:av_512_align2 = SI_SPILL_AV512_RESTORE %stack.0, $sgpr32, 0, implicit $exec :: (load (s512) from %stack.0, align 4, addrspace 5)
-# CHECK-NEXT: S_NOP 0, implicit-def dead early-clobber [[REG1:%[0-9]+]], implicit undef [[RESTORE_0]].sub0_sub1_sub2_sub3, implicit undef [[RESTORE_0]].sub4_sub5_sub6_sub7
-# CHECK-NEXT: undef [[SPLIT0:%[0-9]+]].sub2_sub3:av_512_align2 = COPY undef [[REG1]].sub2_sub3 {
-# CHECK-NEXT: internal [[SPLIT0]].sub0:av_512_align2 = COPY undef [[REG1]].sub0
+# CHECK-NEXT: [[RESTORE_0:%[0-9]+]]:av_512_align2 = SI_SPILL_AV512_RESTORE %stack.0, $sgpr32, 0, implicit $exec :: (load (s512) from %stack.0, align 4, addrspace 5)
+# CHECK-NEXT: S_NOP 0, implicit-def early-clobber $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit [[RESTORE_0]].sub0_sub1_sub2_sub3, implicit [[RESTORE_0]].sub4_sub5_sub6_sub7
+# CHECK-NEXT: undef [[SPLIT0:%[0-9]+]].sub2_sub3:av_512_align2 = COPY undef $vgpr2_vgpr3 {
+# CHECK-NEXT: internal [[SPLIT0]].sub0:av_512_align2 = COPY undef $vgpr0
# CHECK-NEXT: }
# CHECK-NEXT: undef [[SPLIT1:%[0-9]+]].sub2_sub3:av_512_align2 = COPY [[SPLIT0]].sub2_sub3 {
# CHECK-NEXT: internal [[SPLIT1]].sub0:av_512_align2 = COPY [[SPLIT0]].sub0
diff --git a/llvm/test/CodeGen/AMDGPU/issue48473.mir b/llvm/test/CodeGen/AMDGPU/issue48473.mir
index c5cb1445adca0..507b299e79a86 100644
--- a/llvm/test/CodeGen/AMDGPU/issue48473.mir
+++ b/llvm/test/CodeGen/AMDGPU/issue48473.mir
@@ -43,7 +43,7 @@
# %25 to $sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
# CHECK-LABEL: name: issue48473
-# CHECK: S_NOP 0, implicit undef $sgpr0_sgpr1_sgpr2_sgpr3, implicit killed undef $sgpr12_sgpr13_sgpr14_sgpr15, implicit killed undef $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23, implicit killed undef $sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31, implicit killed undef $sgpr84_sgpr85_sgpr86_sgpr87, implicit killed undef $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43, implicit killed undef $sgpr4_sgpr5_sgpr6_sgpr7, implicit killed undef $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51, implicit killed undef $sgpr88_sgpr89_sgpr90_sgpr91, implicit killed undef $sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83, implicit undef $sgpr0_sgpr1_sgpr2_sgpr3, implicit killed undef $sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59, implicit killed undef $sgpr92_sgpr93_sgpr94_sgpr95, implicit killed undef $sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75, implicit undef $sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75, implicit killed undef $sgpr96_sgpr97_sgpr98_sgpr99, implicit killed undef $sgpr8_sgpr9_sgpr10_sgpr11, implicit killed undef $sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
+# CHECK: S_NOP 0, implicit killed undef $sgpr0_sgpr1_sgpr2_sgpr3, implicit killed undef $sgpr12_sgpr13_sgpr14_sgpr15, implicit killed undef $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23, implicit killed undef $sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31, implicit killed undef $sgpr84_sgpr85_sgpr86_sgpr87, implicit killed undef $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43, implicit killed undef $sgpr4_sgpr5_sgpr6_sgpr7, implicit killed undef $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51, implicit killed undef $sgpr88_sgpr89_sgpr90_sgpr91, implicit killed undef $sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83, implicit undef $sgpr0_sgpr1_sgpr2_sgpr3, implicit killed undef $sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59, implicit killed undef $sgpr92_sgpr93_sgpr94_sgpr95, implicit killed undef $sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75, implicit undef $sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75, implicit killed undef $sgpr96_sgpr97_sgpr98_sgpr99, implicit killed undef $sgpr8_sgpr9_sgpr10_sgpr11, implicit killed undef $sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
---
name: issue48473
diff --git a/llvm/test/CodeGen/AMDGPU/register-killed-error-after-alloc-failure0.mir b/llvm/test/CodeGen/AMDGPU/register-killed-error-after-alloc-failure0.mir
index e2a839ea9d24e..306788b0bf5ed 100644
--- a/llvm/test/CodeGen/AMDGPU/register-killed-error-after-alloc-failure0.mir
+++ b/llvm/test/CodeGen/AMDGPU/register-killed-error-after-alloc-failure0.mir
@@ -14,23 +14,23 @@
# ERR: error: <unknown>:0:0: ran out of registers during register allocation
# GREEDY: SI_SPILL_V256_SAVE undef $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-# GREEDY-NEXT: SI_SPILL_V512_SAVE undef $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19
-# GREEDY-NEXT: SI_SPILL_V128_SAVE undef $vgpr0_vgpr1_vgpr2_vgpr3
+# GREEDY-NEXT: SI_SPILL_V512_SAVE killed undef $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19, %stack.1, $sgpr32, 0, implicit $exec :: (store (s512) into %stack.1, align 4, addrspace 5)
+# GREEDY-NEXT: SI_SPILL_V128_SAVE killed undef $vgpr0_vgpr1_vgpr2_vgpr3
-# GREEDY: dead $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19 = SI_SPILL_V512_RESTORE
-# GREEDY: dead $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = SI_SPILL_V256_RESTORE
-# GREEDY: S_NOP 0, implicit undef $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19, implicit undef $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, implicit undef $vgpr0_vgpr1_vgpr2_vgpr3
+# GREEDY: $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19 = SI_SPILL_V512_RESTORE
+# GREEDY: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = SI_SPILL_V256_RESTORE
+# GREEDY: S_NOP 0, implicit killed undef $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19, implicit undef $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, implicit killed undef $vgpr0_vgpr1_vgpr2_vgpr3
# GREEDY: S_NOP 0, implicit killed undef $vgpr20_vgpr21
# BASIC: SI_SPILL_V128_SAVE undef $vgpr0_vgpr1_vgpr2_vgpr3
# BASIC: SI_SPILL_V256_SAVE killed undef $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23
-# BASIC: SI_SPILL_V512_SAVE undef $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+# BASIC: SI_SPILL_V512_SAVE killed undef $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, %stack.0, $sgpr32, 0, implicit $exec :: (store (s512) into %stack.0, align 4, addrspace 5)
# BASIC: SI_SPILL_V64_SAVE killed undef $vgpr0_vgpr1, %stack.{{[0-9]+}}, $sgpr32, 0, implicit $exec :: (store (s64) into %stack.{{[0-9]+}}, align 4, addrspace 5)
-# BASIC: dead $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = SI_SPILL_V512_RESTORE
+# BASIC: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = SI_SPILL_V512_RESTORE %stack.0, $sgpr32, 0, implicit $exec :: (load (s512) from %stack.0, align 4, addrspace 5)
# BASIC: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23 = SI_SPILL_V256_RESTORE
-# BASIC: dead $vgpr0_vgpr1_vgpr2_vgpr3 = SI_SPILL_V128_RESTORE
-# BASIC: S_NOP 0, implicit undef $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit killed undef $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23, implicit undef $vgpr0_vgpr1_vgpr2_vgpr3
+# BASIC: $vgpr0_vgpr1_vgpr2_vgpr3 = SI_SPILL_V128_RESTORE
+# BASIC: S_NOP 0, implicit killed undef $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit killed undef $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23, implicit undef $vgpr0_vgpr1_vgpr2_vgpr3
# BASIC: $vgpr0_vgpr1 = SI_SPILL_V64_RESTORE
--- |
|
We could now delete FailedRegAlloc since the special case handling is now local to the allocation passes. We're still using it to avoid repeated errors, so that would need some new local state.
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.
This fixes an assert after allocation failure.
Rather than collecting failed virtual registers and hacking
on the uses after the fact, directly hack on the uses and rewrite
the registers to the dummy assignment immediately.
Previously we were bypassing LiveRegMatrix and directly assigning
in the VirtRegMap. This resulted in inconsistencies where illegal
overlapping assignments were missing. Rather than try to hack in
some system to manage these in LiveRegMatrix (i.e. hacking around
cases with invalid iterators), avoid this by directly using the
physreg. This should also allow removal of special casing in
virtregrewriter for failed allocations.