[MCP] Enhance MCP copy Instruction removal for special case(reapply) #74239
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
In pr 70778, when we clobbering a register (Def), we also locate and remove the entry in the Copy Maps that indicates Src defined Def. If the "DefRegs" record is empty after deletion, we directly remove the corresponding SrcCopy entry from the Copy Maps. However, since the SrcCopy may also have documented other Def' records, directly removing it is unsafe. We need to enhance the criteria for removing empty records to ensure safety.
|
@llvm/pr-subscribers-llvm-regalloc @llvm/pr-subscribers-backend-x86 Author: Vettel (LWenH) ChangesThis patch try to reimplement #70778, and fix issue: #73512. In #70778 , when we clobbering a register (Def), we will also locate and remove the entry in the Copy Maps that indicates Src defined Def. If the "DefRegs" is being empty after removal, we will directly remove the corresponding SrcCopy entry from the Copy Maps. However, since the SrcCopy may also have referred the Src as the destination of another copy instruction, directly removing it is unsafe. We need to enhance the criteria for removing empty records to ensure safety. Patch is 26.60 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/74239.diff 13 Files Affected:
diff --git a/llvm/lib/CodeGen/MachineCopyPropagation.cpp b/llvm/lib/CodeGen/MachineCopyPropagation.cpp
index a032b31a1fc7c..51e944d0279f2 100644
--- a/llvm/lib/CodeGen/MachineCopyPropagation.cpp
+++ b/llvm/lib/CodeGen/MachineCopyPropagation.cpp
@@ -175,8 +175,46 @@ class CopyTracker {
if (MachineInstr *MI = I->second.MI) {
std::optional<DestSourcePair> CopyOperands =
isCopyInstr(*MI, TII, UseCopyInstr);
- markRegsUnavailable({CopyOperands->Destination->getReg().asMCReg()},
- TRI);
+
+ MCRegister Def = CopyOperands->Destination->getReg().asMCReg();
+ MCRegister Src = CopyOperands->Source->getReg().asMCReg();
+
+ markRegsUnavailable(Def, TRI);
+
+ // Since we clobber the destination of a copy, the semantic of Src's
+ // "DefRegs" to contain Def is no longer effectual. We will also need
+ // to remove the record from the copy maps that indicates Src defined
+ // Def. Failing to do so might cause the target to miss some
+ // opportunities to further eliminate redundant copy instructions.
+ // Consider the following sequence during the
+ // ForwardCopyPropagateBlock procedure:
+ // L1: r0 = COPY r9 <- TrackMI
+ // L2: r0 = COPY r8 <- TrackMI (Remove r9 defined r0 from tracker)
+ // L3: use r0 <- Remove L2 from MaybeDeadCopies
+ // L4: early-clobber r9 <- Clobber r9 (L2 is still valid in tracker)
+ // L5: r0 = COPY r8 <- Remove NopCopy
+ for (MCRegUnit SrcUnit : TRI.regunits(Src)) {
+ auto SrcCopy = Copies.find(SrcUnit);
+ if (SrcCopy != Copies.end() && SrcCopy->second.LastSeenUseInCopy) {
+ // If SrcCopy defines multiple values, we only need
+ // to erase the record for Def in DefRegs.
+ for (auto itr = SrcCopy->second.DefRegs.begin();
+ itr != SrcCopy->second.DefRegs.end(); itr++) {
+ if (*itr == Def) {
+ SrcCopy->second.DefRegs.erase(itr);
+ // If DefReg becomes empty after removal, we can remove the
+ // SrcCopy from the tracker's copy maps. We only remove those
+ // entries solely record the Def is defined by Src. If an
+ // entry also contains the definition record of other Def'
+ // registers, it cannot be cleared.
+ if (SrcCopy->second.DefRegs.empty() && !SrcCopy->second.MI) {
+ Copies.erase(SrcCopy);
+ }
+ break;
+ }
+ }
+ }
+ }
}
// Now we can erase the copy.
Copies.erase(I);
diff --git a/llvm/test/CodeGen/AMDGPU/mcp-implicit-clobber.mir b/llvm/test/CodeGen/AMDGPU/mcp-implicit-clobber.mir
new file mode 100644
index 0000000000000..6e613243e38c5
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/mcp-implicit-clobber.mir
@@ -0,0 +1,26 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 4
+# RUN: llc -march=amdgcn -mcpu=gfx900 %s -o - -run-pass machine-cp -verify-machineinstrs | FileCheck %s
+
+# The MachineCopyPropagation Pass should not treat the subsequent
+# instruction "$sgpr2_sgpr3 = COPY $sgpr6_sgpr7" as a NopCopy.
+# For detailed information, please refer to issue 73512.
+---
+name: foo
+body: |
+ bb.0.entry:
+ liveins: $sgpr4_sgpr5, $sgpr6_sgpr7
+
+ ; CHECK-LABEL: name: foo
+ ; CHECK: liveins: $sgpr4_sgpr5, $sgpr6_sgpr7
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: $sgpr2_sgpr3 = COPY $sgpr6_sgpr7
+ ; CHECK-NEXT: S_NOP 0, implicit-def $sgpr0
+ ; CHECK-NEXT: $sgpr2_sgpr3 = COPY $sgpr6_sgpr7
+ ; CHECK-NEXT: S_NOP 0, implicit $sgpr2_sgpr3
+ $sgpr2_sgpr3 = COPY $sgpr6_sgpr7
+ $sgpr0 = COPY $sgpr3
+ S_NOP 0, implicit-def $sgpr0
+ $sgpr3 = COPY killed $sgpr5
+ $sgpr2_sgpr3 = COPY $sgpr6_sgpr7
+ S_NOP 0, implicit $sgpr2_sgpr3
+...
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-nearbyint-vp.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-nearbyint-vp.ll
index d9958f4aae350..5407eadb160bd 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-nearbyint-vp.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-nearbyint-vp.ll
@@ -637,7 +637,6 @@ define <32 x double> @vp_nearbyint_v32f64(<32 x double> %va, <32 x i1> %m, i32 z
; CHECK-NEXT: vl8r.v v24, (a0) # Unknown-size Folded Reload
; CHECK-NEXT: vfabs.v v16, v24, v0.t
; CHECK-NEXT: vsetvli zero, zero, e64, m8, ta, mu
-; CHECK-NEXT: vmv1r.v v0, v1
; CHECK-NEXT: vmflt.vf v1, v16, fa5, v0.t
; CHECK-NEXT: frflags a0
; CHECK-NEXT: vsetvli zero, zero, e64, m8, ta, ma
diff --git a/llvm/test/CodeGen/X86/shift-i128.ll b/llvm/test/CodeGen/X86/shift-i128.ll
index 1fe8d834dbcdd..4fbe05cd1b2f2 100644
--- a/llvm/test/CodeGen/X86/shift-i128.ll
+++ b/llvm/test/CodeGen/X86/shift-i128.ll
@@ -347,7 +347,6 @@ define void @test_lshr_v2i128(<2 x i128> %x, <2 x i128> %a, ptr nocapture %r) no
; i686-NEXT: movl %edx, %ecx
; i686-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
; i686-NEXT: shrdl %cl, %eax, (%esp) # 4-byte Folded Spill
-; i686-NEXT: movl %edx, %ecx
; i686-NEXT: shrl %cl, %esi
; i686-NEXT: movl {{[0-9]+}}(%esp), %ecx
; i686-NEXT: movl %esi, 28(%ecx)
@@ -489,7 +488,6 @@ define void @test_ashr_v2i128(<2 x i128> %x, <2 x i128> %a, ptr nocapture %r) no
; i686-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
; i686-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
; i686-NEXT: shrdl %cl, %esi, %ebx
-; i686-NEXT: movl %edx, %ecx
; i686-NEXT: sarl %cl, %ebp
; i686-NEXT: movl {{[0-9]+}}(%esp), %ecx
; i686-NEXT: movl %ebp, 28(%ecx)
@@ -623,11 +621,9 @@ define void @test_shl_v2i128(<2 x i128> %x, <2 x i128> %a, ptr nocapture %r) nou
; i686-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
; i686-NEXT: shll %cl, %edi
; i686-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
-; i686-NEXT: movl %ecx, %edi
; i686-NEXT: shldl %cl, %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
; i686-NEXT: negl %ebp
; i686-NEXT: movl 64(%esp,%ebp), %esi
-; i686-NEXT: movl %edi, %ecx
; i686-NEXT: # kill: def $cl killed $cl killed $ecx
; i686-NEXT: movl (%esp), %edi # 4-byte Reload
; i686-NEXT: shldl %cl, %edi, %esi
diff --git a/llvm/test/CodeGen/X86/shift-i256.ll b/llvm/test/CodeGen/X86/shift-i256.ll
index 0e4e706669300..e1466aebf4225 100644
--- a/llvm/test/CodeGen/X86/shift-i256.ll
+++ b/llvm/test/CodeGen/X86/shift-i256.ll
@@ -78,7 +78,6 @@ define void @shift1(i256 %x, i256 %a, ptr nocapture %r) nounwind readnone {
; CHECK-NEXT: movl %eax, %ecx
; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
; CHECK-NEXT: shrdl %cl, %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
-; CHECK-NEXT: movl %eax, %ecx
; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
; CHECK-NEXT: shrdl %cl, %edx, (%esp) # 4-byte Folded Spill
; CHECK-NEXT: movl 28(%esp,%ebp), %edx
diff --git a/llvm/test/CodeGen/X86/smulo-128-legalisation-lowering.ll b/llvm/test/CodeGen/X86/smulo-128-legalisation-lowering.ll
index abab313f4b12e..b2b5bcc5b44b2 100644
--- a/llvm/test/CodeGen/X86/smulo-128-legalisation-lowering.ll
+++ b/llvm/test/CodeGen/X86/smulo-128-legalisation-lowering.ll
@@ -1201,7 +1201,7 @@ define zeroext i1 @smuloi256(i256 %v1, i256 %v2, ptr %res) {
; X86-NEXT: movl %edx, %ebp
; X86-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) ## 4-byte Spill
; X86-NEXT: movl %eax, %ebx
-; X86-NEXT: addl %ebp, %ebx
+; X86-NEXT: addl %edx, %ebx
; X86-NEXT: adcl $0, %ebp
; X86-NEXT: movl %ecx, %eax
; X86-NEXT: movl %ecx, %esi
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-7.ll
index 893bff29b21e4..74347f3881aa9 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-7.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-7.ll
@@ -14828,7 +14828,7 @@ define void @load_i16_stride7_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm13[0],xmm0[1],xmm13[2,3,4,5],xmm0[6],xmm13[7]
; AVX512DQ-SLOW-NEXT: vmovdqa %ymm12, %ymm7
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm13 = ymm12[0,1],ymm8[2,3],ymm12[4,5],ymm8[6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %ymm16, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm12[0,1,0,1]
; AVX512DQ-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1,2],ymm14[3],ymm13[4,5,6,7,8,9,10],ymm14[11],ymm13[12,13,14,15]
@@ -15583,7 +15583,7 @@ define void @load_i16_stride7_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX512DQ-FAST-NEXT: vextracti128 $1, %ymm0, %xmm12
; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm12[0],xmm0[1],xmm12[2,3,4,5],xmm0[6],xmm12[7]
; AVX512DQ-FAST-NEXT: vpblendd {{.*#+}} ymm12 = ymm4[0],ymm15[1],ymm4[2,3],ymm15[4],ymm4[5,6,7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm4, %ymm17
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm16, %ymm17
; AVX512DQ-FAST-NEXT: vmovdqa %ymm15, %ymm13
; AVX512DQ-FAST-NEXT: vextracti128 $1, %ymm12, %xmm14
; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm12 = xmm12[0],xmm14[1],xmm12[2,3,4,5],xmm14[6],xmm12[7]
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i64-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i64-stride-7.ll
index 7d9c056716cee..0db749af9a196 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i64-stride-7.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i64-stride-7.ll
@@ -4196,7 +4196,7 @@ define void @load_i64_stride7_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, pt
; AVX512BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpermt2q %zmm1, %zmm4, %zmm10
; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm9
-; AVX512BW-NEXT: vpermt2q %zmm8, %zmm29, %zmm9
+; AVX512BW-NEXT: vpermt2q %zmm5, %zmm29, %zmm9
; AVX512BW-NEXT: vpermt2q %zmm0, %zmm4, %zmm8
; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm1
; AVX512BW-NEXT: vpermt2q %zmm12, %zmm7, %zmm3
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-5.ll
index 4d5c8fe891e81..7947dd0ff373c 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-5.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-5.ll
@@ -2476,7 +2476,6 @@ define void @load_i8_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,7,7,7]
; SSE-NEXT: psllq $48, %xmm0
; SSE-NEXT: packuswb %xmm1, %xmm0
-; SSE-NEXT: movdqa %xmm7, %xmm4
; SSE-NEXT: movdqa %xmm7, %xmm1
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT: pandn %xmm5, %xmm1
@@ -2533,7 +2532,7 @@ define void @load_i8_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; SSE-NEXT: pandn %xmm1, %xmm2
; SSE-NEXT: movdqa %xmm8, %xmm1
; SSE-NEXT: pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
-; SSE-NEXT: movdqa %xmm4, %xmm0
+; SSE-NEXT: movdqa %xmm7, %xmm0
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT: pandn %xmm4, %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-7.ll
index 0e3f8121db1de..cf1b0f2bc9860 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-7.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-7.ll
@@ -1024,8 +1024,8 @@ define void @load_i8_stride7_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; SSE-NEXT: movdqa {{.*#+}} xmm14 = [65535,0,65535,65535,0,65535,65535,65535]
; SSE-NEXT: movdqa %xmm9, %xmm7
; SSE-NEXT: pand %xmm14, %xmm7
-; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; SSE-NEXT: movdqa %xmm6, %xmm15
+; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; SSE-NEXT: movdqa %xmm5, %xmm15
; SSE-NEXT: pand %xmm14, %xmm15
; SSE-NEXT: movdqa %xmm11, %xmm3
; SSE-NEXT: pandn %xmm8, %xmm3
@@ -2148,7 +2148,6 @@ define void @load_i8_stride7_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; SSE-NEXT: movdqa %xmm5, %xmm9
; SSE-NEXT: pand %xmm13, %xmm9
; SSE-NEXT: por %xmm0, %xmm9
-; SSE-NEXT: movdqa %xmm6, %xmm3
; SSE-NEXT: movdqa %xmm6, %xmm0
; SSE-NEXT: pand %xmm13, %xmm0
; SSE-NEXT: pandn %xmm10, %xmm13
@@ -2185,7 +2184,7 @@ define void @load_i8_stride7_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; SSE-NEXT: pandn %xmm3, %xmm2
+; SSE-NEXT: pandn %xmm6, %xmm2
; SSE-NEXT: por %xmm10, %xmm2
; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa {{.*#+}} xmm7 = [65535,0,65535,65535,65535,65535,65535,65535]
@@ -5451,19 +5450,19 @@ define void @load_i8_stride7_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pand %xmm14, %xmm6
; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; SSE-NEXT: movdqa %xmm0, %xmm3
+; SSE-NEXT: movdqa %xmm14, %xmm3
; SSE-NEXT: movdqa %xmm11, %xmm6
; SSE-NEXT: pandn %xmm11, %xmm3
; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; SSE-NEXT: pand %xmm0, %xmm5
+; SSE-NEXT: pand %xmm14, %xmm5
; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa %xmm2, %xmm3
-; SSE-NEXT: pand %xmm0, %xmm3
+; SSE-NEXT: pand %xmm14, %xmm3
; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pandn %xmm1, %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
@@ -9965,7 +9964,6 @@ define void @load_i8_stride7_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, %ymm23, %ymm3, %ymm6
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm13, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, %xmm14
; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm11 = xmm0[2,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm11[0],xmm3[1],xmm11[1],xmm3[2],xmm11[2],xmm3[3],xmm11[3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm8, %zmm8
@@ -9994,12 +9992,12 @@ define void @load_i8_stride7_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %ymm5, %ymm25, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm5 = xmm14[3,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm5 = xmm0[3,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm7 = xmm13[5,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm11, %zmm5
; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm10, %zmm3, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm14, %xmm7
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm0, %xmm7
; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm9 = xmm13[6,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm2, %zmm2
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-8.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-8.ll
index 0c2df82fd1be5..f2133b9e42d30 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-8.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-8.ll
@@ -11212,7 +11212,6 @@ define void @load_i8_stride8_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; AVX512F-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512F-SLOW-NEXT: vpshufb %xmm1, %xmm7, %xmm9
; AVX512F-SLOW-NEXT: vpshufb %xmm1, %xmm5, %xmm15
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm5, %xmm23
; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512F-SLOW-NEXT: vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512F-SLOW-NEXT: vpsrlq $32, %zmm17, %zmm9
@@ -11289,7 +11288,6 @@ define void @load_i8_stride8_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm30, %xmm10
; AVX512F-SLOW-NEXT: vpshufb %xmm3, %xmm10, %xmm9
; AVX512F-SLOW-NEXT: vpshufb %xmm3, %xmm12, %xmm15
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm12, %xmm31
; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512F-SLOW-NEXT: vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512F-SLOW-NEXT: vinserti128 $1, %xmm9, %ymm0, %ymm9
@@ -11302,7 +11300,7 @@ define void @load_i8_stride8_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512F-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512F-SLOW-NEXT: vpshufb %xmm1, %xmm6, %xmm9
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm23, %xmm11
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm21, %xmm11
; AVX512F-SLOW-NEXT: vpshufb %xmm1, %xmm11, %xmm15
; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512F-SLOW-NEXT: vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
index 5934b80893ce3..2e4f293587064 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
@@ -1343,10 +1343,9 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[0,1],xmm3[3,3]
; SSE-NEXT: movdqa %xmm15, %xmm10
; SSE-NEXT: punpckhwd {{.*#+}} xmm10 = xmm10[4],xmm5[4],xmm10[5],xmm5[5],xmm10[6],xmm5[6],xmm10[7],xmm5[7]
-; SSE-NEXT: movdqa %xmm5, %xmm1
; SSE-NEXT: punpcklwd {{.*#+}} xmm15 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[3,3,3,3,4,5,6,7]
+; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm5[3,3,3,3,4,5,6,7]
; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm9[0,2]
; SSE-NEXT: andps %xmm8, %xmm1
; SSE-NEXT: orps %xmm6, %xmm1
diff --git a/llvm/test/CodeGen/X86/wide-scalar-shift-legalization.ll b/llvm/test/CodeGen/X86/wide-scalar-shift-legalization.ll
index 24475360cbbc4..f84131dfc8797 100644
--- a/llvm/test/CodeGen/X86/wide-scalar-shift-legalization.ll
+++ b/llvm/t...
[truncated]
|
|
Bug Fix Description for #73512: In issue #73512, consider the following instruction sequences during the ForwardCopyPropagateBlock procudure: I tried to debug the routine of ForwardCopyPropagateBlock, and assume that we have the following register units for amdgpu registers (just an assumption for easily explain the root cause of this bug)： When we after processing L2(trackCopy for first two instructions), we have the following records in the copy maps:
But when we processing L3, we need to clobber the register sgpr0 (it's an implicit-def register). We directly remove the record(2, 3) , since it's DefRegs is being empty after the removal of the record that indicating sgpr3 defined sgpr0 :
Then when we reach L4, we need to mark L1 as unavailable during the propagation procedure. In this patch, we need to make sure that the although the SrcCopy's DefRegs is being empty after the "cleaning up" operation, it can't be directly removed from copy maps if its' MI is not NULL. |
|
This seems to at least work for the test cases that failed downstream after #70778 I'll try to run some more testing with this new patch over night. |
Can't say that I know much about this pass and the code diff. But I haven't seen any new failures when running tests this patch, so it's fine with me in that sense. |
Seems like a corner case in MIR code pattern leading to this bug:-), thank you for your time to test this patch further. |
|
Can you change the description to be a revert of the revert (or reapply)? I would also find it easier to read this if it were split into the original revert plus the version with the bug fix and test |
LWenH
changed the title
LWenH
changed the title
Sure, I separated the description for the original patch and the bug fix patch(this patch), and tried to make it more easy to read:-) |
|
Resolve X86 conflict test files. |
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.
=============================== Original Patch==================================
Due to the current mechanism of the MCP Pass, when a register is clobbered, let's say this register is Def, the record in the Tracker that indicates Src defined Def is still remain. This will lead to incorrect invalidations in the Tracker during the subsequent propagation process, thereby preventing the elimination of subsequent NopCopy instructions. In above example, when we clobber r0 at L2, the record in the tracker where r9 defines r0 still exists. Therefore, when we clobber the r9 register at L4, it causes an invalidation of L2 in the Tracker, leading to the inability to eliminate at L5.
In this patch, when we clobber a register Def, we also locate and remove the relative record from the tracker that indicating previous Src defined Def. Since such record is no longer effectual during the propagation procedure.
Please see the C++ test case generated code for vector.body in MIR: https://gcc.godbolt.org/z/nK4oMaWv5
================================BugFix(this patch)================================
This patch try to reimplement #70778, and fix issue: #73512.
In #70778 , when we clobbering a register (Def), we will also locate and remove the entry in the Copy Maps(Tracker) that indicates Src defined Def. If the "DefRegs" is being empty after removal, we will directly remove the corresponding SrcCopy entry from the Copy Maps.
However, since the SrcCopy may also have referred the Src as the destination of another copy instruction, directly removing it is unsafe. We need to enhance the criteria for removing empty records to ensure safety.