[llvm] 7f4a910 - [RISCV] Add Zilsd/Zclsd support to RISCVMakeCompressible. (#140136)

via llvm-commits llvm-commits at lists.llvm.org
Fri May 16 09:36:43 PDT 2025


Author: Craig Topper
Date: 2025-05-16T09:36:39-07:00
New Revision: 7f4a9105df26a1c481d6b6d390a9d24e8ea987d1

URL: https://github.com/llvm/llvm-project/commit/7f4a9105df26a1c481d6b6d390a9d24e8ea987d1
DIFF: https://github.com/llvm/llvm-project/commit/7f4a9105df26a1c481d6b6d390a9d24e8ea987d1.diff

LOG: [RISCV] Add Zilsd/Zclsd support to RISCVMakeCompressible. (#140136)

Added: 
    llvm/test/CodeGen/RISCV/make-compressible-zilsd.mir

Modified: 
    llvm/lib/Target/RISCV/RISCVMakeCompressible.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/RISCV/RISCVMakeCompressible.cpp b/llvm/lib/Target/RISCV/RISCVMakeCompressible.cpp
index 5453753fa4579..1e2bdb10aa810 100644
--- a/llvm/lib/Target/RISCV/RISCVMakeCompressible.cpp
+++ b/llvm/lib/Target/RISCV/RISCVMakeCompressible.cpp
@@ -116,7 +116,9 @@ static unsigned log2LdstWidth(unsigned Opcode) {
   case RISCV::FSW:
     return 2;
   case RISCV::LD:
+  case RISCV::LD_RV32:
   case RISCV::SD:
+  case RISCV::SD_RV32:
   case RISCV::FLD:
   case RISCV::FSD:
     return 3;
@@ -144,7 +146,9 @@ static unsigned offsetMask(unsigned Opcode) {
   case RISCV::FLW:
   case RISCV::FSW:
   case RISCV::LD:
+  case RISCV::LD_RV32:
   case RISCV::SD:
+  case RISCV::SD_RV32:
   case RISCV::FLD:
   case RISCV::FSD:
     return maskTrailingOnes<unsigned>(5U);
@@ -184,7 +188,8 @@ static bool isCompressedReg(Register Reg) {
          RISCV::GPRF16CRegClass.contains(Reg) ||
          RISCV::GPRF32CRegClass.contains(Reg) ||
          RISCV::FPR32CRegClass.contains(Reg) ||
-         RISCV::FPR64CRegClass.contains(Reg);
+         RISCV::FPR64CRegClass.contains(Reg) ||
+         RISCV::GPRPairCRegClass.contains(Reg);
 }
 
 // Return true if MI is a load for which there exists a compressed version.
@@ -203,6 +208,8 @@ static bool isCompressibleLoad(const MachineInstr &MI) {
   case RISCV::LW_INX:
   case RISCV::LD:
     return STI.hasStdExtCOrZca();
+  case RISCV::LD_RV32:
+    return STI.hasStdExtZclsd();
   case RISCV::FLW:
     return !STI.is64Bit() && STI.hasStdExtCOrZcfOrZce();
   case RISCV::FLD:
@@ -225,6 +232,8 @@ static bool isCompressibleStore(const MachineInstr &MI) {
   case RISCV::SW_INX:
   case RISCV::SD:
     return STI.hasStdExtCOrZca();
+  case RISCV::SD_RV32:
+    return STI.hasStdExtZclsd();
   case RISCV::FSW:
     return !STI.is64Bit() && STI.hasStdExtCOrZcfOrZce();
   case RISCV::FSD:
@@ -342,8 +351,10 @@ static Register analyzeCompressibleUses(MachineInstr &FirstMI,
     RCToScavenge = &RISCV::FPR32CRegClass;
   else if (RISCV::FPR64RegClass.contains(RegImm.Reg))
     RCToScavenge = &RISCV::FPR64CRegClass;
+  else if (RISCV::GPRPairRegClass.contains(RegImm.Reg))
+    RCToScavenge = &RISCV::GPRPairCRegClass;
   else
-    return RISCV::NoRegister;
+    return Register();
 
   RegScavenger RS;
   RS.enterBasicBlockEnd(MBB);
@@ -400,6 +411,7 @@ bool RISCVMakeCompressibleOpt::runOnMachineFunction(MachineFunction &Fn) {
 
   const RISCVSubtarget &STI = Fn.getSubtarget<RISCVSubtarget>();
   const RISCVInstrInfo &TII = *STI.getInstrInfo();
+  const RISCVRegisterInfo &TRI = *STI.getRegisterInfo();
 
   // This optimization only makes sense if compressed instructions are emitted.
   if (!STI.hasStdExtCOrZca())
@@ -438,7 +450,20 @@ bool RISCVMakeCompressibleOpt::runOnMachineFunction(MachineFunction &Fn) {
         BuildMI(MBB, MI, MI.getDebugLoc(), TII.get(RISCV::PseudoMV_FPR32INX),
                 NewReg)
             .addReg(RegImm.Reg);
+      } else if (RISCV::GPRPairRegClass.contains(RegImm.Reg)) {
+        assert(RegImm.Imm == 0);
+        BuildMI(MBB, MI, MI.getDebugLoc(), TII.get(RISCV::ADDI),
+                TRI.getSubReg(NewReg, RISCV::sub_gpr_even))
+            .addReg(TRI.getSubReg(RegImm.Reg, RISCV::sub_gpr_even))
+            .addImm(0);
+        BuildMI(MBB, MI, MI.getDebugLoc(), TII.get(RISCV::ADDI),
+                TRI.getSubReg(NewReg, RISCV::sub_gpr_odd))
+            .addReg(TRI.getSubReg(RegImm.Reg, RISCV::sub_gpr_odd))
+            .addImm(0);
       } else {
+        assert((RISCV::FPR32RegClass.contains(RegImm.Reg) ||
+                RISCV::FPR64RegClass.contains(RegImm.Reg)) &&
+               "Expected FP register class");
         // If we are looking at replacing an FPR register we don't expect to
         // have any offset. The only compressible FP instructions with an offset
         // are loads and stores, for which the offset applies to the GPR operand

diff  --git a/llvm/test/CodeGen/RISCV/make-compressible-zilsd.mir b/llvm/test/CodeGen/RISCV/make-compressible-zilsd.mir
new file mode 100644
index 0000000000000..c5ac599d8d53f
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/make-compressible-zilsd.mir
@@ -0,0 +1,299 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -o - %s -mtriple=riscv32 -mattr=+zilsd,+zclsd,+zdinx -simplify-mir \
+# RUN:   -run-pass=riscv-make-compressible | FileCheck --check-prefixes=RV32 %s
+--- |
+  define void @store_common_value_double(ptr %a, ptr %b, ptr %c, i32 %d, double %e, double %f) #0 {
+  entry:
+    store double %f, ptr %a, align 8
+    store double %f, ptr %b, align 8
+    store double %f, ptr %c, align 8
+    ret void
+  }
+
+  define void @store_common_ptr_double(double %a, double %b, double %d, ptr %p) #0 {
+  entry:
+    store volatile double %a, ptr %p, align 8
+    store volatile double %b, ptr %p, align 8
+    store volatile double %b, ptr %p, align 8
+    ret void
+  }
+
+  define void @load_common_ptr_double(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, ptr %g) #0 {
+  entry:
+    %0 = load double, ptr %g, align 8
+    %arrayidx1 = getelementptr inbounds { double, double, i32 }, ptr %g, i32 0, i32 1
+    %1 = load double, ptr %arrayidx1, align 8
+    %arrayidx2 = getelementptr inbounds { double, double, i32 }, ptr %g, i32 0, i32 2
+    %2 = load i32, ptr %arrayidx2, align 8
+    tail call void @load_common_ptr_double_1(double %0, double %1, i32 %2)
+    ret void
+  }
+
+  declare void @load_common_ptr_double_1(double, double, double) #0
+
+  define void @store_large_offset_double(ptr %p, i32 %dummy, double %a, double %b, double %c) #0 {
+  entry:
+    %0 = getelementptr inbounds double, ptr %p, i32 100
+    store volatile double %a, ptr %0, align 8
+    %1 = getelementptr inbounds double, ptr %p, i32 101
+    store volatile double %b, ptr %1, align 8
+    %2 = getelementptr inbounds double, ptr %p, i32 102
+    store volatile double %b, ptr %2, align 8
+    ret void
+  }
+
+  define void @load_large_offset_double(i32 %a, i32 %b, i32 %c, i32 %d, ptr %p) #0 {
+  entry:
+    %arrayidx = getelementptr inbounds { [102 x double], i32 }, ptr %p, i32 0, i32 0, i32 100
+    %0 = load double, ptr %arrayidx, align 8
+    %arrayidx1 = getelementptr inbounds { [102 x double], i32 }, ptr %p, i32 0, i32 0, i32 101
+    %1 = load double, ptr %arrayidx1, align 8
+    %arrayidx2 = getelementptr inbounds { [102 x double], i32 }, ptr %p, i32 0, i32 1
+    %2 = load i32, ptr %arrayidx2, align 8
+    tail call void @load_large_offset_double_1(double %0, double %1, i32 %2)
+    ret void
+  }
+
+  declare void @load_large_offset_double_1(double, double) #0
+
+  define void @store_common_value_double_no_opt(ptr %a, i32 %b, double %c, double %d, double %e) #0 {
+  entry:
+    store double %e, ptr %a, align 8
+    ret void
+  }
+
+  define void @store_common_ptr_double_no_opt(double %a, i32 %b, i32 %c, i32 %d, i32 %e, ptr %p) #0 {
+  entry:
+    store volatile double %a, ptr %p, align 8
+    ret void
+  }
+
+  define double @load_common_ptr_double_no_opt(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, ptr %g) #0 {
+  entry:
+    %0 = load double, ptr %g, align 8
+    ret double %0
+  }
+
+  define void @store_large_offset_double_no_opt(ptr %p, double %a, double %b) #0 {
+  entry:
+    %0 = getelementptr inbounds double, ptr %p, i32 100
+    store volatile double %a, ptr %0, align 8
+    %1 = getelementptr inbounds double, ptr %p, i32 101
+    store volatile double %b, ptr %1, align 8
+    ret void
+  }
+
+  define { double, double } @load_large_offset_double_no_opt(ptr %p) #0 {
+  entry:
+    %arrayidx = getelementptr inbounds double, ptr %p, i32 100
+    %0 = load double, ptr %arrayidx, align 8
+    %arrayidx1 = getelementptr inbounds double, ptr %p, i32 101
+    %1 = load double, ptr %arrayidx1, align 8
+    %2 = insertvalue { double, double } undef, double %0, 0
+    %3 = insertvalue { double, double } %2, double %1, 1
+    ret { double, double } %3
+  }
+
+  attributes #0 = { minsize "target-features"="+zilsd,+zdinx" }
+...
+---
+name:            store_common_value_double
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10, $x11, $x12, $x16, $x17
+
+    ; RV32-LABEL: name: store_common_value_double
+    ; RV32: liveins: $x10, $x11, $x12, $x16, $x17
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: $x14 = ADDI $x16, 0
+    ; RV32-NEXT: $x15 = ADDI $x17, 0
+    ; RV32-NEXT: SD_RV32 $x14_x15, killed renamable $x10, 0 :: (store (s64) into %ir.a)
+    ; RV32-NEXT: SD_RV32 $x14_x15, killed renamable $x11, 0 :: (store (s64) into %ir.b)
+    ; RV32-NEXT: SD_RV32 killed $x14_x15, killed renamable $x12, 0 :: (store (s64) into %ir.c)
+    ; RV32-NEXT: PseudoRET
+    SD_RV32 renamable $x16_x17, killed renamable $x10, 0 :: (store (s64) into %ir.a)
+    SD_RV32 renamable $x16_x17, killed renamable $x11, 0 :: (store (s64) into %ir.b)
+    SD_RV32 killed renamable $x16_x17, killed renamable $x12, 0 :: (store (s64) into %ir.c)
+    PseudoRET
+
+...
+---
+name:            store_common_ptr_double
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10, $x11, $x12, $x13, $x14, $x15, $x16
+
+    ; RV32-LABEL: name: store_common_ptr_double
+    ; RV32: liveins: $x10, $x11, $x12, $x13, $x14, $x15, $x16
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: $x14 = ADDI $x16, 0
+    ; RV32-NEXT: SD_RV32 killed renamable $x10_x11, $x14, 0 :: (volatile store (s64) into %ir.p)
+    ; RV32-NEXT: SD_RV32 renamable $x12_x13, $x14, 0 :: (volatile store (s64) into %ir.p)
+    ; RV32-NEXT: SD_RV32 killed renamable $x12_x13, killed $x14, 0 :: (volatile store (s64) into %ir.p)
+    ; RV32-NEXT: PseudoRET
+    SD_RV32 killed renamable $x10_x11, renamable $x16, 0 :: (volatile store (s64) into %ir.p)
+    SD_RV32 renamable $x12_x13, renamable $x16, 0 :: (volatile store (s64) into %ir.p)
+    SD_RV32 killed renamable $x12_x13, killed renamable $x16, 0 :: (volatile store (s64) into %ir.p)
+    PseudoRET
+
+...
+---
+name:            load_common_ptr_double
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x16
+
+    ; RV32-LABEL: name: load_common_ptr_double
+    ; RV32: liveins: $x16
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: $x15 = ADDI $x16, 0
+    ; RV32-NEXT: renamable $x10_x11 = LD_RV32 $x15, 0 :: (load (s64) from %ir.g)
+    ; RV32-NEXT: renamable $x12_x13 = LD_RV32 $x15, 8 :: (load (s64) from %ir.arrayidx1)
+    ; RV32-NEXT: renamable $x14 = LW killed $x15, 16 :: (load (s32) from %ir.arrayidx2, align 8)
+    ; RV32-NEXT: PseudoTAIL target-flags(riscv-call) @load_common_ptr_double_1, csr_ilp32_lp64, implicit $x2, implicit $x10, implicit $x11, implicit $x12, implicit $x13, implicit $x14
+    renamable $x10_x11 = LD_RV32 renamable $x16, 0 :: (load (s64) from %ir.g)
+    renamable $x12_x13 = LD_RV32 renamable $x16, 8 :: (load (s64) from %ir.arrayidx1)
+    renamable $x14 = LW killed renamable $x16, 16 :: (load (s32) from %ir.arrayidx2, align 8)
+    PseudoTAIL target-flags(riscv-call) @load_common_ptr_double_1, csr_ilp32_lp64, implicit $x2, implicit $x10, implicit $x11, implicit $x12, implicit $x13, implicit $x14
+
+...
+---
+name:            store_large_offset_double
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10, $x12, $x13, $x14, $x15
+
+    ; RV32-LABEL: name: store_large_offset_double
+    ; RV32: liveins: $x10, $x12, $x13, $x14, $x15
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: $x11 = ADDI $x10, 768
+    ; RV32-NEXT: SD_RV32 killed renamable $x12_x13, $x11, 32 :: (volatile store (s64) into %ir.0)
+    ; RV32-NEXT: SD_RV32 renamable $x14_x15, $x11, 40 :: (volatile store (s64) into %ir.1)
+    ; RV32-NEXT: SD_RV32 killed renamable $x14_x15, killed $x11, 48 :: (volatile store (s64) into %ir.2)
+    ; RV32-NEXT: PseudoRET
+    SD_RV32 killed renamable $x12_x13, renamable $x10, 800 :: (volatile store (s64) into %ir.0)
+    SD_RV32 renamable $x14_x15, renamable $x10, 808 :: (volatile store (s64) into %ir.1)
+    SD_RV32 killed renamable $x14_x15, killed renamable $x10, 816 :: (volatile store (s64) into %ir.2)
+    PseudoRET
+
+...
+---
+name:            load_large_offset_double
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x14
+
+    ; RV32-LABEL: name: load_large_offset_double
+    ; RV32: liveins: $x14
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: $x15 = ADDI $x14, 768
+    ; RV32-NEXT: renamable $x10_x11 = LD_RV32 $x15, 32 :: (load (s64) from %ir.arrayidx)
+    ; RV32-NEXT: renamable $x12_x13 = LD_RV32 $x15, 40 :: (load (s64) from %ir.arrayidx1)
+    ; RV32-NEXT: renamable $x14 = LW killed $x15, 48 :: (load (s32) from %ir.arrayidx2, align 8)
+    ; RV32-NEXT: PseudoTAIL target-flags(riscv-call) @load_large_offset_double_1, csr_ilp32_lp64, implicit $x2, implicit $x10, implicit $x11, implicit $x12, implicit $x13, implicit $x14
+    renamable $x10_x11 = LD_RV32 renamable $x14, 800 :: (load (s64) from %ir.arrayidx)
+    renamable $x12_x13 = LD_RV32 renamable $x14, 808 :: (load (s64) from %ir.arrayidx1)
+    renamable $x14 = LW killed renamable $x14, 816 :: (load (s32) from %ir.arrayidx2, align 8)
+    PseudoTAIL target-flags(riscv-call) @load_large_offset_double_1, csr_ilp32_lp64, implicit $x2, implicit $x10, implicit $x11, implicit $x12, implicit $x13, implicit $x14
+...
+---
+name:            store_common_value_double_no_opt
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10, $x16, $x17
+
+    ; RV32-LABEL: name: store_common_value_double_no_opt
+    ; RV32: liveins: $x10, $x16, $x17
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: SD_RV32 killed renamable $x16_x17, killed renamable $x10, 0 :: (store (s64) into %ir.a)
+    ; RV32-NEXT: PseudoRET
+    SD_RV32 killed renamable $x16_x17, killed renamable $x10, 0 :: (store (s64) into %ir.a)
+    PseudoRET
+
+...
+---
+name:            store_common_ptr_double_no_opt
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10, $x11, $x16
+
+    ; RV32-LABEL: name: store_common_ptr_double_no_opt
+    ; RV32: liveins: $x10, $x11, $x16
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: SD_RV32 killed renamable $x10_x11, killed renamable $x16, 0 :: (volatile store (s64) into %ir.p)
+    ; RV32-NEXT: PseudoRET
+    SD_RV32 killed renamable $x10_x11, killed renamable $x16, 0 :: (volatile store (s64) into %ir.p)
+    PseudoRET
+
+...
+---
+name:            load_common_ptr_double_no_opt
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x16
+
+    ; RV32-LABEL: name: load_common_ptr_double_no_opt
+    ; RV32: liveins: $x16
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: renamable $x10_x11 = LD_RV32 killed renamable $x16, 0 :: (load (s64) from %ir.g)
+    ; RV32-NEXT: PseudoRET implicit $x10, implicit $x11
+    renamable $x10_x11 = LD_RV32 killed renamable $x16, 0 :: (load (s64) from %ir.g)
+    PseudoRET implicit $x10, implicit $x11
+
+...
+---
+name:            store_large_offset_double_no_opt
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10, $x11, $x12, $x13, $x14
+
+    ; RV32-LABEL: name: store_large_offset_double_no_opt
+    ; RV32: liveins: $x10, $x11, $x12, $x13, $x14
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: $x15 = ADDI $x14, 0
+    ; RV32-NEXT: $x17 = ADDI $x12, 0
+    ; RV32-NEXT: $x16 = ADDI $x11, 0
+    ; RV32-NEXT: SD_RV32 killed renamable $x16_x17, renamable $x10, 800 :: (volatile store (s64) into %ir.0)
+    ; RV32-NEXT: $x14 = ADDI $x13, 0
+    ; RV32-NEXT: SD_RV32 killed renamable $x14_x15, killed renamable $x10, 808 :: (volatile store (s64) into %ir.1)
+    ; RV32-NEXT: PseudoRET
+    $x15 = ADDI $x14, 0
+    $x17 = ADDI $x12, 0
+    $x16 = ADDI $x11, 0
+    SD_RV32 killed renamable $x16_x17, renamable $x10, 800 :: (volatile store (s64) into %ir.0)
+    $x14 = ADDI $x13, 0
+    SD_RV32 killed renamable $x14_x15, killed renamable $x10, 808 :: (volatile store (s64) into %ir.1)
+    PseudoRET
+
+...
+---
+name:            load_large_offset_double_no_opt
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10
+
+    ; RV32-LABEL: name: load_large_offset_double_no_opt
+    ; RV32: liveins: $x10
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: renamable $x14_x15 = LD_RV32 renamable $x10, 800 :: (load (s64) from %ir.arrayidx)
+    ; RV32-NEXT: renamable $x12_x13 = LD_RV32 killed renamable $x10, 808 :: (load (s64) from %ir.arrayidx1)
+    ; RV32-NEXT: $x10 = ADDI renamable $x14, 0
+    ; RV32-NEXT: $x11 = ADDI killed renamable $x15, 0
+    ; RV32-NEXT: PseudoRET implicit $x10, implicit $x11, implicit $x12, implicit $x13
+    renamable $x14_x15 = LD_RV32 renamable $x10, 800 :: (load (s64) from %ir.arrayidx)
+    renamable $x12_x13 = LD_RV32 killed renamable $x10, 808 :: (load (s64) from %ir.arrayidx1)
+    $x10 = ADDI renamable $x14, 0
+    $x11 = ADDI killed renamable $x15, 0
+    PseudoRET implicit $x10, implicit $x11, implicit $x12, implicit $x13
+
+...


        


More information about the llvm-commits mailing list