Merge pull request #21400 from BruceForstall/FixArm32FloatRange

Fix arm32 local variable references

7 files changed, 66 insertions, 52 deletions

diff --git a/src/jit/compiler.cpp b/src/jit/compiler.cpp
index 43acab3ca8..8093a7defc 100644
--- a/src/jit/compiler.cpp
+++ b/src/jit/compiler.cpp
@@ -4200,7 +4200,8 @@ bool Compiler::compRsvdRegCheck(FrameLayoutState curState)
     // Always do the layout even if returning early. Callers might
     // depend on us to do the layout.
     unsigned frameSize = lvaFrameSize(curState);
-    JITDUMP("\ncompRsvdRegCheck\n"
+    JITDUMP("\n"
+            "compRsvdRegCheck\n"
             "  frame size  = %6d\n"
             "  compArgSize = %6d\n",
             frameSize, compArgSize);
diff --git a/src/jit/compiler.h b/src/jit/compiler.h
index a79d77e4e2..7b91d39e2a 100644
--- a/src/jit/compiler.h
+++ b/src/jit/compiler.h
@@ -3055,7 +3055,7 @@ public:
 #endif
 
 #ifdef _TARGET_ARM_
-    int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset);
+    int lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset, bool isFloatUsage);
 #else
     int lvaFrameAddress(int varNum, bool* pFPbased);
 #endif
diff --git a/src/jit/compiler.hpp b/src/jit/compiler.hpp
index 8461b1f2a0..29a79f862b 100644
--- a/src/jit/compiler.hpp
+++ b/src/jit/compiler.hpp
@@ -2066,25 +2066,37 @@ inline int Compiler::lvaCachedGenericContextArgOffset()
     return lvaCachedGenericContextArgOffs;
 }
 
-/*****************************************************************************
- *
- *  Return the stack framed offset of the given variable; set *FPbased to
- *  true if the variable is addressed off of FP, false if it's addressed
- *  off of SP. Note that 'varNum' can be a negated spill-temporary var index.
- *
- *  mustBeFPBased - strong about whether the base reg is FP. But it is also
- *  strong about not being FPBased after FINAL_FRAME_LAYOUT. i.e.,
- *  it enforces SP based.
- *
- *  addrModeOffset - is the addressing mode offset, for example: v02 + 0x10
- *  So, V02 itself is at offset sp + 0x10 and then addrModeOffset is what gets
- *  added beyond that.
- */
-
+//------------------------------------------------------------------------
+// lvaFrameAddress: Determine the stack frame offset of the given variable,
+// and how to generate an address to that stack frame.
+//
+// Arguments:
+//    varNum         - The variable to inquire about. Positive for user variables
+//                     or arguments, negative for spill-temporaries.
+//    mustBeFPBased  - [_TARGET_ARM_ only] True if the base register must be FP.
+//                     After FINAL_FRAME_LAYOUT, if false, it also requires SP base register.
+//    pBaseReg       - [_TARGET_ARM_ only] Out arg. *pBaseReg is set to the base
+//                     register to use.
+//    addrModeOffset - [_TARGET_ARM_ only] The mode offset within the variable that we need to address.
+//                     For example, for a large struct local, and a struct field reference, this will be the offset
+//                     of the field. Thus, for V02 + 0x28, if V02 itself is at offset SP + 0x10
+//                     then addrModeOffset is what gets added beyond that, here 0x28.
+//    isFloatUsage   - [_TARGET_ARM_ only] True if the instruction being generated is a floating
+//                     point instruction. This requires using floating-point offset restrictions.
+//                     Note that a variable can be non-float, e.g., struct, but accessed as a
+//                     float local field.
+//    pFPbased       - [non-_TARGET_ARM_] Out arg. Set *FPbased to true if the
+//                     variable is addressed off of FP, false if it's addressed
+//                     off of SP.
+//
+// Return Value:
+//    Returns the variable offset from the given base register.
+//
 inline
 #ifdef _TARGET_ARM_
     int
-    Compiler::lvaFrameAddress(int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset)
+    Compiler::lvaFrameAddress(
+        int varNum, bool mustBeFPBased, regNumber* pBaseReg, int addrModeOffset, bool isFloatUsage)
 #else
     int
     Compiler::lvaFrameAddress(int varNum, bool* pFPbased)
@@ -2092,17 +2104,15 @@ inline
 {
     assert(lvaDoneFrameLayout != NO_FRAME_LAYOUT);
 
-    int       offset;
-    bool      FPbased;
-    bool      fConservative = false;
-    var_types type          = TYP_UNDEF;
+    int  varOffset;
+    bool FPbased;
+    bool fConservative = false;
     if (varNum >= 0)
     {
         LclVarDsc* varDsc;
 
         assert((unsigned)varNum < lvaCount);
         varDsc               = lvaTable + varNum;
-        type                 = varDsc->TypeGet();
         bool isPrespilledArg = false;
 #if defined(_TARGET_ARM_) && defined(PROFILING_SUPPORTED)
         isPrespilledArg = varDsc->lvIsParam && compIsProfilerHookNeeded() &&
@@ -2146,7 +2156,7 @@ inline
         }
 #endif // DEBUG
 
-        offset = varDsc->lvStkOffs;
+        varOffset = varDsc->lvStkOffs;
     }
     else // Its a spill-temp
     {
@@ -2160,8 +2170,7 @@ inline
                 tmpDsc = codeGen->regSet.tmpFindNum(varNum, RegSet::TEMP_USAGE_USED);
             }
             assert(tmpDsc != nullptr);
-            offset = tmpDsc->tdTempOffs();
-            type   = tmpDsc->tdTempType();
+            varOffset = tmpDsc->tdTempOffs();
         }
         else
         {
@@ -2188,7 +2197,6 @@ inline
             //   :                         :
             // ---------------------------------------------------
 
-            type          = compFloatingPointUsed ? TYP_FLOAT : TYP_INT;
             fConservative = true;
             if (!FPbased)
             {
@@ -2199,7 +2207,7 @@ inline
 #else
                 int outGoingArgSpaceSize = 0;
 #endif
-                offset = outGoingArgSpaceSize + max(-varNum * TARGET_POINTER_SIZE, (int)lvaGetMaxSpillTempSize());
+                varOffset = outGoingArgSpaceSize + max(-varNum * TARGET_POINTER_SIZE, (int)lvaGetMaxSpillTempSize());
             }
             else
             {
@@ -2207,9 +2215,9 @@ inline
                 CLANG_FORMAT_COMMENT_ANCHOR;
 
 #ifdef _TARGET_ARM_
-                offset = codeGen->genCallerSPtoInitialSPdelta() - codeGen->genCallerSPtoFPdelta();
+                varOffset = codeGen->genCallerSPtoInitialSPdelta() - codeGen->genCallerSPtoFPdelta();
 #else
-                offset                   = -(codeGen->genTotalFrameSize());
+                varOffset                = -(codeGen->genTotalFrameSize());
 #endif
             }
         }
@@ -2231,19 +2239,20 @@ inline
             // we have already selected the instruction. MinOpts will always reserve R10, so
             // for MinOpts always use SP-based offsets, using R10 as necessary, for simplicity.
 
-            int spOffset           = fConservative ? compLclFrameSize : offset + codeGen->genSPtoFPdelta();
-            int actualOffset       = spOffset + addrModeOffset;
-            int encodingLimitUpper = varTypeIsFloating(type) ? 0x3FC : 0xFFF;
-            int encodingLimitLower = varTypeIsFloating(type) ? -0x3FC : -0xFF;
+            int spVarOffset        = fConservative ? compLclFrameSize : varOffset + codeGen->genSPtoFPdelta();
+            int actualSPOffset     = spVarOffset + addrModeOffset;
+            int actualFPOffset     = varOffset + addrModeOffset;
+            int encodingLimitUpper = isFloatUsage ? 0x3FC : 0xFFF;
+            int encodingLimitLower = isFloatUsage ? -0x3FC : -0xFF;
 
             // Use SP-based encoding. During encoding, we'll pick the best encoding for the actual offset we have.
-            if (opts.MinOpts() || (actualOffset <= encodingLimitUpper))
+            if (opts.MinOpts() || (actualSPOffset <= encodingLimitUpper))
             {
-                offset    = spOffset;
+                varOffset = spVarOffset;
                 *pBaseReg = compLocallocUsed ? REG_SAVED_LOCALLOC_SP : REG_SPBASE;
             }
             // Use Frame Pointer (R11)-based encoding.
-            else if ((encodingLimitLower <= offset) && (offset <= encodingLimitUpper))
+            else if ((encodingLimitLower <= actualFPOffset) && (actualFPOffset <= encodingLimitUpper))
             {
                 *pBaseReg = REG_FPBASE;
             }
@@ -2252,7 +2261,7 @@ inline
             // the "reserved register", which will get used during encoding.
             else
             {
-                offset    = spOffset;
+                varOffset = spVarOffset;
                 *pBaseReg = compLocallocUsed ? REG_SAVED_LOCALLOC_SP : REG_SPBASE;
             }
         }
@@ -2265,7 +2274,7 @@ inline
     *pFPbased                            = FPbased;
 #endif
 
-    return offset;
+    return varOffset;
 }
 
 inline bool Compiler::lvaIsParameter(unsigned varNum)
diff --git a/src/jit/emit.cpp b/src/jit/emit.cpp
index 655f0e2bc8..4a3afc1ab0 100644
--- a/src/jit/emit.cpp
+++ b/src/jit/emit.cpp
@@ -6734,7 +6734,8 @@ target_ssize_t emitter::emitGetInsSC(instrDesc* id)
         regNumber baseReg;
         int       offs = id->idAddr()->iiaLclVar.lvaOffset();
 #if defined(_TARGET_ARM_)
-        int adr = emitComp->lvaFrameAddress(varNum, id->idIsLclFPBase(), &baseReg, offs);
+        int adr =
+            emitComp->lvaFrameAddress(varNum, id->idIsLclFPBase(), &baseReg, offs, CodeGen::instIsFP(id->idIns()));
         int dsp = adr + offs;
         if ((id->idIns() == INS_sub) || (id->idIns() == INS_subw))
             dsp = -dsp;
diff --git a/src/jit/emitarm.cpp b/src/jit/emitarm.cpp
index 1b8d2af77d..8098c60a41 100644
--- a/src/jit/emitarm.cpp
+++ b/src/jit/emitarm.cpp
@@ -3474,7 +3474,8 @@ void emitter::emitIns_R_S(instruction ins, emitAttr attr, regNumber reg1, int va
     int      disp;
     unsigned undisp;
 
-    base = emitComp->lvaFrameAddress(varx, emitComp->funCurrentFunc()->funKind != FUNC_ROOT, &reg2, offs);
+    base = emitComp->lvaFrameAddress(varx, emitComp->funCurrentFunc()->funKind != FUNC_ROOT, &reg2, offs,
+                                     CodeGen::instIsFP(ins));
 
     disp   = base + offs;
     undisp = unsigned_abs(disp);
@@ -3495,7 +3496,7 @@ void emitter::emitIns_R_S(instruction ins, emitAttr attr, regNumber reg1, int va
         else
         {
             regNumber rsvdReg = codeGen->rsGetRsvdReg();
-            emitIns_genStackOffset(rsvdReg, varx, offs);
+            emitIns_genStackOffset(rsvdReg, varx, offs, /* isFloatUsage */ true);
             emitIns_R_R(INS_add, EA_4BYTE, rsvdReg, reg2);
             emitIns_R_R_I(ins, attr, reg1, rsvdReg, 0);
             return;
@@ -3519,7 +3520,7 @@ void emitter::emitIns_R_S(instruction ins, emitAttr attr, regNumber reg1, int va
         {
             // Load disp into a register
             regNumber rsvdReg = codeGen->rsGetRsvdReg();
-            emitIns_genStackOffset(rsvdReg, varx, offs);
+            emitIns_genStackOffset(rsvdReg, varx, offs, /* isFloatUsage */ false);
             fmt = IF_T2_E0;
         }
     }
@@ -3545,7 +3546,7 @@ void emitter::emitIns_R_S(instruction ins, emitAttr attr, regNumber reg1, int va
         {
             // Load disp into a register
             regNumber rsvdReg = codeGen->rsGetRsvdReg();
-            emitIns_genStackOffset(rsvdReg, varx, offs);
+            emitIns_genStackOffset(rsvdReg, varx, offs, /* isFloatUsage */ false);
             emitIns_R_R_R(ins, attr, reg1, reg2, rsvdReg);
             return;
         }
@@ -3583,13 +3584,14 @@ void emitter::emitIns_R_S(instruction ins, emitAttr attr, regNumber reg1, int va
 }
 
 // generate the offset of &varx + offs into a register
-void emitter::emitIns_genStackOffset(regNumber r, int varx, int offs)
+void emitter::emitIns_genStackOffset(regNumber r, int varx, int offs, bool isFloatUsage)
 {
     regNumber regBase;
     int       base;
     int       disp;
 
-    base = emitComp->lvaFrameAddress(varx, emitComp->funCurrentFunc()->funKind != FUNC_ROOT, &regBase, offs);
+    base =
+        emitComp->lvaFrameAddress(varx, emitComp->funCurrentFunc()->funKind != FUNC_ROOT, &regBase, offs, isFloatUsage);
     disp = base + offs;
 
     emitIns_R_S(INS_movw, EA_4BYTE, r, varx, offs);
@@ -3633,7 +3635,8 @@ void emitter::emitIns_S_R(instruction ins, emitAttr attr, regNumber reg1, int va
     int      disp;
     unsigned undisp;
 
-    base = emitComp->lvaFrameAddress(varx, emitComp->funCurrentFunc()->funKind != FUNC_ROOT, &reg2, offs);
+    base = emitComp->lvaFrameAddress(varx, emitComp->funCurrentFunc()->funKind != FUNC_ROOT, &reg2, offs,
+                                     CodeGen::instIsFP(ins));
 
     disp   = base + offs;
     undisp = unsigned_abs(disp);
@@ -3654,7 +3657,7 @@ void emitter::emitIns_S_R(instruction ins, emitAttr attr, regNumber reg1, int va
         else
         {
             regNumber rsvdReg = codeGen->rsGetRsvdReg();
-            emitIns_genStackOffset(rsvdReg, varx, offs);
+            emitIns_genStackOffset(rsvdReg, varx, offs, /* isFloatUsage */ true);
             emitIns_R_R(INS_add, EA_4BYTE, rsvdReg, reg2);
             emitIns_R_R_I(ins, attr, reg1, rsvdReg, 0);
             return;
@@ -3676,7 +3679,7 @@ void emitter::emitIns_S_R(instruction ins, emitAttr attr, regNumber reg1, int va
     {
         // Load disp into a register
         regNumber rsvdReg = codeGen->rsGetRsvdReg();
-        emitIns_genStackOffset(rsvdReg, varx, offs);
+        emitIns_genStackOffset(rsvdReg, varx, offs, /* isFloatUsage */ false);
         fmt = IF_T2_E0;
     }
     assert((fmt == IF_T1_J2) || (fmt == IF_T2_E0) || (fmt == IF_T2_H0) || (fmt == IF_T2_VLDST) || (fmt == IF_T2_K1));
@@ -6363,7 +6366,7 @@ size_t emitter::emitOutputInstr(insGroup* ig, instrDesc* id, BYTE** dp)
         int       varNum = id->idAddr()->iiaLclVar.lvaVarNum();
         unsigned  ofs    = AlignDown(id->idAddr()->iiaLclVar.lvaOffset(), TARGET_POINTER_SIZE);
         regNumber regBase;
-        int       adr = emitComp->lvaFrameAddress(varNum, true, &regBase, ofs);
+        int adr = emitComp->lvaFrameAddress(varNum, true, &regBase, ofs, /* isFloatUsage */ false); // no float GC refs
         if (id->idGCref() != GCT_NONE)
         {
             emitGCvarLiveUpd(adr + ofs, varNum, id->idGCref(), dst);
diff --git a/src/jit/emitarm.h b/src/jit/emitarm.h
index fe24007ef9..1e79af0fed 100644
--- a/src/jit/emitarm.h
+++ b/src/jit/emitarm.h
@@ -267,7 +267,7 @@ void emitIns_C(instruction ins, emitAttr attr, CORINFO_FIELD_HANDLE fdlHnd, int
 
 void emitIns_S(instruction ins, emitAttr attr, int varx, int offs);
 
-void emitIns_genStackOffset(regNumber r, int varx, int offs);
+void emitIns_genStackOffset(regNumber r, int varx, int offs, bool isFloatUsage);
 
 void emitIns_S_R(instruction ins, emitAttr attr, regNumber ireg, int varx, int offs);
 
diff --git a/src/jit/lclvars.cpp b/src/jit/lclvars.cpp
index 339f7a5ff7..e58130bd5b 100644
--- a/src/jit/lclvars.cpp
+++ b/src/jit/lclvars.cpp
@@ -6722,7 +6722,7 @@ void Compiler::lvaDumpFrameLocation(unsigned lclNum)
     regNumber baseReg;
 
 #ifdef _TARGET_ARM_
-    offset = lvaFrameAddress(lclNum, compLocallocUsed, &baseReg, 0);
+    offset = lvaFrameAddress(lclNum, compLocallocUsed, &baseReg, 0, /* isFloatUsage */ false);
 #else
     bool EBPbased;
     offset  = lvaFrameAddress(lclNum, &EBPbased);