Minor optimizations.

src/llvm/lib/Target/W65816/W65816ISelLowering.cpp

@@ -1336,6 +1336,14 @@ SDValue W65816TargetLowering::LowerSELECT_CC(SDValue Op,
   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
   SDLoc DL(Op);
 
+  // Both arms the same value: no select.  The i32 splits below make these
+  // -- a 0/1 boolean widened to i32 has 0 in both high halves -- and the
+  // DAG combiner never sees them, as the legalizer lowers each new i16
+  // SELECT_CC at once.  Left alone, each is a CMP, a branch to the next
+  // instruction and a spilled PHI.
+  if (TVal == FVal)
+    return TVal;
+
   // i32 / i64 SELECT_CC.  An i8/i16 result selects straight on the CMP/SBC
   // chain's flags (or on the equivalent i16 compare).  An i32 result needs
   // two per-half selects and a glue value has only one consumer, so it

src/llvm/lib/Target/W65816/W65816RegisterInfo.td

@@ -50,10 +50,13 @@ def P   : W65816Reg<8, "p">,   DwarfRegNum<[8]>;
 // registers with cheap LDA/STA dp inter-register moves.  This
 // relieves pressure on the single Acc16 register (A) so greedy
 // regalloc can succeed on functions with multiple simultaneously-
-// live i16 vregs.  Caller-save: callees may freely overwrite them,
+// live i16 vregs.  IMG0..IMG7 are caller-save: a call clobbers them,
-// so regalloc spills around any call that might touch them.  Their
+// so regalloc spills around any call.  IMG8..IMG15 are callee-saved:
-// HWEncoding is never emitted (asmprinter translates IMGn references
+// W65816ImgCalleeSave saves and restores them in any function that
-// into LDA/STA dp with the right address).
+// writes them, and callers keep values in them across calls, so hand
+// asm called from C must not overwrite them without doing the same.
+// Their HWEncoding is never emitted (asmprinter translates IMGn
+// references into LDA/STA dp with the right address).
 //
 // Layout: IMG0..IMG7 at $D0..$DE (legacy slot block); IMG8..IMG15
 // at $C0..$CE.  Avoid stepping on user DP allocations below $C0.

src/llvm/test/CodeGen/W65816/fp-cmp-const-int.ll

@@ -70,6 +70,7 @@ define i16 @f_fabs_olt(float %x) {
 ; CHECK-NOT:   jsl
 ; CHECK:       and #0x7fff
 ; CHECK:       cmp #0x4020
+; CHECK-NEXT:  lda #0x1
 ; CHECK-NEXT:  bcc
 ; CHECK:       rtl
   %a = call float @llvm.fabs.f32(float %x)
@@ -84,6 +85,7 @@ define i16 @f_ogt0(float %x) {
 ; CHECK:       adc #0xffff
 ; CHECK:       adc #0xffff
 ; CHECK:       cmp #0x7f80
+; CHECK-NEXT:  lda #0x1
 ; CHECK-NEXT:  bcc
 ; CHECK:       rtl
   %c = fcmp ogt float %x, 0.0

src/llvm/test/CodeGen/W65816/select-same-arms.ll

@@ -0,0 +1,47 @@
+; A select whose two arms are the same value is that value: no compare, no
+; branch.  A 0/1 boolean widened to i32 (a returned bool, a zext to i32) is
+; split into a select per half, and the high half's arms are both 0.  The
+; legalizer lowers each new i16 SELECT_CC at once, so the DAG combiner's
+; same-arms fold never saw it: each one left a second compare of the value,
+; a branch to the next instruction and a PHI spilled to a slot.
+;
+; RUN: llc -mtriple=w65816 -O2 < %s | FileCheck %s
+
+target triple = "w65816"
+
+@keys = external global [60 x i8]
+
+; One compare of the loaded byte, one branch on it.
+define zeroext i1 @keyDown(i16 zeroext %key) {
+; CHECK-LABEL: keyDown:
+; CHECK:       lda keys, x
+; CHECK:       bne .LBB0_3
+; CHECK-NOT:   b{{eq|ne}}
+; CHECK:       rtl
+entry:
+  %cmp = icmp ugt i16 %key, 59
+  br i1 %cmp, label %return, label %if.end
+
+if.end:
+  %idx = zext nneg i16 %key to i32
+  %p = getelementptr inbounds i8, ptr @keys, i32 %idx
+  %v = load i8, ptr %p, align 1
+  %t = icmp ne i8 %v, 0
+  br label %return
+
+return:
+  %r = phi i1 [ %t, %if.end ], [ false, %entry ]
+  ret i1 %r
+}
+
+; The high half is the constant, the low half the one select.
+define i32 @isZero32(i16 %a) {
+; CHECK-LABEL: isZero32:
+; CHECK:       beq .LBB1_2
+; CHECK-NOT:   b{{eq|ne}}
+; CHECK:       ldx #0x0
+; CHECK:       rtl
+  %c = icmp eq i16 %a, 0
+  %r = zext i1 %c to i32
+  ret i32 %r
+}