Add a t8jccd_long encoding recipe for brz.b1 and brnz.b1 in 32-bit mode.

The register allocator can't handle branches with constrained register
operands, and the brz.b1/brnz.b1 instructions only have the t8jccd_abcd
in 32-bit mode where no REX prefixes are possible.

This adds a worst case encoding for those cases where a b1 value lives
in a non-ABCD register.

cranelift/filetests/isa/intel/binary32.cton

@@ -1,5 +1,6 @@
 ; binary emission of 32-bit code.
 test binemit
+set is_compressed
 isa intel haswell
 
 ; The binary encodings can be verified with the command:
@@ -393,3 +394,30 @@ ebb1:
 ebb2:
     trap user0                                  ; bin: 0f 0b
 }
+
+; Special branch encodings only for I32 mode.
+function %special_branches() {
+ebb0:
+    [-,%rcx]            v1 = iconst.i32 1
+    [-,%rsi]            v2 = iconst.i32 2
+    [-,%rdi]            v3 = icmp eq v1, v2
+    [-,%rbx]            v4 = icmp ugt v1, v2
+
+    ; asm: testl $0xff, %edi
+    ; asm: je ebb1
+    brz v3, ebb1                                ; bin: f7 c7 000000ff 0f 84 00000015
+    ; asm: testb %bl, %bl
+    ; asm: je ebb1
+    brz v4, ebb1                                ; bin: 84 db 74 11
+    ; asm: testl $0xff, %edi
+    ; asm: jne ebb1
+    brnz v3, ebb1                               ; bin: f7 c7 000000ff 0f 85 00000005
+    ; asm: testb %bl, %bl
+    ; asm: jne ebb1
+    brnz v4, ebb1                               ; bin: 84 db 75 01
+
+    return
+
+ebb1:
+    return
+}

lib/cretonne/meta/isa/intel/encodings.py

@@ -292,6 +292,12 @@ enc_i32_i64(base.brnz, r.tjccd, 0x85)
 
 # Branch on a b1 value in a register only looks at the low 8 bits. See also
 # bint encodings below.
+#
+# Start with the worst-case encoding for I32 only. The register allocator can't
+# handle a branch with an ABCD-constrained operand.
+I32.enc(base.brz.b1, *r.t8jccd_long(0x84))
+I32.enc(base.brnz.b1, *r.t8jccd_long(0x85))
+
 enc_both(base.brz.b1, r.t8jccb_abcd, 0x74)
 enc_both(base.brz.b1, r.t8jccd_abcd, 0x84)
 enc_both(base.brnz.b1, r.t8jccb_abcd, 0x75)

lib/cretonne/meta/isa/intel/recipes.py

@@ -848,6 +848,25 @@ t8jccd_abcd = TailRecipe(
         disp4(destination, func, sink);
         ''')
 
+# Worst case test-and-branch recipe for brz.b1 and brnz.b1 in 32-bit mode.
+# The register allocator can't handle a branch instruction with constrained
+# operands like the t8jccd_abcd above. This variant can accept the b1 opernd in
+# any register, but is is larger because it uses a 32-bit test instruction with
+# a 0xff immediate.
+t8jccd_long = TailRecipe(
+        't8jccd_long', Branch, size=5 + 6, ins=GPR, outs=(),
+        branch_range=32,
+        emit='''
+        // test32 r, 0xff.
+        PUT_OP((bits & 0xff00) | 0xf7, rex1(in_reg0), sink);
+        modrm_r_bits(in_reg0, bits, sink);
+        sink.put4(0xff);
+        // Jcc instruction.
+        sink.put1(0x0f);
+        sink.put1(bits as u8);
+        disp4(destination, func, sink);
+        ''')
+
 # Comparison that produces a `b1` result in a GPR.
 #
 # This is a macro of a `cmp` instruction followed by a `setCC` instruction.