@@ -208,6 +208,22 @@ static const char *opcodes[] = {
[OP_SET_BE] = "setbe",
[OP_SET_AE] = "setae",
+ /* floating-point comparison */
+ [OP_FCMP_ORD] = "fcmpord",
+ [OP_FCMP_OEQ] = "fcmpoeq",
+ [OP_FCMP_ONE] = "fcmpone",
+ [OP_FCMP_OLE] = "fcmpole",
+ [OP_FCMP_OGE] = "fcmpoge",
+ [OP_FCMP_OLT] = "fcmpolt",
+ [OP_FCMP_OGT] = "fcmpogt",
+ [OP_FCMP_UEQ] = "fcmpueq",
+ [OP_FCMP_UNE] = "fcmpune",
+ [OP_FCMP_ULE] = "fcmpule",
+ [OP_FCMP_UGE] = "fcmpuge",
+ [OP_FCMP_ULT] = "fcmpult",
+ [OP_FCMP_UGT] = "fcmpugt",
+ [OP_FCMP_UNO] = "fcmpuno",
+
/* Uni */
[OP_NOT] = "not",
[OP_NEG] = "neg",
@@ -437,6 +453,7 @@ const char *show_instruction(struct instruction *insn)
show_pseudo(insn->src));
break;
case OP_BINARY ... OP_BINARY_END:
+ case OP_FPCMP ... OP_FPCMP_END:
case OP_BINCMP ... OP_BINCMP_END:
buf += sprintf(buf, "%s <- %s, %s", show_pseudo(insn->target), show_pseudo(insn->src1), show_pseudo(insn->src2));
break;
@@ -1465,10 +1482,10 @@ static pseudo_t linearize_compare(struct entrypoint *ep, struct expression *expr
[SPECIAL_UNSIGNED_LTE] = OP_SET_BE,
[SPECIAL_UNSIGNED_GTE] = OP_SET_AE,
};
-
+ int op = opcode_float(cmpop[expr->op], expr->right->ctype);
pseudo_t src1 = linearize_expression(ep, expr->left);
pseudo_t src2 = linearize_expression(ep, expr->right);
- pseudo_t dst = add_binary_op(ep, expr->ctype, cmpop[expr->op], src1, src2);
+ pseudo_t dst = add_binary_op(ep, expr->ctype, op, src1, src2);
return dst;
}
@@ -170,6 +170,24 @@ enum opcode {
OP_OR_BOOL,
OP_BINARY_END = OP_OR_BOOL,
+ /* floating-point comparison */
+ OP_FPCMP,
+ OP_FCMP_ORD = OP_FPCMP,
+ OP_FCMP_OEQ,
+ OP_FCMP_ONE,
+ OP_FCMP_OLE,
+ OP_FCMP_OGE,
+ OP_FCMP_OLT,
+ OP_FCMP_OGT,
+ OP_FCMP_UEQ,
+ OP_FCMP_UNE,
+ OP_FCMP_ULE,
+ OP_FCMP_UGE,
+ OP_FCMP_ULT,
+ OP_FCMP_UGT,
+ OP_FCMP_UNO,
+ OP_FPCMP_END = OP_FCMP_UNO,
+
/* Binary comparison */
OP_BINCMP,
OP_SET_EQ = OP_BINCMP,
@@ -66,6 +66,7 @@ static void track_instruction_usage(struct basic_block *bb, struct instruction *
/* Binary */
case OP_BINARY ... OP_BINARY_END:
+ case OP_FPCMP ... OP_FPCMP_END:
case OP_BINCMP ... OP_BINCMP_END:
USES(src1); USES(src2); DEFINES(target);
break;
@@ -23,14 +23,32 @@
#include "linearize.h"
const struct opcode_table opcode_table[OP_LAST] = {
- [OP_SET_EQ] = { .negate = OP_SET_NE, .swap = OP_SET_EQ, },
- [OP_SET_NE] = { .negate = OP_SET_EQ, .swap = OP_SET_NE, },
- [OP_SET_LT] = { .negate = OP_SET_GE, .swap = OP_SET_GT, },
- [OP_SET_LE] = { .negate = OP_SET_GT, .swap = OP_SET_GE, },
- [OP_SET_GE] = { .negate = OP_SET_LT, .swap = OP_SET_LE, },
- [OP_SET_GT] = { .negate = OP_SET_LE, .swap = OP_SET_LT, },
- [OP_SET_B ] = { .negate = OP_SET_AE, .swap = OP_SET_A , },
- [OP_SET_BE] = { .negate = OP_SET_A , .swap = OP_SET_AE, },
- [OP_SET_AE] = { .negate = OP_SET_B , .swap = OP_SET_BE, },
- [OP_SET_A ] = { .negate = OP_SET_BE, .swap = OP_SET_B , },
+ [OP_SET_EQ] = { .negate = OP_SET_NE, .swap = OP_SET_EQ, .to_float = OP_FCMP_OEQ, },
+ [OP_SET_NE] = { .negate = OP_SET_EQ, .swap = OP_SET_NE, .to_float = OP_FCMP_UNE, },
+ [OP_SET_LT] = { .negate = OP_SET_GE, .swap = OP_SET_GT, .to_float = OP_FCMP_OLT, },
+ [OP_SET_LE] = { .negate = OP_SET_GT, .swap = OP_SET_GE, .to_float = OP_FCMP_OLE, },
+ [OP_SET_GE] = { .negate = OP_SET_LT, .swap = OP_SET_LE, .to_float = OP_FCMP_OGE, },
+ [OP_SET_GT] = { .negate = OP_SET_LE, .swap = OP_SET_LT, .to_float = OP_FCMP_OGT, },
+ [OP_SET_B ] = { .negate = OP_SET_AE, .swap = OP_SET_A , .to_float = OP_FCMP_OLT, },
+ [OP_SET_BE] = { .negate = OP_SET_A , .swap = OP_SET_AE, .to_float = OP_FCMP_OLE, },
+ [OP_SET_AE] = { .negate = OP_SET_B , .swap = OP_SET_BE, .to_float = OP_FCMP_OGE, },
+ [OP_SET_A ] = { .negate = OP_SET_BE, .swap = OP_SET_B , .to_float = OP_FCMP_OGT, },
+
+ [OP_FCMP_ORD] = { .negate = OP_FCMP_UNO, .swap = OP_FCMP_ORD, },
+ [OP_FCMP_UNO] = { .negate = OP_FCMP_ORD, .swap = OP_FCMP_UNO, },
+
+ [OP_FCMP_OEQ] = { .negate = OP_FCMP_UNE, .swap = OP_FCMP_OEQ, },
+ [OP_FCMP_ONE] = { .negate = OP_FCMP_UEQ, .swap = OP_FCMP_ONE, },
+ [OP_FCMP_UEQ] = { .negate = OP_FCMP_ONE, .swap = OP_FCMP_UEQ, },
+ [OP_FCMP_UNE] = { .negate = OP_FCMP_OEQ, .swap = OP_FCMP_UNE, },
+
+ [OP_FCMP_OLT] = { .negate = OP_FCMP_UGE, .swap = OP_FCMP_OGT, },
+ [OP_FCMP_OLE] = { .negate = OP_FCMP_UGT, .swap = OP_FCMP_OGE, },
+ [OP_FCMP_OGE] = { .negate = OP_FCMP_ULT, .swap = OP_FCMP_OLE, },
+ [OP_FCMP_OGT] = { .negate = OP_FCMP_ULE, .swap = OP_FCMP_OLT, },
+
+ [OP_FCMP_ULT] = { .negate = OP_FCMP_OGE, .swap = OP_FCMP_UGT, },
+ [OP_FCMP_ULE] = { .negate = OP_FCMP_OGT, .swap = OP_FCMP_UGE, },
+ [OP_FCMP_UGE] = { .negate = OP_FCMP_OLT, .swap = OP_FCMP_ULE, },
+ [OP_FCMP_UGT] = { .negate = OP_FCMP_OLE, .swap = OP_FCMP_ULT, },
};
@@ -1,10 +1,20 @@
#ifndef OPCODE_H
#define OPCODE_H
+#include "symbol.h"
extern const struct opcode_table {
int negate:8;
int swap:8;
+ int to_float:8;
} opcode_table[];
+
+static inline int opcode_float(int opcode, struct symbol *type)
+{
+ if (!type || !is_float_type(type))
+ return opcode;
+ return opcode_table[opcode].to_float;
+}
+
#endif
@@ -429,7 +429,7 @@ static int simplify_seteq_setne(struct instruction *insn, long long value)
inverse = (insn->opcode == OP_SET_NE) == value;
opcode = def->opcode;
switch (opcode) {
- case OP_BINCMP ... OP_BINCMP_END:
+ case OP_FPCMP ... OP_BINCMP_END:
// Convert:
// setcc.n %t <- %a, %b
// setne.m %r <- %t, $0
@@ -492,17 +492,20 @@ static LLVMValueRef calc_gep(LLVMBuilderRef builder, LLVMValueRef base, LLVMValu
static LLVMRealPredicate translate_fop(int opcode)
{
static const LLVMRealPredicate trans_tbl[] = {
- [OP_SET_EQ] = LLVMRealOEQ,
- [OP_SET_NE] = LLVMRealUNE,
- [OP_SET_LE] = LLVMRealOLE,
- [OP_SET_GE] = LLVMRealOGE,
- [OP_SET_LT] = LLVMRealOLT,
- [OP_SET_GT] = LLVMRealOGT,
- /* Are these used with FP? */
- [OP_SET_B] = LLVMRealOLT,
- [OP_SET_A] = LLVMRealOGT,
- [OP_SET_BE] = LLVMRealOLE,
- [OP_SET_AE] = LLVMRealOGE,
+ [OP_FCMP_ORD] = LLVMRealORD,
+ [OP_FCMP_OEQ] = LLVMRealOEQ,
+ [OP_FCMP_ONE] = LLVMRealONE,
+ [OP_FCMP_OLE] = LLVMRealOLE,
+ [OP_FCMP_OGE] = LLVMRealOGE,
+ [OP_FCMP_OLT] = LLVMRealOLT,
+ [OP_FCMP_OGT] = LLVMRealOGT,
+ [OP_FCMP_UEQ] = LLVMRealUEQ,
+ [OP_FCMP_UNE] = LLVMRealUNE,
+ [OP_FCMP_ULE] = LLVMRealULE,
+ [OP_FCMP_UGE] = LLVMRealUGE,
+ [OP_FCMP_ULT] = LLVMRealULT,
+ [OP_FCMP_UGT] = LLVMRealUGT,
+ [OP_FCMP_UNO] = LLVMRealUNO,
};
return trans_tbl[opcode];
@@ -1029,7 +1032,7 @@ static void output_insn(struct function *fn, struct instruction *insn)
case OP_BINARY ... OP_BINARY_END:
output_op_binary(fn, insn);
break;
- case OP_BINCMP ... OP_BINCMP_END:
+ case OP_FPCMP ... OP_BINCMP_END:
output_op_compare(fn, insn);
break;
case OP_SEL:
new file mode 100644
@@ -0,0 +1,123 @@
+extern double g;
+
+int fcmp_eq(double a) { return (g == a); }
+int fcmp_ne(double a) { return (g != a); }
+
+int fcmp_gt(double a) { return (g > a); }
+int fcmp_ge(double a) { return (g >= a); }
+int fcmp_le(double a) { return (g <= a); }
+int fcmp_lt(double a) { return (g < a); }
+
+int nfcmp_ne(double a) { return !(g == a); }
+int nfcmp_eq(double a) { return !(g != a); }
+
+int nfcmp_le(double a) { return !(g > a); }
+int nfcmp_lt(double a) { return !(g >= a); }
+int nfcmp_gt(double a) { return !(g <= a); }
+int nfcmp_ge(double a) { return !(g < a); }
+
+/*
+ * check-name: canonical-cmp
+ * check-command: test-linearize -Wno-decl $file
+ *
+ * check-output-exclude: \$123,
+ *
+ * check-output-start
+fcmp_eq:
+.L0:
+ <entry-point>
+ load.64 %r1 <- 0[g]
+ fcmpoeq.32 %r3 <- %r1, %arg1
+ ret.32 %r3
+
+
+fcmp_ne:
+.L2:
+ <entry-point>
+ load.64 %r5 <- 0[g]
+ fcmpune.32 %r7 <- %r5, %arg1
+ ret.32 %r7
+
+
+fcmp_gt:
+.L4:
+ <entry-point>
+ load.64 %r9 <- 0[g]
+ fcmpogt.32 %r11 <- %r9, %arg1
+ ret.32 %r11
+
+
+fcmp_ge:
+.L6:
+ <entry-point>
+ load.64 %r13 <- 0[g]
+ fcmpoge.32 %r15 <- %r13, %arg1
+ ret.32 %r15
+
+
+fcmp_le:
+.L8:
+ <entry-point>
+ load.64 %r17 <- 0[g]
+ fcmpole.32 %r19 <- %r17, %arg1
+ ret.32 %r19
+
+
+fcmp_lt:
+.L10:
+ <entry-point>
+ load.64 %r21 <- 0[g]
+ fcmpolt.32 %r23 <- %r21, %arg1
+ ret.32 %r23
+
+
+nfcmp_ne:
+.L12:
+ <entry-point>
+ load.64 %r25 <- 0[g]
+ fcmpune.32 %r28 <- %r25, %arg1
+ ret.32 %r28
+
+
+nfcmp_eq:
+.L14:
+ <entry-point>
+ load.64 %r30 <- 0[g]
+ fcmpoeq.32 %r33 <- %r30, %arg1
+ ret.32 %r33
+
+
+nfcmp_le:
+.L16:
+ <entry-point>
+ load.64 %r35 <- 0[g]
+ fcmpule.32 %r38 <- %r35, %arg1
+ ret.32 %r38
+
+
+nfcmp_lt:
+.L18:
+ <entry-point>
+ load.64 %r40 <- 0[g]
+ fcmpult.32 %r43 <- %r40, %arg1
+ ret.32 %r43
+
+
+nfcmp_gt:
+.L20:
+ <entry-point>
+ load.64 %r45 <- 0[g]
+ fcmpugt.32 %r48 <- %r45, %arg1
+ ret.32 %r48
+
+
+nfcmp_ge:
+.L22:
+ <entry-point>
+ load.64 %r50 <- 0[g]
+ fcmpuge.32 %r53 <- %r50, %arg1
+ ret.32 %r53
+
+
+ * check-output-end
+ */
Comparision of floating-point values can't be done like for integral values because of the possibility to have NaNs which can't be ordered with normal values or even between themselves. The real difference appears once there is any "reasoning" done with the result of the comparison. For example, once NaNs are taken in account: "!(a < b)" and "(a >= b)" are not the same. In fact the usual comparison operators must be reinterpreted as implicitely first testing if any of the operand is a Nan and return 'false' if it is the case. Thus "a < b" becomes "!isnan(a) && !isnan(b) && (a < b)". If we need to negate the comparison we get "!(a < b)" which naturally becomes "isnan(a) || isnan(b) || (a >= b)". We thus need two sets of operators for comparison of floats: one for the "ordered" values (only true if neither operand is a Nan) and one for the "values" (also true if either operand is a NaN). A negation of the comparison switch from one of the set to the other. So, introduce another set of instructions for the comparison of floats. Note: the C standard requires that: *) "x == x" is false if x is a NaN, *) "x != x" is true if x is a NaN, and this is coherent with "x != x" <-> "!(x == x)". Signed-off-by: Luc Van Oostenryck <luc.vanoostenryck@gmail.com> --- linearize.c | 21 ++++++- linearize.h | 18 ++++++ liveness.c | 1 + opcode.c | 38 ++++++++---- opcode.h | 10 ++++ simplify.c | 2 +- sparse-llvm.c | 27 +++++---- validation/optim/canonical-fcmp.c | 123 ++++++++++++++++++++++++++++++++++++++ 8 files changed, 215 insertions(+), 25 deletions(-) create mode 100644 validation/optim/canonical-fcmp.c