From 050fc8b27a852007f8bb667999e1c8cfd31f90e1 Mon Sep 17 00:00:00 2001 From: Roger Sayle Date: Tue, 25 Aug 2020 19:02:45 +0100 Subject: [PATCH] hppa: PR middle-end/87256: Improved hppa_rtx_costs avoids synth_mult madness. This is my proposed fix to PR middle-end/87256 where synth_mult takes an unreasonable amount of CPU time determining an optimal sequence of instructions to perform multiplication by (large) integer constants on hppa. One workaround proposed in bugzilla, is to increase the hash table used to cache/reuse intermediate results. This helps but is a workaround for the (hidden) underlying problem. The real issue is that the hppa_rtx_costs function is providing wildly inaccurate values (estimates) to the middle-end. For example, (p*q)+(r*s) would appear to be cheaper than a single multiplication. Another example is that "(ashiftrt:di regA regB)" is claimed to be only be COST_N_INSNS(1) when in fact the hppa backend actually generates slightly more than a single instruction. It turns out that simply tightening up the logic in hppa_rtx_costs to return more reasonable values, dramatically reduces the number of recursive invocations in synth_mult for the test case in PR87256, and presumably also produces faster code (that should be observable in benchmarks). 2020-08-25 Roger Sayle gcc/ChangeLog PR middle-end/87256 * config/pa/pa.c (hppa_rtx_costs_shadd_p): New helper function to check for coefficients supported by shNadd and shladd,l. (hppa_rtx_costs): Rewrite to avoid using estimates based upon FACTOR and enable recursing deeper into RTL expressions. --- gcc/config/pa/pa.c | 172 ++++++++++++++++++++++++++++++++++++++++++----------- 1 file changed, 138 insertions(+), 34 deletions(-) diff --git a/gcc/config/pa/pa.c b/gcc/config/pa/pa.c index 07d3287..cb88852 100644 --- a/gcc/config/pa/pa.c +++ b/gcc/config/pa/pa.c @@ -1492,6 +1492,33 @@ hppa_address_cost (rtx X, machine_mode mode ATTRIBUTE_UNUSED, } } +/* Return true if X represents a (possibly non-canonical) shNadd pattern. + The machine mode of X is known to be SImode or DImode. */ + +static bool +hppa_rtx_costs_shadd_p (rtx x) +{ + if (GET_CODE (x) != PLUS + || !REG_P (XEXP (x, 1))) + return false; + rtx op0 = XEXP (x, 0); + if (GET_CODE (op0) == ASHIFT + && CONST_INT_P (XEXP (op0, 1)) + && REG_P (XEXP (op0, 0))) + { + unsigned HOST_WIDE_INT x = UINTVAL (XEXP (op0, 1)); + return x == 1 || x == 2 || x == 3; + } + if (GET_CODE (op0) == MULT + && CONST_INT_P (XEXP (op0, 1)) + && REG_P (XEXP (op0, 0))) + { + unsigned HOST_WIDE_INT x = UINTVAL (XEXP (op0, 1)); + return x == 2 || x == 4 || x == 8; + } + return false; +} + /* Compute a (partial) cost for rtx X. Return true if the complete cost has been computed, and false if subexpressions should be scanned. In either case, *TOTAL contains the cost result. */ @@ -1499,15 +1526,16 @@ hppa_address_cost (rtx X, machine_mode mode ATTRIBUTE_UNUSED, static bool hppa_rtx_costs (rtx x, machine_mode mode, int outer_code, int opno ATTRIBUTE_UNUSED, - int *total, bool speed ATTRIBUTE_UNUSED) + int *total, bool speed) { - int factor; int code = GET_CODE (x); switch (code) { case CONST_INT: - if (INTVAL (x) == 0) + if (outer_code == SET) + *total = COSTS_N_INSNS (1); + else if (INTVAL (x) == 0) *total = 0; else if (INT_14_BITS (x)) *total = 1; @@ -1530,32 +1558,35 @@ hppa_rtx_costs (rtx x, machine_mode mode, int outer_code, && outer_code != SET) *total = 0; else - *total = 8; + *total = 8; return true; case MULT: if (GET_MODE_CLASS (mode) == MODE_FLOAT) { *total = COSTS_N_INSNS (3); - return true; } - - /* A mode size N times larger than SImode needs O(N*N) more insns. */ - factor = GET_MODE_SIZE (mode) / 4; - if (factor == 0) - factor = 1; - - if (TARGET_PA_11 && !TARGET_DISABLE_FPREGS && !TARGET_SOFT_FLOAT) - *total = factor * factor * COSTS_N_INSNS (8); + else if (mode == DImode) + { + if (TARGET_PA_11 && !TARGET_DISABLE_FPREGS && !TARGET_SOFT_FLOAT) + *total = COSTS_N_INSNS (32); + else + *total = COSTS_N_INSNS (80); + } else - *total = factor * factor * COSTS_N_INSNS (20); - return true; + { + if (TARGET_PA_11 && !TARGET_DISABLE_FPREGS && !TARGET_SOFT_FLOAT) + *total = COSTS_N_INSNS (8); + else + *total = COSTS_N_INSNS (20); + } + return REG_P (XEXP (x, 0)) && REG_P (XEXP (x, 1)); case DIV: if (GET_MODE_CLASS (mode) == MODE_FLOAT) { *total = COSTS_N_INSNS (14); - return true; + return false; } /* FALLTHRU */ @@ -1563,34 +1594,107 @@ hppa_rtx_costs (rtx x, machine_mode mode, int outer_code, case MOD: case UMOD: /* A mode size N times larger than SImode needs O(N*N) more insns. */ - factor = GET_MODE_SIZE (mode) / 4; - if (factor == 0) - factor = 1; - - *total = factor * factor * COSTS_N_INSNS (60); - return true; + if (mode == DImode) + *total = COSTS_N_INSNS (240); + else + *total = COSTS_N_INSNS (60); + return REG_P (XEXP (x, 0)) && REG_P (XEXP (x, 1)); case PLUS: /* this includes shNadd insns */ case MINUS: if (GET_MODE_CLASS (mode) == MODE_FLOAT) + *total = COSTS_N_INSNS (3); + else if (mode == DImode) { - *total = COSTS_N_INSNS (3); - return true; + if (TARGET_64BIT) + { + *total = COSTS_N_INSNS (1); + /* Handle shladd,l instructions. */ + if (hppa_rtx_costs_shadd_p (x)) + return true; + } + else + *total = COSTS_N_INSNS (2); } - - /* A size N times larger than UNITS_PER_WORD needs N times as - many insns, taking N times as long. */ - factor = GET_MODE_SIZE (mode) / UNITS_PER_WORD; - if (factor == 0) - factor = 1; - *total = factor * COSTS_N_INSNS (1); - return true; + else + { + *total = COSTS_N_INSNS (1); + /* Handle shNadd instructions. */ + if (hppa_rtx_costs_shadd_p (x)) + return true; + } + return REG_P (XEXP (x, 0)) + && (REG_P (XEXP (x, 1)) + || CONST_INT_P (XEXP (x, 1))); case ASHIFT: + if (mode == DImode) + { + if (TARGET_64BIT) + *total = COSTS_N_INSNS (3); + else if (REG_P (XEXP (x, 0)) && CONST_INT_P (XEXP (x, 1))) + { + *total = COSTS_N_INSNS (2); + return true; + } + else if (speed) + *total = COSTS_N_INSNS (13); + else + *total = COSTS_N_INSNS (18); + } + else if (TARGET_64BIT) + *total = COSTS_N_INSNS (4); + else + *total = COSTS_N_INSNS (2); + return REG_P (XEXP (x, 0)) + && (REG_P (XEXP (x, 1)) + || CONST_INT_P (XEXP (x, 1))); + case ASHIFTRT: + if (mode == DImode) + { + if (TARGET_64BIT) + *total = COSTS_N_INSNS (3); + else if (REG_P (XEXP (x, 0)) && CONST_INT_P (XEXP (x, 1))) + { + *total = COSTS_N_INSNS (2); + return true; + } + else if (speed) + *total = COSTS_N_INSNS (14); + else + *total = COSTS_N_INSNS (19); + } + else if (TARGET_64BIT) + *total = COSTS_N_INSNS (4); + else + *total = COSTS_N_INSNS (2); + return REG_P (XEXP (x, 0)) + && (REG_P (XEXP (x, 1)) + || CONST_INT_P (XEXP (x, 1))); + case LSHIFTRT: - *total = COSTS_N_INSNS (1); - return true; + if (mode == DImode) + { + if (TARGET_64BIT) + *total = COSTS_N_INSNS (2); + else if (REG_P (XEXP (x, 0)) && CONST_INT_P (XEXP (x, 1))) + { + *total = COSTS_N_INSNS (2); + return true; + } + else if (speed) + *total = COSTS_N_INSNS (12); + else + *total = COSTS_N_INSNS (15); + } + else if (TARGET_64BIT) + *total = COSTS_N_INSNS (3); + else + *total = COSTS_N_INSNS (2); + return REG_P (XEXP (x, 0)) + && (REG_P (XEXP (x, 1)) + || CONST_INT_P (XEXP (x, 1))); default: return false; -- 2.7.4